Appendix B — Further papers
For each chapter, a longer list of papers grouped by theme: classic papers, reviews, methods and recent developments. They are not required reading, but they are the papers behind the text of the chapter and a good place to go deeper. Each paper is given with its title, a DOI link and, where available, its abstract.
B.1 Sequence searching
Further papers for Chapter 1.
Papers marked ★ were mandatory in earlier reading lists. In the HTML version, click Abstract to show the abstract.
B.1.1 Reviews — start here
Pearson (2013) — an introduction to similarity searching, homology inference and statistics; the best single starting point.
An Introduction to Sequence Similarity (“Homology”) Searching. Pearson (2013). Current Protocols in Bioinformatics. doi:10.1002/0471250953.bi0301s42
Abstract
Sequence similarity searching, typically with BLAST, is the most widely used and most reliable strategy for characterizing newly determined sequences. Sequence similarity searches can identify "homologous" proteins or genes by detecting excess similarity- statistically significant similarity that reflects common ancestry. This unit provides an overview of the inference of homology from significant similarity, and introduces other units in this chapter that provide more details on effective strategies for identifying homologs.
Eddy (2004a) — “What is a hidden Markov model?” (two pages).
What is a hidden Markov model?. Eddy (2004). Nature Biotechnology. doi:10.1038/nbt1004-1315
Eddy (1998) — profile hidden Markov models.
Profile hidden Markov models.. Eddy (1998). Bioinformatics. doi:10.1093/bioinformatics/14.9.755
Abstract
The recent literature on profile hidden Markov model (profile HMM) methods and software is reviewed. Profile HMMs turn a multiple sequence alignment into a position-specific scoring system suitable for searching databases for remotely homologous sequences. Profile HMM analyses complement standard pairwise comparison methods for large-scale sequence analysis. Several software implementations and two large libraries of profile HMMs of common protein domains are available. HMM methods performed comparably to threading methods in the CASP2 structure prediction exercise.
Lambert et al. (2003) — review of common sequence alignment methods.
Review of common sequence alignment methods: clues to enhance reliability. Lambert et al. (2003). Current Genomics.
Bepler and Berger (2021) — learning the protein language: evolution, structure and function.
Learning the protein language: Evolution, structure, and function. Bepler & Berger (2021). Cell Systems. doi:10.1016/j.cels.2021.05.017
Abstract
Language models have recently emerged as a powerful machine-learning approach for distilling information from massive protein sequence databases. From readily available sequence data alone, these models discover evolutionary, structural, and functional organization across protein space. Using language models, we can encode amino-acid sequences into distributed vector representations that capture their structural and functional properties, as well as evaluate the evolutionary fitness of sequence variants. We discuss recent advances in protein language modeling and their applications to downstream protein property prediction problems. We then consider how these models can be enriched with prior biological knowledge and introduce an approach for encoding protein structural knowledge into the learned representations. The knowledge distilled by these models allows us to improve downstream function prediction through transfer learning. Deep protein language models are revolutionizing protein biology. They suggest new ways to approach protein and therapeutic design. However, further developments are needed to encode strong biological priors into protein language models and to increase their accessibility to the broader community.
B.1.2 Homology
Reeck et al. (1987) — “homology” in proteins and nucleic acids: a terminology muddle and a way out of it.
“Homology” in proteins and nucleic acids: A terminology muddle and a way out of it. Reeck et al. (1987). Cell. doi:10.1016/0092-8674(87)90322-9
Doolittle (1981) — similar amino acid sequences: chance or common ancestry?
Similar Amino Acid Sequences: Chance or Common Ancestry?. Doolittle (1981). Science. doi:10.1126/science.7280687
Abstract
The systemic comparison of every newly determined amino acid sequence with all other known sequences may allow a complete reconstruction of the evolutionary events leading to contemporary proteins. But sometimes the surviving similarities are so vague that even computer-based sequence comparisons procedures are unable to validate relationships. In other cases similar sequences may appear in totally alien proteins as a result of mere chance or, occasionally, by the convergent evolution of sequences with special properties.
Rost (1999) — the twilight zone of protein sequence alignments.
Twilight zone of protein sequence alignments. Rost (1999). Protein Engineering, Design and Selection. doi:10.1093/protein/12.2.85
Abstract
Sequence alignments unambiguously distinguish between protein pairs of similar and non-similar structure when the pairwise sequence identity is high (>40% for long alignments). The signal gets blurred in the twilight zone of 20-35% sequence identity. Here, more than a million sequence alignments were analysed between protein pairs of known structures to re-define a line distinguishing between true and false positives for low levels of similarity. Four results stood out. (i) The transition from the safe zone of sequence alignment into the twilight zone is described by an explosion of false negatives. More than 95% of all pairs detected in the twilight zone had different structures. More precisely, above a cut-off roughly corresponding to 30% sequence identity, 90% of the pairs were homologous; below 25% less than 10% were. (ii) Whether or not sequence homology implied structural identity depended crucially on the alignment length. For example, if 10 residues were similar in an alignment of length 16 (>60%), structural similarity could not be inferred. (iii) The 'more similar than identical' rule (discarding all pairs for which percentage similarity was lower than percentage identity) reduced false positives significantly. (iv) Using intermediate sequences for finding links between more distant families was almost as successful: pairs were predicted to be homologous when the respective sequence families had proteins in common. All findings are applicable to automatic database searches.
Fitch (1970) — distinguishing homologous from analogous proteins (orthology and paralogy defined).
Distinguishing Homologous from Analogous Proteins. Fitch (1970). Systematic Zoology. doi:10.2307/2412448
B.1.3 Pairwise alignment and substitution matrices
Bellman (1954) — the theory of dynamic programming.
The theory of dynamic programming. Bellman (1954). Bulletin of the American Mathematical Society. doi:10.1090/s0002-9904-1954-09848-8
Needleman and Wunsch (1970) — global alignment by dynamic programming.
A general method applicable to the search for similarities in the amino acid sequence of two proteins. Needleman & Wunsch (1970). Journal of Molecular Biology. doi:10.1016/0022-2836(70)90057-4
Gotoh (1982) — affine gap penalties in quadratic time.
An improved algorithm for matching biological sequences. Gotoh (1982). Journal of Molecular Biology. doi:10.1016/0022-2836(82)90398-9
Myers and Miller (1988) — optimal alignments in linear space.
Optimal alignments in linear space. Myers & Miller (1988). Bioinformatics. doi:10.1093/bioinformatics/4.1.11
Abstract
Space, not time, is often the limiting factor when computing optimal sequence alignments, and a number of recent papers in the biology literature have proposed space-saving strategies. However, a 1975 computer science paper by Hirschberg presented a method that is superior to the new proposals, both in theory and in practice. The goal of this paper is to give Hirschberg's idea the visibility it deserves by developing a linear-space version of Gotoh's algorithm, which accommodates affine gap penalties. A portable C-software package implementing this algorithm is available on the BIONET free of charge.
Dayhoff et al. (1978) — PAM matrices.
A model of evolutionary change in proteins. Dayhoff et al. (1978). Atlas of Protein Sequence and Structure.
Henikoff and Henikoff (1992) — BLOSUM matrices.
Amino acid substitution matrices from protein blocks.. Henikoff & Henikoff (1992). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.89.22.10915
Abstract
Methods for alignment of protein sequences typically measure similarity by using a substitution matrix with scores for all possible exchanges of one amino acid with another. The most widely used matrices are based on the Dayhoff model of evolutionary rates. Using a different approach, we have derived substitution matrices from about 2000 blocks of aligned sequence segments characterizing more than 500 groups of related proteins. This led to marked improvements in alignments and in searches using queries from each of the groups.
Altschul (1991) — substitution matrices from an information-theoretic perspective.
Amino acid substitution matrices from an information theoretic perspective. Altschul (1991). Journal of Molecular Biology. doi:10.1016/0022-2836(91)90193-a
Abstract
Protein sequence alignments have become an important tool for molecular biologists. Local alignments are frequently constructed with the aid of a "substitution score matrix" that specifies a score for aligning each pair of amino acid residues. Over the years, many different substitution matrices have been proposed, based on a wide variety of rationales. Statistical results, however, demonstrate that any such matrix is implicitly a "log-odds" matrix, with a specific target distribution for aligned pairs of amino acid residues. In the light of information theory, it is possible to express the scores of a substitution matrix in bits and to see that different matrices are better adapted to different purposes. The most widely used matrix for protein sequence comparison has been the PAM-250 matrix. It is argued that for database searches the PAM-120 matrix generally is more appropriate, while for comparing two specific proteins with suspected homology the PAM-200 matrix is indicated. Examples discussed include the lipocalins, human alpha 1 B-glycoprotein, the cystic fibrosis transmembrane conductance regulator and the globins.
Eddy (2004b) — where did the BLOSUM62 matrix come from?
Where did the BLOSUM62 alignment score matrix come from?. Eddy (2004). Nature Biotechnology. doi:10.1038/nbt0804-1035
Abstract
Many sequence alignment programs use the BLOSUM62 score matrix to score pairs of aligned residues. Where did BLOSUM62 come from?
B.1.4 Heuristic search and statistics
Pearson and Lipman (1988) — FASTA.
Improved tools for biological sequence comparison.. Pearson & Lipman (1988). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.85.8.2444
Abstract
We have developed three computer programs for comparisons of protein and DNA sequences. They can be used to search sequence data bases, evaluate similarity scores, and identify periodic structures based on local sequence similarity. The FASTA program is a more sensitive derivative of the FASTP program, which can be used to search protein or DNA sequence data bases and can compare a protein sequence to a DNA sequence data base by translating the DNA data base as it is searched. FASTA includes an additional step in the calculation of the initial pairwise similarity score that allows multiple regions of similarity to be joined to increase the score of related sequences. The RDF2 program can be used to evaluate the significance of similarity scores using a shuffling method that preserves local sequence composition. The LFASTA program can display all the regions of local similarity between two sequences with scores greater than a threshold, using the same scoring parameters and a similar alignment algorithm; these local similarities can be displayed as a "graphic matrix" plot or as individual alignments. In addition, these programs have been generalized to allow comparison of DNA or protein sequences based on a variety of alternative scoring matrices.
Altschul et al. (1990) — the original BLAST.
Basic local alignment search tool. Altschul et al. (1990). Journal of Molecular Biology. doi:10.1016/s0022-2836(05)80360-2
Abstract
A new approach to rapid sequence comparison, basic local alignment search tool (BLAST), directly approximates alignments that optimize a measure of local similarity, the maximal segment pair (MSP) score. Recent mathematical results on the stochastic properties of MSP scores allow an analysis of the performance of this method as well as the statistical significance of alignments it generates. The basic algorithm is simple and robust; it can be implemented in a number of ways and applied in a variety of contexts including straightforward DNA and protein sequence database searches, motif searches, gene identification searches, and in the analysis of multiple regions of similarity in long DNA sequences. In addition to its flexibility and tractability to mathematical analysis, BLAST is an order of magnitude faster than existing sequence comparison tools of comparable sensitivity.
Karlin and Altschul (1990) — statistical significance of local alignment scores.
Methods for assessing the statistical significance of molecular sequence features by using general scoring schemes.. Karlin & Altschul (1990). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.87.6.2264
Abstract
An unusual pattern in a nucleic acid or protein sequence or a region of strong similarity shared by two or more sequences may have biological significance. It is therefore desirable to know whether such a pattern can have arisen simply by chance. To identify interesting sequence patterns, appropriate scoring values can be assigned to the individual residues of a single sequence or to sets of residues when several sequences are compared. For single sequences, such scores can reflect biophysical properties such as charge, volume, hydrophobicity, or secondary structure potential; for multiple sequences, they can reflect nucleotide or amino acid similarity measured in a wide variety of ways. Using an appropriate random model, we present a theory that provides precise numerical formulas for assessing the statistical significance of any region with high aggregate score. A second class of results describes the composition of high-scoring segments. In certain contexts, these permit the choice of scoring systems which are "optimal" for distinguishing biologically relevant patterns. Examples are given of applications of the theory to a variety of protein sequences, highlighting segments with unusual biological features. These include distinctive charge regions in transcription factors and protooncogene products, pronounced hydrophobic segments in various receptor and transport proteins, and statistically significant subalignments involving the recently characterized cystic fibrosis gene.
Kent (2002) — BLAT, for fast genome-scale alignment.
BLAT —The BLAST -Like Alignment Tool. Kent (2002). Genome Research. doi:10.1101/gr.229202
Abstract
Analyzing vertebrate genomes requires rapid mRNA/DNA and cross-species protein alignments. A new tool, BLAT, is more accurate and 500 times faster than popular existing tools for mRNA/DNA alignments and 50 times faster for protein alignments at sensitivity settings typically used when comparing vertebrate sequences. BLAT's speed stems from an index of all nonoverlapping K-mers in the genome. This index fits inside the RAM of inexpensive computers, and need only be computed once for each genome assembly. BLAT has several major stages. It uses the index to find regions in the genome likely to be homologous to the query sequence. It performs an alignment between homologous regions. It stitches together these aligned regions (often exons) into larger alignments (typically genes). Finally, BLAT revisits small internal exons possibly missed at the first stage and adjusts large gap boundaries that have canonical splice sites where feasible. This paper describes how BLAT was optimized. Effects on speed and sensitivity are explored for various K-mer sizes, mismatch schemes, and number of required index matches. BLAT is compared with other alignment programs on various test sets and then used in several genome-wide applications. http://genome.ucsc.edu hosts a web-based BLAT server for the human genome.
Buchfink et al. (2021) — DIAMOND, BLAST-level sensitivity at much higher speed.
Sensitive protein alignments at tree-of-life scale using DIAMOND. Buchfink et al. (2021). Nature Methods. doi:10.1038/s41592-021-01101-x
Abstract
We are at the beginning of a genomic revolution in which all known species are planned to be sequenced. Accessing such data for comparative analyses is crucial in this new age of data-driven biology. Here, we introduce an improved version of DIAMOND that greatly exceeds previous search performances and harnesses supercomputing to perform tree-of-life scale protein alignments in hours, while matching the sensitivity of the gold standard BLASTP.
Steinegger and Söding (2018) — Linclust: clustering huge sequence sets in linear time.
Clustering huge protein sequence sets in linear time. Steinegger & Söding (2018). Nature Communications. doi:10.1038/s41467-018-04964-5
Abstract
Metagenomic datasets contain billions of protein sequences that could greatly enhance large-scale functional annotation and structure prediction. Utilizing this enormous resource would require reducing its redundancy by similarity clustering. However, clustering hundreds of millions of sequences is impractical using current algorithms because their runtimes scale as the input set size N times the number of clusters K, which is typically of similar order as N, resulting in runtimes that increase almost quadratically with N. We developed Linclust, the first clustering algorithm whose runtime scales as N, independent of K. It can also cluster datasets several times larger than the available main memory. We cluster 1.6 billion metagenomic sequence fragments in 10 h on a single server to 50% sequence identity, >1000 times faster than has been possible before. Linclust will help to unlock the great wealth contained in metagenomic and genomic sequence databases.
Barrio-Hernandez et al. (2023) — clustering the AlphaFold database with Foldseek.
Clustering predicted structures at the scale of the known protein universe. Barrio-Hernandez et al. (2023). Nature. doi:10.1038/s41586-023-06510-w
Abstract
Proteins are key to all cellular processes and their structure is important in understanding their function and evolution. Sequence-based predictions of protein structures have increased in accuracy 1 , and over 214 million predicted structures are available in the AlphaFold database 2 . However, studying protein structures at this scale requires highly efficient methods. Here, we developed a structural-alignment-based clustering algorithm-Foldseek cluster-that can cluster hundreds of millions of structures. Using this method, we have clustered all of the structures in the AlphaFold database, identifying 2.30 million non-singleton structural clusters, of which 31% lack annotations representing probable previously undescribed structures. Clusters without annotation tend to have few representatives covering only 4% of all proteins in the AlphaFold database. Evolutionary analysis suggests that most clusters are ancient in origin but 4% seem to be species specific, representing lower-quality predictions or examples of de novo gene birth. We also show how structural comparisons can be used to predict domain families and their relationships, identifying examples of remote structural similarity. On the basis of these analyses, we identify several examples of human immune-related proteins with putative remote homology in prokaryotic species, illustrating the value of this resource for studying protein function and evolution across the tree of life.
Ferragina and Manzini (2000) — the FM-index.
Opportunistic data structures with applications. Ferragina & Manzini (2000). Proceedings 41st Annual Symposium on Foundations of Computer Science. doi:10.1109/sfcs.2000.892127
Abstract
We address the issue of compressing and indexing data. We devise a data structure whose space occupancy is a function of the entropy of the underlying data set. We call the data structure opportunistic since its space occupancy is decreased when the input is compressible and this space reduction is achieved at no significant slowdown in the query performance. More precisely, its space occupancy is optimal in an information-content sense because text T[1,u] is stored using O(H/sub k/(T))+o(1) bits per input symbol in the worst case, where H/sub k/(T) is the kth order empirical entropy of T (the bound holds for any fixed k). Given an arbitrary string P[1,p], the opportunistic data structure allows to search for the occurrences of P in T in O(p+occlog/sup /spl epsiv//u) time (for any fixed /spl epsiv/>0). If data are uncompressible we achieve the best space bound currently known (Grossi and Vitter, 2000); on compressible data our solution improves the succinct suffix array of (Grossi and Vitter, 2000) and the classical suffix tree and suffix array data structures either in space or in query time or both. We also study our opportunistic data structure in a dynamic setting and devise a variant achieving effective search and update time bounds. Finally, we show how to plug our opportunistic data structure into the Glimpse tool (Manber and Wu, 1994). The result is an indexing tool which achieves sublinear space and sublinear query time complexity.
B.1.5 Multiple sequence alignment and motifs
Feng and Doolittle (1987) — progressive alignment.
Progressive sequence alignment as a prerequisite to correct phylogenetic trees. Feng & Doolittle (1987). Journal of Molecular Evolution. doi:10.1007/bf02603120
Abstract
A progressive alignment method is described that utilizes the Needleman and Wunsch pairwise alignment algorithm iteratively to achieve the multiple alignment of a set of protein sequences and to construct an evolutionary tree depicting their relationship. The sequences are assumed a priori to share a common ancestor, and the trees are constructed from difference matrices derived directly from the multiple alignment. The thrust of the method involves putting more trust in the comparison of recently diverged sequences than in those evolved in the distant past. In particular, this rule is followed: "once a gap, always a gap." The method has been applied to three sets of protein sequences: 7 superoxide dismutases, 11 globins, and 9 tyrosine kinase-like sequences. Multiple alignments and phylogenetic trees for these sets of sequences were determined and compared with trees derived by conventional pairwise treatments. In several instances, the progressive method led to trees that appeared to be more in line with biological expectations than were trees obtained by more commonly used methods.
Thompson et al. (1994) — ClustalW.
CLUSTAL W: improving the sensitivity of progressive multiple sequence alignment through sequence weighting, position-specific gap penalties and weight matrix choice. Thompson et al. (1994). Nucleic Acids Research. doi:10.1093/nar/22.22.4673
Abstract
The sensitivity of the commonly used progressive multiple sequence alignment method has been greatly improved for the alignment of divergent protein sequences. Firstly, individual weights are assigned to each sequence in a partial alignment in order to down-weight near-duplicate sequences and up-weight the most divergent ones. Secondly, amino acid substitution matrices are varied at different alignment stages according to the divergence of the sequences to be aligned. Thirdly, residue-specific gap penalties and locally reduced gap penalties in hydrophilic regions encourage new gaps in potential loop regions rather than regular secondary structure. Fourthly, positions in early alignments where gaps have been opened receive locally reduced gap penalties to encourage the opening up of new gaps at these positions. These modifications are incorporated into a new program, CLUSTAL W which is freely available.
Sievers et al. (2011) — Clustal Omega.
Fast, scalable generation of high-quality protein multiple sequence alignments using Clustal Omega. Sievers et al. (2011). Molecular Systems Biology. doi:10.1038/msb.2011.75
Abstract
Multiple sequence alignments are fundamental to many sequence analysis methods. Most alignments are computed using the progressive alignment heuristic. These methods are starting to become a bottleneck in some analysis pipelines when faced with data sets of the size of many thousands of sequences. Some methods allow computation of larger data sets while sacrificing quality, and others produce high-quality alignments, but scale badly with the number of sequences. In this paper, we describe a new program called Clustal Omega, which can align virtually any number of protein sequences quickly and that delivers accurate alignments. The accuracy of the package on smaller test cases is similar to that of the high-quality aligners. On larger data sets, Clustal Omega outperforms other packages in terms of execution time and quality. Clustal Omega also has powerful features for adding sequences to and exploiting information in existing alignments, making use of the vast amount of precomputed information in public databases like Pfam.
Notredame et al. (2000) — T-Coffee (consistency).
T-coffee: a novel method for fast and accurate multiple sequence alignment. Notredame et al. (2000). Journal of Molecular Biology. doi:10.1006/jmbi.2000.4042
Abstract
We describe a new method (T-Coffee) for multiple sequence alignment that provides a dramatic improvement in accuracy with a modest sacrifice in speed as compared to the most commonly used alternatives. The method is broadly based on the popular progressive approach to multiple alignment but avoids the most serious pitfalls caused by the greedy nature of this algorithm. With T-Coffee we pre-process a data set of all pair-wise alignments between the sequences. This provides us with a library of alignment information that can be used to guide the progressive alignment. Intermediate alignments are then based not only on the sequences to be aligned next but also on how all of the sequences align with each other. This alignment information can be derived from heterogeneous sources such as a mixture of alignment programs and/or structure superposition. Here, we illustrate the power of the approach by using a combination of local and global pair-wise alignments to generate the library. The resulting alignments are significantly more reliable, as determined by comparison with a set of 141 test cases, than any of the popular alternatives that we tried. The improvement, especially clear with the more difficult test cases, is always visible, regardless of the phylogenetic spread of the sequences in the tests.
Edgar (2004) — MUSCLE.
MUSCLE: multiple sequence alignment with high accuracy and high throughput. Edgar (2004). Nucleic Acids Research. doi:10.1093/nar/gkh340
Abstract
We describe MUSCLE, a new computer program for creating multiple alignments of protein sequences. Elements of the algorithm include fast distance estimation using kmer counting, progressive alignment using a new profile function we call the log-expectation score, and refinement using tree-dependent restricted partitioning. The speed and accuracy of MUSCLE are compared with T-Coffee, MAFFT and CLUSTALW on four test sets of reference alignments: BAliBASE, SABmark, SMART and a new benchmark, PREFAB. MUSCLE achieves the highest, or joint highest, rank in accuracy on each of these sets. Without refinement, MUSCLE achieves average accuracy statistically indistinguishable from T-Coffee and MAFFT, and is the fastest of the tested methods for large numbers of sequences, aligning 5000 sequences of average length 350 in 7 min on a current desktop computer. The MUSCLE program, source code and PREFAB test data are freely available at http://www.drive5. com/muscle.
Katoh and Standley (2013) — MAFFT.
MAFFT Multiple Sequence Alignment Software Version 7: Improvements in Performance and Usability. Katoh & Standley (2013). Molecular Biology and Evolution. doi:10.1093/molbev/mst010
Abstract
We report a major update of the MAFFT multiple sequence alignment program. This version has several new features, including options for adding unaligned sequences into an existing alignment, adjustment of direction in nucleotide alignment, constrained alignment and parallel processing, which were implemented after the previous major update. This report shows actual examples to explain how these features work, alone and in combination. Some examples incorrectly aligned by MAFFT are also shown to clarify its limitations. We discuss how to avoid misalignments, and our ongoing efforts to overcome such limitations.
Lawrence et al. (1993) — Gibbs sampling for subtle motifs.
Detecting Subtle Sequence Signals: a Gibbs Sampling Strategy for Multiple Alignment. Lawrence et al. (1993). Science. doi:10.1126/science.8211139
Abstract
A wealth of protein and DNA sequence data is being generated by genome projects and other sequencing efforts. A crucial barrier to deciphering these sequences and understanding the relations among them is the difficulty of detecting subtle local residue patterns common to multiple sequences. Such patterns frequently reflect similar molecular structures and biological properties. A mathematical definition of this "local multiple alignment" problem suitable for full computer automation has been used to develop a new and sensitive algorithm, based on the statistical method of iterative sampling. This algorithm finds an optimized local alignment model for N sequences in N-linear time, requiring only seconds on current workstations, and allows the simultaneous detection and optimization of multiple patterns and pattern repeats. The method is illustrated as applied to helix-turn-helix proteins, lipocalins, and prenyltransferases.
Bailey and Elkan (1994) — MEME (expectation maximisation).
Fitting a mixture model by expectation maximization to discover motifs in biopolymers. Bailey & Elkan (1994). Proceedings of the Second International Conference on Intelligent Systems for Molecular Biology (ISMB).
Schneider and Stephens (1990) — sequence logos.
Sequence logos: a new way to display consensus sequences. Schneider & Stephens (1990). Nucleic Acids Research. doi:10.1093/nar/18.20.6097
Abstract
A graphical method is presented for displaying the patterns in a set of aligned sequences. The characters representing the sequence are stacked on top of each other for each position in the aligned sequences. The height of each letter is made proportional to its frequency, and the letters are sorted so the most common one is on top. The height of the entire stack is then adjusted to signify the information content of the sequences at that position. From these 'sequence logos', one can determine not only the consensus sequence but also the relative frequency of bases and the information content (measured in bits) at every position in a site or sequence. The logo displays both significant residues and subtle sequence patterns.
B.1.7 RNA homology search
Eddy and Durbin (1994) — covariance models for RNA sequence analysis.
RNA sequence analysis using covariance models. Eddy & Durbin (1994). Nucleic Acids Research. doi:10.1093/nar/22.11.2079
Abstract
We describe a general approach to several RNA sequence analysis problems using probabilistic models that flexibly describe the secondary structure and primary sequence consensus of an RNA sequence family. We call these models 'covariance models'. A covariance model of tRNA sequences is an extremely sensitive and discriminative tool for searching for additional tRNAs and tRNA-related sequences in sequence databases. A model can be built automatically from an existing sequence alignment. We also describe an algorithm for learning a model and hence a consensus secondary structure from initially unaligned example sequences and no prior structural information. Models trained on unaligned tRNA examples correctly predict tRNA secondary structure and produce high-quality multiple alignments. The approach may be applied to any family of small RNA sequences.
Nawrocki and Eddy (2013) — Infernal 1.1: 100-fold faster RNA homology searches.
Infernal 1.1: 100-fold faster RNA homology searches. Nawrocki & Eddy (2013). Bioinformatics. doi:10.1093/bioinformatics/btt509
Abstract
Infernal builds probabilistic profiles of the sequence and secondary structure of an RNA family called covariance models (CMs) from structurally annotated multiple sequence alignments given as input. Infernal uses CMs to search for new family members in sequence databases and to create potentially large multiple sequence alignments. Version 1.1 of Infernal introduces a new filter pipeline for RNA homology search based on accelerated profile hidden Markov model (HMM) methods and HMM-banded CM alignment methods. This enables ∼100-fold acceleration over the previous version and ∼10 000-fold acceleration over exhaustive non-filtered CM searches. Availability Source code, documentation and the benchmark are downloadable from http://infernal.janelia.org. Infernal is freely licensed under the GNU GPLv3 and should be portable to any POSIX-compliant operating system, including Linux and Mac OS/X. Documentation includes a user's guide with a tutorial, a discussion of file formats and user options and additional details on methods implemented in the software. Contact nawrockie@janelia.hhmi.org
Griffiths-Jones et al. (2003), Kalvari et al. (2020) — the Rfam database.
Rfam: an RNA family database. Griffiths-Jones et al. (2003). Nucleic Acids Research. doi:10.1093/nar/gkg006
Abstract
Rfam is a collection of multiple sequence alignments and covariance models representing non-coding RNA families. Rfam is available on the web in the UK at http://www.sanger.ac.uk/Software/Rfam/ and in the US at http://rfam.wustl.edu/. These websites allow the user to search a query sequence against a library of covariance models, and view multiple sequence alignments and family annotation. The database can also be downloaded in flatfile form and searched locally using the INFERNAL package (http://infernal.wustl.edu/). The first release of Rfam (1.0) contains 25 families, which annotate over 50 000 non-coding RNA genes in the taxonomic divisions of the EMBL nucleotide database.
Rfam 14: expanded coverage of metagenomic, viral and microRNA families. Kalvari et al. (2020). Nucleic Acids Research. doi:10.1093/nar/gkaa1047
Abstract
Rfam is a database of RNA families where each of the 3444 families is represented by a multiple sequence alignment of known RNA sequences and a covariance model that can be used to search for additional members of the family. Recent developments have involved expert collaborations to improve the quality and coverage of Rfam data, focusing on microRNAs, viral and bacterial RNAs. We have completed the first phase of synchronising microRNA families in Rfam and miRBase, creating 356 new Rfam families and updating 40. We established a procedure for comprehensive annotation of viral RNA families starting with Flavivirus and Coronaviridae RNAs. We have also increased the coverage of bacterial and metagenome-based RNA families from the ZWD database. These developments have enabled a significant growth of the database, with the addition of 759 new families in Rfam 14. To facilitate further community contribution to Rfam, expert users are now able to build and submit new families using the newly developed Rfam Cloud family curation system. New Rfam website features include a new sequence similarity search powered by RNAcentral, as well as search and visualisation of families with pseudoknots. Rfam is freely available at https://rfam.org.
Nussinov and Jacobson (1980), Zuker and Stiegler (1981) — RNA secondary structure by dynamic programming.
Fast algorithm for predicting the secondary structure of single-stranded RNA.. Nussinov & Jacobson (1980). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.77.11.6309
Abstract
A computer method is presented for finding the most stable secondary structures in long single-stranded RNAs. It is 1-2 orders of magnitude faster than existing codes. The time required for its application increases as N3 for a chain N nucleotides long. As many as 1000 nucleotides can be searched in a single run. The approach is systematic and builds an optimal structure in a straightforward inductive procedure based on an exact mathematical algorithm. Two simple half-matrices are constructed and the best folded form is read directly from the second matrix by a simple back-tracking procedure. The program utilizes published values for base-pairing energies to compute one structure with the lowest free energy.
Optimal computer folding of large RNA sequences using thermodynamics and auxiliary information. Zuker & Stiegler (1981). Nucleic Acids Research. doi:10.1093/nar/9.1.133
Abstract
This paper presents a new computer method for folding an RNA molecule that finds a conformation of minimum free energy using published values of stacking and destabilizing energies. It is based on a dynamic programming algorithm from applied mathematics, and is much more efficient, faster, and can fold larger molecules than procedures which have appeared up to now in the biological literature. Its power is demonstrated in the folding of a 459 nucleotide immunoglobulin gamma 1 heavy chain messenger RNA fragment. We go beyond the basic method to show how to incorporate additional information into the algorithm. This includes data on chemical reactivity and enzyme susceptibility. We illustrate this with the folding of two large fragments from the 16S ribosomal RNA of Escherichia coli.
Sankoff (1985) — simultaneous folding and alignment.
Simultaneous Solution of the RNA Folding, Alignment and Protosequence Problems. Sankoff (1985). SIAM Journal on Applied Mathematics. doi:10.1137/0145048
Abstract
The alignment of finite sequences, the inference of ribonucleic acid secondary structures (folding), and the reconstruction of ancestral sequences on a phylogenetic tree, are three problems which have dynamic programming solutions, which we formulate in a common mathematical framework. Combining the objective functions for alignment (parsimony, or minimal mutations) and folding (free energy), we present an algorithm which solves all three problems simultaneously for a set of N sequences of length n in time proportional to $n^{3N} $ and storage $n^{2N} $. Incorporating a “cutting corners” constraint against biologically unlikely alignments reduces these requirements so that they are proportional to $n^{3} $ and $n^{2} $, respectively, for fixed N.
Rivas et al. (2016) — a statistical test for conserved RNA structure.
A statistical test for conserved RNA structure shows lack of evidence for structure in lncRNAs. Rivas et al. (2016). Nature Methods. doi:10.1038/nmeth.4066
Abstract
Many functional RNAs have an evolutionarily conserved secondary structure. Conservation of RNA base pairing induces pairwise covariations in sequence alignments. We developed a computational method, R-scape (RNA Structural Covariation Above Phylogenetic Expectation), that quantitatively tests whether covariation analysis supports the presence of a conserved RNA secondary structure. R-scape analysis finds no statistically significant support for proposed secondary structures of the long noncoding RNAs HOTAIR, SRA, and Xist.
Lowe and Eddy (1997) — tRNAscan-SE.
tRNAscan-SE: A Program for Improved Detection of Transfer RNA Genes in Genomic Sequence. Lowe & Eddy (1997). Nucleic Acids Research. doi:10.1093/nar/25.5.955
Abstract
We describe a program, tRNAscan-SE, which identifies 99-100% of transfer RNA genes in DNA sequence while giving less than one false positive per 15 gigabases. Two previously described tRNA detection programs are used as fast, first-pass prefilters to identify candidate tRNAs, which are then analyzed by a highly selective tRNA covariance model. This work represents a practical application of RNA covariance models, which are general, probabilistic secondary structure profiles based on stochastic context-free grammars. tRNAscan-SE searches at approximately 30 000 bp/s. Additional extensions to tRNAscan-SE detect unusual tRNA homologues such as selenocysteine tRNAs, tRNA-derived repetitive elements and tRNA pseudogenes.
B.1.8 Alignment-free comparison
Zielezinski et al. (2017) — review of alignment-free sequence comparison.
Alignment-free sequence comparison: benefits, applications, and tools. Zielezinski et al. (2017). Genome Biology. doi:10.1186/s13059-017-1319-7
Abstract
Alignment-free sequence analyses have been applied to problems ranging from whole-genome phylogeny to the classification of protein families, identification of horizontally transferred genes, and detection of recombined sequences. The strength of these methods makes them particularly useful for next-generation sequencing data processing and analysis. However, many researchers are unclear about how these methods work, how they compare to alignment-based methods, and what their potential is for use for their research. We address these questions and provide a guide to the currently available alignment-free sequence analysis tools.
Ondov et al. (2016) — Mash: genome distances with MinHash.
Mash: fast genome and metagenome distance estimation using MinHash. Ondov et al. (2016). Genome Biology. doi:10.1186/s13059-016-0997-x
Abstract
Mash extends the MinHash dimensionality-reduction technique to include a pairwise mutation distance and P value significance test, enabling the efficient clustering and search of massive sequence collections. Mash reduces large sequences and sequence sets to small, representative sketches, from which global mutation distances can be rapidly estimated. We demonstrate several use cases, including the clustering of all 54,118 NCBI RefSeq genomes in 33 CPU h; real-time database search using assembled or unassembled Illumina, Pacific Biosciences, and Oxford Nanopore data; and the scalable clustering of hundreds of metagenomic samples by composition. Mash is freely released under a BSD license ( https://github.com/marbl/mash ).
Wood and Salzberg (2014) — Kraken: metagenomic classification with \(k\)-mers.
Kraken: ultrafast metagenomic sequence classification using exact alignments. Wood & Salzberg (2014). Genome Biology. doi:10.1186/gb-2014-15-3-r46
Abstract
Kraken is an ultrafast and highly accurate program for assigning taxonomic labels to metagenomic DNA sequences. Previous programs designed for this task have been relatively slow and computationally expensive, forcing researchers to use faster abundance estimation programs, which only classify small subsets of metagenomic data. Using exact alignment of k-mers, Kraken achieves classification accuracy comparable to the fastest BLAST program. In its fastest mode, Kraken classifies 100 base pair reads at a rate of over 4.1 million reads per minute, 909 times faster than Megablast and 11 times faster than the abundance estimation program MetaPhlAn. Kraken is available at http://ccb.jhu.edu/software/kraken/.
C. Jain et al. (2018) — ANI and prokaryotic species boundaries.
High throughput ANI analysis of 90K prokaryotic genomes reveals clear species boundaries. Jain et al. (2018). Nature Communications. doi:10.1038/s41467-018-07641-9
Abstract
A fundamental question in microbiology is whether there is continuum of genetic diversity among genomes, or clear species boundaries prevail instead. Whole-genome similarity metrics such as Average Nucleotide Identity (ANI) help address this question by facilitating high resolution taxonomic analysis of thousands of genomes from diverse phylogenetic lineages. To scale to available genomes and beyond, we present FastANI, a new method to estimate ANI using alignment-free approximate sequence mapping. FastANI is accurate for both finished and draft genomes, and is up to three orders of magnitude faster compared to alignment-based approaches. We leverage FastANI to compute pairwise ANI values among all prokaryotic genomes available in the NCBI database. Our results reveal clear genetic discontinuity, with 99.8% of the total 8 billion genome pairs analyzed conforming to >95% intra-species and <83% inter-species ANI values. This discontinuity is manifested with or without the most frequently sequenced species, and is robust to historic additions in the genome databases.
B.1.9 Benchmarking
Brenner et al. (1998) — assessing sequence comparison methods with structurally identified distant relationships.
Assessing sequence comparison methods with reliable structurally identified distant evolutionary relationships. Brenner et al. (1998). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.95.11.6073
Abstract
Pairwise sequence comparison methods have been assessed using proteins whose relationships are known reliably from their structures and functions, as described in the SCOP database [Murzin, A. G., Brenner, S. E., Hubbard, T. & Chothia C. (1995) J. Mol. Biol. 247, 536-540]. The evaluation tested the programs BLAST [Altschul, S. F., Gish, W., Miller, W., Myers, E. W. & Lipman, D. J. (1990). J. Mol. Biol. 215, 403-410], WU-BLAST2 [Altschul, S. F. & Gish, W. (1996) Methods Enzymol. 266, 460-480], FASTA [Pearson, W. R. & Lipman, D. J. (1988) Proc. Natl. Acad. Sci. USA 85, 2444-2448], and SSEARCH [Smith, T. F. & Waterman, M. S. (1981) J. Mol. Biol. 147, 195-197] and their scoring schemes. The error rate of all algorithms is greatly reduced by using statistical scores to evaluate matches rather than percentage identity or raw scores. The E-value statistical scores of SSEARCH and FASTA are reliable: the number of false positives found in our tests agrees well with the scores reported. However, the P-values reported by BLAST and WU-BLAST2 exaggerate significance by orders of magnitude. SSEARCH, FASTA ktup = 1, and WU-BLAST2 perform best, and they are capable of detecting almost all relationships between proteins whose sequence identities are >30%. For more distantly related proteins, they do much less well; only one-half of the relationships between proteins with 20-30% identity are found. Because many homologs have low sequence similarity, most distant relationships cannot be detected by any pairwise comparison method; however, those which are identified may be used with confidence.
Lindahl and Elofsson (2000) — identification of related proteins at family, superfamily and fold level.
Identification of related proteins on family, superfamily and fold level. Lindahl & Elofsson (2000). Journal of Molecular Biology. doi:10.1006/jmbi.1999.3377
Abstract
Proteins might have considerable structural similarities even when no evolutionary relationship of their sequences can be detected. This property is often referred to as the proteins sharing only a "fold". Of course, there are also sequences of common origin in each fold, called a "superfamily", and in them groups of sequences with clear similarities, designated "family". Developing algorithms to reliably identify proteins related at any level is one of the most important challenges in the fast growing field of bioinformatics today. However, it is not at all certain that a method proficient at finding sequence similarities performs well at the other levels, or vice versa.Here, we have compared the performance of various search methods on these different levels of similarity. As expected, we show that it becomes much harder to detect proteins as their sequences diverge. For family related sequences the best method gets 75% of the top hits correct. When the sequences differ but the proteins belong to the same superfamily this drops to 29%, and in the case of proteins with only fold similarity it is as low as 15%. We have made a more complete analysis of the performance of different algorithms than earlier studies, also including threading methods in the comparison. Using this method a more detailed picture emerges, showing multiple sequence information to improve detection on the two closer levels of relationship. We have also compared the different methods of including this information in prediction algorithms. For lower specificities, the best scheme to use is a linking method connecting proteins through an intermediate hit. For higher specificities, better performance is obtained by PSI-BLAST and some procedures using hidden Markov models. We also show that a threading method, THREADER, performs significantly better than any other method at fold recognition.
Gribskov and Robinson (1996) — ROC analysis of sequence matching.
Use of receiver operating characteristic (ROC) analysis to evaluate sequence matching. Gribskov & Robinson (1996). Computers & Chemistry. doi:10.1016/s0097-8485(96)80004-0
Abstract
In this paper, we borrow the idea of the receiver operating characteristic (ROC) from clinical medicine and demonstrate its application to sequence comparison. The ROC includes elements of both sensitivity and specificity, and is a quantitative measure of the usefulness of a diagnostic. The ROC is used in this work to investigate the effects of scoring table and gap penalties on database searches. Studies on three families of proteins, 4Fe-4S ferredoxins, lysR bacterial regulatory proteins, and bacterial RNA polymerase sigma-factors lead to the following conclusions: sequence families are quite idiosyncratic, but the best PAM distance for database searches using the Smith-Waterman method is somewhat larger than predicted by theoretical methods, about 200 PAM. The length independent gap penalty (gap initiation penalty) is quite important, but shows a broad peak at values of about 20-24. The length dependent gap penalty (gap extension penalty) is almost irrelevant suggesting that successful database searches rely only to a limited degree on gapped alignments. Taken together, these observations lead to the conclusion that the optimal conditions for alignments and database searches are not, and should not be expected to be, the same.
B.1.10 Domains
Sonnhammer et al. (1997) — Pfam: domain families from seed alignments.
Pfam: A comprehensive database of protein domain families based on seed alignments. Sonnhammer et al. (1997). Proteins: Structure, Function, and Genetics. doi:10.1002/(sici)1097-0134(199707)28:3<405::aid-prot10>3.0.co;2-l
Abstract
Databases of multiple sequence alignments are a valuable aid to protein sequence classification and analysis. One of the main challenges when constructing such a database is to simultaneously satisfy the conflicting demands of completeness on the one hand and quality of alignment and domain definitions on the other. The latter properties are best dealt with by manual approaches, whereas completeness in practice is only amenable to automatic methods. Herein we present a database based on hidden Markov model profiles (HMMs), which combines high quality and completeness. Our database, Pfam, consists of parts A and B. Pfam-A is curated and contains well-characterized protein domain families with high quality alignments, which are maintained by using manually checked seed alignments and HMMs to find and align all members. Pfam-B contains sequence families that were generated automatically by applying the Domainer algorithm to cluster and align the remaining protein sequences after removal of Pfam-A domains. By using Pfam, a large number of previously unannotated proteins from the Caenorhabditis elegans genome project were classified. We have also identified many novel family memberships in known proteins, including new kazal, Fibronectin type III, and response regulator receiver domains. Pfam-A families have permanent accession numbers and form a library of HMMs available for searching and automatic annotation of new protein sequences.
Mistry et al. (2020), Paysan-Lafosse et al. (2024) — Pfam today, including Pfam-N.
Pfam: The protein families database in 2021. Mistry et al. (2020). Nucleic Acids Research. doi:10.1093/nar/gkaa913
Abstract
The Pfam database is a widely used resource for classifying protein sequences into families and domains. Since Pfam was last described in this journal, over 350 new families have been added in Pfam 33.1 and numerous improvements have been made to existing entries. To facilitate research on COVID-19, we have revised the Pfam entries that cover the SARS-CoV-2 proteome, and built new entries for regions that were not covered by Pfam. We have reintroduced Pfam-B which provides an automatically generated supplement to Pfam and contains 136 730 novel clusters of sequences that are not yet matched by a Pfam family. The new Pfam-B is based on a clustering by the MMseqs2 software. We have compared all of the regions in the RepeatsDB to those in Pfam and have started to use the results to build and refine Pfam repeat families. Pfam is freely available for browsing and download at http://pfam.xfam.org/.
The Pfam protein families database: embracing AI/ML. Paysan-Lafosse et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae997
Abstract
The Pfam protein families database is a comprehensive collection of protein domains and families used for genome annotation and protein structure and function analysis (https://www.ebi.ac.uk/interpro/). This update describes major developments in Pfam since 2020, including decommissioning the Pfam website and integration with InterPro, harmonization with the ECOD structural classification, and expanded curation of metagenomic, microprotein and repeat-containing families. We highlight how AlphaFold structure predictions are being leveraged to refine domain boundaries and identify new domains. New families discovered through large-scale sequence similarity analysis of AlphaFold models are described. We also detail the development of Pfam-N, which uses deep learning to expand family coverage, achieving an 8.8% increase in UniProtKB coverage compared to standard Pfam. We discuss plans for more frequent Pfam releases integrated with InterPro and the potential for artificial intelligence to further assist curation. Despite recent advances, many protein families remain to be classified, and Pfam continues working toward comprehensive coverage of the protein universe.
Bileschi et al. (2022) — deep learning to annotate the protein universe (Pfam-N).
Using deep learning to annotate the protein universe. Bileschi et al. (2022). Nature Biotechnology. doi:10.1038/s41587-021-01179-w
Abstract
Understanding the relationship between amino acid sequence and protein function is a long-standing challenge with far-reaching scientific and translational implications. State-of-the-art alignment-based techniques cannot predict function for one-third of microbial protein sequences, hampering our ability to exploit data from diverse organisms. Here, we train deep learning models to accurately predict functional annotations for unaligned amino acid sequences across rigorous benchmark assessments built from the 17,929 families of the protein families database Pfam. The models infer known patterns of evolutionary substitutions and learn representations that accurately cluster sequences from unseen families. Combining deep models with existing methods significantly improves remote homology detection, suggesting that the deep models learn complementary information. This approach extends the coverage of Pfam by >9.5%, exceeding additions made over the last decade, and predicts function for 360 human reference proteome proteins with no previous Pfam annotation. These results suggest that deep learning models will be a core component of future protein annotation tools.
Lau et al. (2024) — TED: the Encyclopedia of Domains in the AlphaFold database.
Exploring structural diversity across the protein universe with The Encyclopedia of Domains. Lau et al. (2024). Science. doi:10.1126/science.adq4946
Abstract
The AlphaFold Protein Structure Database (AFDB) contains more than 214 million predicted protein structures composed of domains, which are independently folding units found in multiple structural and functional contexts. Identifying domains can enable many functional and evolutionary analyses but has remained challenging because of the sheer scale of the data. Using deep learning methods, we have detected and classified every domain in the AFDB, producing The Encyclopedia of Domains. We detected nearly 365 million domains, over 100 million more than can be found by sequence methods, covering more than 1 million taxa. Reassuringly, 77% of the nonredundant domains are similar to known superfamilies, greatly expanding representation of their domain space. We uncovered more than 10,000 new structural interactions between superfamilies and thousands of new folds across the fold space continuum.
B.1.11 Protein language models
Rives et al. (2021) — ESM-1b: biological structure and function emerge from unsupervised learning.
Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Rives et al. (2021). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.2016239118
Abstract
In the field of artificial intelligence, a combination of scale in data and model capacity enabled by unsupervised learning has led to major advances in representation learning and statistical generation. In the life sciences, the anticipated growth of sequencing promises unprecedented data on natural sequence diversity. Protein language modeling at the scale of evolution is a logical step toward predictive and generative artificial intelligence for biology. To this end, we use unsupervised learning to train a deep contextual language model on 86 billion amino acids across 250 million protein sequences spanning evolutionary diversity. The resulting model contains information about biological properties in its representations. The representations are learned from sequence data alone. The learned representation space has a multiscale organization reflecting structure from the level of biochemical properties of amino acids to remote homology of proteins. Information about secondary and tertiary structure is encoded in the representations and can be identified by linear projections. Representation learning produces features that generalize across a range of applications, enabling state-of-the-art supervised prediction of mutational effect and secondary structure and improving state-of-the-art features for long-range contact prediction.
Elnaggar et al. (2022) — ProtTrans.
ProtTrans: Toward Understanding the Language of Life Through Self-Supervised Learning. Elnaggar et al. (2022). IEEE Transactions on Pattern Analysis and Machine Intelligence. doi:10.1109/tpami.2021.3095381
Abstract
Computational biology and bioinformatics provide vast data gold-mines from protein sequences, ideal for Language Models (LMs) taken from Natural Language Processing (NLP). These LMs reach for new prediction frontiers at low inference costs. Here, we trained two auto-regressive models (Transformer-XL, XLNet) and four auto-encoder models (BERT, Albert, Electra, T5) on data from UniRef and BFD containing up to 393 billion amino acids. The protein LMs (pLMs) were trained on the Summit supercomputer using 5616 GPUs and TPU Pod up-to 1024 cores. Dimensionality reduction revealed that the raw pLM-embeddings from unlabeled data captured some biophysical features of protein sequences. We validated the advantage of using the embeddings as exclusive input for several subsequent tasks: (1) a per-residue (per-token) prediction of protein secondary structure (3-state accuracy Q3=81%-87%); (2) per-protein (pooling) predictions of protein sub-cellular location (ten-state accuracy: Q10=81%) and membrane versus water-soluble (2-state accuracy Q2=91%). For secondary structure, the most informative embeddings (ProtT5) for the first time outperformed the state-of-the-art without multiple sequence alignments (MSAs) or evolutionary information thereby bypassing expensive database searches. Taken together, the results implied that pLMs learned some of the grammar of the language of life. All our models are available through https://github.com/agemagician/ProtTrans.
Lin et al. (2023) — ESM-2 and ESMFold.
Evolutionary-scale prediction of atomic-level protein structure with a language model. Lin et al. (2023). Science. doi:10.1126/science.ade2574
Abstract
Recent advances in machine learning have leveraged evolutionary information in multiple sequence alignments to predict protein structure. We demonstrate direct inference of full atomic-level protein structure from primary sequence using a large language model. As language models of protein sequences are scaled up to 15 billion parameters, an atomic-resolution picture of protein structure emerges in the learned representations. This results in an order-of-magnitude acceleration of high-resolution structure prediction, which enables large-scale structural characterization of metagenomic proteins. We apply this capability to construct the ESM Metagenomic Atlas by predicting structures for >617 million metagenomic protein sequences, including >225 million that are predicted with high confidence, which gives a view into the vast breadth and diversity of natural proteins.
Hamamsy et al. (2023) — TM-Vec: remote homology from sequence embeddings.
Protein remote homology detection and structural alignment using deep learning. Hamamsy et al. (2023). Nature Biotechnology. doi:10.1038/s41587-023-01917-2
Abstract
Exploiting sequence-structure-function relationships in biotechnology requires improved methods for aligning proteins that have low sequence similarity to previously annotated proteins. We develop two deep learning methods to address this gap, TM-Vec and DeepBLAST. TM-Vec allows searching for structure-structure similarities in large sequence databases. It is trained to accurately predict TM-scores as a metric of structural similarity directly from sequence pairs without the need for intermediate computation or solution of structures. Once structurally similar proteins have been identified, DeepBLAST can structurally align proteins using only sequence information by identifying structurally homologous regions between proteins. It outperforms traditional sequence alignment methods and performs similarly to structure-based alignment methods. We show the merits of TM-Vec and DeepBLAST on a variety of datasets, including better identification of remotely homologous proteins compared with state-of-the-art sequence alignment and structure prediction methods.
Hong et al. (2024) — dense retrieval for fast, sensitive homologue detection.
Fast, sensitive detection of protein homologs using deep dense retrieval. Hong et al. (2024). Nature Biotechnology. doi:10.1038/s41587-024-02353-6
Abstract
The identification of protein homologs in large databases using conventional methods, such as protein sequence comparison, often misses remote homologs. Here, we offer an ultrafast, highly sensitive method, dense homolog retriever (DHR), for detecting homologs on the basis of a protein language model and dense retrieval techniques. Its dual-encoder architecture generates different embeddings for the same protein sequence and easily locates homologs by comparing these representations. Its alignment-free nature improves speed and the protein language model incorporates rich evolutionary and structural information within DHR embeddings. DHR achieves a >10% increase in sensitivity compared to previous methods and a >56% increase in sensitivity at the superfamily level for samples that are challenging to identify using alignment-based approaches. It is up to 22 times faster than traditional methods such as PSI-BLAST and DIAMOND and up to 28,700 times faster than HMMER. The new remote homologs exclusively found by DHR are useful for revealing connections between well-characterized proteins and improving our knowledge of protein evolution, structure and function.
Kaminski et al. (2023) — pLM-BLAST.
pLM-BLAST: distant homology detection based on direct comparison of sequence representations from protein language models. Kaminski et al. (2023). Bioinformatics. doi:10.1093/bioinformatics/btad579
Abstract
Motivation The detection of homology through sequence comparison is a typical first step in the study of protein function and evolution. In this work, we explore the applicability of protein language models to this task. Results We introduce pLM-BLAST, a tool inspired by BLAST, that detects distant homology by comparing single-sequence representations (embeddings) derived from a protein language model, ProtT5. Our benchmarks reveal that pLM-BLAST maintains a level of accuracy on par with HHsearch for both highly similar sequences (with >50% identity) and markedly divergent sequences (with Availability and implementation pLM-BLAST is accessible via the MPI Bioinformatics Toolkit as a web server for searching precomputed databases (https://toolkit.tuebingen.mpg.de/tools/plmblast). It is also available as a standalone tool for building custom databases and performing batch searches (https://github.com/labstructbioinf/pLM-BLAST).
Liu et al. (2024) — PLMSearch.
PLMSearch: Protein language model powers accurate and fast sequence search for remote homology. Liu et al. (2024). Nature Communications. doi:10.1038/s41467-024-46808-5
Abstract
Homologous protein search is one of the most commonly used methods for protein annotation and analysis. Compared to structure search, detecting distant evolutionary relationships from sequences alone remains challenging. Here we propose PLMSearch (Protein Language Model), a homologous protein search method with only sequences as input. PLMSearch uses deep representations from a pre-trained protein language model and trains the similarity prediction model with a large number of real structure similarity. This enables PLMSearch to capture the remote homology information concealed behind the sequences. Extensive experimental results show that PLMSearch can search millions of query-target protein pairs in seconds like MMseqs2 while increasing the sensitivity by more than threefold, and is comparable to state-of-the-art structure search methods. In particular, unlike traditional sequence search methods, PLMSearch can recall most remote homology pairs with dissimilar sequences but similar structures. PLMSearch is freely available at https://dmiip.sjtu.edu.cn/PLMSearch .
B.2 Databases
Further papers for Chapter 2.
In the HTML version, click Abstract to show the abstract.
B.2.1 General
Rigden and Fernández (2025) — the 2026 NAR database issue.
The 2026 Nucleic Acids Research database issue and the online molecular biology database collection. Rigden & Fernández (2025). Nucleic Acids Research. doi:10.1093/nar/gkaf1427
Abstract
The 2026 Nucleic Acids Research database issue has 182 papers from across biology and neighbouring fields. Eighty-four of these papers describe new databases, while 86 are updates on databases that have previously appeared here. Twelve more papers cover databases most recently published elsewhere. New nucleic acid databases include NapRNAdb for noncapped RNA and GlycoRNAdb. Protein structure is covered by updates from wwPDB members and the AlphaFold Database; SMART, PROSITE, and eggNOG cover domains and families. The Open Enzyme Database and QSproteome are new community-orientated initiatives. JoGo covers hierarchically named and contextualised human haplotypes in the issue's first Breakthrough paper; So3D provides genuinely 3D spatial transcriptomics in the other. Foundational databases Genenames.org and Gene Ontology also provide updates. The Database Issue is freely available on the Nucleic Acids Research website (https://academic.oup.com/nar). At the NAR online Molecular Biology Database Collection (http://www.oxfordjournals.org/nar/database/c/), over the past year, 899 entries were reviewed, 96 new resources added, and 319 discontinued URLs removed, bringing the total number of databases to 2173.
Sayers et al. (2024) — database resources of NCBI.
Database resources of the National Center for Biotechnology Information in 2025. Sayers et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae979
Abstract
The National Center for Biotechnology Information (NCBI) provides online information resources for biology, including the GenBank® nucleic acid sequence repository and the PubMed® repository of citations and abstracts published in life science journals. NCBI provides search and retrieval operations for most of these data from 31 distinct repositories and knowledgebases. The E-utilities serve as the programming interface for most of these. Resources receiving significant updates in the past year include PubMed, PubMed Central, Bookshelf, the NIH Comparative Genomics Resource, BLAST, Sequence Read Archive, Taxonomy, iCn3D, Conserved Domain Database, Pathogen Detection, antimicrobial resistance resources and PubChem. These resources can be accessed through the NCBI home page at https://www.ncbi.nlm.nih.gov.
B.2.2 Sequences and genomes
Arita et al. (2020) — the International Nucleotide Sequence Database Collaboration.
The international nucleotide sequence database collaboration. Arita et al. (2020). Nucleic Acids Research. doi:10.1093/nar/gkaa967
Abstract
The International Nucleotide Sequence Database Collaboration (INSDC; http://www.insdc.org/) has been the core infrastructure for collecting and providing nucleotide sequence data and metadata for >30 years. Three partner organizations, the DNA Data Bank of Japan (DDBJ) at the National Institute of Genetics in Mishima, Japan; the European Nucleotide Archive (ENA) at the European Molecular Biology Laboratory's European Bioinformatics Institute (EMBL-EBI) in Hinxton, UK; and GenBank at National Center for Biotechnology Information (NCBI), National Library of Medicine, National Institutes of Health in Bethesda, Maryland, USA have been collaboratively maintaining the INSDC for the benefit of not only science but all types of community worldwide.
O’Leary et al. (2015) — RefSeq.
Reference sequence (RefSeq) database at NCBI: current status, taxonomic expansion, and functional annotation. O’Leary et al. (2015). Nucleic Acids Research. doi:10.1093/nar/gkv1189
Abstract
The RefSeq project at the National Center for Biotechnology Information (NCBI) maintains and curates a publicly available database of annotated genomic, transcript, and protein sequence records (http://www.ncbi.nlm.nih.gov/refseq/). The RefSeq project leverages the data submitted to the International Nucleotide Sequence Database Collaboration (INSDC) against a combination of computation, manual curation, and collaboration to produce a standard set of stable, non-redundant reference sequences. The RefSeq project augments these reference sequences with current knowledge including publications, functional features and informative nomenclature. The database currently represents sequences from more than 55,000 organisms (>4800 viruses, >40,000 prokaryotes and >10,000 eukaryotes; RefSeq release 71), ranging from a single record to complete genomes. This paper summarizes the current status of the viral, prokaryotic, and eukaryotic branches of the RefSeq project, reports on improvements to data access and details efforts to further expand the taxonomic representation of the collection. We also highlight diverse functional curation initiatives that support multiple uses of RefSeq data including taxonomic validation, genome annotation, comparative genomics, and clinical testing. We summarize our approach to utilizing available RNA-Seq and other data types in our manual curation process for vertebrate, plant, and other species, and describe a new direction for prokaryotic genomes and protein name management.
Suzek et al. (2014) — UniRef clusters.
UniRef clusters: a comprehensive and scalable alternative for improving sequence similarity searches. Suzek et al. (2014). Bioinformatics. doi:10.1093/bioinformatics/btu739
Abstract
Motivation UniRef databases provide full-scale clustering of UniProtKB sequences and are utilized for a broad range of applications, particularly similarity-based functional annotation. Non-redundancy and intra-cluster homogeneity in UniRef were recently improved by adding a sequence length overlap threshold. Our hypothesis is that these improvements would enhance the speed and sensitivity of similarity searches and improve the consistency of annotation within clusters. Results Intra-cluster molecular function consistency was examined by analysis of Gene Ontology terms. Results show that UniRef clusters bring together proteins of identical molecular function in more than 97% of the clusters, implying that clusters are useful for annotation and can also be used to detect annotation inconsistencies. To examine coverage in similarity results, BLASTP searches against UniRef50 followed by expansion of the hit lists with cluster members demonstrated advantages compared with searches against UniProtKB sequences; the searches are concise (∼7 times shorter hit list before expansion), faster (∼6 times) and more sensitive in detection of remote similarities (>96% recall at e-value <0.0001). Our results support the use of UniRef clusters as a comprehensive and scalable alternative to native sequence databases for similarity searches and reinforces its reliability for use in functional annotation.
Richardson et al. (2022) — MGnify.
MGnify: the microbiome sequence data analysis resource in 2023. Richardson et al. (2022). Nucleic Acids Research. doi:10.1093/nar/gkac1080
Abstract
The MGnify platform (https://www.ebi.ac.uk/metagenomics) facilitates the assembly, analysis and archiving of microbiome-derived nucleic acid sequences. The platform provides access to taxonomic assignments and functional annotations for nearly half a million analyses covering metabarcoding, metatranscriptomic, and metagenomic datasets, which are derived from a wide range of different environments. Over the past 3 years, MGnify has not only grown in terms of the number of datasets contained but also increased the breadth of analyses provided, such as the analysis of long-read sequences. The MGnify protein database now exceeds 2.4 billion non-redundant sequences predicted from metagenomic assemblies. This collection is now organised into a relational database making it possible to understand the genomic context of the protein through navigation back to the source assembly and sample metadata, marking a major improvement. To extend beyond the functional annotations already provided in MGnify, we have applied deep learning-based annotation methods. The technology underlying MGnify's Application Programming Interface (API) and website has been upgraded, and we have enabled the ability to perform downstream analysis of the MGnify data through the introduction of a coupled Jupyter Lab environment.
Dyer et al. (2024) — Ensembl 2025.
Ensembl 2025. Dyer et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae1071
Abstract
Ensembl (www.ensembl.org) is an open platform integrating publicly available genomics data across the tree of life with a focus on eukaryotic species related to human health, agriculture and biodiversity. This year has seen a continued expansion in the number of species represented, with >4800 eukaryotic and >31 300 prokaryotic genomes available. The new Ensembl site, currently in beta, has continued to develop, currently holding >2700 eukaryotic genome assemblies. The new site provides genome, gene, transcript, homology and variation views, and will replace the current Rapid Release site; this represents a key step towards provision of a single integrated Ensembl site. Additional activities have included developing improved regulatory annotation for human, mouse and agricultural species, and expanding the Ensembl Variant Effect Predictor tool. To learn more about Ensembl, help and documentation are available along with an extensive training program that can be accessed via our training pages.
Perez et al. (2024) — the UCSC Genome Browser.
The UCSC Genome Browser database: 2025 update. Perez et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae974
Abstract
The UCSC Genome Browser (https://genome.ucsc.edu) is a widely utilized web-based tool for visualization and analysis of genomic data, encompassing over 4000 assemblies from diverse organisms. Since its release in 2001, it has become an essential resource for genomics and bioinformatics research. Annotation data available on Genome Browser includes both internally created and maintained tracks as well as custom tracks and track hubs provided by the research community. This last year's updates include over 25 new annotation tracks such as the gnomAD 4.1 track on the human GRCh38/hg38 assembly, the addition of three new public hubs, and significant expansions to the Genome Archive[GenArk) system for interacting with the enormous variety of assemblies. We have also made improvements to our interface, including updates to the browser graphic page, such as a new popup dialog feature that now displays item details without requiring navigation away from the main Genome Browser page. GenePred tracks have been upgraded with right-click options for zooming and precise navigation, along with enhanced mouseOver functions. Additional improvements include a new grouping feature for track hubs and hub description info links. A new tutorial focusing on Clinical Genetics has also been added to the UCSC Genome Browser.
B.2.3 Protein families and structures
Sonnhammer et al. (1997), Mistry et al. (2020) — Pfam.
Pfam: A comprehensive database of protein domain families based on seed alignments. Sonnhammer et al. (1997). Proteins: Structure, Function, and Genetics. doi:10.1002/(sici)1097-0134(199707)28:3<405::aid-prot10>3.0.co;2-l
Abstract
Databases of multiple sequence alignments are a valuable aid to protein sequence classification and analysis. One of the main challenges when constructing such a database is to simultaneously satisfy the conflicting demands of completeness on the one hand and quality of alignment and domain definitions on the other. The latter properties are best dealt with by manual approaches, whereas completeness in practice is only amenable to automatic methods. Herein we present a database based on hidden Markov model profiles (HMMs), which combines high quality and completeness. Our database, Pfam, consists of parts A and B. Pfam-A is curated and contains well-characterized protein domain families with high quality alignments, which are maintained by using manually checked seed alignments and HMMs to find and align all members. Pfam-B contains sequence families that were generated automatically by applying the Domainer algorithm to cluster and align the remaining protein sequences after removal of Pfam-A domains. By using Pfam, a large number of previously unannotated proteins from the Caenorhabditis elegans genome project were classified. We have also identified many novel family memberships in known proteins, including new kazal, Fibronectin type III, and response regulator receiver domains. Pfam-A families have permanent accession numbers and form a library of HMMs available for searching and automatic annotation of new protein sequences.
Pfam: The protein families database in 2021. Mistry et al. (2020). Nucleic Acids Research. doi:10.1093/nar/gkaa913
Abstract
The Pfam database is a widely used resource for classifying protein sequences into families and domains. Since Pfam was last described in this journal, over 350 new families have been added in Pfam 33.1 and numerous improvements have been made to existing entries. To facilitate research on COVID-19, we have revised the Pfam entries that cover the SARS-CoV-2 proteome, and built new entries for regions that were not covered by Pfam. We have reintroduced Pfam-B which provides an automatically generated supplement to Pfam and contains 136 730 novel clusters of sequences that are not yet matched by a Pfam family. The new Pfam-B is based on a clustering by the MMseqs2 software. We have compared all of the regions in the RepeatsDB to those in Pfam and have started to use the results to build and refine Pfam repeat families. Pfam is freely available for browsing and download at http://pfam.xfam.org/.
Berman et al. (2000) — the Protein Data Bank.
The Protein Data Bank. Berman et al. (2000). Nucleic Acids Research. doi:10.1093/nar/28.1.235
Abstract
The Protein Data Bank (PDB; http://www.rcsb.org/pdb/ ) is the single worldwide archive of structural data of biological macromolecules. This paper describes the goals of the PDB, the systems in place for data deposition and access, how to obtain further information, and near-term plans for the future development of the resource.
Turner et al. (2023) — the Electron Microscopy Data Bank.
EMDB—the Electron Microscopy Data Bank. Turner et al. (2023). Nucleic Acids Research. doi:10.1093/nar/gkad1019
Abstract
The Electron Microscopy Data Bank (EMDB) is the global public archive of three-dimensional electron microscopy (3DEM) maps of biological specimens derived from transmission electron microscopy experiments. As of 2021, EMDB is managed by the Worldwide Protein Data Bank consortium (wwPDB; wwpdb.org) as a wwPDB Core Archive, and the EMDB team is a core member of the consortium. Today, EMDB houses over 30 000 entries with maps containing macromolecules, complexes, viruses, organelles and cells. Herein, we provide an overview of the rapidly growing EMDB archive, including its current holdings, recent updates, and future plans.
Murzin et al. (1995), Sillitoe et al. (2020), Cheng et al. (2014), Lau et al. (2024) — SCOP, CATH, ECOD and TED.
SCOP: A structural classification of proteins database for the investigation of sequences and structures. Murzin et al. (1995). Journal of Molecular Biology. doi:10.1016/s0022-2836(05)80134-2
CATH: increased structural coverage of functional space. Sillitoe et al. (2020). Nucleic Acids Research. doi:10.1093/nar/gkaa1079
Abstract
CATH (https://www.cathdb.info) identifies domains in protein structures from wwPDB and classifies these into evolutionary superfamilies, thereby providing structural and functional annotations. There are two levels: CATH-B, a daily snapshot of the latest domain structures and superfamily assignments, and CATH+, with additional derived data, such as predicted sequence domains, and functionally coherent sequence subsets (Functional Families or FunFams). The latest CATH+ release, version 4.3, significantly increases coverage of structural and sequence data, with an addition of 65,351 fully-classified domains structures (+15%), providing 500 238 structural domains, and 151 million predicted sequence domains (+59%) assigned to 5481 superfamilies. The FunFam generation pipeline has been re-engineered to cope with the increased influx of data. Three times more sequences are captured in FunFams, with a concomitant increase in functional purity, information content and structural coverage. FunFam expansion increases the structural annotations provided for experimental GO terms (+59%). We also present CATH-FunVar web-pages displaying variations in protein sequences and their proximity to known or predicted functional sites. We present two case studies (1) putative cancer drivers and (2) SARS-CoV-2 proteins. Finally, we have improved links to and from CATH including SCOP, InterPro, Aquaria and 2DProt.
ECOD: An Evolutionary Classification of Protein Domains. Cheng et al. (2014). PLoS Computational Biology. doi:10.1371/journal.pcbi.1003926
Abstract
Understanding the evolution of a protein, including both close and distant relationships, often reveals insight into its structure and function. Fast and easy access to such up-to-date information facilitates research. We have developed a hierarchical evolutionary classification of all proteins with experimentally determined spatial structures, and presented it as an interactive and updatable online database. ECOD (Evolutionary Classification of protein Domains) is distinct from other structural classifications in that it groups domains primarily by evolutionary relationships (homology), rather than topology (or "fold"). This distinction highlights cases of homology between domains of differing topology to aid in understanding of protein structure evolution. ECOD uniquely emphasizes distantly related homologs that are difficult to detect, and thus catalogs the largest number of evolutionary links among structural domain classifications. Placing distant homologs together underscores the ancestral similarities of these proteins and draws attention to the most important regions of sequence and structure, as well as conserved functional sites. ECOD also recognizes closer sequence-based relationships between protein domains. Currently, approximately 100,000 protein structures are classified in ECOD into 9,000 sequence families clustered into close to 2,000 evolutionary groups. The classification is assisted by an automated pipeline that quickly and consistently classifies weekly releases of PDB structures and allows for continual updates. This synchronization with PDB uniquely distinguishes ECOD among all protein classifications. Finally, we present several case studies of homologous proteins not recorded in other classifications, illustrating the potential of how ECOD can be used to further biological and evolutionary studies.
Exploring structural diversity across the protein universe with The Encyclopedia of Domains. Lau et al. (2024). Science. doi:10.1126/science.adq4946
Abstract
The AlphaFold Protein Structure Database (AFDB) contains more than 214 million predicted protein structures composed of domains, which are independently folding units found in multiple structural and functional contexts. Identifying domains can enable many functional and evolutionary analyses but has remained challenging because of the sheer scale of the data. Using deep learning methods, we have detected and classified every domain in the AFDB, producing The Encyclopedia of Domains. We detected nearly 365 million domains, over 100 million more than can be found by sequence methods, covering more than 1 million taxa. Reassuringly, 77% of the nonredundant domains are similar to known superfamilies, greatly expanding representation of their domain space. We uncovered more than 10,000 new structural interactions between superfamilies and thousands of new folds across the fold space continuum.
Varadi et al. (2021), Varadi et al. (2023) — the AlphaFold database.
AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Varadi et al. (2021). Nucleic Acids Research. doi:10.1093/nar/gkab1061
Abstract
The AlphaFold Protein Structure Database (AlphaFold DB, https://alphafold.ebi.ac.uk) is an openly accessible, extensive database of high-accuracy protein-structure predictions. Powered by AlphaFold v2.0 of DeepMind, it has enabled an unprecedented expansion of the structural coverage of the known protein-sequence space. AlphaFold DB provides programmatic access to and interactive visualization of predicted atomic coordinates, per-residue and pairwise model-confidence estimates and predicted aligned errors. The initial release of AlphaFold DB contains over 360,000 predicted structures across 21 model-organism proteomes, which will soon be expanded to cover most of the (over 100 million) representative sequences from the UniRef90 data set.
AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences. Varadi et al. (2023). Nucleic Acids Research. doi:10.1093/nar/gkad1011
Abstract
The AlphaFold Database Protein Structure Database (AlphaFold DB, https://alphafold.ebi.ac.uk) has significantly impacted structural biology by amassing over 214 million predicted protein structures, expanding from the initial 300k structures released in 2021. Enabled by the groundbreaking AlphaFold2 artificial intelligence (AI) system, the predictions archived in AlphaFold DB have been integrated into primary data resources such as PDB, UniProt, Ensembl, InterPro and MobiDB. Our manuscript details subsequent enhancements in data archiving, covering successive releases encompassing model organisms, global health proteomes, Swiss-Prot integration, and a host of curated protein datasets. We detail the data access mechanisms of AlphaFold DB, from direct file access via FTP to advanced queries using Google Cloud Public Datasets and the programmatic access endpoints of the database. We also discuss the improvements and services added since its initial release, including enhancements to the Predicted Aligned Error viewer, customisation options for the 3D viewer, and improvements in the search engine of AlphaFold DB.
Lin et al. (2023) — ESMFold and the ESM Metagenomic Atlas.
Evolutionary-scale prediction of atomic-level protein structure with a language model. Lin et al. (2023). Science. doi:10.1126/science.ade2574
Abstract
Recent advances in machine learning have leveraged evolutionary information in multiple sequence alignments to predict protein structure. We demonstrate direct inference of full atomic-level protein structure from primary sequence using a large language model. As language models of protein sequences are scaled up to 15 billion parameters, an atomic-resolution picture of protein structure emerges in the learned representations. This results in an order-of-magnitude acceleration of high-resolution structure prediction, which enables large-scale structural characterization of metagenomic proteins. We apply this capability to construct the ESM Metagenomic Atlas by predicting structures for >617 million metagenomic protein sequences, including >225 million that are predicted with high confidence, which gives a view into the vast breadth and diversity of natural proteins.
Kim et al. (2024) — BFVD, predicted viral protein structures.
BFVD—a large repository of predicted viral protein structures. Kim et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae1119
Abstract
The AlphaFold Protein Structure Database (AFDB) is the largest repository of accurately predicted structures with taxonomic labels. Despite providing predictions for over 214 million UniProt entries, the AFDB does not cover viral sequences, severely limiting their study. To address this, we created the Big Fantastic Virus Database (BFVD), a repository of 351 242 protein structures predicted by applying ColabFold to the viral sequence representatives of the UniRef30 clusters. By utilizing homology searches across two petabases of assembled sequencing data, we improved 36% of these structure predictions beyond ColabFold's initial results. BFVD holds a unique repertoire of protein structures as over 62% of its entries show no or low structural similarity to existing repositories. We demonstrate how a substantial fraction of bacteriophage proteins, which remained unannotated based on their sequences, can be matched with similar structures from BFVD. In that, BFVD is on par with the AFDB, while holding nearly three orders of magnitude fewer structures. BFVD is an important virus-specific expansion to protein structure repositories, offering new opportunities to advance viral research. BFVD can be freely downloaded at bfvd.steineggerlab.workers.dev and queried using Foldseek and UniProt labels at bfvd.foldseek.com.
B.2.4 Function and pathways
Ashburner et al. (2000) — the original Gene Ontology paper.
Gene Ontology: tool for the unification of biology. Ashburner et al. (2000). Nature Genetics. doi:10.1038/75556
Kanehisa et al. (2024) — KEGG.
KEGG: biological systems database as a model of the real world. Kanehisa et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae909
Abstract
KEGG (https://www.kegg.jp/) is a database resource for representation and analysis of biological systems. Pathway maps are the primary dataset in KEGG representing systemic functions of the cell and the organism in terms of molecular interaction and reaction networks. The KEGG Orthology (KO) system is a mechanism for linking genes and proteins to pathway maps and other molecular networks. Each KO is a generic gene identifier and each pathway map is created as a network of KO nodes. This architecture enables KEGG pathway mapping to uncover systemic features from KO assigned genomes and metagenomes. Additional roles of KOs include characterization of conserved genes and conserved units of genes in organism groups, which can be done by taxonomy mapping. A new tool has been developed for identifying conserved gene orders in chromosomes, in which gene orders are treated as sequences of KOs. Furthermore, a new dataset called VOG (virus ortholog group) is computationally generated from virus proteins and expanded to proteins of cellular organisms, allowing gene orders to be compared as VOG sequences as well. Together with these datasets and analysis tools, new types of pathway maps are being developed to present a global view of biological processes involving multiple organism groups.
Milacic et al. (2023) — Reactome.
The Reactome Pathway Knowledgebase 2024. Milacic et al. (2023). Nucleic Acids Research. doi:10.1093/nar/gkad1025
Abstract
The Reactome Knowledgebase (https://reactome.org), an Elixir and GCBR core biological data resource, provides manually curated molecular details of a broad range of normal and disease-related biological processes. Processes are annotated as an ordered network of molecular transformations in a single consistent data model. Reactome thus functions both as a digital archive of manually curated human biological processes and as a tool for discovering functional relationships in data such as gene expression profiles or somatic mutation catalogs from tumor cells. Here we review progress towards annotation of the entire human proteome, targeted annotation of disease-causing genetic variants of proteins and of small-molecule drugs in a pathway context, and towards supporting explicit annotation of cell- and tissue-specific pathways. Finally, we briefly discuss issues involved in making Reactome more fully interoperable with other related resources such as the Gene Ontology and maintaining the resulting community resource network.
De Paolis Kaluza et al. (2026) — the fifth CAFA challenge.
Advances in Protein Function Prediction from the Fifth CAFA Challenge. De Paolis Kaluza et al. (2026). bioRxiv (preprint). doi:10.64898/2026.04.27.716980
Abstract
The Critical Assessment of Functional Annotation (CAFA) is a long-standing community effort to independently assess computational methods for protein function prediction, to highlight well-performing methodologies, to identify bottlenecks in the field, and to provide a forum for the dissemination of results and exchange of ideas. In its fifth round (CAFA5) of triennial challenges, a partnership with Kaggle Inc. facilitated participation from a large community of data scientists and computational biologists through a competitive prospective challenge on the crowdsourcing platform. In this work, we present an in-depth analysis of the submitted predictions and report improvements in accuracy over all methods from the previous CAFA challenges. We further introduce a new evaluation setting for proteins with pre-existing (incomplete) annotations and identify the need for methods that better leverage existing annotations to predict those that will be discovered later. Finally, we characterize the prospective evaluation framework by examining performance on a strict set of unpublished annotations and across intermediate database releases. Our results indicate that recent developments in the field, such as the availability of protein language models and accurately predicted 3D structures, as well as the growth of experimental annotations through biocuration, have all contributed to performance improvements. 1
B.2.5 Interactions
Toro et al. (2021) — IntAct.
The IntAct database: efficient access to fine-grained molecular interaction data. del Toro et al. (2021). Nucleic Acids Research. doi:10.1093/nar/gkab1006
Abstract
The IntAct molecular interaction database (https://www.ebi.ac.uk/intact) is a curated resource of molecular interactions, derived from the scientific literature and from direct data depositions. As of August 2021, IntAct provides more than one million binary interactions, curated by twelve global partners of the International Molecular Exchange consortium, for which the IntAct database provides a shared curation and dissemination platform. The IMEx curation policy has always emphasised a fine-grained data and curation model, aiming to capture the relevant experimental detail essential for the interpretation of the provided molecular interaction data. Here, we present recent curation focus and progress, as well as a completely redeveloped website which presents IntAct data in a much more user-friendly and detailed way.
Oughtred et al. (2020) — BioGRID.
The BioGRID database: A comprehensive biomedical resource of curated protein, genetic, and chemical interactions. Oughtred et al. (2020). Protein Science. doi:10.1002/pro.3978
Abstract
The BioGRID (Biological General Repository for Interaction Datasets, thebiogrid.org) is an open-access database resource that houses manually curated protein and genetic interactions from multiple species including yeast, worm, fly, mouse, and human. The ~1.93 million curated interactions in BioGRID can be used to build complex networks to facilitate biomedical discoveries, particularly as related to human health and disease. All BioGRID content is curated from primary experimental evidence in the biomedical literature, and includes both focused low-throughput studies and large high-throughput datasets. BioGRID also captures protein post-translational modifications and protein or gene interactions with bioactive small molecules including many known drugs. A built-in network visualization tool combines all annotations and allows users to generate network graphs of protein, genetic and chemical interactions. In addition to general curation across species, BioGRID undertakes themed curation projects in specific aspects of cellular regulation, for example the ubiquitin-proteasome system, as well as specific disease areas, such as for the SARS-CoV-2 virus that causes COVID-19 severe acute respiratory syndrome. A recent extension of BioGRID, named the Open Repository of CRISPR Screens (ORCS, orcs.thebiogrid.org), captures single mutant phenotypes and genetic interactions from published high throughput genome-wide CRISPR/Cas9-based genetic screens. BioGRID-ORCS contains datasets for over 1,042 CRISPR screens carried out to date in human, mouse and fly cell lines. The biomedical research community can freely access all BioGRID data through the web interface, standardized file downloads, or via model organism databases and partner meta-databases.
Meldal et al. (2021) — Complex Portal.
Complex Portal 2022: new curation frontiers. Meldal et al. (2021). Nucleic Acids Research. doi:10.1093/nar/gkab991
Abstract
The Complex Portal (www.ebi.ac.uk/complexportal) is a manually curated, encyclopaedic database of macromolecular complexes with known function from a range of model organisms. It summarizes complex composition, topology and function along with links to a large range of domain-specific resources (i.e. wwPDB, EMDB and Reactome). Since the last update in 2019, we have produced a first draft complexome for Escherichia coli, maintained and updated that of Saccharomyces cerevisiae, added over 40 coronavirus complexes and increased the human complexome to over 1100 complexes that include approximately 200 complexes that act as targets for viral proteins or are part of the immune system. The display of protein features in ComplexViewer has been improved and the participant table is now colour-coordinated with the nodes in ComplexViewer. Community collaboration has expanded, for example by contributing to an analysis of putative transcription cofactors and providing data accessible to semantic web tools through Wikidata which is now populated with manually curated Complex Portal content through a new bot. Our data license is now CC0 to encourage data reuse. Users are encouraged to get in touch, provide us with feedback and send curation requests through the 'Support' link.
Buzzao et al. (2024) — FunCoup 6.
FunCoup 6: advancing functional association networks across species with directed links and improved user experience. Buzzao et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae1021
Abstract
FunCoup 6 (https://funcoup.org) represents a significant advancement in global functional association networks, aiming to provide researchers with a comprehensive view of the functional coupling interactome. This update introduces novel methodologies and integrated tools for improved network inference and analysis. Major new developments in FunCoup 6 include vastly expanding the coverage of gene regulatory links, a new framework for bin-free Bayesian training and a new website. FunCoup 6 integrates a new tool for disease and drug target module identification using the TOPAS algorithm. To expand the utility of the resource for biomedical research, it incorporates pathway enrichment analysis using the ANUBIX and EASE algorithms. The unique comparative interactomics analysis in FunCoup provides insights of network conservation, now allowing users to align orthologs only or query each species network independently. Bin-free training was applied to 23 primary species, and in addition, networks were generated for all remaining 618 species in InParanoiDB 9. Accompanying these advancements, FunCoup 6 features a new redesigned website, together with updated API functionalities, and represents a pivotal step forward in functional genomics research, offering unique capabilities for exploring the complex landscape of protein interactions.
Fischer et al. (2025) — hu.MAP 3.0.
hu.MAP3.0: atlas of human protein complexes by integration of >25,000 proteomic experiments. Fischer et al. (2025). Molecular Systems Biology. doi:10.1038/s44320-025-00121-5
Abstract
Macromolecular protein complexes carry out most cellular functions. Unfortunately, we lack the subunit composition for many human protein complexes. To address this gap we integrated >25,000 mass spectrometry experiments using a machine learning approach to identify >15,000 human protein complexes. We show our map of protein complexes is highly accurate and more comprehensive than previous maps, placing nearly 70% of human proteins into their physical contexts. We globally characterize our complexes using mass spectrometry based protein covariation data (ProteomeHD.2) and identify covarying complexes suggesting common functional associations. hu.MAP3.0 generates testable functional hypotheses for 472 uncharacterized proteins which we support using AlphaFold modeling. Additionally, we use AlphaFold modeling to identify 5871 mutually exclusive proteins in hu.MAP3.0 complexes suggesting complexes serve different functional roles depending on their subunit composition. We identify expression as the primary way cells and organisms relieve the conflict of mutually exclusive subunits. Finally, we import our complexes to EMBL-EBI's Complex Portal ( https://www.ebi.ac.uk/complexportal/home ) and provide complexes through our hu.MAP3.0 web interface ( https://humap3.proteincomplexes.org/ ). We expect our resource to be highly impactful to the broader research community.
B.2.6 Expression, proteomics and medical data
Clough et al. (2023) — GEO.
NCBI GEO: archive for gene expression and epigenomics data sets: 23-year update. Clough et al. (2023). Nucleic Acids Research. doi:10.1093/nar/gkad965
Abstract
The Gene Expression Omnibus (GEO) is an international public repository that archives gene expression and epigenomics data sets generated by next-generation sequencing and microarray technologies. Data are typically submitted to GEO by researchers in compliance with widespread journal and funder mandates to make generated data publicly accessible. The resource handles raw data files, processed data files and descriptive metadata for over 200 000 studies and 6.5 million samples, all of which are indexed, searchable and downloadable. Additionally, GEO offers web-based tools that facilitate analysis and visualization of differential gene expression. This article presents the current status and recent advancements in GEO, including the generation of consistently computed gene expression count matrices for thousands of RNA-seq studies, and new interactive graphical plots in GEO2R that help users identify differentially expressed genes and assess data set quality. The GEO repository is built and maintained by the National Center for Biotechnology Information (NCBI), a division of the National Library of Medicine (NLM), and is publicly accessible at https://www.ncbi.nlm.nih.gov/geo/.
Perez-Riverol et al. (2024) — PRIDE.
The PRIDE database at 20 years: 2025 update. Perez-Riverol et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae1011
Abstract
The PRoteomics IDEntifications (PRIDE) database (https://www.ebi.ac.uk/pride/) is the world's leading mass spectrometry (MS)-based proteomics data repository and one of the founding members of the ProteomeXchange consortium. This manuscript summarizes the developments in PRIDE resources and related tools for the last three years. The number of submitted datasets to PRIDE Archive (the archival component of PRIDE) has reached on average around 534 datasets per month. This has been possible thanks to continuous improvements in infrastructure such as a new file transfer protocol for very large datasets (Globus), a new data resubmission pipeline and an automatic dataset validation process. Additionally, we will highlight novel activities such as the availability of the PRIDE chatbot (based on the use of open-source Large Language Models), and our work to improve support for MS crosslinking datasets. Furthermore, we will describe how we have increased our efforts to reuse, reanalyze and disseminate high-quality proteomics data into added-value resources such as UniProt, Ensembl and Expression Atlas.
Uhlén et al. (2015) — the Human Protein Atlas.
Tissue-based map of the human proteome. Uhlén et al. (2015). Science. doi:10.1126/science.1260419
Abstract
Resolving the molecular details of proteome variation in the different tissues and organs of the human body will greatly increase our knowledge of human biology and disease. Here, we present a map of the human tissue proteome based on an integrated omics approach that involves quantitative transcriptomics at the tissue and organ level, combined with tissue microarray-based immunohistochemistry, to achieve spatial localization of proteins down to the single-cell level. Our tissue-based analysis detected more than 90% of the putative protein-coding genes. We used this approach to explore the human secretome, the membrane proteome, the druggable proteome, the cancer proteome, and the metabolic functions in 32 different tissues and organs. All the data are integrated in an interactive Web-based database that allows exploration of individual proteins, as well as navigation of global expression patterns, in all major tissues and organs in the human body.
Chen et al. (2023) — gnomAD.
A genomic mutational constraint map using variation in 76,156 human genomes. Chen et al. (2023). Nature. doi:10.1038/s41586-023-06045-0
Abstract
The depletion of disruptive variation caused by purifying natural selection (constraint) has been widely used to investigate protein-coding genes underlying human disorders 1-4 , but attempts to assess constraint for non-protein-coding regions have proved more difficult. Here we aggregate, process and release a dataset of 76,156 human genomes from the Genome Aggregation Database (gnomAD)-the largest public open-access human genome allele frequency reference dataset-and use it to build a genomic constraint map for the whole genome (genomic non-coding constraint of haploinsufficient variation (Gnocchi)). We present a refined mutational model that incorporates local sequence context and regional genomic features to detect depletions of variation. As expected, the average constraint for protein-coding sequences is stronger than that for non-coding regions. Within the non-coding genome, constrained regions are enriched for known regulatory elements and variants that are implicated in complex human diseases and traits, facilitating the triangulation of biological annotation, disease association and natural selection to non-coding DNA analysis. More constrained regulatory elements tend to regulate more constrained protein-coding genes, which in turn suggests that non-coding constraint can aid the identification of constrained genes that are as yet unrecognized by current gene constraint metrics. We demonstrate that this genome-wide constraint map improves the identification and interpretation of functional human genetic variation.
Landrum et al. (2017) — ClinVar.
ClinVar: improving access to variant interpretations and supporting evidence. Landrum et al. (2017). Nucleic Acids Research. doi:10.1093/nar/gkx1153
Abstract
ClinVar (https://www.ncbi.nlm.nih.gov/clinvar/) is a freely available, public archive of human genetic variants and interpretations of their significance to disease, maintained at the National Institutes of Health. Interpretations of the clinical significance of variants are submitted by clinical testing laboratories, research laboratories, expert panels and other groups. ClinVar aggregates data by variant-disease pairs, and by variant (or set of variants). Data aggregated by variant are accessible on the website, in an improved set of variant call format files and as a new comprehensive XML report. ClinVar recently started accepting submissions that are focused primarily on providing phenotypic information for individuals who have had genetic testing. Submissions may come from clinical providers providing their own interpretation of the variant ('provider interpretation') or from groups such as patient registries that primarily provide phenotypic information from patients ('phenotyping only'). ClinVar continues to make improvements to its search and retrieval functions. Several new fields are now indexed for more precise searching, and filters allow the user to narrow down a large set of search results.
Amberger et al. (2018) — OMIM.
OMIM.org: leveraging knowledge across phenotype–gene relationships. Amberger et al. (2018). Nucleic Acids Research. doi:10.1093/nar/gky1151
Abstract
For over 50 years Mendelian Inheritance in Man has chronicled the collective knowledge of the field of medical genetics. It initially cataloged the known X-linked, autosomal recessive and autosomal dominant inherited disorders, but grew to be the primary repository of curated information on both genes and genetic phenotypes and the relationships between them. Each phenotype and gene is given a separate entry assigned a stable, unique identifier. The entries contain structured summaries of new and important information based on expert review of the biomedical literature. OMIM.org provides interactive access to the knowledge repository, including genomic coordinate searches of the gene map, views of genetic heterogeneity of phenotypes in Phenotypic Series, and side-by-side comparisons of clinical synopses. OMIM.org also supports computational queries via a robust API. All entries have extensive targeted links to other genomic resources and additional references. Updates to OMIM can be found on the update list or followed through the MIMmatch service. Updated user guides and tutorials are available on the website. As of September 2018, OMIM had over 24,600 entries, and the OMIM Morbid Map Scorecard had 6,259 molecularized phenotypes connected to 3,961 genes.
Sollis et al. (2022) — the GWAS Catalog.
The NHGRI-EBI GWAS Catalog: knowledgebase and deposition resource. Sollis et al. (2022). Nucleic Acids Research. doi:10.1093/nar/gkac1010
Abstract
The NHGRI-EBI GWAS Catalog (www.ebi.ac.uk/gwas) is a FAIR knowledgebase providing detailed, structured, standardised and interoperable genome-wide association study (GWAS) data to >200 000 users per year from academic research, healthcare and industry. The Catalog contains variant-trait associations and supporting metadata for >45 000 published GWAS across >5000 human traits, and >40 000 full P-value summary statistics datasets. Content is curated from publications or acquired via author submission of prepublication summary statistics through a new submission portal and validation tool. GWAS data volume has vastly increased in recent years. We have updated our software to meet this scaling challenge and to enable rapid release of submitted summary statistics. The scope of the repository has expanded to include additional data types of high interest to the community, including sequencing-based GWAS, gene-based analyses and copy number variation analyses. Community outreach has increased the number of shared datasets from under-represented traits, e.g. cancer, and we continue to contribute to awareness of the lack of population diversity in GWAS. Interoperability of the Catalog has been enhanced through links to other resources including the Polygenic Score Catalog and the International Mouse Phenotyping Consortium, refinements to GWAS trait annotation, and the development of a standard format for GWAS data.
Weinstein et al. (2013) — TCGA Pan-Cancer.
The Cancer Genome Atlas Pan-Cancer analysis project. Weinstein et al. (2013). Nature Genetics. doi:10.1038/ng.2764
Abstract
The Cancer Genome Atlas (TCGA) Research Network has profiled and analyzed large numbers of human tumors to discover molecular aberrations at the DNA, RNA, protein and epigenetic levels. The resulting rich data provide a major opportunity to develop an integrated picture of commonalities, differences and emergent themes across tumor lineages. The Pan-Cancer initiative compares the first 12 tumor types profiled by TCGA. Analysis of the molecular aberrations and their functional roles across tumor types will teach us how to extend therapies effective in one cancer type to others with a similar genomic profile.
Sondka et al. (2023) — COSMIC.
COSMIC: a curated database of somatic variants and clinical data for cancer. Sondka et al. (2023). Nucleic Acids Research. doi:10.1093/nar/gkad986
Abstract
The Catalogue Of Somatic Mutations In Cancer (COSMIC), https://cancer.sanger.ac.uk/cosmic, is an expert-curated knowledgebase providing data on somatic variants in cancer, supported by a comprehensive suite of tools for interpreting genomic data, discerning the impact of somatic alterations on disease, and facilitating translational research. The catalogue is accessed and used by thousands of cancer researchers and clinicians daily, allowing them to quickly access information from an immense pool of data curated from over 29 thousand scientific publications and large studies. Within the last 4 years, COSMIC has substantially expanded its utility by adding new resources: the Mutational Signatures catalogue, the Cancer Mutation Census, and Actionability. To improve data accessibility and interoperability, somatic variants have received stable genomic identifiers that are associated with their genomic coordinates in GRCh37 and GRCh38, and new export files with reduced data redundancy have been made available for download.
B.3 Evolution
Further papers for Chapter 3.
Papers marked ★ were mandatory in earlier reading lists. In the HTML version, click Abstract to show the abstract.
B.3.1 The origin of life
Miller (1953), Parker et al. (2011) — the Miller–Urey experiment and its reanalysis.
A Production of Amino Acids Under Possible Primitive Earth Conditions. Miller (1953). Science. doi:10.1126/science.117.3046.528
Primordial synthesis of amines and amino acids in a 1958 Miller H 2 S-rich spark discharge experiment. Parker et al. (2011). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1019191108
Abstract
Archived samples from a previously unreported 1958 Stanley Miller electric discharge experiment containing hydrogen sulfide (H(2)S) were recently discovered and analyzed using high-performance liquid chromatography and time-of-flight mass spectrometry. We report here the detection and quantification of primary amine-containing compounds in the original sample residues, which were produced via spark discharge using a gaseous mixture of H(2)S, CH(4), NH(3), and CO(2). A total of 23 amino acids and 4 amines, including 7 organosulfur compounds, were detected in these samples. The major amino acids with chiral centers are racemic within the accuracy of the measurements, indicating that they are not contaminants introduced during sample storage. This experiment marks the first synthesis of sulfur amino acids from spark discharge experiments designed to imitate primordial environments. The relative yield of some amino acids, in particular the isomers of aminobutyric acid, are the highest ever found in a spark discharge experiment. The simulated primordial conditions used by Miller may serve as a model for early volcanic plume chemistry and provide insight to the possible roles such plumes may have played in abiotic organic synthesis. Additionally, the overall abundances of the synthesized amino acids in the presence of H(2)S are very similar to the abundances found in some carbonaceous meteorites, suggesting that H(2)S may have played an important role in prebiotic reactions in early solar system environments.
Oró (1960) — prebiotic synthesis of adenine.
Synthesis of adenine from ammonium cyanide. Oró (1960). Biochemical and Biophysical Research Communications. doi:10.1016/0006-291x(60)90138-8
Kruger et al. (1982), Guerrier-Takada et al. (1983) — the discovery of ribozymes.
Self-splicing RNA: Autoexcision and autocyclization of the ribosomal RNA intervening sequence of tetrahymena. Kruger et al. (1982). Cell. doi:10.1016/0092-8674(82)90414-7
Abstract
In the macronuclear rRNA genes of Tetrahymena thermophila, a 413 bp intervening sequence (IVS) interrupts the 26S rRNA-coding region. A restriction fragment of the rDNA containing the IVS and portions of the adjacent rRNA sequences (exons) was inserted downstream from the lac UV5 promoter in a recombinant plasmid. Transcription of this template by purified Escherichia coli RNA polymerase in vitro produced a shortened version of the pre-rRNA, which was then deproteinized. When incubated with monovalent and divalent cations and a guanosine factor, this RNA underwent splicing. The reactions that were characterized included the precise excision of the IVS, attachment of guanosine to the 5' end of the IVS, covalent cyclization of the IVS and ligation of the exons. We conclude that splicing activity is intrinsic to the structure of the RNA, and that enzymes, small nuclear RNAs and folding of the pre-rRNA into an RNP are unnecessary for these reactions. We propose that the IVS portion of the RNA has several enzyme-like properties that enable it to break and reform phosphodiester bonds. The finding of autocatalytic rearrangements of RNA molecules has implications for the mechanism and the evolution of other reactions that involve RNA.
The RNA moiety of ribonuclease P is the catalytic subunit of the enzyme. Guerrier-Takada et al. (1983). Cell. doi:10.1016/0092-8674(83)90117-4
Abstract
The RNA moieties of ribonuclease P purified from both E. coli (M1 RNA) and B. subtilis (P-RNA) can cleave tRNA precursor molecules in buffers containing either 60 mM Mg2+ or 10 mM Mg2+ plus 1 mM spermidine. The RNA acts as a true catalyst under these conditions whereas the protein moieties of the enzymes alone show no catalytic activity. However, in buffers containing 5-10 mM Mg2+ (in the absence of spermidine) both kinds of subunits are required for enzymatic activity, as shown previously. In the presence of low concentrations of Mg2+, in vitro, the RNA and protein subunits from one species can complement subunits from the other species in reconstitution experiments. When the precursor to E. coli 4.5S RNA is used as a substrate, only the enzyme complexes formed with M1 RNA from E. coli and the protein moieties from either bacterial species are active.
Gilbert (1986) — the RNA world named.
Origin of life: The RNA world. Gilbert (1986). Nature. doi:10.1038/319618a0
Joyce (2002), Orgel (2004), Higgs and Lehman (2014) — reviews of the RNA world.
The antiquity of RNA-based evolution. Joyce (2002). Nature. doi:10.1038/418214a
Abstract
All life that is known to exist on Earth today and all life for which there is evidence in the geological record seems to be of the same form--one based on DNA genomes and protein enzymes. Yet there are strong reasons to conclude that DNA- and protein-based life was preceded by a simpler life form based primarily on RNA. This earlier era is referred to as the 'RNA world', during which the genetic information resided in the sequence of RNA molecules and the phenotype derived from the catalytic properties of RNA.
Prebiotic Chemistry and the Origin of the RNA World. Orgel (2004). Critical Reviews in Biochemistry and Molecular Biology. doi:10.1080/10409230490460765
Abstract
The demonstration that ribosomal peptide synthesis is a ribozyme-catalyzed reaction makes it almost certain that there was once an RNA World. The central problem for origin-of-life studies, therefore, is to understand how a protein-free RNA World became established on the primitive Earth. We first review the literature on the prebiotic synthesis of the nucleotides, the nonenzymatic synthesis and copying of polynucleotides, and the selection of ribozyme catalysts of a kind that might have facilitated polynucleotide replication. This leads to a brief outline of the Molecular Biologists' Dream, an optimistic scenario for the origin of the RNA World. In the second part of the review we point out the many unresolved problems presented by the Molecular Biologists' Dream. This in turn leads to a discussion of genetic systems simpler than RNA that might have "invented" RNA. Finally, we review studies of prebiotic membrane formation.
The RNA World: molecular cooperation at the origins of life. Higgs & Lehman (2014). Nature Reviews Genetics. doi:10.1038/nrg3841
Abstract
The RNA World concept posits that there was a period of time in primitive Earth's history - about 4 billion years ago - when the primary living substance was RNA or something chemically similar. In the past 50 years, this idea has gone from speculation to a prevailing idea. In this Review, we summarize the key logic behind the RNA World and describe some of the most important recent advances that have been made to support and expand this logic. We also discuss the ways in which molecular cooperation involving RNAs would facilitate the emergence and early evolution of life. The immediate future of RNA World research should be a very dynamic one.
Ban et al. (2000) — the ribosome is a ribozyme.
The Complete Atomic Structure of the Large Ribosomal Subunit at 2.4 Å Resolution. Ban et al. (2000). Science. doi:10.1126/science.289.5481.905
Abstract
The large ribosomal subunit catalyzes peptide bond formation and binds initiation, termination, and elongation factors. We have determined the crystal structure of the large ribosomal subunit from Haloarcula marismortui at 2.4 angstrom resolution, and it includes 2833 of the subunit's 3045 nucleotides and 27 of its 31 proteins. The domains of its RNAs all have irregular shapes and fit together in the ribosome like the pieces of a three-dimensional jigsaw puzzle to form a large, monolithic structure. Proteins are abundant everywhere on its surface except in the active site where peptide bond formation occurs and where it contacts the small subunit. Most of the proteins stabilize the structure by interacting with several RNA domains, often using idiosyncratically folded extensions that reach into the subunit's interior.
Powner et al. (2009), Sutherland (2015) — prebiotic synthesis of nucleotides.
Synthesis of activated pyrimidine ribonucleotides in prebiotically plausible conditions. Powner et al. (2009). Nature. doi:10.1038/nature08013
Abstract
At some stage in the origin of life, an informational polymer must have arisen by purely chemical means. According to one version of the 'RNA world' hypothesis this polymer was RNA, but attempts to provide experimental support for this have failed. In particular, although there has been some success demonstrating that 'activated' ribonucleotides can polymerize to form RNA, it is far from obvious how such ribonucleotides could have formed from their constituent parts (ribose and nucleobases). Ribose is difficult to form selectively, and the addition of nucleobases to ribose is inefficient in the case of purines and does not occur at all in the case of the canonical pyrimidines. Here we show that activated pyrimidine ribonucleotides can be formed in a short sequence that bypasses free ribose and the nucleobases, and instead proceeds through arabinose amino-oxazoline and anhydronucleoside intermediates. The starting materials for the synthesis-cyanamide, cyanoacetylene, glycolaldehyde, glyceraldehyde and inorganic phosphate-are plausible prebiotic feedstock molecules, and the conditions of the synthesis are consistent with potential early-Earth geochemical models. Although inorganic phosphate is only incorporated into the nucleotides at a late stage of the sequence, its presence from the start is essential as it controls three reactions in the earlier stages by acting as a general acid/base catalyst, a nucleophilic catalyst, a pH buffer and a chemical buffer. For prebiotic reaction sequences, our results highlight the importance of working with mixed chemical systems in which reactants for a particular reaction step can also control other steps.
The Origin of Life—Out of the Blue. Sutherland (2015). Angewandte Chemie International Edition. doi:10.1002/anie.201506585
Abstract
Either to sustain autotrophy, or as a prelude to heterotrophy, organic synthesis from an environmentally available C1 feedstock molecule is crucial to the origin of life. Recent findings augment key literature results and suggest that hydrogen cyanide--"Blausäure"--was that feedstock.
Lincoln and Joyce (2009) — self-sustained replication of an RNA enzyme.
Self-Sustained Replication of an RNA Enzyme. Lincoln & Joyce (2009). Science. doi:10.1126/science.1167856
Abstract
An RNA enzyme that catalyzes the RNA-templated joining of RNA was converted to a format whereby two enzymes catalyze each other's synthesis from a total of four oligonucleotide substrates. These cross-replicating RNA enzymes undergo self-sustained exponential amplification in the absence of proteins or other biological materials. Amplification occurs with a doubling time of about 1 hour and can be continued indefinitely. Populations of various cross-replicating enzymes were constructed and allowed to compete for a common pool of substrates, during which recombinant replicators arose and grew to dominate the population. These replicating RNA enzymes can serve as an experimental model of a genetic system. Many such model systems could be constructed, allowing different selective outcomes to be related to the underlying properties of the genetic system.
Wächtershäuser (1988), Martin and Russell (2003) — metabolism first, and hydrothermal vents.
Before enzymes and templates: theory of surface metabolism. Wächtershäuser (1988). Microbiological Reviews. doi:10.1128/mr.52.4.452-484.1988
On the origins of cells: a hypothesis for the evolutionary transitions from abiotic geochemistry to chemoautotrophic prokaryotes, and from prokaryotes to nucleated cells. Martin & Russell (2003). Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences. doi:10.1098/rstb.2002.1183
Abstract
All life is organized as cells. Physical compartmentation from the environment and self-organization of self-contained redox reactions are the most conserved attributes of living things, hence inorganic matter with such attributes would be life's most likely forebear. We propose that life evolved in structured iron monosulphide precipitates in a seepage site hydrothermal mound at a redox, pH and temperature gradient between sulphide-rich hydrothermal fluid and iron(II)-containing waters of the Hadean ocean floor. The naturally arising, three-dimensional compartmentation observed within fossilized seepage-site metal sulphide precipitates indicates that these inorganic compartments were the precursors of cell walls and membranes found in free-living prokaryotes. The known capability of FeS and NiS to catalyse the synthesis of the acetyl-methylsulphide from carbon monoxide and methylsulphide, constituents of hydrothermal fluid, indicates that pre-biotic syntheses occurred at the inner surfaces of these metal-sulphide-walled compartments, which furthermore restrained reacted products from diffusion into the ocean, providing sufficient concentrations of reactants to forge the transition from geochemistry to biochemistry. The chemistry of what is known as the RNA-world could have taken place within these naturally forming, catalyticwalled compartments to give rise to replicating systems. Sufficient concentrations of precursors to support replication would have been synthesized in situ geochemically and biogeochemically, with FeS (and NiS) centres playing the central catalytic role. The universal ancestor we infer was not a free-living cell, but rather was confined to the naturally chemiosmotic, FeS compartments within which the synthesis of its constituents occurred. The first free-living cells are suggested to have been eubacterial and archaebacterial chemoautotrophs that emerged more than 3.8 Gyr ago from their inorganic confines. We propose that the emergence of these prokaryotic lineages from inorganic confines occurred independently, facilitated by the independent origins of membrane-lipid biosynthesis: isoprenoid ether membranes in the archaebacterial and fatty acid ester membranes in the eubacterial lineage. The eukaryotes, all of which are ancestrally heterotrophs and possess eubacterial lipids, are suggested to have arisen ca. 2 Gyr ago through symbiosis involving an autotrophic archaebacterial host and a heterotrophic eubacterial symbiont, the common ancestor of mitochondria and hydrogenosomes. The attributes shared by all prokaryotes are viewed as inheritances from their confined universal ancestor. The attributes that distinguish eubacteria and archaebacteria, yet are uniform within the groups, are viewed as relics of their phase of differentiation after divergence from the non-free-living universal ancestor and before the origin of the free-living chemoautotrophic lifestyle. The attributes shared by eukaryotes with eubacteria and archaebacteria, respectively, are viewed as inheritances via symbiosis. The attributes unique to eukaryotes are viewed as inventions specific to their lineage. The origin of the eukaryotic endomembrane system and nuclear membrane are suggested to be the fortuitous result of the expression of genes for eubacterial membrane lipid synthesis by an archaebacterial genetic apparatus in a compartment that was not fully prepared to accommodate such compounds, resulting in vesicles of eubacterial lipids that accumulated in the cytosol around their site of synthesis. Under these premises, the most ancient divide in the living world is that between eubacteria and archaebacteria, yet the steepest evolutionary grade is that between prokaryotes and eukaryotes.
Bell et al. (2015), Dodd et al. (2017) — the earliest evidence for life.
Potentially biogenic carbon preserved in a 4.1 billion-year-old zircon. Bell et al. (2015). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1517557112
Abstract
Evidence of life on Earth is manifestly preserved in the rock record. However, the microfossil record only extends to ∼ 3.5 billion years (Ga), the chemofossil record arguably to ∼ 3.8 Ga, and the rock record to 4.0 Ga. Detrital zircons from Jack Hills, Western Australia range in age up to nearly 4.4 Ga. From a population of over 10,000 Jack Hills zircons, we identified one >3.8-Ga zircon that contains primary graphite inclusions. Here, we report carbon isotopic measurements on these inclusions in a concordant, 4.10 ± 0.01-Ga zircon. We interpret these inclusions as primary due to their enclosure in a crack-free host as shown by transmission X-ray microscopy and their crystal habit. Their δ(13)CPDB of -24 ± 5‰ is consistent with a biogenic origin and may be evidence that a terrestrial biosphere had emerged by 4.1 Ga, or ∼ 300 My earlier than has been previously proposed.
Evidence for early life in Earth’s oldest hydrothermal vent precipitates. Dodd et al. (2017). Nature. doi:10.1038/nature21377
Abstract
Although it is not known when or where life on Earth began, some of the earliest habitable environments may have been submarine-hydrothermal vents. Here we describe putative fossilized microorganisms that are at least 3,770 million and possibly 4,280 million years old in ferruginous sedimentary rocks, interpreted as seafloor-hydrothermal vent-related precipitates, from the Nuvvuagittuq belt in Quebec, Canada. These structures occur as micrometre-scale haematite tubes and filaments with morphologies and mineral assemblages similar to those of filamentous microorganisms from modern hydrothermal vent precipitates and analogous microfossils in younger rocks. The Nuvvuagittuq rocks contain isotopically light carbon in carbonate and carbonaceous material, which occurs as graphitic inclusions in diagenetic carbonate rosettes, apatite blades intergrown among carbonate rosettes and magnetite-haematite granules, and is associated with carbonate in direct contact with the putative microfossils. Collectively, these observations are consistent with an oxidized biomass and provide evidence for biological activity in submarine-hydrothermal environments more than 3,770 million years ago.
Des Marais et al. (2008) — the NASA astrobiology roadmap.
The NASA Astrobiology Roadmap. Des Marais et al. (2008). Astrobiology. doi:10.1089/ast.2008.0819
Abstract
The NASA Astrobiology Roadmap provides guidance for research and technology development across the NASA enterprises that encompass the space, Earth, and biological sciences. The ongoing development of astrobiology roadmaps embodies the contributions of diverse scientists and technologists from government, universities, and private institutions. The Roadmap addresses three basic questions: how does life begin and evolve, does life exist elsewhere in the universe, and what is the future of life on Earth and beyond? Seven Science Goals outline the following key domains of investigation: understanding the nature and distribution of habitable environments in the universe, exploring for habitable environments and life in our own Solar System, understanding the emergence of life, determining how early life on Earth interacted and evolved with its changing environment, understanding the evolutionary mechanisms and environmental limits of life, determining the principles that will shape life in the future, and recognizing signatures of life on other worlds and on early Earth. For each of these goals, Science Objectives outline more specific high priority efforts for the next three to five years. These eighteen objectives are being integrated with NASA strategic planning.
Mayor and Queloz (1995) — the first planet around a Sun-like star.
A Jupiter-mass companion to a solar-type star. Mayor & Queloz (1995). Nature. doi:10.1038/378355a0
B.3.2 Concepts and the neutral theory
King and Jukes (1969) — non-Darwinian evolution.
Non-Darwinian Evolution: Most evolutionary change in proteins may be due to neutral mutations and genetic drift.. King & Jukes (1969). Science. doi:10.1126/science.164.3881.788
Ohta (1973) — the nearly neutral theory.
Slightly Deleterious Mutant Substitutions in Evolution. Ohta (1973). Nature. doi:10.1038/246096a0
Pauling and Zuckerkandl (1963), Langley and Fitch (1974) — the molecular clock and its constancy.
Chemical paleogenetics: molecular ``restoration studies'' of extinct forms of life. Pauling & Zuckerkandl (1963). Acta Chemica Scandinavica.
An examination of the constancy of the rate of molecular evolution. Langley & Fitch (1974). Journal of Molecular Evolution.
Smith and Haigh (1974) — the hitchhiking effect.
The hitch-hiking effect of a favourable gene. Smith & Haigh (1974). Genetical Research. doi:10.1017/s0016672300014634
Abstract
When a selectively favourable gene substitution occurs in a population, changes in gene frequencies will occur at closely linked loci. In the case of a neutral polymorphism, average heterozygosity will be reduced to an extent which varies with distance from the substituted locus. The aggregate effect of substitution on neutral polymorphism is estimated; in populations of total size 10 6 or more (and perhaps of 10 4 or more), this effect will be more important than that of random fixation. This may explain why the extent of polymorphism in natural populations does not vary as much as one would expect from a consideration of the equilibrium between mutation and random fixation in populations of different sizes. For a selectively maintained polymorphism at a linked locus, this process will only be important in the long run if it leads to complete fixation. If the selective coefficients at the linked locus are small compared to those at the substituted locus, it is shown that the probability of complete fixation at the linked locus is approximately exp (− Nc ), where c is the recombinant fraction and N the population size. It follows that in a large population a selective substitution can occur in a cistron without eliminating a selectively maintained polymorphism in the same cistron.
Charlesworth (2009) — effective population size and molecular evolution.
Effective population size and patterns of molecular evolution and variation. Charlesworth (2009). Nature Reviews Genetics. doi:10.1038/nrg2526
Abstract
The effective size of a population, N(e), determines the rate of change in the composition of a population caused by genetic drift, which is the random sampling of genetic variants in a finite population. N(e) is crucial in determining the level of variability in a population, and the effectiveness of selection relative to drift. This article reviews the properties of N(e) in a variety of different situations of biological interest, and the factors that influence it. In particular, the action of selection means that N(e) varies across the genome, and advances in genomic techniques are giving new insights into how selection shapes N(e).
Lynch and Conery (2003), Lynch et al. (2016) — genome complexity and the drift barrier.
The Origins of Genome Complexity. Lynch & Conery (2003). Science. doi:10.1126/science.1089370
Abstract
Complete genomic sequences from diverse phylogenetic lineages reveal notable increases in genome complexity from prokaryotes to multicellular eukaryotes. The changes include gradual increases in gene number, resulting from the retention of duplicate genes, and more abrupt increases in the abundance of spliceosomal introns and mobile genetic elements. We argue that many of these modifications emerged passively in response to the long-term population-size reductions that accompanied increases in organism size. According to this model, much of the restructuring of eukaryotic genomes was initiated by nonadaptive processes, and this in turn provided novel substrates for the secondary evolution of phenotypic complexity by natural selection. The enormous long-term effective population sizes of prokaryotes may impose a substantial barrier to the evolution of complex genomes and morphologies.
Genetic drift, selection and the evolution of the mutation rate. Lynch et al. (2016). Nature Reviews Genetics. doi:10.1038/nrg.2016.104
Abstract
As one of the few cellular traits that can be quantified across the tree of life, DNA-replication fidelity provides an excellent platform for understanding fundamental evolutionary processes. Furthermore, because mutation is the ultimate source of all genetic variation, clarifying why mutation rates vary is crucial for understanding all areas of biology. A potentially revealing hypothesis for mutation-rate evolution is that natural selection primarily operates to improve replication fidelity, with the ultimate limits to what can be achieved set by the power of random genetic drift. This drift-barrier hypothesis is consistent with comparative measures of mutation rates, provides a simple explanation for the existence of error-prone polymerases and yields a formal counter-argument to the view that selection fine-tunes gene-specific mutation rates.
Papers on phylogeny and the tree of life are listed in Chapter 4.
B.3.3 Homology, orthology and duplication
★ Altenhoff et al. (2019) — inferring orthology and paralogy.
Inferring Orthology and Paralogy. Altenhoff et al. (2019). Evolutionary Genomics. doi:10.1007/978-1-4939-9074-0_5
Abstract
The distinction between orthologs and paralogs, genes that started diverging by speciation versus duplication, is relevant in a wide range of contexts, most notably phylogenetic tree inference and protein function annotation. In this chapter, we provide an overview of the methods used to infer orthology and paralogy. We survey both graph-based approaches (and their various grouping strategies) and tree-based approaches, which solve the more general problem of gene/species tree reconciliation. We discuss conceptual differences among the various orthology inference methods and databases and examine the difficult issue of verifying and benchmarking orthology predictions. Finally, we review typical applications of orthologous genes, groups, and reconciled trees and conclude with thoughts on future methodological developments.
Fitch (1970), Fitch (2000) — homology defined.
Distinguishing Homologous from Analogous Proteins. Fitch (1970). Systematic Zoology. doi:10.2307/2412448
Homology. Fitch (2000). Trends in Genetics. doi:10.1016/s0168-9525(00)02005-9
Abstract
There are many problems relating to defining the terminology used to describe various biological relationships and getting agreement on which definitions are best. Here, I examine 15 terminological problems, all of which are current, and all of which relate to the usage of homology and its associated terms. I suggest a set of definitions that are intended to be totally consistent among themselves and also as consistent as possible with most current usage.
Tatusov et al. (1997) — COGs.
A Genomic Perspective on Protein Families. Tatusov et al. (1997). Science. doi:10.1126/science.278.5338.631
Abstract
In order to extract the maximum amount of information from the rapidly accumulating genome sequences, all conserved genes need to be classified according to their homologous relationships. Comparison of proteins encoded in seven complete genomes from five major phylogenetic lineages and elucidation of consistent patterns of sequence similarities allowed the delineation of 720 clusters of orthologous groups (COGs). Each COG consists of individual orthologous proteins or orthologous sets of paralogs from at least three lineages. Orthologs typically have the same function, allowing transfer of functional information from one member to an entire COG. This relation automatically yields a number of functional predictions for poorly characterized genomes. The COGs comprise a framework for functional and evolutionary genome analysis.
Remm et al. (2001) — InParanoid.
Automatic clustering of orthologs and in-paralogs from pairwise species comparisons. Remm et al. (2001). Journal of Molecular Biology. doi:10.1006/jmbi.2000.5197
Abstract
Orthologs are genes in different species that originate from a single gene in the last common ancestor of these species. Such genes have often retained identical biological roles in the present-day organisms. It is hence important to identify orthologs for transferring functional information between genes in different organisms with a high degree of reliability. For example, orthologs of human proteins are often functionally characterized in model organisms. Unfortunately, orthology analysis between human and e.g. invertebrates is often complex because of large numbers of paralogs within protein families. Paralogs that predate the species split, which we call out-paralogs, can easily be confused with true orthologs. Paralogs that arose after the species split, which we call in-paralogs, however, are bona fide orthologs by definition. Orthologs and in-paralogs are typically detected with phylogenetic methods, but these are slow and difficult to automate. Automatic clustering methods based on two-way best genome-wide matches on the other hand, have so far not separated in-paralogs from out-paralogs effectively. We present a fully automatic method for finding orthologs and in-paralogs from two species. Ortholog clusters are seeded with a two-way best pairwise match, after which an algorithm for adding in-paralogs is applied. The method bypasses multiple alignments and phylogenetic trees, which can be slow and error-prone steps in classical ortholog detection. Still, it robustly detects complex orthologous relationships and assigns confidence values for both orthologs and in-paralogs. The program, called INPARANOID, was tested on all completely sequenced eukaryotic genomes. To assess the quality of INPARANOID results, ortholog clusters were generated from a dataset of worm and mammalian transmembrane proteins, and were compared to clusters derived by manual tree-based ortholog detection methods. This study led to the identification with a high degree of confidence of over a dozen novel worm-mammalian ortholog assignments that were previously undetected because of shortcomings of phylogenetic methods.A WWW server that allows searching for orthologs between human and several fully sequenced genomes is installed at http://www.cgb.ki.se/inparanoid/. This is the first comprehensive resource with orthologs of all fully sequenced eukaryotic genomes. Programs and tables of orthology assignments are available from the same location.
Li et al. (2003) — OrthoMCL.
OrthoMCL: Identification of Ortholog Groups for Eukaryotic Genomes. Li et al. (2003). Genome Research. doi:10.1101/gr.1224503
Abstract
The identification of orthologous groups is useful for genome annotation, studies on gene/protein evolution, comparative genomics, and the identification of taxonomically restricted sequences. Methods successfully exploited for prokaryotic genome analysis have proved difficult to apply to eukaryotes, however, as larger genomes may contain multiple paralogous genes, and sequence information is often incomplete. OrthoMCL provides a scalable method for constructing orthologous groups across multiple eukaryotic taxa, using a Markov Cluster algorithm to group (putative) orthologs and paralogs. This method performs similarly to the INPARANOID algorithm when applied to two genomes, but can be extended to cluster orthologs from multiple species. OrthoMCL clusters are coherent with groups identified by EGO, but improved recognition of "recent" paralogs permits overlapping EGO groups representing the same gene to be merged. Comparison with previously assigned EC annotations suggests a high degree of reliability, implying utility for automated eukaryotic genome annotation. OrthoMCL has been applied to the proteome data set from seven publicly available genomes (human, fly, worm, yeast, Arabidopsis, the malaria parasite Plasmodium falciparum, and Escherichia coli). A Web interface allows queries based on individual genes or user-defined phylogenetic patterns (http://www.cbil.upenn.edu/gene-family). Analysis of clusters incorporating P. falciparum genes identifies numerous enzymes that were incompletely annotated in first-pass annotation of the parasite genome.
Huerta-Cepas et al. (2018) — eggNOG.
eggNOG 5.0: a hierarchical, functionally and phylogenetically annotated orthology resource based on 5090 organisms and 2502 viruses. Huerta-Cepas et al. (2018). Nucleic Acids Research. doi:10.1093/nar/gky1085
Abstract
eggNOG is a public database of orthology relationships, gene evolutionary histories and functional annotations. Here, we present version 5.0, featuring a major update of the underlying genome sets, which have been expanded to 4445 representative bacteria and 168 archaea derived from 25 038 genomes, as well as 477 eukaryotic organisms and 2502 viral proteomes that were selected for diversity and filtered by genome quality. In total, 4.4M orthologous groups (OGs) distributed across 379 taxonomic levels were computed together with their associated sequence alignments, phylogenies, HMM models and functional descriptors. Precomputed evolutionary analysis provides fine-grained resolution of duplication/speciation events within each OG. Our benchmarks show that, despite doubling the amount of genomes, the quality of orthology assignments and functional annotations (80% coverage) has persisted without significant changes across this update. Finally, we improved eggNOG online services for fast functional annotation and orthology prediction of custom genomics or metagenomics datasets. All precomputed data are publicly available for downloading or via API queries at http://eggnog.embl.de.
Altenhoff et al. (2016) — benchmarking orthology inference.
Standardized benchmarking in the quest for orthologs. Altenhoff et al. (2016). Nature Methods. doi:10.1038/nmeth.3830
Abstract
Achieving high accuracy in orthology inference is essential for many comparative, evolutionary and functional genomic analyses, yet the true evolutionary history of genes is generally unknown and orthologs are used for very different applications across phyla, requiring different precision-recall trade-offs. As a result, it is difficult to assess the performance of orthology inference methods. Here, we present a community effort to establish standards and an automated web-based service to facilitate orthology benchmarking. Using this service, we characterize 15 well-established inference methods and resources on a battery of 20 different benchmarks. Standardized benchmarking provides a way for users to identify the most effective methods for the problem at hand, sets a minimum requirement for new tools and resources, and guides the development of more accurate orthology inference methods.
Nehrt et al. (2011), Altenhoff et al. (2012) — the ortholog conjecture debate.
Testing the Ortholog Conjecture with Comparative Functional Genomic Data from Mammals. Nehrt et al. (2011). PLoS Computational Biology. doi:10.1371/journal.pcbi.1002073
Abstract
A common assumption in comparative genomics is that orthologous genes share greater functional similarity than do paralogous genes (the "ortholog conjecture"). Many methods used to computationally predict protein function are based on this assumption, even though it is largely untested. Here we present the first large-scale test of the ortholog conjecture using comparative functional genomic data from human and mouse. We use the experimentally derived functions of more than 8,900 genes, as well as an independent microarray dataset, to directly assess our ability to predict function using both orthologs and paralogs. Both datasets show that paralogs are often a much better predictor of function than are orthologs, even at lower sequence identities. Among paralogs, those found within the same species are consistently more functionally similar than those found in a different species. We also find that paralogous pairs residing on the same chromosome are more functionally similar than those on different chromosomes, perhaps due to higher levels of interlocus gene conversion between these pairs. In addition to offering implications for the computational prediction of protein function, our results shed light on the relationship between sequence divergence and functional divergence. We conclude that the most important factor in the evolution of function is not amino acid sequence, but rather the cellular context in which proteins act.
Resolving the Ortholog Conjecture: Orthologs Tend to Be Weakly, but Significantly, More Similar in Function than Paralogs. Altenhoff et al. (2012). PLoS Computational Biology. doi:10.1371/journal.pcbi.1002514
Abstract
The function of most proteins is not determined experimentally, but is extrapolated from homologs. According to the "ortholog conjecture", or standard model of phylogenomics, protein function changes rapidly after duplication, leading to paralogs with different functions, while orthologs retain the ancestral function. We report here that a comparison of experimentally supported functional annotations among homologs from 13 genomes mostly supports this model. We show that to analyze GO annotation effectively, several confounding factors need to be controlled: authorship bias, variation of GO term frequency among species, variation of background similarity among species pairs, and propagated annotation bias. After controlling for these biases, we observe that orthologs have generally more similar functional annotations than paralogs. This is especially strong for sub-cellular localization. We observe only a weak decrease in functional similarity with increasing sequence divergence. These findings hold over a large diversity of species; notably orthologs from model organisms such as E. coli, yeast or mouse have conserved function with human proteins.
Ohno et al. (1968) — evolution by gene duplication.
Evolution from fish to mammals by gene duplication. Ohno et al. (1968). Hereditas.
Force et al. (1999) — subfunctionalisation (the DDC model).
Preservation of Duplicate Genes by Complementary, Degenerative Mutations. Force et al. (1999). Genetics. doi:10.1093/genetics/151.4.1531
Abstract
The origin of organismal complexity is generally thought to be tightly coupled to the evolution of new gene functions arising subsequent to gene duplication. Under the classical model for the evolution of duplicate genes, one member of the duplicated pair usually degenerates within a few million years by accumulating deleterious mutations, while the other duplicate retains the original function. This model further predicts that on rare occasions, one duplicate may acquire a new adaptive function, resulting in the preservation of both members of the pair, one with the new function and the other retaining the old. However, empirical data suggest that a much greater proportion of gene duplicates is preserved than predicted by the classical model. Here we present a new conceptual framework for understanding the evolution of duplicate genes that may help explain this conundrum. Focusing on the regulatory complexity of eukaryotic genes, we show how complementary degenerative mutations in different regulatory elements of duplicated genes can facilitate the preservation of both duplicates, thereby increasing long-term opportunities for the evolution of new gene functions. The duplication-degeneration-complementation (DDC) model predicts that (1) degenerative mutations in regulatory elements can increase rather than reduce the probability of duplicate gene preservation and (2) the usual mechanism of duplicate gene preservation is the partitioning of ancestral functions rather than the evolution of new functions. We present several examples (including analysis of a new engrailed gene in zebrafish) that appear to be consistent with the DDC model, and we suggest several analytical and experimental approaches for determining whether the complementary loss of gene subfunctions or the acquisition of novel functions are likely to be the primary mechanisms for the preservation of gene duplicates. For a newly duplicated paralog, survival depends on the outcome of the race between entropic decay and chance acquisition of an advantageous regulatory mutation. Sidow 1996(p. 717) On one hand, it may fix an advantageous allele giving it a slightly different, and selectable, function from its original copy. This initial fixation provides substantial protection against future fixation of null mutations, allowing additional mutations to accumulate that refine functional differentiation. Alternatively, a duplicate locus can instead first fix a null allele, becoming a pseudogene. Walsh 1995 (p. 426) Duplicated genes persist only if mutations create new and essential protein functions, an event that is predicted to occur rarely. Nadeau and Sankoff 1997 (p. 1259) Thus overall, with complex metazoans, the major mechanism for retention of ancient gene duplicates would appear to have been the acquisition of novel expression sites for developmental genes, with its accompanying opportunity for new gene roles underlying the progressive extension of development itself. Cooke et al. 1997 (p. 362)
Innan and Kondrashov (2010) — models of duplicate gene evolution.
The evolution of gene duplications: classifying and distinguishing between models. Innan & Kondrashov (2010). Nature Reviews Genetics. doi:10.1038/nrg2689
Abstract
Gene duplications and their subsequent divergence play an important part in the evolution of novel gene functions. Several models for the emergence, maintenance and evolution of gene copies have been proposed. However, a clear consensus on how gene duplications are fixed and maintained in genomes is lacking. Here, we present a comprehensive classification of the models that are relevant to all stages of the evolution of gene duplications. Each model predicts a unique combination of evolutionary dynamics and functional properties. Setting out these predictions is an important step towards identifying the main mechanisms that are involved in the evolution of gene duplications.
Papp et al. (2003) — dosage sensitivity.
Dosage sensitivity and the evolution of gene families in yeast. Papp et al. (2003). Nature. doi:10.1038/nature01771
Abstract
According to what we term the balance hypothesis, an imbalance in the concentration of the subcomponents of a protein-protein complex can be deleterious. If so, there are two consequences: first, both underexpression and overexpression of protein complex subunits should lower fitness, and second, the accuracy of transcriptional co-regulation of subunits should reflect the deleterious consequences of imbalance. Here we show that all these predictions are upheld in yeast (Saccharomyces cerevisiae). This supports the hypothesis that dominance is a by-product of physiology and metabolism rather than the result of selection to mask the deleterious effects of mutations. Beyond this, single-gene duplication of protein subunits is expected to be harmful, as this, too, leads to imbalance. As then expected, we find that members of large gene families are rarely involved in complexes. The balance hypothesis therefore provides a single theoretical framework for understanding components both of dominance and of gene family size.
Wolfe and Shields (1997), Kellis et al. (2004) — whole-genome duplication in yeast.
Molecular evidence for an ancient duplication of the entire yeast genome. Wolfe & Shields (1997). Nature. doi:10.1038/42711
Abstract
Gene duplication is an important source of evolutionary novelty. Most duplications are of just a single gene, but Ohno proposed that whole-genome duplication (polyploidy) is an important evolutionary mechanism. Many duplicate genes have been found in Saccharomyces cerevisiae, and these often seem to be phenotypically redundant. Here we show that the arrangement of duplicated genes in the S. cerevisiae genome is consistent with Ohno's hypothesis. We propose a model in which this species is a degenerate tetraploid resulting from a whole-genome duplication that occurred after the divergence of Saccharomyces from Kluyveromyces. Only a small fraction of the genes were subsequently retained in duplicate (most were deleted), and gene order was rearranged by many reciprocal translocations between chromosomes. Protein pairs derived from this duplication event make up 13% of all yeast proteins, and include pairs of transcription factors, protein kinases, myosins, cyclins and pheromones. Tetraploidy may have facilitated the evolution of anaerobic fermentation in Saccharomyces.
Proof and evolutionary analysis of ancient genome duplication in the yeast Saccharomyces cerevisiae. Kellis et al. (2004). Nature. doi:10.1038/nature02424
Abstract
Whole-genome duplication followed by massive gene loss and specialization has long been postulated as a powerful mechanism of evolutionary innovation. Recently, it has become possible to test this notion by searching complete genome sequence for signs of ancient duplication. Here, we show that the yeast Saccharomyces cerevisiae arose from ancient whole-genome duplication, by sequencing and analysing Kluyveromyces waltii, a related yeast species that diverged before the duplication. The two genomes are related by a 1:2 mapping, with each region of K. waltii corresponding to two regions of S. cerevisiae, as expected for whole-genome duplication. This resolves the long-standing controversy on the ancestry of the yeast genome, and makes it possible to study the fate of duplicated genes directly. Strikingly, 95% of cases of accelerated evolution involve only one member of a gene pair, providing strong support for a specific model of evolution, and allowing us to distinguish ancestral and derived functions.
Dehal and Boore (2005) — two rounds of genome duplication in vertebrates.
Two Rounds of Whole Genome Duplication in the Ancestral Vertebrate. Dehal & Boore (2005). PLoS Biology. doi:10.1371/journal.pbio.0030314
Abstract
The hypothesis that the relatively large and complex vertebrate genome was created by two ancient, whole genome duplications has been hotly debated, but remains unresolved. We reconstructed the evolutionary relationships of all gene families from the complete gene sets of a tunicate, fish, mouse, and human, and then determined when each gene duplicated relative to the evolutionary tree of the organisms. We confirmed the results of earlier studies that there remains little signal of these events in numbers of duplicated genes, gene tree topology, or the number of genes per multigene family. However, when we plotted the genomic map positions of only the subset of paralogous genes that were duplicated prior to the fish-tetrapod split, their global physical organization provides unmistakable evidence of two distinct genome duplication events early in vertebrate evolution indicated by clear patterns of four-way paralogous regions covering a large part of the human genome. Our results highlight the potential for these large-scale genomic events to have driven the evolutionary success of the vertebrate lineage.
B.3.4 Protein evolution
Apic et al. (2001), Björklund et al. (2005) — domain combinations and rearrangements.
Domain combinations in archaeal, eubacterial and eukaryotic proteomes. Apic et al. (2001). Journal of Molecular Biology. doi:10.1006/jmbi.2001.4776
Abstract
There is a limited repertoire of domain families that are duplicated and combined in different ways to form the set of proteins in a genome. Proteins are gene products, and at the level of genes, duplication, recombination, fusion and fission are the processes that produce new genes. We attempt to gain an overview of these processes by studying the evolutionary units in proteins, domains, in the protein sequences of 40 genomes. The domain and superfamily definitions in the Structural Classification of Proteins Database are used, so that we can view all pairs of adjacent domains in genome sequences in terms of their superfamily combinations. We find 783 out of the 859 superfamilies in SCOP in these genomes, and the 783 families occur in 1307 pairwise combinations. Most families are observed in combination with one or two other families, while a few families are very versatile in their combinatorial behaviour; 209 families do not make combinations with other families. This type of pattern can be described as a scale-free network. We also study the N to C-terminal orientation of domain pairs and domain repeats. The phylogenetic distribution of domain combinations is surveyed, to establish the extent of common and kingdom-specific combinations. Of the kingdom-specific combinations, significantly more combinations consist of families present in all three kingdoms than of families present in one or two kingdoms. Hence, we are led to conclude that recombination between common families, as compared to the invention of new families and recombination among these, has also been a major contribution to the evolution of kingdom-specific and species-specific functions in organisms in all three kingdoms. Finally, we compare the set of the domain combinations in the genomes to those in the RCSB Protein Data Bank, and discuss the implications for structural genomics.
Domain Rearrangements in Protein Evolution. Björklund et al. (2005). Journal of Molecular Biology. doi:10.1016/j.jmb.2005.08.067
Abstract
Most eukaryotic proteins are multi-domain proteins that are created from fusions of genes, deletions and internal repetitions. An investigation of such evolutionary events requires a method to find the domain architecture from which each protein originates. Therefore, we defined a novel measure, domain distance, which is calculated as the number of domains that differ between two domain architectures. Using this measure the evolutionary events that distinguish a protein from its closest ancestor have been studied and it was found that indels are more common than internal repetition and that the exchange of a domain is rare. Indels and repetitions are common at both the N and C-terminals while they are rare between domains. The evolution of the majority of multi-domain proteins can be explained by the stepwise insertions of single domains, with the exception of repeats that sometimes are duplicated several domains in tandem. We show that domain distances agree with sequence similarity and semantic similarity based on gene ontology annotations. In addition, we demonstrate the use of the domain distance measure to build evolutionary trees. Finally, the evolution of multi-domain proteins is exemplified by a closer study of the evolution of two protein families, non-receptor tyrosine kinases and RhoGEFs.
Kummerfeld and Teichmann (2005) — fusion and fission.
Relative rates of gene fusion and fission in multi-domain proteins. Kummerfeld & Teichmann (2005). Trends in Genetics. doi:10.1016/j.tig.2004.11.007
Abstract
During evolution genes can produce more complex proteins by gene fusion or less complex proteins by gene fission. Considering proteins from 131 completely sequenced genomes from all three kingdoms of life, we identified 2869 groups of multi-domain proteins as a single protein in certain organisms and as two or more smaller proteins with equivalent domain architectures in other organisms. We found that fusion events are approximately four times more common than fission events, and we established that, in most cases, any particular fusion or fission event only occurred once during the course of evolution.
Levy et al. (2008) — assembly of complexes reflects evolution.
Assembly reflects evolution of protein complexes. Levy et al. (2008). Nature. doi:10.1038/nature06942
Abstract
A homomer is formed by self-interacting copies of a protein unit. This is functionally important, as in allostery, and structurally crucial because mis-assembly of homomers is implicated in disease. Homomers are widespread, with 50-70% of proteins with a known quaternary state assembling into such structures. Despite their prevalence, their role in the evolution of cellular machinery and the potential for their use in the design of new molecular machines, little is known about the mechanisms that drive formation of homomers at the level of evolution and assembly in the cell. Here we present an analysis of over 5,000 unique atomic structures and show that the quaternary structure of homomers is conserved in over 70% of protein pairs sharing as little as 30% sequence identity. Where quaternary structure is not conserved among the members of a protein family, a detailed investigation revealed well-defined evolutionary pathways by which proteins transit between different quaternary structure types. Furthermore, we show by perturbing subunit interfaces within complexes and by mass spectrometry analysis, that the (dis)assembly pathway mimics the evolutionary pathway. These data represent a molecular analogy to Haeckel's evolutionary paradigm of embryonic development, where an intermediate in the assembly of a complex represents a form that appeared in its own evolutionary history. Our model of self-assembly allows reliable prediction of evolution and assembly of a complex solely from its crystal structure.
Björklund et al. (2006) — expansion of domain repeats.
Expansion of Protein Domain Repeats. Björklund et al. (2006). PLoS Computational Biology. doi:10.1371/journal.pcbi.0020114
Abstract
Many proteins, especially in eukaryotes, contain tandem repeats of several domains from the same family. These repeats have a variety of binding properties and are involved in protein-protein interactions as well as binding to other ligands such as DNA and RNA. The rapid expansion of protein domain repeats is assumed to have evolved through internal tandem duplications. However, the exact mechanisms behind these tandem duplications are not well-understood. Here, we have studied the evolution, function, protein structure, gene structure, and phylogenetic distribution of domain repeats. For this purpose we have assigned Pfam-A domain families to 24 proteomes with more sensitive domain assignments in the repeat regions. These assignments confirmed previous findings that eukaryotes, and in particular vertebrates, contain a much higher fraction of proteins with repeats compared with prokaryotes. The internal sequence similarity in each protein revealed that the domain repeats are often expanded through duplications of several domains at a time, while the duplication of one domain is less common. Many of the repeats appear to have been duplicated in the middle of the repeat region. This is in strong contrast to the evolution of other proteins that mainly works through additions of single domains at either terminus. Further, we found that some domain families show distinct duplication patterns, e.g., nebulin domains have mainly been expanded with a unit of seven domains at a time, while duplications of other domain families involve varying numbers of domains. Finally, no common mechanism for the expansion of all repeats could be detected. We found that the duplication patterns show no dependence on the size of the domains. Further, repeat expansion in some families can possibly be explained by shuffling of exons. However, exon shuffling could not have created all repeats.
Kajava (2012) — tandem repeats in proteins.
Tandem repeats in proteins: From sequence to structure. Kajava (2012). Journal of Structural Biology. doi:10.1016/j.jsb.2011.08.009
Abstract
The bioinformatics analysis of proteins containing tandem repeats requires special computer programs and databases, since the conventional approaches predominantly developed for globular domains have limited success. Here, I survey bioinformatics tools which have been developed recently for identification and proteome-wide analysis of protein repeats. The last few years have also been marked by an emergence of new 3D structures of these proteins. Appraisal of the known structures and their classification uncovers a straightforward relationship between their architecture and the length of the repetitive units. This relationship and the repetitive character of structural folds suggest rules for better prediction of the 3D structures of such proteins. Furthermore, bioinformatics approaches combined with low resolution structural data, from biophysical techniques, especially, the recently emerged cryo-electron microscopy, lead to reliable prediction of the protein repeat structures and their mode of binding with partners within molecular complexes. This hybrid approach can actively be used for structural and functional annotations of proteomes.
Brown et al. (2011), Lee et al. (2014), Light et al. (2013) — evolution of disordered regions.
Evolution and disorder. Brown et al. (2011). Current Opinion in Structural Biology. doi:10.1016/j.sbi.2011.02.005
Abstract
The evolution of disordered proteins or regions of proteins differs from that of ordered proteins because of the differences in their sequence composition, intramolecular contacts, and function. Recent assessments of disordered protein evolution at the sequence, structural, and functional levels support this hypothesis. Disordered proteins have a different pattern of accepted point mutations, exhibit higher rates of insertions and deletions, and generally, but not always, evolve more rapidly than ordered proteins. Even with these high rates of sequence evolution, a few examples have shown that disordered proteins maintain their flexibility under physiological conditions, and it is hypothesized that they maintain specific structural ensembles.
Classification of Intrinsically Disordered Regions and Proteins. van der Lee et al. (2014). Chemical Reviews. doi:10.1021/cr400525m
Abstract
Characterization of unannotated and uncharacterized protein segments is expected to lead to the discovery of novel functions as well as provide important insights into existing biological processes. In addition, it is likely to shed new light on molecular mechanisms of diseases that are not yet fully understood. The classical concept implies that protein sequence defines structure, which in turn determines function; that is, function can be inferred from the sequence and its structure. Even when protein sequences diverge during evolution, for example, after gene duplication, the overall fold of their structures remains roughly the same. Therefore, structural similarity between proteins can reveal distant evolutionary relationships that are not easily detectable using sequence-based methods.
Protein Expansion Is Primarily due to Indels in Intrinsically Disordered Regions. Light et al. (2013). Molecular Biology and Evolution. doi:10.1093/molbev/mst157
Abstract
Proteins evolve not only through point mutations but also by insertion and deletion events, which affect the length of the protein. It is well known that such indel events most frequently occur in surface-exposed loops. However, detailed analysis of indel events in distantly related and fast-evolving proteins is hampered by the difficulty involved in correctly aligning such sequences. Here, we circumvent this problem by first only analyzing homologous proteins based on length variation rather than pairwise alignments. Using this approach, we find a surprisingly strong relationship between difference in length and difference in the number of intrinsically disordered residues, where up to three quarters of the length variation can be explained by changes in the number of intrinsically disordered residues. Further, we find that disorder is common in both insertions and deletions. A more detailed analysis reveals that indel events do not induce disorder but rather that already disordered regions accrue indels, suggesting that there is a lowered selective pressure for indels to occur within intrinsically disordered regions.
Göbel et al. (1994), Weigt et al. (2009), Marks et al. (2011) — correlated mutations and coevolution.
Correlated mutations and residue contacts in proteins. Göbel et al. (1994). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.340180402
Abstract
The maintenance of protein function and structure constrains the evolution of amino acid sequences. This fact can be exploited to interpret correlated mutations observed in a sequence family as an indication of probable physical contact in three dimensions. Here we present a simple and general method to analyze correlations in mutational behavior between different positions in a multiple sequence alignment. We then use these correlations to predict contact maps for each of 11 protein families and compare the result with the contacts determined by crystallography. For the most strongly correlated residue pairs predicted to be in contact, the prediction accuracy ranges from 37 to 68% and the improvement ratio relative to a random prediction from 1.4 to 5.1. Predicted contact maps can be used as input for the calculation of protein tertiary structure, either from sequence information alone or in combination with experimental information.
Identification of direct residue contacts in protein–protein interaction by message passing. Weigt et al. (2009). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.0805923106
Abstract
Understanding the molecular determinants of specificity in protein-protein interaction is an outstanding challenge of postgenome biology. The availability of large protein databases generated from sequences of hundreds of bacterial genomes enables various statistical approaches to this problem. In this context covariance-based methods have been used to identify correlation between amino acid positions in interacting proteins. However, these methods have an important shortcoming, in that they cannot distinguish between directly and indirectly correlated residues. We developed a method that combines covariance analysis with global inference analysis, adopted from use in statistical physics. Applied to a set of >2,500 representatives of the bacterial two-component signal transduction system, the combination of covariance with global inference successfully and robustly identified residue pairs that are proximal in space without resorting to ad hoc tuning parameters, both for heterointeractions between sensor kinase (SK) and response regulator (RR) proteins and for homointeractions between RR proteins. The spectacular success of this approach illustrates the effectiveness of the global inference approach in identifying direct interaction based on sequence information alone. We expect this method to be applicable soon to interaction surfaces between proteins present in only 1 copy per genome as the number of sequenced genomes continues to expand. Use of this method could significantly increase the potential targets for therapeutic intervention, shed light on the mechanism of protein-protein interaction, and establish the foundation for the accurate prediction of interacting protein partners.
Protein 3D Structure Computed from Evolutionary Sequence Variation. Marks et al. (2011). PLoS ONE. doi:10.1371/journal.pone.0028766
Abstract
The evolutionary trajectory of a protein through sequence space is constrained by its function. Collections of sequence homologs record the outcomes of millions of evolutionary experiments in which the protein evolves according to these constraints. Deciphering the evolutionary record held in these sequences and exploiting it for predictive and engineering purposes presents a formidable challenge. The potential benefit of solving this challenge is amplified by the advent of inexpensive high-throughput genomic sequencing.In this paper we ask whether we can infer evolutionary constraints from a set of sequence homologs of a protein. The challenge is to distinguish true co-evolution couplings from the noisy set of observed correlations. We address this challenge using a maximum entropy model of the protein sequence, constrained by the statistics of the multiple sequence alignment, to infer residue pair couplings. Surprisingly, we find that the strength of these inferred couplings is an excellent predictor of residue-residue proximity in folded structures. Indeed, the top-scoring residue couplings are sufficiently accurate and well-distributed to define the 3D protein fold with remarkable accuracy.We quantify this observation by computing, from sequence alone, all-atom 3D structures of fifteen test proteins from different fold classes, ranging in size from 50 to 260 residues, including a G-protein coupled receptor. These blinded inferences are de novo, i.e., they do not use homology modeling or sequence-similar fragments from known structures. The co-evolution signals provide sufficient information to determine accurate 3D protein structure to 2.7-4.8 Å C(α)-RMSD error relative to the observed structure, over at least two-thirds of the protein (method called EVfold, details at http://EVfold.org). This discovery provides insight into essential interactions constraining protein evolution and will facilitate a comprehensive survey of the universe of protein structures, new strategies in protein and drug design, and the identification of functional genetic variants in normal and disease genomes.
Barbosa-Morais et al. (2012), Merkin et al. (2012), Tress et al. (2017) — evolution and function of alternative splicing.
The Evolutionary Landscape of Alternative Splicing in Vertebrate Species. Barbosa-Morais et al. (2012). Science. doi:10.1126/science.1230612
Abstract
How species with similar repertoires of protein-coding genes differ so markedly at the phenotypic level is poorly understood. By comparing organ transcriptomes from vertebrate species spanning ~350 million years of evolution, we observed significant differences in alternative splicing complexity between vertebrate lineages, with the highest complexity in primates. Within 6 million years, the splicing profiles of physiologically equivalent organs diverged such that they are more strongly related to the identity of a species than they are to organ type. Most vertebrate species-specific splicing patterns are cis-directed. However, a subset of pronounced splicing changes are predicted to remodel protein interactions involving trans-acting regulators. These events likely further contributed to the diversification of splicing and other transcriptomic changes that underlie phenotypic differences among vertebrate species.
Evolutionary Dynamics of Gene and Isoform Regulation in Mammalian Tissues. Merkin et al. (2012). Science. doi:10.1126/science.1228186
Abstract
Most mammalian genes produce multiple distinct messenger RNAs through alternative splicing, but the extent of splicing conservation is not clear. To assess tissue-specific transcriptome variation across mammals, we sequenced complementary DNA from nine tissues from four mammals and one bird in biological triplicate, at unprecedented depth. We find that while tissue-specific gene expression programs are largely conserved, alternative splicing is well conserved in only a subset of tissues and is frequently lineage-specific. Thousands of previously unknown, lineage-specific, and conserved alternative exons were identified; widely conserved alternative exons had signatures of binding by MBNL, PTB, RBFOX, STAR, and TIA family splicing factors, implicating them as ancestral mammalian splicing regulators. Our data also indicate that alternative splicing often alters protein phosphorylatability, delimiting the scope of kinase signaling.
Alternative Splicing May Not Be the Key to Proteome Complexity. Tress et al. (2017). Trends in Biochemical Sciences. doi:10.1016/j.tibs.2016.08.008
Abstract
Alternative splicing is commonly believed to be a major source of cellular protein diversity. However, although many thousands of alternatively spliced transcripts are routinely detected in RNA-seq studies, reliable large-scale mass spectrometry-based proteomics analyses identify only a small fraction of annotated alternative isoforms. The clearest finding from proteomics experiments is that most human genes have a single main protein isoform, while those alternative isoforms that are identified tend to be the most biologically plausible: those with the most cross-species conservation and those that do not compromise functional domains. Indeed, most alternative exons do not seem to be under selective pressure, suggesting that a large majority of predicted alternative transcripts may not even be translated into proteins.
Drummond et al. (2005) — why highly expressed proteins evolve slowly.
Why highly expressed proteins evolve slowly. Drummond et al. (2005). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.0504070102
Abstract
Much recent work has explored molecular and population-genetic constraints on the rate of protein sequence evolution. The best predictor of evolutionary rate is expression level, for reasons that have remained unexplained. Here, we hypothesize that selection to reduce the burden of protein misfolding will favor protein sequences with increased robustness to translational missense errors. Pressure for translational robustness increases with expression level and constrains sequence evolution. Using several sequenced yeast genomes, global expression and protein abundance data, and sets of paralogs traceable to an ancient whole-genome duplication in yeast, we rule out several confounding effects and show that expression level explains roughly half the variation in Saccharomyces cerevisiae protein evolutionary rates. We examine causes for expression's dominant role and find that genome-wide tests favor the translational robustness explanation over existing hypotheses that invoke constraints on function or translational efficiency. Our results suggest that proteins evolve at rates largely unrelated to their functions and can explain why highly expressed proteins evolve slowly across the tree of life.
Taverna and Goldstein (2002) — robustness and marginal stability.
Why are proteins so robust to site mutations?. Taverna & Goldstein (2002). Journal of Molecular Biology. doi:10.1006/jmbi.2001.5226
Abstract
There have been repeated observations that proteins are surprisingly robust to site mutations, enduring significant numbers of substitutions with little change in structure, stability, or function. These results are almost paradoxical in light of what is known about random heteropolymers and the sensitivity of their properties to seemingly trivial mutations. To address this discrepancy, the preservation of biological protein properties in the presence of mutation has been interpreted as indicating the independence of selective pressure on such properties. Such results also lead to the prediction that de novo protein design should be relatively easy, in contrast to what is observed. Here, we use a computational model with lattice proteins to demonstrate how this robustness can result from population dynamics during the evolutionary process. As a result, sequence plasticity may be a characteristic of evolutionarily derived proteins and not necessarily a property of designed proteins. This suggests that this robustness must be re-interpreted in evolutionary terms, and has consequences for our understanding of both in vivo and in vitro protein evolution.
Doolittle (2010) — the roots of bioinformatics in protein evolution.
The Roots of Bioinformatics in Protein Evolution. Doolittle (2010). PLoS Computational Biology. doi:10.1371/journal.pcbi.1000875
B.3.5 Selection
Yang and Bielawski (2000) — statistical methods for detecting molecular adaptation.
Statistical methods for detecting molecular adaptation. Yang & Bielawski (2000). Trends in Ecology & Evolution. doi:10.1016/s0169-5347(00)01994-7
Abstract
The past few years have seen the development of powerful statistical methods for detecting adaptive molecular evolution. These methods compare synonymous and nonsynonymous substitution rates in protein-coding genes, and regard a nonsynonymous rate elevated above the synonymous rate as evidence for darwinian selection. Numerous cases of molecular adaptation are being identified in various systems from viruses to humans. Although previous analyses averaging rates over sites and time have little power, recent methods designed to detect positive selection at individual sites and lineages have been successful. Here, we summarize recent statistical methods for detecting molecular adaptation, and discuss their limitations and possible improvements.
Nei and Gojobori (1986), Goldman and Yang (1994) — counting methods and codon models.
Simple methods for estimating the numbers of synonymous and nonsynonymous nucleotide substitutions. Nei & Gojobori (1986). Molecular Biology and Evolution. doi:10.1093/oxfordjournals.molbev.a040410
Abstract
Two simple methods for estimating the numbers of synonymous and nonsynonymous nucleotide substitutions are presented. Although they give no weights to different types of codon substitutions, these methods give essentially the same results as those obtained by Miyata and Yasunaga's and by Li et al.'s methods. Computer simulation indicates that estimates of synonymous substitutions obtained by the two methods are quite accurate unless the number of nucleotide substitutions per site is very large. It is shown that all available methods tend to give an underestimate of the number of nonsynonymous substitutions when the number is large.
A codon-based model of nucleotide substitution for protein-coding DNA sequences. Goldman & Yang (1994). Molecular Biology and Evolution. doi:10.1093/oxfordjournals.molbev.a040153
Abstract
A codon-based model for the evolution of protein-coding DNA sequences is presented for use in phylogenetic estimation. A Markov process is used to describe substitutions between codons. Transition/transversion rate bias and codon usage bias are allowed in the model, and selective restraints at the protein level are accommodated using physicochemical distances between the amino acids coded for by the codons. Analyses of two data sets suggest that the new codon-based model can provide a better fit to data than can nucleotide-based models and can produce more reliable estimates of certain biologically important measures such as the transition/transversion rate ratio and the synonymous/nonsynonymous substitution rate ratio.
Nielsen and Yang (1998), Zhang (2005) — site and branch-site models.
Likelihood Models for Detecting Positively Selected Amino Acid Sites and Applications to the HIV-1 Envelope Gene. Nielsen & Yang (1998). Genetics. doi:10.1093/genetics/148.3.929
Abstract
Several codon-based models for the evolution of protein-coding DNA sequences are developed that account for varying selection intensity among amino acid sites. The "neutral model" assumes two categories of sites at which amino acid replacements are either neutral or deleterious. The "positive-selection model" assumes an additional category of positively selected sites at which nonsynonymous substitutions occur at a higher rate than synonymous ones. This model is also used to identify target sites for positive selection. The models are applied to a data set of the V3 region of the HIV-1 envelope gene, sequenced at different years after the infection of one patient. The results provide strong support for variable selection intensity among amino acid sites The neutral model is rejected in favor of the positive-selection model, indicating the operation of positive selection in the region. Positively selected sites are found in both the V3 region and the flanking regions.
Evaluation of an Improved Branch-Site Likelihood Method for Detecting Positive Selection at the Molecular Level. Zhang (2005). Molecular Biology and Evolution. doi:10.1093/molbev/msi237
Abstract
Detecting positive Darwinian selection at the DNA sequence level has been a subject of considerable interest. However, positive selection is difficult to detect because it often operates episodically on a few amino acid sites, and the signal may be masked by negative selection. Several methods have been developed to test positive selection that acts on given branches (branch methods) or on a subset of sites (site methods). Recently, Yang, Z., and R. Nielsen (2002. Codon-substitution models for detecting molecular adaptation at individual sites along specific lineages. Mol. Biol. Evol. 19:908-917) developed likelihood ratio tests (LRTs) based on branch-site models to detect positive selection that affects a small number of sites along prespecified lineages. However, computer simulations suggested that the tests were sensitive to the model assumptions and were unable to distinguish between relaxation of selective constraint and positive selection (Zhang, J. 2004. Frequent false detection of positive selection by the likelihood method with branch-site models. Mol. Biol. Evol. 21:1332-1339). Here, we describe a modified branch-site model and use it to construct two LRTs, called branch-site tests 1 and 2. We applied the new tests to reanalyze several real data sets and used computer simulation to examine the performance of the two tests by examining their false-positive rate, power, and robustness. We found that test 1 was unable to distinguish relaxed constraint from positive selection affecting the lineages of interest, while test 2 had acceptable false-positive rates and appeared robust against violations of model assumptions. As test 2 is a direct test of positive selection on the lineages of interest, it is referred to as the branch-site test of positive selection and is recommended for use in real data analysis. The test appeared conservative overall, but exhibited better power in detecting positive selection than the branch-based test. Bayes empirical Bayes identification of amino acid sites under positive selection along the foreground branches was found to be reliable, but lacked power.
Yang (2007), Pond et al. (2004) — PAML and HyPhy.
PAML 4: Phylogenetic Analysis by Maximum Likelihood. Yang (2007). Molecular Biology and Evolution. doi:10.1093/molbev/msm088
Abstract
PAML, currently in version 4, is a package of programs for phylogenetic analyses of DNA and protein sequences using maximum likelihood (ML). The programs may be used to compare and test phylogenetic trees, but their main strengths lie in the rich repertoire of evolutionary models implemented, which can be used to estimate parameters in models of sequence evolution and to test interesting biological hypotheses. Uses of the programs include estimation of synonymous and nonsynonymous rates (d(N) and d(S)) between two protein-coding DNA sequences, inference of positive Darwinian selection through phylogenetic comparison of protein-coding genes, reconstruction of ancestral genes and proteins for molecular restoration studies of extinct life forms, combined analysis of heterogeneous data sets from multiple gene loci, and estimation of species divergence times incorporating uncertainties in fossil calibrations. This note discusses some of the major applications of the package, which includes example data sets to demonstrate their use. The package is written in ANSI C, and runs under Windows, Mac OSX, and UNIX systems. It is available at -- (http://abacus.gene.ucl.ac.uk/software/paml.html).
HyPhy: hypothesis testing using phylogenies. Pond et al. (2004). Bioinformatics. doi:10.1093/bioinformatics/bti079
Abstract
Unlabelled The HyPhypackage is designed to provide a flexible and unified platform for carrying out likelihood-based analyses on multiple alignments of molecular sequence data, with the emphasis on studies of rates and patterns of sequence evolution. Availability http://www.hyphy.org Contact muse@stat.ncsu.edu Supplementary information HyPhydocumentation and tutorials are available at http://www.hyphy.org.
McDonald and Kreitman (1991) — the McDonald–Kreitman test.
Adaptive protein evolution at the Adh locus in Drosophila. McDonald & Kreitman (1991). Nature. doi:10.1038/351652a0
Abstract
Proteins often differ in amino-acid sequence across species. This difference has evolved by the accumulation of neutral mutations by random drift, the fixation of adaptive mutations by selection, or a mixture of the two. Here we propose a simple statistical test of the neutral protein evolution hypothesis based on a comparison of the number of amino-acid replacement substitutions to synonymous substitutions in the coding region of a locus. If the observed substitutions are neutral, the ratio of replacement to synonymous fixed differences between species should be the same as the ratio of replacement to synonymous polymorphisms within species. DNA sequence data on the Adh locus (encoding alcohol dehydrogenase, EC 1.1.1.1) in three species in the Drosophila melanogaster species subgroup do not fit this expectation; instead, there are more fixed replacement differences between species than expected. We suggest that these excess replacement substitutions result from adaptive fixation of selectively advantageous mutations.
Kryazhimskiy and Plotkin (2008) — the population genetics of dN/dS.
The Population Genetics of dN/dS. Kryazhimskiy & Plotkin (2008). PLoS Genetics. doi:10.1371/journal.pgen.1000304
Abstract
Evolutionary pressures on proteins are often quantified by the ratio of substitution rates at non-synonymous and synonymous sites. The dN/dS ratio was originally developed for application to distantly diverged sequences, the differences among which represent substitutions that have fixed along independent lineages. Nevertheless, the dN/dS measure is often applied to sequences sampled from a single population, the differences among which represent segregating polymorphisms. Here, we study the expected dN/dS ratio for samples drawn from a single population under selection, and we find that in this context, dN/dS is relatively insensitive to the selection coefficient. Moreover, the hallmark signature of positive selection over divergent lineages, dN/dS>1, is violated within a population. For population samples, the relationship between selection and dN/dS does not follow a monotonic function, and so it may be impossible to infer selection pressures from dN/dS. These results have significant implications for the interpretation of dN/dS measurements among population-genetic samples.
B.3.6 Genome evolution
Ochman et al. (2000), Koonin et al. (2001) — horizontal gene transfer.
Lateral gene transfer and the nature of bacterial innovation. Ochman et al. (2000). Nature. doi:10.1038/35012500
Abstract
Unlike eukaryotes, which evolve principally through the modification of existing genetic information, bacteria have obtained a significant proportion of their genetic diversity through the acquisition of sequences from distantly related organisms. Horizontal gene transfer produces extremely dynamic genomes in which substantial amounts of DNA are introduced into and deleted from the chromosome. These lateral transfers have effectively changed the ecological and pathogenic character of bacterial species.
Horizontal Gene Transfer in Prokaryotes: Quantification and Classification. Koonin et al. (2001). Annual Review of Microbiology. doi:10.1146/annurev.micro.55.1.709
Abstract
Comparative analysis of bacterial, archaeal, and eukaryotic genomes indicates that a significant fraction of the genes in the prokaryotic genomes have been subject to horizontal transfer. In some cases, the amount and source of horizontal gene transfer can be linked to an organism's lifestyle. For example, bacterial hyperthermophiles seem to have exchanged genes with archaea to a greater extent than other bacteria, whereas transfer of certain classes of eukaryotic genes is most common in parasitic and symbiotic bacteria. Horizontal transfer events can be classified into distinct categories of acquisition of new genes, acquisition of paralogs of existing genes, and xenologous gene displacement whereby a gene is displaced by a horizontally transferred ortholog from another lineage (xenolog). Each of these types of horizontal gene transfer is common among prokaryotes, but their relative contributions differ in different lineages. The fixation and long-term persistence of horizontally transferred genes suggests that they confer a selective advantage on the recipient organism. In most cases, the nature of this advantage remains unclear, but detailed examination of several cases of acquisition of eukaryotic genes by bacteria seems to reveal the evolutionary forces involved. Examples include isoleucyl-tRNA synthetases whose acquisition from eukaryotes by several bacteria is linked to antibiotic resistance, ATP/ADP translocases acquired by intracellular parasitic bacteria, Chlamydia and Rickettsia, apparently from plants, and proteases that may be implicated in chlamydial pathogenesis.
W. F. Doolittle (1999) — the tree of life questioned.
Phylogenetic Classification and the Universal Tree. Doolittle (1999). Science. doi:10.1126/science.284.5423.2124
Abstract
From comparative analyses of the nucleotide sequences of genes encoding ribosomal RNAs and several proteins, molecular phylogeneticists have constructed a "universal tree of life," taking it as the basis for a "natural" hierarchical classification of all living things. Although confidence in some of the tree's early branches has recently been shaken, new approaches could still resolve many methodological uncertainties. More challenging is evidence that most archaeal and bacterial genomes (and the inferred ancestral eukaryotic nuclear genome) contain genes from multiple sources. If "chimerism" or "lateral gene transfer" cannot be dismissed as trivial in extent or limited to special categories of genes, then no hierarchical universal classification can be taken as natural. Molecular phylogeneticists will have failed to find the "true tree," not because their methods are inadequate or because they have chosen the wrong genes, but because the history of life cannot properly be represented as a tree. However, taxonomies based on molecular sequences will remain indispensable, and understanding of the evolutionary process will ultimately be enriched, not impoverished.
Tettelin et al. (2005) — the pan-genome.
Genome analysis of multiple pathogenic isolates of Streptococcus agalactiae : Implications for the microbial “pan-genome”. Tettelin et al. (2005). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.0506758102
Abstract
The development of efficient and inexpensive genome sequencing methods has revolutionized the study of human bacterial pathogens and improved vaccine design. Unfortunately, the sequence of a single genome does not reflect how genetic variability drives pathogenesis within a bacterial species and also limits genome-wide screens for vaccine candidates or for antimicrobial targets. We have generated the genomic sequence of six strains representing the five major disease-causing serotypes of Streptococcus agalactiae, the main cause of neonatal infection in humans. Analysis of these genomes and those available in databases showed that the S. agalactiae species can be described by a pan-genome consisting of a core genome shared by all isolates, accounting for approximately 80% of any single genome, plus a dispensable genome consisting of partially shared and strain-specific genes. Mathematical extrapolation of the data suggests that the gene reservoir available for inclusion in the S. agalactiae pan-genome is vast and that unique genes will continue to be identified even after sequencing hundreds of genomes.
Huynen and Nimwegen (1998), Qian et al. (2001) — power-law distributions of family sizes.
The frequency distribution of gene family sizes in complete genomes. Huynen & van Nimwegen (1998). Molecular Biology and Evolution. doi:10.1093/oxfordjournals.molbev.a025959
Abstract
We compare the frequency distribution of gene family sizes in the complete genomes of six bacteria (Escherichia coli, Haemophilus influenzae, Helicobacter pylori, Mycoplasma genitalium, Mycoplasma pneumoniae, and Synechocystis sp. PCC6803), two Archaea (Methanococcus jannaschii and Methanobacterium thermoautotrophicum), one eukaryote (Saccharomyces cerevisiae), the vaccinia virus, and the bacteriophage T4. The sizes of the gene families versus their frequencies show power-law distributions that tend to become flatter (have a larger exponent) as the number of genes in the genome increases. Power-law distributions generally occur as the limit distribution of a multiplicative stochastic process with a boundary constraint. We discuss various models that can account for a multiplicative process determining the sizes of gene families in the genome. In particular, we argue that, in order to explain the observed distributions, gene families have to behave in a coherent fashion within the genome; i.e., the probabilities of duplications of genes within a gene family are not independent of each other. Likewise, the probabilities of deletions of genes within a gene family are not independent of each other.
Protein family and fold occurrence in genomes: power-law behaviour and evolutionary model. Qian et al. (2001). Journal of Molecular Biology. doi:10.1006/jmbi.2001.5079
Abstract
Global surveys of genomes measure the usage of essential molecular parts, defined here as protein families, superfamilies or folds, in different organisms. Based on surveys of the first 20 completely sequenced genomes, we observe that the occurrence of these parts follows a power-law distribution. That is, the number of distinct parts (F) with a given genomic occurrence (V) decays as F=aV(-b), with a few parts occurring many times and most occurring infrequently. For a given organism, the distributions of families, superfamilies and folds are nearly identical, and this is reflected in the size of the decay exponent b. Moreover, the exponent varies between different organisms, with those of smaller genomes displaying a steeper decay (i.e. larger b). Clearly, the power law indicates a preference to duplicate genes that encode for molecular parts which are already common. Here, we present a minimal, but biologically meaningful model that accurately describes the observed power law. Although the model performs equally well for all three protein classes, we focus on the occurrence of folds in preference to families and superfamilies. This is because folds are comparatively insensitive to the effects of point mutations that can cause a family member to diverge beyond detectable similarity. In the model, genomes evolve through two basic operations: (i) duplication of existing genes; (ii) net flow of new genes. The flow term is closely related to the exponent b and can accommodate considerable gene loss; however, we demonstrate that the observed data is reproduced best with a net inflow, i.e. with more gene gain than loss. Moreover, we show that prokaryotes have much higher rates of gene acquisition than eukaryotes, probably reflecting lateral transfer. A further natural outcome from our model is an estimation of the fold composition of the initial genome, which potentially relates to the common ancestor for modern organisms. Supplementary material pertaining to this work is available from www.partslist.org/powerlaw.
Koonin et al. (2002), Nimwegen (2003) — models and scaling laws.
The structure of the protein universe and genome evolution. Koonin et al. (2002). Nature. doi:10.1038/nature01256
Abstract
Despite the practically unlimited number of possible protein sequences, the number of basic shapes in which proteins fold seems not only to be finite, but also to be relatively small, with probably no more than 10,000 folds in existence. Moreover, the distribution of proteins among these folds is highly non-homogeneous -- some folds and superfamilies are extremely abundant, but most are rare. Protein folds and families encoded in diverse genomes show similar size distributions with notable mathematical properties, which also extend to the number of connections between domains in multidomain proteins. All these distributions follow asymptotic power laws, such as have been identified in a wide variety of biological and physical systems, and which are typically associated with scale-free networks. These findings suggest that genome evolution is driven by extremely general mechanisms based on the preferential attachment principle.
Scaling laws in the functional content of genomes. van Nimwegen (2003). Trends in Genetics. doi:10.1016/s0168-9525(03)00203-8
Koonin (2011) — are there laws of genome evolution?
Are There Laws of Genome Evolution?. Koonin (2011). PLoS Computational Biology. doi:10.1371/journal.pcbi.1002173
Abstract
Research in quantitative evolutionary genomics and systems biology led to the discovery of several universal regularities connecting genomic and molecular phenomic variables. These universals include the log-normal distribution of the evolutionary rates of orthologous genes; the power law-like distributions of paralogous family size and node degree in various biological networks; the negative correlation between a gene's sequence evolution rate and expression level; and differential scaling of functional classes of genes with genome size. The universals of genome evolution can be accounted for by simple mathematical models similar to those used in statistical physics, such as the birth-death-innovation model. These models do not explicitly incorporate selection; therefore, the observed universal regularities do not appear to be shaped by selection but rather are emergent properties of gene ensembles. Although a complete physical theory of evolutionary biology is inconceivable, the universals of genome evolution might qualify as "laws of evolutionary genomics" in the same sense "law" is understood in modern physics.
B.4 Phylogeny
Further papers for Chapter 4.
Papers marked ★ were mandatory in earlier reading lists. In the HTML version, click Abstract to show the abstract.
B.4.1 Reviews and books
Yang and Rannala (2012) — molecular phylogenetics: principles and practice.
Molecular phylogenetics: principles and practice. Yang & Rannala (2012). Nature Reviews Genetics. doi:10.1038/nrg3186
Abstract
Phylogenies are important for addressing various biological questions such as relationships among species or genes, the origin and spread of viral infection and the demographic changes and migration patterns of species. The advancement of sequencing technologies has taken phylogenetic analysis to a new height. Phylogenies have permeated nearly every branch of biology, and the plethora of phylogenetic methods and software packages that are now available may seem daunting to an experimental biologist. Here, we review the major methods of phylogenetic analysis, including parsimony, distance, likelihood and Bayesian methods. We discuss their strengths and weaknesses and provide guidance for their use.
Felsenstein (2004) — Inferring Phylogenies, the standard textbook, with much of the history.
Inferring Phylogenies. Felsenstein (2004). Sinauer Associates.
Delsuc et al. (2005) — phylogenomics and the reconstruction of the tree of life.
Phylogenomics and the reconstruction of the tree of life. Delsuc et al. (2005). Nature Reviews Genetics. doi:10.1038/nrg1603
Abstract
As more complete genomes are sequenced, phylogenetic analysis is entering a new era - that of phylogenomics. One branch of this expanding field aims to reconstruct the evolutionary history of organisms on the basis of the analysis of their genomes. Recent studies have demonstrated the power of this approach, which has the potential to provide answers to several fundamental evolutionary questions. However, challenges for the future have also been revealed. The very nature of the evolutionary history of organisms and the limitations of current phylogenetic reconstruction methods mean that part of the tree of life might prove difficult, if not impossible, to resolve with confidence.
Steel and Penny (2000) — parsimony, likelihood and the role of models.
Parsimony, Likelihood, and the Role of Models in Molecular Phylogenetics. Steel & Penny (2000). Molecular Biology and Evolution. doi:10.1093/oxfordjournals.molbev.a026364
Abstract
Methods such as maximum parsimony (MP) are frequently criticized as being statistically unsound and not being based on any "model." On the other hand, advocates of MP claim that maximum likelihood (ML) has some fundamental problems. Here, we explore the connection between the different versions of MP and ML methods, particularly in light of recent theoretical results. We describe links between the two methods--for example, we describe how MP can be regarded as an ML method when there is no common mechanism between sites (such as might occur with morphological data and certain forms of molecular data). In the process, we clarify certain historical points of disagreement between proponents of the two methodologies, including a discussion of several forms of the ML optimality criterion. We also describe some additional results that shed light on how much needs to be assumed about underlying models of sequence evolution in order to successfully reconstruct evolutionary trees.
Steenwyk et al. (2023) — incongruence in the phylogenomics era.
Incongruence in the phylogenomics era. Steenwyk et al. (2023). Nature Reviews Genetics. doi:10.1038/s41576-023-00620-x
Abstract
Genome-scale data and the development of novel statistical phylogenetic approaches have greatly aided the reconstruction of a broad sketch of the tree of life and resolved many of its branches. However, incongruence - the inference of conflicting evolutionary histories - remains pervasive in phylogenomic data, hampering our ability to reconstruct and interpret the tree of life. Biological factors, such as incomplete lineage sorting, horizontal gene transfer, hybridization, introgression, recombination and convergent molecular evolution, can lead to gene phylogenies that differ from the species tree. In addition, analytical factors, including stochastic, systematic and treatment errors, can drive incongruence. Here, we review these factors, discuss methodological advances to identify and handle incongruence, and highlight avenues for future research.
B.4.2 History and classic methods
Hennig (1965) — phylogenetic systematics (cladistics).
Phylogenetic Systematics. Hennig (1965). Annual Review of Entomology. doi:10.1146/annurev.en.10.010165.000525
Fitch and Margoliash (1967) — one of the first molecular trees.
Construction of Phylogenetic Trees: A method based on mutation distances as estimated from cytochrome c sequences is of general applicability.. Fitch & Margoliash (1967). Science. doi:10.1126/science.155.3760.279
Fitch (1971) — minimum change on a given topology (parsimony).
Toward Defining the Course of Evolution: Minimum Change for a Specific Tree Topology. Fitch (1971). Systematic Zoology. doi:10.2307/2412116
Sankoff (1975) — weighted parsimony.
Minimal Mutation Trees of Sequences. Sankoff (1975). SIAM Journal on Applied Mathematics. doi:10.1137/0128004
Abstract
Given a finite tree, some of whose vertices are identified with given finite sequences, we show how to construct sequences for all the remaining vertices simultaneously, so as to minimize the total edge-length of the tree. Edge-length is calculated by a metric whose biological significance is the mutational distance between two sequences.
Felsenstein (1978) — when parsimony is positively misleading: long-branch attraction.
Cases in which Parsimony or Compatibility Methods will be Positively Misleading. Felsenstein (1978). Systematic Biology. doi:10.1093/sysbio/27.4.401
Abstract
For some simple three- and four-species cases involving a character with two states, it is determined under what conditions several methods of phylogenetic inference will fail to converge to the true phylogeny as more and more data are accumulated. The methods are the Camin-Sokal parsimony method, the compatibility method, and Farris's unrooted Wagner tree parsimony method. In all cases the conditions for this failure (which is the failure to be statistically consistent) are essentially that parallel changes exceed informative, nonparallel changes. It is possible for these methods to be inconsistent even when change is improbable a priori, provided that evolutionary rates in different lineages are sufficiently unequal. It is by extension of this approach that we may provide a sound methodology for evaluating methods of phylogenetic inference.
★ Felsenstein (1985) — confidence limits on phylogenies: the bootstrap.
CONFIDENCE LIMITS ON PHYLOGENIES: AN APPROACH USING THE BOOTSTRAP. Felsenstein (1985). Evolution. doi:10.1111/j.1558-5646.1985.tb00420.x
Abstract
The recently-developed statistical method known as the "bootstrap" can be used to place confidence intervals on phylogenies. It involves resampling points from one's own data, with replacement, to create a series of bootstrap samples of the same size as the original data. Each of these is analyzed, and the variation among the resulting estimates taken to indicate the size of the error involved in making estimates from the original data. In the case of phylogenies, it is argued that the proper method of resampling is to keep all of the original species while sampling characters with replacement, under the assumption that the characters have been independently drawn by the systematist and have evolved independently. Majority-rule consensus trees can be used to construct a phylogeny showing all of the inferred monophyletic groups that occurred in a majority of the bootstrap samples. If a group shows up 95% of the time or more, the evidence for it is taken to be statistically significant. Existing computer programs can be used to analyze different bootstrap samples by using weights on the characters, the weight of a character being how many times it was drawn in bootstrap sampling. When all characters are perfectly compatible, as envisioned by Hennig, bootstrap sampling becomes unnecessary; the bootstrap method would show significant evidence for a group if it is defined by three or more characters.
Studier and Keppler (1988), Gascuel (2006) — what neighbour joining does and why it works.
A note on the neighbor-joining algorithm of Saitou and Nei. Studier & Keppler (1988). Molecular Biology and Evolution. doi:10.1093/oxfordjournals.molbev.a040527
Neighbor-Joining Revealed. Gascuel (2006). Molecular Biology and Evolution. doi:10.1093/molbev/msl072
Abstract
It is nearly 20 years since the landmark paper (Saitou and Nei 1987) in Molecular Biology and Evolution introducing Neighbor-Joining (NJ). The method has become the most widely used method for building phylogenetic trees from distances, and the original paper has been cited about 13,000 times (Science Citation Index). Yet the question "what does the NJ method seek to do?" has until recently proved somewhat elusive, leading to some imprecise claims and misunderstanding. However, a rigorous answer to this question has recently been provided by further mathematical investigation, and the purpose of this note is to highlight these results and their significance for interpreting NJ. The origins of this story lie in a paper by Pauplin (2000) though its continuation has unfolded in more mathematically inclined literature. Our aim here is to make these findings more widely accessible.
Desper and Gascuel (2002) — FastME and the minimum-evolution criterion.
Fast and Accurate Phylogeny Reconstruction Algorithms Based on the Minimum-Evolution Principle. Desper & Gascuel (2002). Journal of Computational Biology. doi:10.1089/106652702761034136
Abstract
The Minimum Evolution (ME) approach to phylogeny estimation has been shown to be statistically consistent when it is used in conjunction with ordinary least-squares (OLS) fitting of a metric to a tree structure. The traditional approach to using ME has been to start with the Neighbor Joining (NJ) topology for a given matrix and then do a topological search from that starting point. The first stage requires O(n(3)) time, where n is the number of taxa, while the current implementations of the second are in O(p n(3)) or more, where p is the number of swaps performed by the program. In this paper, we examine a greedy approach to minimum evolution which produces a starting topology in O(n(2)) time. Moreover, we provide an algorithm that searches for the best topology using nearest neighbor interchanges (NNIs), where the cost of doing p NNIs is O(n(2) + p n), i.e., O(n(2)) in practice because p is always much smaller than n. The Greedy Minimum Evolution (GME) algorithm, when used in combination with NNIs, produces trees which are fairly close to NJ trees in terms of topological accuracy. We also examine ME under a balanced weighting scheme, where sibling subtrees have equal weight, as opposed to the standard "unweighted" OLS, where all taxa have the same weight so that the weight of a subtree is equal to the number of its taxa. The balanced minimum evolution scheme (BME) runs slower than the OLS version, requiring O(n(2) x diam(T)) operations to build the starting tree and O(p n x diam(T)) to perform the NNIs, where diam(T) is the topological diameter of the output tree. In the usual Yule-Harding distribution on phylogenetic trees, the diameter expectation is in log(n), so our algorithms are in practice faster that NJ. Moreover, this BME scheme yields a very significant improvement over NJ and other distance-based algorithms, especially with large trees, in terms of topological accuracy.
Rannala and Yang (1996), Huelsenbeck et al. (2001) — Bayesian inference of phylogeny.
Probability distribution of molecular evolutionary trees: A new method of phylogenetic inference. Rannala & Yang (1996). Journal of Molecular Evolution. doi:10.1007/bf02338839
Abstract
A new method is presented for inferring evolutionary trees using nucleotide sequence data. The birth-death process is used as a model of speciation and extinction to specify the prior distribution of phylogenies and branching times. Nucleotide substitution is modeled by a continuous-time Markov process. Parameters of the branching model and the substitution model are estimated by maximum likelihood. The posterior probabilities of different phylogenies are calculated and the phylogeny with the highest posterior probability is chosen as the best estimate of the evolutionary relationship among species. We refer to this as the maximum posterior probability (MAP) tree. The posterior probability provides a natural measure of the reliability of the estimated phylogeny. Two example data sets are analyzed to infer the phylogenetic relationship of human, chimpanzee, gorilla, and orangutan. The best trees estimated by the new method are the same as those from the maximum likelihood analysis of separate topologies, but the posterior probabilities are quite different from the bootstrap proportions. The results of the method are found to be insensitive to changes in the rate parameter of the branching process.
Bayesian Inference of Phylogeny and Its Impact on Evolutionary Biology. Huelsenbeck et al. (2001). Science. doi:10.1126/science.1065889
Abstract
As a discipline, phylogenetics is becoming transformed by a flood of molecular data. These data allow broad questions to be asked about the history of life, but also present difficult statistical and computational problems. Bayesian inference of phylogeny brings a new perspective to a number of outstanding issues in evolutionary biology, including the analysis of large phylogenetic trees and complex evolutionary models and the detection of the footprint of natural selection in DNA sequences.
B.4.3 Substitution models and support
Jukes and Cantor (1969), Kimura (1980), Hasegawa et al. (1985) — the classical nucleotide models.
Evolution of Protein Molecules. Jukes & Cantor (1969). Mammalian Protein Metabolism. doi:10.1016/b978-1-4832-3211-9.50009-7
A simple method for estimating evolutionary rates of base substitutions through comparative studies of nucleotide sequences. Kimura (1980). Journal of Molecular Evolution. doi:10.1007/bf01731581
Abstract
Some simple formulae were obtained which enable us to estimate evolutionary distances in terms of the number of nucleotide substitutions (and, also, the evolutionary rates when the divergence times are known). In comparing a pair of nucleotide sequences, we distinguish two types of differences; if homologous sites are occupied by different nucleotide bases but both are purines or both pyrimidines, the difference is called type I (or "transition" type), while, if one of the two is a purine and the other is a pyrimidine, the difference is called type II (or "transversion" type). Letting P and Q be respectively the fractions of nucleotide sites showing type I and type II differences between two sequences compared, then the evolutionary distance per site is K = -(1/2) ln [(1-2P-Q) square root of 1-2Q]. The evolutionary rate per year is then given by k = K/(2T), where T is the time since the divergence of the two sequences. If only the third codon positions are compared, the synonymous component of the evolutionary base substitutions per site is estimated by K'S = -(1/2) ln (1-2P-Q). Also, formulae for standard errors were obtained. Some examples were worked out using reported globin sequences to show that synonymous substitutions occur at much higher rates than amino acid-altering substitutions in evolution.
Dating of the human-ape splitting by a molecular clock of mitochondrial DNA. Hasegawa et al. (1985). Journal of Molecular Evolution. doi:10.1007/bf02101694
Abstract
A new statistical method for estimating divergence dates of species from DNA sequence data by a molecular clock approach is developed. This method takes into account effectively the information contained in a set of DNA sequence data. The molecular clock of mitochondrial DNA (mtDNA) was calibrated by setting the date of divergence between primates and ungulates at the Cretaceous-Tertiary boundary (65 million years ago), when the extinction of dinosaurs occurred. A generalized least-squares method was applied in fitting a model to mtDNA sequence data, and the clock gave dates of 92.3 +/- 11.7, 13.3 +/- 1.5, 10.9 +/- 1.2, 3.7 +/- 0.6, and 2.7 +/- 0.6 million years ago (where the second of each pair of numbers is the standard deviation) for the separation of mouse, gibbon, orangutan, gorilla, and chimpanzee, respectively, from the line leading to humans. Although there is some uncertainty in the clock, this dating may pose a problem for the widely believed hypothesis that the pipedal creature Australopithecus afarensis, which lived some 3.7 million years ago at Laetoli in Tanzania and at Hadar in Ethiopia, was ancestral to man and evolved after the human-ape splitting. Another likelier possibility is that mtDNA was transferred through hybridization between a proto-human and a proto-chimpanzee after the former had developed bipedalism.
Yang (1994) — gamma-distributed rates across sites.
Maximum likelihood phylogenetic estimation from DNA sequences with variable rates over sites: Approximate methods. Yang (1994). Journal of Molecular Evolution. doi:10.1007/bf00160154
Abstract
Two approximate methods are proposed for maximum likelihood phylogenetic estimation, which allow variable rates of substitution across nucleotide sites. Three data sets with quite different characteristics were analyzed to examine empirically the performance of these methods. The first, called the "discrete gamma model," uses several categories of rates to approximate the gamma distribution, with equal probability for each category. The mean of each category is used to represent all the rates falling in the category. The performance of this method is found to be quite good, and four such categories appear to be sufficient to produce both an optimum, or near-optimum fit by the model to the data, and also an acceptable approximation to the continuous distribution. The second method, called "fixed-rates model", classifies sites into several classes according to their rates predicted assuming the star tree. Sites in different classes are then assumed to be evolving at these fixed rates when other tree topologies are evaluated. Analyses of the data sets suggest that this method can produce reasonable results, but it seems to share some properties of a least-squares pairwise comparison; for example, interior branch lengths in nonbest trees are often found to be zero. The computational requirements of the two methods are comparable to that of Felsenstein's (1981, J Mol Evol 17:368-376) model, which assumes a single rate for all the sites.
Whelan and Goldman (2001), Le and Gascuel (2008) — empirical protein models (WAG, LG).
A General Empirical Model of Protein Evolution Derived from Multiple Protein Families Using a Maximum-Likelihood Approach. Whelan & Goldman (2001). Molecular Biology and Evolution. doi:10.1093/oxfordjournals.molbev.a003851
Abstract
Phylogenetic inference from amino acid sequence data uses mainly empirical models of amino acid replacement and is therefore dependent on those models. Two of the more widely used models, the Dayhoff and JTT models, are estimated using similar methods that can utilize large numbers of sequences from many unrelated protein families but are somewhat unsatisfactory because they rely on assumptions that may lead to systematic error and discard a large amount of the information within the sequences. The alternative method of maximum-likelihood estimation may utilize the information in the sequence data more efficiently and suffers from no systematic error, but it has previously been applicable to relatively few sequences related by a single phylogenetic tree. Here, we combine the best attributes of these two methods using an approximate maximum-likelihood method. We implemented this approach to estimate a new model of amino acid replacement from a database of globular protein sequences comprising 3,905 amino acid sequences split into 182 protein families. While the new model has an overall structure similar to those of other commonly used models, there are significant differences. The new model outperforms the Dayhoff and JTT models with respect to maximum-likelihood values for a large majority of the protein families in our database. This suggests that it provides a better overall fit to the evolutionary process in globular proteins and may lead to more accurate phylogenetic tree estimates. Potentially, this matrix, and the methods used to generate it, may also be useful in other areas of research, such as biological sequence database searching, sequence alignment, and protein structure prediction, for which an accurate description of amino acid replacement is required.
An Improved General Amino Acid Replacement Matrix. Le & Gascuel (2008). Molecular Biology and Evolution. doi:10.1093/molbev/msn067
Abstract
Amino acid replacement matrices are an essential basis of protein phylogenetics. They are used to compute substitution probabilities along phylogeny branches and thus the likelihood of the data. They are also essential in protein alignment. A number of replacement matrices and methods to estimate these matrices from protein alignments have been proposed since the seminal work of Dayhoff et al. (1972). An important advance was achieved by Whelan and Goldman (2001) and their WAG matrix, thanks to an efficient maximum likelihood estimation approach that accounts for the phylogenies of sequences within each training alignment. We further refine this method by incorporating the variability of evolutionary rates across sites in the matrix estimation and using a much larger and diverse database than BRKALN, which was used to estimate WAG. To estimate our new matrix (called LG after the authors), we use an adaptation of the XRATE software and 3,912 alignments from Pfam, comprising approximately 50,000 sequences and approximately 6.5 million residues overall. To evaluate the LG performance, we use an independent sample consisting of 59 alignments from TreeBase and randomly divide Pfam alignments into 3,412 training and 500 test alignments. The comparison with WAG and JTT shows a clear likelihood improvement. With TreeBase, we find that 1) the average Akaike information criterion gain per site is 0.25 and 0.42, when compared with WAG and JTT, respectively; 2) LG is significantly better than WAG for 38 alignments (among 59), and significantly worse with 2 alignments only; and 3) tree topologies inferred with LG, WAG, and JTT frequently differ, indicating that using LG impacts not only the likelihood value but also the output tree. Results with the test alignments from Pfam are analogous. LG and a PHYML implementation can be downloaded from http://atgc.lirmm.fr/LG.
Lartillot and Philippe (2004) — the CAT site-heterogeneous mixture model.
A Bayesian Mixture Model for Across-Site Heterogeneities in the Amino-Acid Replacement Process. Lartillot & Philippe (2004). Molecular Biology and Evolution. doi:10.1093/molbev/msh112
Abstract
Most current models of sequence evolution assume that all sites of a protein evolve under the same substitution process, characterized by a 20 x 20 substitution matrix. Here, we propose to relax this assumption by developing a Bayesian mixture model that allows the amino-acid replacement pattern at different sites of a protein alignment to be described by distinct substitution processes. Our model, named CAT, assumes the existence of distinct processes (or classes) differing by their equilibrium frequencies over the 20 residues. Through the use of a Dirichlet process prior, the total number of classes and their respective amino-acid profiles, as well as the affiliations of each site to a given class, are all free variables of the model. In this way, the CAT model is able to adapt to the complexity actually present in the data, and it yields an estimate of the substitutional heterogeneity through the posterior mean number of classes. We show that a significant level of heterogeneity is present in the substitution patterns of proteins, and that the standard one-matrix model fails to account for this heterogeneity. By evaluating the Bayes factor, we demonstrate that the standard model is outperformed by CAT on all of the data sets which we analyzed. Altogether, these results suggest that the complexity of the pattern of substitution of real sequences is better captured by the CAT model, offering the possibility of studying its impact on phylogenetic reconstruction and its connections with structure-function determinants.
Posada and Crandall (1998), Kalyaanamoorthy et al. (2017) — model selection (ModelTest, ModelFinder).
MODELTEST: testing the model of DNA substitution.. Posada & Crandall (1998). Bioinformatics. doi:10.1093/bioinformatics/14.9.817
Abstract
The program MODELTEST uses log likelihood scores to establish the model of DNA evolution that best fits the data. Availability The MODELTEST package, including the source code and some documentation is available at http://bioag.byu. edu/zoology/crandall_lab/modeltest.html.
ModelFinder: fast model selection for accurate phylogenetic estimates. Kalyaanamoorthy et al. (2017). Nature Methods. doi:10.1038/nmeth.4285
Abstract
Model-based molecular phylogenetics plays an important role in comparisons of genomic data, and model selection is a key step in all such analyses. We present ModelFinder, a fast model-selection method that greatly improves the accuracy of phylogenetic estimates by incorporating a model of rate heterogeneity across sites not previously considered in this context and by allowing concurrent searches of model space and tree space.
Hoang et al. (2017), Anisimova and Gascuel (2006), Lemoine et al. (2018) — fast and alternative measures of support.
UFBoot2: Improving the Ultrafast Bootstrap Approximation. Hoang et al. (2017). Molecular Biology and Evolution. doi:10.1093/molbev/msx281
Abstract
The standard bootstrap (SBS), despite being computationally intensive, is widely used in maximum likelihood phylogenetic analyses. We recently proposed the ultrafast bootstrap approximation (UFBoot) to reduce computing time while achieving more unbiased branch supports than SBS under mild model violations. UFBoot has been steadily adopted as an efficient alternative to SBS and other bootstrap approaches. Here, we present UFBoot2, which substantially accelerates UFBoot and reduces the risk of overestimating branch supports due to polytomies or severe model violations. Additionally, UFBoot2 provides suitable bootstrap resampling strategies for phylogenomic data. UFBoot2 is 778 times (median) faster than SBS and 8.4 times (median) faster than RAxML rapid bootstrap on tested data sets. UFBoot2 is implemented in the IQ-TREE software package version 1.6 and freely available at http://www.iqtree.org.
Approximate Likelihood-Ratio Test for Branches: A Fast, Accurate, and Powerful Alternative. Anisimova & Gascuel (2006). Systematic Biology. doi:10.1080/10635150600755453
Abstract
We revisit statistical tests for branches of evolutionary trees reconstructed upon molecular data. A new, fast, approximate likelihood-ratio test (aLRT) for branches is presented here as a competitive alternative to nonparametric bootstrap and Bayesian estimation of branch support. The aLRT is based on the idea of the conventional LRT, with the null hypothesis corresponding to the assumption that the inferred branch has length 0. We show that the LRT statistic is asymptotically distributed as a maximum of three random variables drawn from the chi(0)2 + chi(1)2 distribution. The new aLRT of interior branch uses this distribution for significance testing, but the test statistic is approximated in a slightly conservative but practical way as 2(l1- l2), i.e., double the difference between the maximum log-likelihood values corresponding to the best tree and the second best topological arrangement around the branch of interest. Such a test is fast because the log-likelihood value l2 is computed by optimizing only over the branch of interest and the four adjacent branches, whereas other parameters are fixed at their optimal values corresponding to the best ML tree. The performance of the new test was studied on simulated 4-, 12-, and 100-taxon data sets with sequences of different lengths. The aLRT is shown to be accurate, powerful, and robust to certain violations of model assumptions. The aLRT is implemented within the algorithm used by the recent fast maximum likelihood tree estimation program PHYML (Guindon and Gascuel, 2003).
Renewing Felsenstein’s phylogenetic bootstrap in the era of big data. Lemoine et al. (2018). Nature. doi:10.1038/s41586-018-0043-0
Abstract
Felsenstein's application of the bootstrap method to evolutionary trees is one of the most cited scientific papers of all time. The bootstrap method, which is based on resampling and replications, is used extensively to assess the robustness of phylogenetic inferences. However, increasing numbers of sequences are now available for a wide variety of species, and phylogenies based on hundreds or thousands of taxa are becoming routine. With phylogenies of this size Felsenstein's bootstrap tends to yield very low supports, especially on deep branches. Here we propose a new version of the phylogenetic bootstrap in which the presence of inferred branches in replications is measured using a gradual 'transfer' distance rather than the binary presence or absence index used in Felsenstein's original version. The resulting supports are higher and do not induce falsely supported branches. The application of our method to large mammal, HIV and simulated datasets reveals their phylogenetic signals, whereas Felsenstein's bootstrap fails to do so.
Shimodaira (2002) — the AU test for comparing trees.
An Approximately Unbiased Test of Phylogenetic Tree Selection. Shimodaira (2002). Systematic Biology. doi:10.1080/10635150290069913
Abstract
An approximately unbiased (AU) test that uses a newly devised multiscale bootstrap technique was developed for general hypothesis testing of regions in an attempt to reduce test bias. It was applied to maximum-likelihood tree selection for obtaining the confidence set of trees. The AU test is based on the theory of Efron et al. (Proc. Natl. Acad. Sci. USA 93:13429-13434; 1996), but the new method provides higher-order accuracy yet simpler implementation. The AU test, like the Shimodaira-Hasegawa (SH) test, adjusts the selection bias overlooked in the standard use of the bootstrap probability and Kishino-Hasegawa tests. The selection bias comes from comparing many trees at the same time and often leads to overconfidence in the wrong trees. The SH test, though safe to use, may exhibit another type of bias such that it appears conservative. Here I show that the AU test is less biased than other methods in typical cases of tree selection. These points are illustrated in a simulation study as well as in the analysis of mammalian mitochondrial protein sequences. The theoretical argument provides a simple formula that covers the bootstrap probability test, the Kishino-Hasegawa test, the AU test, and the Zharkikh-Li test. A practical suggestion is provided as to which test should be used under particular circumstances.
B.4.4 Software
Stamatakis (2014) — RAxML.
RAxML version 8: a tool for phylogenetic analysis and post-analysis of large phylogenies. Stamatakis (2014). Bioinformatics. doi:10.1093/bioinformatics/btu033
Abstract
Motivation Phylogenies are increasingly used in all fields of medical and biological research. Moreover, because of the next-generation sequencing revolution, datasets used for conducting phylogenetic analyses grow at an unprecedented pace. RAxML (Randomized Axelerated Maximum Likelihood) is a popular program for phylogenetic analyses of large datasets under maximum likelihood. Since the last RAxML paper in 2006, it has been continuously maintained and extended to accommodate the increasingly growing input datasets and to serve the needs of the user community. Results I present some of the most notable new features and extensions of RAxML, such as a substantial extension of substitution models and supported data types, the introduction of SSE3, AVX and AVX2 vector intrinsics, techniques for reducing the memory requirements of the code and a plethora of operations for conducting post-analyses on sets of trees. In addition, an up-to-date 50-page user manual covering all new RAxML options is available.
Guindon et al. (2010) — PhyML 3.0.
New Algorithms and Methods to Estimate Maximum-Likelihood Phylogenies: Assessing the Performance of PhyML 3.0. Guindon et al. (2010). Systematic Biology. doi:10.1093/sysbio/syq010
Abstract
PhyML is a phylogeny software based on the maximum-likelihood principle. Early PhyML versions used a fast algorithm performing nearest neighbor interchanges to improve a reasonable starting tree topology. Since the original publication (Guindon S., Gascuel O. 2003. A simple, fast and accurate algorithm to estimate large phylogenies by maximum likelihood. Syst. Biol. 52:696-704), PhyML has been widely used (>2500 citations in ISI Web of Science) because of its simplicity and a fair compromise between accuracy and speed. In the meantime, research around PhyML has continued, and this article describes the new algorithms and methods implemented in the program. First, we introduce a new algorithm to search the tree space with user-defined intensity using subtree pruning and regrafting topological moves. The parsimony criterion is used here to filter out the least promising topology modifications with respect to the likelihood function. The analysis of a large collection of real nucleotide and amino acid data sets of various sizes demonstrates the good performance of this method. Second, we describe a new test to assess the support of the data for internal branches of a phylogeny. This approach extends the recently proposed approximate likelihood-ratio test and relies on a nonparametric, Shimodaira-Hasegawa-like procedure. A detailed analysis of real alignments sheds light on the links between this new approach and the more classical nonparametric bootstrap method. Overall, our tests show that the last version (3.0) of PhyML is fast, accurate, stable, and ready to use. A Web server and binary files are available from http://www.atgc-montpellier.fr/phyml/.
Price et al. (2010) — FastTree 2.
FastTree 2 – Approximately Maximum-Likelihood Trees for Large Alignments. Price et al. (2010). PLoS ONE. doi:10.1371/journal.pone.0009490
Abstract
Background We recently described FastTree, a tool for inferring phylogenies for alignments with up to hundreds of thousands of sequences. Here, we describe improvements to FastTree that improve its accuracy without sacrificing scalability. Methodology/principal findings Where FastTree 1 used nearest-neighbor interchanges (NNIs) and the minimum-evolution criterion to improve the tree, FastTree 2 adds minimum-evolution subtree-pruning-regrafting (SPRs) and maximum-likelihood NNIs. FastTree 2 uses heuristics to restrict the search for better trees and estimates a rate of evolution for each site (the "CAT" approximation). Nevertheless, for both simulated and genuine alignments, FastTree 2 is slightly more accurate than a standard implementation of maximum-likelihood NNIs (PhyML 3 with default settings). Although FastTree 2 is not quite as accurate as methods that use maximum-likelihood SPRs, most of the splits that disagree are poorly supported, and for large alignments, FastTree 2 is 100-1,000 times faster. FastTree 2 inferred a topology and likelihood-based local support values for 237,882 distinct 16S ribosomal RNAs on a desktop computer in 22 hours and 5.8 gigabytes of memory. Conclusions/significance FastTree 2 allows the inference of maximum-likelihood phylogenies for huge alignments. FastTree 2 is freely available at http://www.microbesonline.org/fasttree.
Ronquist et al. (2012) — MrBayes 3.2.
MrBayes 3.2: Efficient Bayesian Phylogenetic Inference and Model Choice Across a Large Model Space. Ronquist et al. (2012). Systematic Biology. doi:10.1093/sysbio/sys029
Abstract
Since its introduction in 2001, MrBayes has grown in popularity as a software package for Bayesian phylogenetic inference using Markov chain Monte Carlo (MCMC) methods. With this note, we announce the release of version 3.2, a major upgrade to the latest official release presented in 2003. The new version provides convergence diagnostics and allows multiple analyses to be run in parallel with convergence progress monitored on the fly. The introduction of new proposals and automatic optimization of tuning parameters has improved convergence for many problems. The new version also sports significantly faster likelihood calculations through streaming single-instruction-multiple-data extensions (SSE) and support of the BEAGLE library, allowing likelihood calculations to be delegated to graphics processing units (GPUs) on compatible hardware. Speedup factors range from around 2 with SSE code to more than 50 with BEAGLE for codon problems. Checkpointing across all models allows long runs to be completed even when an analysis is prematurely terminated. New models include relaxed clocks, dating, model averaging across time-reversible substitution models, and support for hard, negative, and partial (backbone) tree constraints. Inference of species trees from gene trees is supported by full incorporation of the Bayesian estimation of species trees (BEST) algorithms. Marginal model likelihoods for Bayes factor tests can be estimated accurately across the entire model space using the stepping stone method. The new version provides more output options than previously, including samples of ancestral states, site rates, site d(N)/d(S) rations, branch rates, and node dates. A wide range of statistics on tree parameters can also be output for visualization in FigTree and compatible software.
Suchard et al. (2018), Bouckaert et al. (2019) — BEAST and BEAST 2.
Bayesian phylogenetic and phylodynamic data integration using BEAST 1.10. Suchard et al. (2018). Virus Evolution. doi:10.1093/ve/vey016
Abstract
The Bayesian Evolutionary Analysis by Sampling Trees (BEAST) software package has become a primary tool for Bayesian phylogenetic and phylodynamic inference from genetic sequence data. BEAST unifies molecular phylogenetic reconstruction with complex discrete and continuous trait evolution, divergence-time dating, and coalescent demographic models in an efficient statistical inference engine using Markov chain Monte Carlo integration. A convenient, cross-platform, graphical user interface allows the flexible construction of complex evolutionary analyses.
BEAST 2.5: An advanced software platform for Bayesian evolutionary analysis. Bouckaert et al. (2019). PLOS Computational Biology. doi:10.1371/journal.pcbi.1006650
Abstract
Elaboration of Bayesian phylogenetic inference methods has continued at pace in recent years with major new advances in nearly all aspects of the joint modelling of evolutionary data. It is increasingly appreciated that some evolutionary questions can only be adequately answered by combining evidence from multiple independent sources of data, including genome sequences, sampling dates, phenotypic data, radiocarbon dates, fossil occurrences, and biogeographic range information among others. Including all relevant data into a single joint model is very challenging both conceptually and computationally. Advanced computational software packages that allow robust development of compatible (sub-)models which can be composed into a full model hierarchy have played a key role in these developments. Developing such software frameworks is increasingly a major scientific activity in its own right, and comes with specific challenges, from practical software design, development and engineering challenges to statistical and conceptual modelling challenges. BEAST 2 is one such computational software platform, and was first announced over 4 years ago. Here we describe a series of major new developments in the BEAST 2 core platform and model hierarchy that have occurred since the first release of the software, culminating in the recent 2.5 release.
Drummond et al. (2006) — relaxed clocks and dating.
Relaxed Phylogenetics and Dating with Confidence. Drummond et al. (2006). PLoS Biology. doi:10.1371/journal.pbio.0040088
Abstract
In phylogenetics, the unrooted model of phylogeny and the strict molecular clock model are two extremes of a continuum. Despite their dominance in phylogenetic inference, it is evident that both are biologically unrealistic and that the real evolutionary process lies between these two extremes. Fortunately, intermediate models employing relaxed molecular clocks have been described. These models open the gate to a new field of "relaxed phylogenetics." Here we introduce a new approach to performing relaxed phylogenetic analysis. We describe how it can be used to estimate phylogenies and divergence times in the face of uncertainty in evolutionary rates and calibration times. Our approach also provides a means for measuring the clocklikeness of datasets and comparing this measure between different genes and phylogenies. We find no significant rate autocorrelation among branches in three large datasets, suggesting that autocorrelated models are not necessarily suitable for these data. In addition, we place these datasets on the continuum of clocklikeness between a strict molecular clock and the alternative unrooted extreme. Finally, we present analyses of 102 bacterial, 106 yeast, 61 plant, 99 metazoan, and 500 primate alignments. From these we conclude that our method is phylogenetically more accurate and precise than the traditional unrooted model while adding the ability to infer a timescale to evolution.
B.4.5 Alignment uncertainty and phylogenomics
Thorne et al. (1991) — an evolutionary model for maximum-likelihood alignment (TKF91).
An evolutionary model for maximum likelihood alignment of DNA sequences. Thorne et al. (1991). Journal of Molecular Evolution. doi:10.1007/bf02193625
Abstract
Most algorithms for the alignment of biological sequences are not derived from an evolutionary model. Consequently, these alignment algorithms lack a strong statistical basis. A maximum likelihood method for the alignment of two DNA sequences is presented. This method is based upon a statistical model of DNA sequence evolution for which we have obtained explicit transition probabilities. The evolutionary model can also be used as the basis of procedures that estimate the evolutionary parameters relevant to a pair of unaligned DNA sequences. A parameter-estimation approach which takes into account all possible alignments between two sequences is introduced; the danger of estimating evolutionary parameters from a single alignment is discussed.
Redelings and Suchard (2005) — joint Bayesian estimation of alignment and phylogeny.
Joint Bayesian Estimation of Alignment and Phylogeny. Redelings & Suchard (2005). Systematic Biology. doi:10.1080/10635150590947041
Abstract
We describe a novel model and algorithm for simultaneously estimating multiple molecular sequence alignments and the phylogenetic trees that relate the sequences. Unlike current techniques that base phylogeny estimates on a single estimate of the alignment, we take alignment uncertainty into account by considering all possible alignments. Furthermore, because the alignment and phylogeny are constructed simultaneously, a guide tree is not needed. This sidesteps the problem in which alignments created by progressive alignment are biased toward the guide tree used to generate them. Joint estimation also allows us to model rate variation between sites when estimating the alignment and to use the evidence in shared insertion/deletions (indels) to group sister taxa in the phylogeny. Our indel model makes use of affine gap penalties and considers indels of multiple letters. We make the simplifying assumption that the indel process is identical on all branches. As a result, the probability of a gap is independent of branch length. We use a Markov chain Monte Carlo (MCMC) method to sample from the posterior of the joint model, estimating the most probable alignment and tree and their support simultaneously. We describe a new MCMC transition kernel that improves our algorithm's mixing efficiency, allowing the MCMC chains to converge even when started from arbitrary alignments. Our software implementation can estimate alignment uncertainty and we describe a method for summarizing this uncertainty in a single plot.
Wong et al. (2008) — alignment uncertainty and genomic analysis.
Alignment Uncertainty and Genomic Analysis. Wong et al. (2008). Science. doi:10.1126/science.1151532
Abstract
The statistical methods applied to the analysis of genomic data do not account for uncertainty in the sequence alignment. Indeed, the alignment is treated as an observation, and all of the subsequent inferences depend on the alignment being correct. This may not have been too problematic for many phylogenetic studies, in which the gene is carefully chosen for, among other things, ease of alignment. However, in a comparative genomics study, the same statistical methods are applied repeatedly on thousands of genes, many of which will be difficult to align. Using genomic data from seven yeast species, we show that uncertainty in the alignment can lead to several problems, including different alignment methods resulting in different conclusions.
Talavera and Castresana (2007), Capella-Gutiérrez et al. (2009) — alignment trimming.
Improvement of Phylogenies after Removing Divergent and Ambiguously Aligned Blocks from Protein Sequence Alignments. Talavera & Castresana (2007). Systematic Biology. doi:10.1080/10635150701472164
Abstract
Alignment quality may have as much impact on phylogenetic reconstruction as the phylogenetic methods used. Not only the alignment algorithm, but also the method used to deal with the most problematic alignment regions, may have a critical effect on the final tree. Although some authors remove such problematic regions, either manually or using automatic methods, in order to improve phylogenetic performance, others prefer to keep such regions to avoid losing any information. Our aim in the present work was to examine whether phylogenetic reconstruction improves after alignment cleaning or not. Using simulated protein alignments with gaps, we tested the relative performance in diverse phylogenetic analyses of the whole alignments versus the alignments with problematic regions removed with our previously developed Gblocks program. We also tested the performance of more or less stringent conditions in the selection of blocks. Alignments constructed with different alignment methods (ClustalW, Mafft, and Probcons) were used to estimate phylogenetic trees by maximum likelihood, neighbor joining, and parsimony. We show that, in most alignment conditions, and for alignments that are not too short, removal of blocks leads to better trees. That is, despite losing some information, there is an increase in the actual phylogenetic signal. Overall, the best trees are obtained by maximum-likelihood reconstruction of alignments cleaned by Gblocks. In general, a relaxed selection of blocks is better for short alignment, whereas a stringent selection is more adequate for longer ones. Finally, we show that cleaned alignments produce better topologies although, paradoxically, with lower bootstrap. This indicates that divergent and problematic alignment regions may lead, when present, to apparently better supported although, in fact, more biased topologies.
trimAl: a tool for automated alignment trimming in large-scale phylogenetic analyses. Capella-Gutiérrez et al. (2009). Bioinformatics. doi:10.1093/bioinformatics/btp348
Abstract
Multiple sequence alignments are central to many areas of bioinformatics. It has been shown that the removal of poorly aligned regions from an alignment increases the quality of subsequent analyses. Such an alignment trimming phase is complicated in large-scale phylogenetic analyses that deal with thousands of alignments. Here, we present trimAl, a tool for automated alignment trimming, which is especially suited for large-scale phylogenetic analyses. trimAl can consider several parameters, alone or in multiple combinations, for selecting the most reliable positions in the alignment. These include the proportion of sequences with a gap, the level of amino acid similarity and, if several alignments for the same set of sequences are provided, the level of consistency across different alignments. Moreover, trimAl can automatically select the parameters to be used in each specific alignment so that the signal-to-noise ratio is optimized. Availability trimAl has been written in C++, it is portable to all platforms. trimAl is freely available for download (http://trimal.cgenomics.org) and can be used online through the Phylemon web server (http://phylemon2.bioinfo.cipf.es/). Supplementary Material is available at http://trimal.cgenomics.org/publications.
Rokas et al. (2003), Jeffroy et al. (2006) — the promise and pitfalls of concatenation.
Genome-scale approaches to resolving incongruence in molecular phylogenies. Rokas et al. (2003). Nature. doi:10.1038/nature02053
Abstract
One of the most pervasive challenges in molecular phylogenetics is the incongruence between phylogenies obtained using different data sets, such as individual genes. To systematically investigate the degree of incongruence, and potential methods for resolving it, we screened the genome sequences of eight yeast species and selected 106 widely distributed orthologous genes for phylogenetic analyses, singly and by concatenation. Our results suggest that data sets consisting of single or a small number of concatenated genes have a significant probability of supporting conflicting topologies. By contrast, analyses of the entire data set of concatenated genes yielded a single, fully resolved species tree with maximum support. Comparable results were obtained with a concatenation of a minimum of 20 genes; substantially more genes than commonly used but a small fraction of any genome. These results have important implications for resolving branches of the tree of life.
Phylogenomics: the beginning of incongruence?. Jeffroy et al. (2006). Trends in Genetics. doi:10.1016/j.tig.2006.02.003
Abstract
Until recently, molecular phylogenies based on a single or few orthologous genes often yielded contradictory results. Using multiple genes in a large concatenation was proposed to end these incongruences. Here we show that single-gene phylogenies often produce incongruences, albeit ones lacking statistically significant support. By contrast, the use of different tree reconstruction methods on different partitions of the concatenated supergene leads to well-resolved, but real (i.e. statistically significant) incongruences. Gathering a large amount of data is not sufficient to produce reliable trees, given the current limitation of tree reconstruction methods, especially when the quality of data is poor. We propose that selecting only data that contain minimal nonphylogenetic signals takes full advantage of phylogenomics and markedly reduces incongruence.
Maddison (1997), Degnan and Rosenberg (2009), Edwards (2009) — gene trees, species trees and the multispecies coalescent.
Gene Trees in Species Trees. Maddison (1997). Systematic Biology. doi:10.1093/sysbio/46.3.523
Abstract
Exploration of the relationship between gene trees and their containing species trees leads to consideration of how to reconstruct species trees from gene trees and of the concept of phylogeny as a cloud of gene histories. When gene copies are sampled from various species, the gene tree relating these copies might disagree with the species phylogeny. This discord can arise from horizontal transfer (including hybridization), lineage sorting, and gene duplication and extinction. Lineage sorting could also be called deep coalescence, the failure of ancestral copies to coalesce (looking backwards in time) into a common ancestral copy until deeper than previous speciation events. These events depend on various factors; for instance, deep coalescence is more likely if the branches of the species tree are short (in generations) and wide (in population size). A similar dependence on process is found in historical biogeography and host-parasite relationships. Each of the processes of discord could yield a different parsimony criterion for reconstructing the species tree from a set of gene trees: with horizontal transfer, choose the species tree that minimizes the number of transfer events; with deep coalescence, choose the tree minimizing the number of extra gene lineages that had to coexist along species lineages; with gene duplication, choose the tree minimizing duplication and/or extinction events. Maximum likelihood methods for reconstructing the species tree are also possible because coalescence theory provides the probability that a particular gene tree would occur given a species tree (with branch lengths and widths specified). In considering these issues, one is provoked to reconsider precisely what is phylogeny. Perhaps it is misleading to view some gene trees as agreeing and other gene trees as disagreeing with the species tree; rather, all of the gene trees are part of the species tree, which can be visualized like a fuzzy statistical distribution, a cloud of gene histories. Alternatively, phylogeny might be (and has been) viewed not as a history of what happened, genetically, but as a history of what could have happened, i.e., a history of changes in the probabilities of inter-breeding.
Gene tree discordance, phylogenetic inference and the multispecies coalescent. Degnan & Rosenberg (2009). Trends in Ecology & Evolution. doi:10.1016/j.tree.2009.01.009
Abstract
The field of phylogenetics is entering a new era in which trees of historical relationships between species are increasingly inferred from multilocus and genomic data. A major challenge for incorporating such large amounts of data into inference of species trees is that conflicting genealogical histories often exist in different genes throughout the genome. Recent advances in genealogical modeling suggest that resolving close species relationships is not quite as simple as applying more data to the problem. Here we discuss the complexities of genealogical discordance and review the issues that new methods for multilocus species tree inference will need to address to account successfully for naturally occurring genomic variability in evolutionary histories.
IS A NEW AND GENERAL THEORY OF MOLECULAR SYSTEMATICS EMERGING?. Edwards (2009). Evolution. doi:10.1111/j.1558-5646.2008.00549.x
Abstract
The advent and maturation of algorithms for estimating species trees-phylogenetic trees that allow gene tree heterogeneity and whose tips represent lineages, populations and species, as opposed to genes-represent an exciting confluence of phylogenetics, phylogeography, and population genetics, and ushers in a new generation of concepts and challenges for the molecular systematist. In this essay I argue that to better deal with the large multilocus datasets brought on by phylogenomics, and to better align the fields of phylogeography and phylogenetics, we should embrace the primacy of species trees, not only as a new and useful practical tool for systematics, but also as a long-standing conceptual goal of systematics that, largely due to the lack of appropriate computational tools, has been eclipsed in the past few decades. I suggest that phylogenies as gene trees are a "local optimum" for systematics, and review recent advances that will bring us to the broader optimum inherent in species trees. In addition to adopting new methods of phylogenetic analysis (and ideally reserving the term "phylogeny" for species trees rather than gene trees), the new paradigm suggests shifts in a number of practices, such as sampling data to maximize not only the number of accumulated sites but also the number of independently segregating genes; routinely using coalescent or other models in computer simulations to allow gene tree heterogeneity; and understanding better the role of concatenation in influencing topologies and confidence in phylogenies. By building on the foundation laid by concepts of gene trees and coalescent theory, and by taking cues from recent trends in multilocus phylogeography, molecular systematics stands to be enriched. Many of the challenges and lessons learned for estimating gene trees will carry over to the challenge of estimating species trees, although adopting the species tree paradigm will clarify many issues (such as the nature of polytomies and the star tree paradox), raise conceptually new challenges, or provide new answers to old questions.
Mirarab et al. (2014), Zhang et al. (2018) — ASTRAL.
ASTRAL: genome-scale coalescent-based species tree estimation. Mirarab et al. (2014). Bioinformatics. doi:10.1093/bioinformatics/btu462
Abstract
Motivation Species trees provide insight into basic biology, including the mechanisms of evolution and how it modifies biomolecular function and structure, biodiversity and co-evolution between genes and species. Yet, gene trees often differ from species trees, creating challenges to species tree estimation. One of the most frequent causes for conflicting topologies between gene trees and species trees is incomplete lineage sorting (ILS), which is modelled by the multi-species coalescent. While many methods have been developed to estimate species trees from multiple genes, some which have statistical guarantees under the multi-species coalescent model, existing methods are too computationally intensive for use with genome-scale analyses or have been shown to have poor accuracy under some realistic conditions. Results We present ASTRAL, a fast method for estimating species trees from multiple genes. ASTRAL is statistically consistent, can run on datasets with thousands of genes and has outstanding accuracy-improving on MP-EST and the population tree from BUCKy, two statistically consistent leading coalescent-based methods. ASTRAL is often more accurate than concatenation using maximum likelihood, except when ILS levels are low or there are too few gene trees. Availability and implementation ASTRAL is available in open source form at https://github.com/smirarab/ASTRAL/. Datasets studied in this article are available at http://www.cs.utexas.edu/users/phylo/datasets/astral. Supplementary information Supplementary data are available at Bioinformatics online.
ASTRAL-III: polynomial time species tree reconstruction from partially resolved gene trees. Zhang et al. (2018). BMC Bioinformatics. doi:10.1186/s12859-018-2129-y
Abstract
Background Evolutionary histories can be discordant across the genome, and such discordances need to be considered in reconstructing the species phylogeny. ASTRAL is one of the leading methods for inferring species trees from gene trees while accounting for gene tree discordance. ASTRAL uses dynamic programming to search for the tree that shares the maximum number of quartet topologies with input gene trees, restricting itself to a predefined set of bipartitions. Results We introduce ASTRAL-III, which substantially improves the running time of ASTRAL-II and guarantees polynomial running time as a function of both the number of species (n) and the number of genes (k). ASTRAL-III limits the bipartition constraint set (X) to grow at most linearly with n and k. Moreover, it handles polytomies more efficiently than ASTRAL-II, exploits similarities between gene trees better, and uses several techniques to avoid searching parts of the search space that are mathematically guaranteed not to include the optimal tree. The asymptotic running time of ASTRAL-III in the presence of polytomies is [Formula: see text] where D=O(nk) is the sum of degrees of all unique nodes in input trees. The running time improvements enable us to test whether contracting low support branches in gene trees improves the accuracy by reducing noise. In extensive simulations, we show that removing branches with very low support (e.g., below 10%) improves accuracy while overly aggressive filtering is harmful. We observe on a biological avian phylogenomic dataset of 14K genes that contracting low support branches greatly improve results. Conclusions ASTRAL-III is a faster version of the ASTRAL method for phylogenetic reconstruction and can scale up to 10,000 species. With ASTRAL-III, low support branches can be removed, resulting in improved accuracy.
Goodman et al. (1979), Szöllősi et al. (2014) — reconciliation and joint gene-tree–species-tree inference.
Fitting the gene lineage into its species lineage, a parsimony strategy illustrated by cladograms constructed from globin sequences. Goodman et al. (1979). Systematic Zoology.
The Inference of Gene Trees with Species Trees. Szöllősi et al. (2014). Systematic Biology. doi:10.1093/sysbio/syu048
Abstract
This article reviews the various models that have been used to describe the relationships between gene trees and species trees. Molecular phylogeny has focused mainly on improving models for the reconstruction of gene trees based on sequence alignments. Yet, most phylogeneticists seek to reveal the history of species. Although the histories of genes and species are tightly linked, they are seldom identical, because genes duplicate, are lost or horizontally transferred, and because alleles can coexist in populations for periods that may span several speciation events. Building models describing the relationship between gene and species trees can thus improve the reconstruction of gene trees when a species tree is known, and vice versa. Several approaches have been proposed to solve the problem in one direction or the other, but in general neither gene trees nor species trees are known. Only a few studies have attempted to jointly infer gene trees and species trees. These models account for gene duplication and loss, transfer or incomplete lineage sorting. Some of them consider several types of events together, but none exists currently that considers the full repertoire of processes that generate gene trees along the species tree. Simulations as well as empirical studies on genomic data show that combining gene tree-species tree models with models of sequence evolution improves gene tree reconstruction. In turn, these better gene trees provide a more reliable basis for studying genome evolution or reconstructing ancestral chromosomes and ancestral gene sequences. We predict that gene tree-species tree methods that can deal with genomic data sets will be instrumental to advancing our understanding of genomic evolution.
Huson and Bryant (2005), Bryant (2003) — phylogenetic networks and Neighbor-Net.
Application of Phylogenetic Networks in Evolutionary Studies. Huson & Bryant (2005). Molecular Biology and Evolution. doi:10.1093/molbev/msj030
Abstract
The evolutionary history of a set of taxa is usually represented by a phylogenetic tree, and this model has greatly facilitated the discussion and testing of hypotheses. However, it is well known that more complex evolutionary scenarios are poorly described by such models. Further, even when evolution proceeds in a tree-like manner, analysis of the data may not be best served by using methods that enforce a tree structure but rather by a richer visualization of the data to evaluate its properties, at least as an essential first step. Thus, phylogenetic networks should be employed when reticulate events such as hybridization, horizontal gene transfer, recombination, or gene duplication and loss are believed to be involved, and, even in the absence of such events, phylogenetic networks have a useful role to play. This article reviews the terminology used for phylogenetic networks and covers both split networks and reticulate networks, how they are defined, and how they can be interpreted. Additionally, the article outlines the beginnings of a comprehensive statistical framework for applying split network methods. We show how split networks can represent confidence sets of trees and introduce a conservative statistical test for whether the conflicting signal in a network is treelike. Finally, this article describes a new program, SplitsTree4, an interactive and comprehensive tool for inferring different types of phylogenetic networks from sequences, distances, and trees.
Neighbor-Net: An Agglomerative Method for the Construction of Phylogenetic Networks. Bryant (2003). Molecular Biology and Evolution. doi:10.1093/molbev/msh018
Abstract
We present Neighbor-Net, a distance based method for constructing phylogenetic networks that is based on the Neighbor-Joining (NJ) algorithm of Saitou and Nei. Neighbor-Net provides a snapshot of the data that can guide more detailed analysis. Unlike split decomposition, Neighbor-Net scales well and can quickly produce detailed and informative networks for several hundred taxa. We illustrate the method by reanalyzing three published data sets: a collection of 110 highly recombinant Salmonella multi-locus sequence typing sequences, the 135 "African Eve" human mitochondrial sequences published by Vigilant et al., and a collection of 12 Archeal chaperonin sequences demonstrating strong evidence for gene conversion. Neighbor-Net is available as part of the SplitsTree4 software package.
B.4.6 The tree of life
Woese and Fox (1977), Woese et al. (1990) — the archaea and the three domains.
Phylogenetic structure of the prokaryotic domain: The primary kingdoms. Woese & Fox (1977). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.74.11.5088
Abstract
A phylogenetic analysis based upon ribosomal RNA sequence characterization reveals that living systems represent one of three aboriginal lines of descent: (i) the eubacteria, comprising all typical bacteria; (ii) the archaebacteria, containing methanogenic bacteria; and (iii) the urkaryotes, now represented in the cytoplasmic component of eukaryotic cells.
Towards a natural system of organisms: proposal for the domains Archaea, Bacteria, and Eucarya.. Woese et al. (1990). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.87.12.4576
Abstract
Molecular structures and sequences are generally more revealing of evolutionary relationships than are classical phenotypes (particularly so among microorganisms). Consequently, the basis for the definition of taxa has progressively shifted from the organismal to the cellular to the molecular level. Molecular comparisons show that life on this planet divides into three primary groupings, commonly known as the eubacteria, the archaebacteria, and the eukaryotes. The three are very dissimilar, the differences that separate them being of a more profound nature than the differences that separate typical kingdoms, such as animals and plants. Unfortunately, neither of the conventionally accepted views of the natural relationships among living systems--i.e., the five-kingdom taxonomy or the eukaryote-prokaryote dichotomy--reflects this primary tripartite division of the living world. To remedy this situation we propose that a formal system of organisms be established in which above the level of kingdom there exists a new taxon called a "domain." Life on this planet would then be seen as comprising three domains, the Bacteria, the Archaea, and the Eucarya, each containing two or more kingdoms. (The Eucarya, for example, contain Animalia, Plantae, Fungi, and a number of others yet to be defined). Although taxonomic structure within the Bacteria and Eucarya is not treated herein, Archaea is formally subdivided into the two kingdoms Euryarchaeota (encompassing the methanogens and their phenotypically diverse relatives) and Crenarchaeota (comprising the relatively tight clustering of extremely thermophilic archaebacteria, whose general phenotype appears to resemble most the ancestral phenotype of the Archaea.
Gogarten et al. (1989), Iwabe et al. (1989) — rooting the tree of life with ancient paralogues.
Evolution of the vacuolar H+-ATPase: implications for the origin of eukaryotes.. Gogarten et al. (1989). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.86.17.6661
Abstract
Active transport across the vacuolar components of the eukaryotic endomembrane system is energized by a specific vacuolar H+-ATPase. The amino acid sequences of the 70- and 60-kDa subunits of the vacuolar H+-ATPase are approximately equal to 25% identical to the beta and alpha subunits, respectively, of the eubacterial-type F0F1-ATPases. We now report that the same vacuolar H+-ATPase subunits are approximately equal to 50% identical to the alpha and beta subunits, respectively, of the sulfur-metabolizing Sulfolobus acidocaldarius, an archaebacterium (Archaeobacterium). Moreover, the homologue of an 88-amino acid stretch near the amino-terminal end of the 70-kDa subunit is absent from the F0F1-ATPase beta subunit but is present in the alpha subunit of Sulfolobus. Since the two types of subunits (alpha and beta subunits; 60- and 70-kDa subunits) are homologous to each other, they must have arisen by a gene duplication that occurred prior to the last common ancestor of the eubacteria, eukaryotes, and Sulfolobus. Thus, the phylogenetic tree of the subunits can be rooted at the site where the gene duplication occurred. The inferred evolutionary tree contains two main branches: a eubacterial branch and an eocyte branch that gave rise to Sulfolobus and the eukaryotic host cell. The implication is that the vacuolar H+-ATPase of eukaryotes arose by the internalization of the plasma membrane H+-ATPase of an archaebacterial-like ancestral cell.
Evolutionary relationship of archaebacteria, eubacteria, and eukaryotes inferred from phylogenetic trees of duplicated genes.. Iwabe et al. (1989). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.86.23.9355
Abstract
All extant organisms are though to be classified into three primary kingdoms, eubacteria, eukaryotes, and archaebacteria. The molecular evolutionary studies on the origin and evolution of archaebacteria to date have been carried out by inferring a molecular phylogenetic tree of the primary kingdoms based on comparison of a single molecule from a variety of extant species. From such comparison, it was not possible to derive the exact evolutionary relationship among the primary kingdoms, because the root of the tree could not be determined uniquely. To overcome this difficulty, we compared a pair of duplicated genes, elongation factors Tu and G, and the alpha and beta subunits of ATPase, which are thought to have diverged by gene duplication before divergence of the primary kingdoms. Using each protein pair, we inferred a composite phylogenetic tree with two clusters corresponding to different proteins, from which the evolutionary relationship of the primary kingdoms is determined uniquely. The inferred composite trees reveal that archaebacteria are more closely related to eukaryotes than to eubacteria for all the cases. By bootstrap resamplings, this relationship is reproduced with probabilities of 0.96, 0.79, 1.0, and 1.0 for elongation factors Tu and G and for ATPase subunits alpha and beta, respectively. There are also several lines of evidence for the close sequence similarity between archaebacteria and eukaryotes. Thus we propose that this tree topology represents the general evolutionary relationship among the three primary kingdoms.
Brown et al. (2015), Parks et al. (2018) — the candidate phyla radiation and GTDB.
Unusual biology across a group comprising more than 15% of domain Bacteria. Brown et al. (2015). Nature. doi:10.1038/nature14486
Abstract
A prominent feature of the bacterial domain is a radiation of major lineages that are defined as candidate phyla because they lack isolated representatives. Bacteria from these phyla occur in diverse environments and are thought to mediate carbon and hydrogen cycles. Genomic analyses of a few representatives suggested that metabolic limitations have prevented their cultivation. Here we reconstructed 8 complete and 789 draft genomes from bacteria representing >35 phyla and documented features that consistently distinguish these organisms from other bacteria. We infer that this group, which may comprise >15% of the bacterial domain, has shared evolutionary history, and describe it as the candidate phyla radiation (CPR). All CPR genomes are small and most lack numerous biosynthetic pathways. Owing to divergent 16S ribosomal RNA (rRNA) gene sequences, 50-100% of organisms sampled from specific phyla would evade detection in typical cultivation-independent surveys. CPR organisms often have self-splicing introns and proteins encoded within their rRNA genes, a feature rarely reported in bacteria. Furthermore, they have unusual ribosome compositions. All are missing a ribosomal protein often absent in symbionts, and specific lineages are missing ribosomal proteins and biogenesis factors considered universal in bacteria. This implies different ribosome structures and biogenesis mechanisms, and underlines unusual biology across a large part of the bacterial domain.
A standardized bacterial taxonomy based on genome phylogeny substantially revises the tree of life. Parks et al. (2018). Nature Biotechnology. doi:10.1038/nbt.4229
Abstract
Taxonomy is an organizing principle of biology and is ideally based on evolutionary relationships among organisms. Development of a robust bacterial taxonomy has been hindered by an inability to obtain most bacteria in pure culture and, to a lesser extent, by the historical use of phenotypes to guide classification. Culture-independent sequencing technologies have matured sufficiently that a comprehensive genome-based taxonomy is now possible. We used a concatenated protein phylogeny as the basis for a bacterial taxonomy that conservatively removes polyphyletic groups and normalizes taxonomic ranks on the basis of relative evolutionary divergence. Under this approach, 58% of the 94,759 genomes comprising the Genome Taxonomy Database had changes to their existing taxonomy. This result includes the description of 99 phyla, including six major monophyletic units from the subdivision of the Proteobacteria, and amalgamation of the Candidate Phyla Radiation into a single phylum. Our taxonomy should enable improved classification of uncultured bacteria and provide a sound basis for ecological and evolutionary studies.
Weiss et al. (2016), Moody et al. (2024) — the nature of LUCA.
The physiology and habitat of the last universal common ancestor. Weiss et al. (2016). Nature Microbiology. doi:10.1038/nmicrobiol.2016.116
Abstract
The concept of a last universal common ancestor of all cells (LUCA, or the progenote) is central to the study of early evolution and life's origin, yet information about how and where LUCA lived is lacking. We investigated all clusters and phylogenetic trees for 6.1 million protein coding genes from sequenced prokaryotic genomes in order to reconstruct the microbial ecology of LUCA. Among 286,514 protein clusters, we identified 355 protein families (∼0.1%) that trace to LUCA by phylogenetic criteria. Because these proteins are not universally distributed, they can shed light on LUCA's physiology. Their functions, properties and prosthetic groups depict LUCA as anaerobic, CO2-fixing, H2-dependent with a Wood-Ljungdahl pathway, N2-fixing and thermophilic. LUCA's biochemistry was replete with FeS clusters and radical reaction mechanisms. Its cofactors reveal dependence upon transition metals, flavins, S-adenosyl methionine, coenzyme A, ferredoxin, molybdopterin, corrins and selenium. Its genetic code required nucleoside modifications and S-adenosyl methionine-dependent methylations. The 355 phylogenies identify clostridia and methanogens, whose modern lifestyles resemble that of LUCA, as basal among their respective domains. LUCA inhabited a geochemically active environment rich in H2, CO2 and iron. The data support the theory of an autotrophic origin of life involving the Wood-Ljungdahl pathway in a hydrothermal setting.
The nature of the last universal common ancestor and its impact on the early Earth system. Moody et al. (2024). Nature Ecology & Evolution. doi:10.1038/s41559-024-02461-1
Abstract
The nature of the last universal common ancestor (LUCA), its age and its impact on the Earth system have been the subject of vigorous debate across diverse disciplines, often based on disparate data and methods. Age estimates for LUCA are usually based on the fossil record, varying with every reinterpretation. The nature of LUCA's metabolism has proven equally contentious, with some attributing all core metabolisms to LUCA, whereas others reconstruct a simpler life form dependent on geochemistry. Here we infer that LUCA lived ~4.2 Ga (4.09-4.33 Ga) through divergence time analysis of pre-LUCA gene duplicates, calibrated using microbial fossils and isotope records under a new cross-bracing implementation. Phylogenetic reconciliation suggests that LUCA had a genome of at least 2.5 Mb (2.49-2.99 Mb), encoding around 2,600 proteins, comparable to modern prokaryotes. Our results suggest LUCA was a prokaryote-grade anaerobic acetogen that possessed an early immune system. Although LUCA is sometimes perceived as living in isolation, we infer LUCA to have been part of an established ecological system. The metabolism of LUCA would have provided a niche for other microbial community members and hydrogen recycling by atmospheric photochemistry could have supported a modestly productive early ecosystem.
Coleman et al. (2021) — rooting the bacterial tree.
A rooted phylogeny resolves early bacterial evolution. Coleman et al. (2021). Science. doi:10.1126/science.abe0511
Abstract
A rooted bacterial tree is necessary to understand early evolution, but the position of the root is contested. Here, we model the evolution of 11,272 gene families to identify the root, extent of horizontal gene transfer (HGT), and the nature of the last bacterial common ancestor (LBCA). Our analyses root the tree between the major clades Terrabacteria and Gracilicutes and suggest that LBCA was a free-living flagellated, rod-shaped double-membraned organism. Contrary to recent proposals, our analyses reject a basal placement of the Candidate Phyla Radiation, which instead branches sister to Chloroflexota within Terrabacteria. While most gene families (92%) have evidence of HGT, overall, two-thirds of gene transmissions have been vertical, suggesting that a rooted tree provides a meaningful frame of reference for interpreting bacterial evolution.
Spang et al. (2015), Zaremba-Niedzwiedzka et al. (2017), Imachi et al. (2020), Eme et al. (2023) — Asgard archaea and the origin of eukaryotes.
Complex archaea that bridge the gap between prokaryotes and eukaryotes. Spang et al. (2015). Nature. doi:10.1038/nature14447
Abstract
The origin of the eukaryotic cell remains one of the most contentious puzzles in modern biology. Recent studies have provided support for the emergence of the eukaryotic host cell from within the archaeal domain of life, but the identity and nature of the putative archaeal ancestor remain a subject of debate. Here we describe the discovery of 'Lokiarchaeota', a novel candidate archaeal phylum, which forms a monophyletic group with eukaryotes in phylogenomic analyses, and whose genomes encode an expanded repertoire of eukaryotic signature proteins that are suggestive of sophisticated membrane remodelling capabilities. Our results provide strong support for hypotheses in which the eukaryotic host evolved from a bona fide archaeon, and demonstrate that many components that underpin eukaryote-specific features were already present in that ancestor. This provided the host with a rich genomic 'starter-kit' to support the increase in the cellular and genomic complexity that is characteristic of eukaryotes.
Asgard archaea illuminate the origin of eukaryotic cellular complexity. Zaremba-Niedzwiedzka et al. (2017). Nature. doi:10.1038/nature21031
Abstract
The origin and cellular complexity of eukaryotes represent a major enigma in biology. Current data support scenarios in which an archaeal host cell and an alphaproteobacterial (mitochondrial) endosymbiont merged together, resulting in the first eukaryotic cell. The host cell is related to Lokiarchaeota, an archaeal phylum with many eukaryotic features. The emergence of the structural complexity that characterizes eukaryotic cells remains unclear. Here we describe the 'Asgard' superphylum, a group of uncultivated archaea that, as well as Lokiarchaeota, includes Thor-, Odin- and Heimdallarchaeota. Asgard archaea affiliate with eukaryotes in phylogenomic analyses, and their genomes are enriched for proteins formerly considered specific to eukaryotes. Notably, thorarchaeal genomes encode several homologues of eukaryotic membrane-trafficking machinery components, including Sec23/24 and TRAPP domains. Furthermore, we identify thorarchaeal proteins with similar features to eukaryotic coat proteins involved in vesicle biogenesis. Our results expand the known repertoire of 'eukaryote-specific' proteins in Archaea, indicating that the archaeal host cell already contained many key components that govern eukaryotic cellular complexity.
Isolation of an archaeon at the prokaryote–eukaryote interface. Imachi et al. (2020). Nature. doi:10.1038/s41586-019-1916-6
Abstract
The origin of eukaryotes remains unclear 1-4 . Current data suggest that eukaryotes may have emerged from an archaeal lineage known as 'Asgard' archaea 5,6 . Despite the eukaryote-like genomic features that are found in these archaea, the evolutionary transition from archaea to eukaryotes remains unclear, owing to the lack of cultured representatives and corresponding physiological insights. Here we report the decade-long isolation of an Asgard archaeon related to Lokiarchaeota from deep marine sediment. The archaeon-'Candidatus Prometheoarchaeum syntrophicum' strain MK-D1-is an anaerobic, extremely slow-growing, small coccus (around 550 nm in diameter) that degrades amino acids through syntrophy. Although eukaryote-like intracellular complexes have been proposed for Asgard archaea 6 , the isolate has no visible organelle-like structure. Instead, Ca. P. syntrophicum is morphologically complex and has unique protrusions that are long and often branching. On the basis of the available data obtained from cultivation and genomics, and reasoned interpretations of the existing literature, we propose a hypothetical model for eukaryogenesis, termed the entangle-engulf-endogenize (also known as E 3 ) model.
Inference and reconstruction of the heimdallarchaeial ancestry of eukaryotes. Eme et al. (2023). Nature. doi:10.1038/s41586-023-06186-2
Abstract
In the ongoing debates about eukaryogenesis-the series of evolutionary events leading to the emergence of the eukaryotic cell from prokaryotic ancestors-members of the Asgard archaea play a key part as the closest archaeal relatives of eukaryotes 1 . However, the nature and phylogenetic identity of the last common ancestor of Asgard archaea and eukaryotes remain unresolved 2-4 . Here we analyse distinct phylogenetic marker datasets of an expanded genomic sampling of Asgard archaea and evaluate competing evolutionary scenarios using state-of-the-art phylogenomic approaches. We find that eukaryotes are placed, with high confidence, as a well-nested clade within Asgard archaea and as a sister lineage to Hodarchaeales, a newly proposed order within Heimdallarchaeia. Using sophisticated gene tree and species tree reconciliation approaches, we show that analogous to the evolution of eukaryotic genomes, genome evolution in Asgard archaea involved significantly more gene duplication and fewer gene loss events compared with other archaea. Finally, we infer that the last common ancestor of Asgard archaea was probably a thermophilic chemolithotroph and that the lineage from which eukaryotes evolved adapted to mesophilic conditions and acquired the genetic potential to support a heterotrophic lifestyle. Our work provides key insights into the prokaryote-to-eukaryote transition and a platform for better understanding the emergence of cellular complexity in eukaryotic cells.
Williams et al. (2019) — the two-domain tree of life.
Phylogenomics provides robust support for a two-domains tree of life. Williams et al. (2019). Nature Ecology & Evolution. doi:10.1038/s41559-019-1040-x
Abstract
Hypotheses about the origin of eukaryotic cells are classically framed within the context of a universal 'tree of life' based on conserved core genes. Vigorous ongoing debate about eukaryote origins is based on assertions that the topology of the tree of life depends on the taxa included and the choice and quality of genomic data analysed. Here we have reanalysed the evidence underpinning those claims and apply more data to the question by using supertree and coalescent methods to interrogate >3,000 gene families in archaea and eukaryotes. We find that eukaryotes consistently originate from within the archaea in a two-domains tree when due consideration is given to the fit between model and data. Our analyses support a close relationship between eukaryotes and Asgard archaea and identify the Heimdallarchaeota as the current best candidate for the closest archaeal relatives of the eukaryotic nuclear lineage.
Burki et al. (2020) — the new tree of eukaryotes.
The New Tree of Eukaryotes. Burki et al. (2020). Trends in Ecology & Evolution. doi:10.1016/j.tree.2019.08.008
Abstract
For 15 years, the eukaryote Tree of Life (eToL) has been divided into five to eight major groupings, known as 'supergroups'. However, the tree has been profoundly rearranged during this time. The new eToL results from the widespread application of phylogenomics and numerous discoveries of major lineages of eukaryotes, mostly free-living heterotrophic protists. The evidence that supports the tree has transitioned from a synthesis of molecular phylogenetics and biological characters to purely molecular phylogenetics. Most current supergroups lack defining morphological or cell-biological characteristics, making the supergroup label even more arbitrary than before. Going forward, the combination of traditional culturing with maturing culture-free approaches and phylogenomics should accelerate the process of completing and resolving the eToL at its deepest levels.
Betts et al. (2018) — dating early evolution with genomes and fossils.
Integrated genomic and fossil evidence illuminates life’s early evolution and eukaryote origin. Betts et al. (2018). Nature Ecology & Evolution. doi:10.1038/s41559-018-0644-x
Abstract
Establishing a unified timescale for the early evolution of Earth and life is challenging and mired in controversy because of the paucity of fossil evidence, the difficulty of interpreting it and dispute over the deepest branching relationships in the tree of life. Surprisingly, it remains perhaps the only episode in the history of life where literal interpretations of the fossil record hold sway, revised with every new discovery and reinterpretation. We derive a timescale of life, combining a reappraisal of the fossil material with new molecular clock analyses. We find the last universal common ancestor of cellular life to have predated the end of late heavy bombardment (>3.9 billion years ago (Ga)). The crown clades of the two primary divisions of life, Eubacteria and Archaebacteria, emerged much later (<3.4 Ga), relegating the oldest fossil evidence for life to their stem lineages. The Great Oxidation Event significantly predates the origin of modern Cyanobacteria, indicating that oxygenic photosynthesis evolved within the cyanobacterial stem lineage. Modern eukaryotes do not constitute a primary lineage of life and emerged late in Earth's history (<1.84 Ga), falsifying the hypothesis that the Great Oxidation Event facilitated their radiation. The symbiotic origin of mitochondria at 2.053-1.21 Ga reflects a late origin of the total-group Alphaproteobacteria to which the free living ancestor of mitochondria belonged.
B.4.7 Hominoid evolution
The Chimpanzee Sequencing and Analysis Consortium (2005), Scally et al. (2012), Prado-Martinez et al. (2013) — the chimpanzee, gorilla and great-ape genomes.
Initial sequence of the chimpanzee genome and comparison with the human genome. Sequencing & Consortium (2005). Nature. doi:10.1038/nature04072
Abstract
Here we present a draft genome sequence of the common chimpanzee (Pan troglodytes). Through comparison with the human genome, we have generated a largely complete catalogue of the genetic differences that have accumulated since the human and chimpanzee species diverged from our common ancestor, constituting approximately thirty-five million single-nucleotide changes, five million insertion/deletion events, and various chromosomal rearrangements. We use this catalogue to explore the magnitude and regional variation of mutational forces shaping these two genomes, and the strength of positive and negative selection acting on their genes. In particular, we find that the patterns of evolution in human and chimpanzee protein-coding genes are highly correlated and dominated by the fixation of neutral and slightly deleterious alleles. We also use the chimpanzee genome as an outgroup to investigate human population genetics and identify signatures of selective sweeps in recent human evolution.
Insights into hominid evolution from the gorilla genome sequence. Scally et al. (2012). Nature. doi:10.1038/nature10842
Abstract
Gorillas are humans' closest living relatives after chimpanzees, and are of comparable importance for the study of human origins and evolution. Here we present the assembly and analysis of a genome sequence for the western lowland gorilla, and compare the whole genomes of all extant great ape genera. We propose a synthesis of genetic and fossil evidence consistent with placing the human-chimpanzee and human-chimpanzee-gorilla speciation events at approximately 6 and 10 million years ago. In 30% of the genome, gorilla is closer to human or chimpanzee than the latter are to each other; this is rarer around coding genes, indicating pervasive selection throughout great ape evolution, and has functional consequences in gene expression. A comparison of protein coding genes reveals approximately 500 genes showing accelerated evolution on each of the gorilla, human and chimpanzee lineages, and evidence for parallel acceleration, particularly of genes involved in hearing. We also compare the western and eastern gorilla species, estimating an average sequence divergence time 1.75 million years ago, but with evidence for more recent genetic exchange and a population bottleneck in the eastern species. The use of the genome sequence in these and future analyses will promote a deeper understanding of great ape biology and evolution.
Great ape genetic diversity and population history. Prado-Martinez et al. (2013). Nature. doi:10.1038/nature12228
Abstract
Most great ape genetic variation remains uncharacterized; however, its study is critical for understanding population history, recombination, selection and susceptibility to disease. Here we sequence to high coverage a total of 79 wild- and captive-born individuals representing all six great ape species and seven subspecies and report 88.8 million single nucleotide polymorphisms. Our analysis provides support for genetically distinct populations within each species, signals of gene flow, and the split of common chimpanzees into two distinct groups: Nigeria-Cameroon/western and central/eastern populations. We find extensive inbreeding in almost all wild populations, with eastern gorillas being the most extreme. Inferred effective population sizes have varied radically over time in different lineages and this appears to have a profound effect on the genetic diversity at, or close to, genes in almost all species. We discover and assign 1,982 loss-of-function variants throughout the human and great ape lineages, determining that the rate of gene loss has not been different in the human branch compared to other internal branches in the great ape phylogeny. This comprehensive catalogue of great ape genome diversity provides a framework for understanding evolution and a resource for more effective management of wild and captive great ape populations.
Yoo et al. (2025) — complete telomere-to-telomere genomes of the great apes.
Complete sequencing of ape genomes. Yoo et al. (2025). Nature. doi:10.1038/s41586-025-08816-3
Abstract
The most dynamic and repetitive regions of great ape genomes have traditionally been excluded from comparative studies 1-3 . Consequently, our understanding of the evolution of our species is incomplete. Here we present haplotype-resolved reference genomes and comparative analyses of six ape species: chimpanzee, bonobo, gorilla, Bornean orangutan, Sumatran orangutan and siamang. We achieve chromosome-level contiguity with substantial sequence accuracy (<1 error in 2.7 megabases) and completely sequence 215 gapless chromosomes telomere-to-telomere. We resolve challenging regions, such as the major histocompatibility complex and immunoglobulin loci, to provide in-depth evolutionary insights. Comparative analyses enabled investigations of the evolution and diversity of regions previously uncharacterized or incompletely studied without bias from mapping to the human reference genome. Such regions include newly minted gene families in lineage-specific segmental duplications, centromeric DNA, acrocentric chromosomes and subterminal heterochromatin. This resource serves as a comprehensive baseline for future evolutionary studies of humans and our closest living ape relatives.
Hobolth et al. (2007) — speciation times and incomplete lineage sorting (CoalHMM).
Genomic Relationships and Speciation Times of Human, Chimpanzee, and Gorilla Inferred from a Coalescent Hidden Markov Model. Hobolth et al. (2007). PLoS Genetics. doi:10.1371/journal.pgen.0030007
Abstract
The genealogical relationship of human, chimpanzee, and gorilla varies along the genome. We develop a hidden Markov model (HMM) that incorporates this variation and relate the model parameters to population genetics quantities such as speciation times and ancestral population sizes. Our HMM is an analytically tractable approximation to the coalescent process with recombination, and in simulations we see no apparent bias in the HMM estimates. We apply the HMM to four autosomal contiguous human-chimp-gorilla-orangutan alignments comprising a total of 1.9 million base pairs. We find a very recent speciation time of human-chimp (4.1 +/- 0.4 million years), and fairly large ancestral effective population sizes (65,000 +/- 30,000 for the human-chimp ancestor and 45,000 +/- 10,000 for the human-chimp-gorilla ancestor). Furthermore, around 50% of the human genome coalesces with chimpanzee after speciation with gorilla. We also consider 250,000 base pairs of X-chromosome alignments and find an effective population size much smaller than 75% of the autosomal effective population sizes. Finally, we find that the rate of transitions between different genealogies correlates well with the region-wide present-day human recombination rate, but does not correlate with the fine-scale recombination rates and recombination hot spots, suggesting that the latter are evolutionarily transient.
Patterson et al. (2006) — complex speciation of humans and chimpanzees.
Genetic evidence for complex speciation of humans and chimpanzees. Patterson et al. (2006). Nature. doi:10.1038/nature04789
Abstract
The genetic divergence time between two species varies substantially across the genome, conveying important information about the timing and process of speciation. Here we develop a framework for studying this variation and apply it to about 20 million base pairs of aligned sequence from humans, chimpanzees, gorillas and more distantly related primates. Human-chimpanzee genetic divergence varies from less than 84% to more than 147% of the average, a range of more than 4 million years. Our analysis also shows that human-chimpanzee speciation occurred less than 6.3 million years ago and probably more recently, conflicting with some interpretations of ancient fossils. Most strikingly, chromosome X shows an extremely young genetic divergence time, close to the genome minimum along nearly its entire length. These unexpected features would be explained if the human and chimpanzee lineages initially diverged, then later exchanged genes before separating permanently.
Scally and Durbin (2012) — the human mutation rate and the dating of human evolution.
Revising the human mutation rate: implications for understanding human evolution. Scally & Durbin (2012). Nature Reviews Genetics. doi:10.1038/nrg3295
Abstract
It is now possible to make direct measurements of the mutation rate in modern humans using next-generation sequencing. These measurements reveal a value that is approximately half of that previously derived from fossil calibration, and this has implications for our understanding of demographic events in human evolution and other aspects of population genetics. Here, we discuss the implications of a lower-than-expected mutation rate in relation to the timescale of human evolution.
B.5 Genomics
Further papers for Chapter 5.
Papers marked ★ were mandatory in earlier reading lists. In the HTML version, click Abstract to show the abstract.
B.5.1 Sequencing technologies
Shendure et al. (2017) — DNA sequencing at 40: past, present and future.
DNA sequencing at 40: past, present and future. Shendure et al. (2017). Nature. doi:10.1038/nature24286
Abstract
This review commemorates the 40th anniversary of DNA sequencing, a period in which we have already witnessed multiple technological revolutions and a growth in scale from a few kilobases to the first human genome, and now to millions of human and a myriad of other genomes. DNA sequencing has been extensively and creatively repurposed, including as a 'counter' for a vast range of molecular phenomena. We predict that in the long view of history, the impact of DNA sequencing will be on a par with that of the microscope.
Maxam and Gilbert (1977), Sanger et al. (1977) — the two original sequencing methods.
A new method for sequencing DNA.. Maxam & Gilbert (1977). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.74.2.560
Abstract
DNA can be sequenced by a chemical procedure that breaks a terminally labeled DNA molecule partially at each repetition of a base. The lengths of the labeled fragments then identify the positions of that base. We describe reactions that cleave DNA preferentially at guanines, at adenines, at cytosines and thymines equally, and at cytosines alone. When the products of these four reactions are resolved by size, by electrophoresis on a polyacrylamide gel, the DNA sequence can be read from the pattern of radioactive bands. The technique will permit sequencing of at least 100 bases from the point of labeling.
DNA sequencing with chain-terminating inhibitors. Sanger et al. (1977). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.74.12.5463
Abstract
A new method for determining nucleotide sequences in DNA is described. It is similar to the "plus and minus" method [Sanger, F. & Coulson, A. R. (1975) J. Mol. Biol. 94, 441-448] but makes use of the 2',3'-dideoxy and arabinonucleoside analogues of the normal deoxynucleoside triphosphates, which act as specific chain-terminating inhibitors of DNA polymerase. The technique has been applied to the DNA of bacteriophage varphiX174 and is more rapid and more accurate than either the plus or the minus method.
Smith et al. (1986) — fluorescence detection in automated sequencing.
Fluorescence detection in automated DNA sequence analysis. Smith et al. (1986). Nature. doi:10.1038/321674a0
Abstract
We have developed a method for the partial automation of DNA sequence analysis. Fluorescence detection of the DNA fragments is accomplished by means of a fluorophore covalently attached to the oligonucleotide primer used in enzymatic DNA sequence analysis. A different coloured fluorophore is used for each of the reactions specific for the bases A, C, G and T. The reaction mixtures are combined and co-electrophoresed down a single polyacrylamide gel tube, the separated fluorescent bands of DNA are detected near the bottom of the tube, and the sequence information is acquired directly by computer.
Ewing et al. (1998) — Phred base calling and quality scores.
Base-Calling of Automated Sequencer Traces Using Phred. I. Accuracy Assessment. Ewing et al. (1998). Genome Research. doi:10.1101/gr.8.3.175
Abstract
The availability of massive amounts of DNA sequence information has begun to revolutionize the practice of biology. As a result, current large-scale sequencing output, while impressive, is not adequate to keep pace with growing demand and, in particular, is far short of what will be required to obtain the 3-billion-base human genome sequence by the target date of 2005. To reach this goal, improved automation will be essential, and it is particularly important that human involvement in sequence data processing be significantly reduced or eliminated. Progress in this respect will require both improved accuracy of the data processing software and reliable accuracy measures to reduce the need for human involvement in error correction and make human review more efficient. Here, we describe one step toward that goal: a base-calling program for automated sequencer traces, phred, with improved accuracy. phred appears to be the first base-calling program to achieve a lower error rate than the ABI software, averaging 40%-50% fewer errors in the data sets examined independent of position in read, machine running conditions, or sequencing chemistry.
Ronaghi et al. (1998), Margulies et al. (2005) — pyrosequencing and 454.
A Sequencing Method Based on Real-Time Pyrophosphate. Ronaghi et al. (1998). Science. doi:10.1126/science.281.5375.363
Genome sequencing in microfabricated high-density picolitre reactors. Margulies et al. (2005). Nature. doi:10.1038/nature03959
Abstract
The proliferation of large-scale DNA-sequencing projects in recent years has driven a search for alternative methods to reduce time and cost. Here we describe a scalable, highly parallel sequencing system with raw throughput significantly greater than that of state-of-the-art capillary electrophoresis instruments. The apparatus uses a novel fibre-optic slide of individual wells and is able to sequence 25 million bases, at 99% or better accuracy, in one four-hour run. To achieve an approximately 100-fold increase in throughput over current Sanger sequencing technology, we have developed an emulsion method for DNA amplification and an instrument for sequencing by synthesis using a pyrosequencing protocol optimized for solid support and picolitre-scale volumes. Here we show the utility, throughput, accuracy and robustness of this system by shotgun sequencing and de novo assembly of the Mycoplasma genitalium genome with 96% coverage at 99.96% accuracy in one run of the machine.
Bentley et al. (2008) — Illumina reversible-terminator chemistry.
Accurate whole human genome sequencing using reversible terminator chemistry. Bentley et al. (2008). Nature. doi:10.1038/nature07517
Abstract
DNA sequence information underpins genetic research, enabling discoveries of important biological or medical benefit. Sequencing projects have traditionally used long (400-800 base pair) reads, but the existence of reference sequences for the human and many other genomes makes it possible to develop new, fast approaches to re-sequencing, whereby shorter reads are compared to a reference to identify intraspecies genetic variation. Here we report an approach that generates several billion bases of accurate nucleotide sequence per experiment at low cost. Single molecules of DNA are attached to a flat surface, amplified in situ and used as templates for synthetic sequencing with fluorescent reversible terminator deoxyribonucleotides. Images of the surface are analysed to generate high-quality sequence. We demonstrate application of this approach to human genome sequencing on flow-sorted X chromosomes and then scale the approach to determine the genome sequence of a male Yoruba from Ibadan, Nigeria. We build an accurate consensus sequence from >30x average depth of paired 35-base reads. We characterize four million single-nucleotide polymorphisms and four hundred thousand structural variants, many of which were previously unknown. Our approach is effective for accurate, rapid and economical whole-genome re-sequencing and many other biomedical applications.
Eid et al. (2009), Wenger et al. (2019) — PacBio single-molecule sequencing and HiFi reads.
Real-Time DNA Sequencing from Single Polymerase Molecules. Eid et al. (2009). Science. doi:10.1126/science.1162986
Abstract
We present single-molecule, real-time sequencing data obtained from a DNA polymerase performing uninterrupted template-directed synthesis using four distinguishable fluorescently labeled deoxyribonucleoside triphosphates (dNTPs). We detected the temporal order of their enzymatic incorporation into a growing DNA strand with zero-mode waveguide nanostructure arrays, which provide optical observation volume confinement and enable parallel, simultaneous detection of thousands of single-molecule sequencing reactions. Conjugation of fluorophores to the terminal phosphate moiety of the dNTPs allows continuous observation of DNA synthesis over thousands of bases without steric hindrance. The data report directly on polymerase dynamics, revealing distinct polymerization states and pause sites corresponding to DNA secondary structure. Sequence data were aligned with the known reference sequence to assay biophysical parameters of polymerization for each template position. Consensus sequences were generated from the single-molecule reads at 15-fold coverage, showing a median accuracy of 99.3%, with no systematic error beyond fluorophore-dependent error rates.
Accurate circular consensus long-read sequencing improves variant detection and assembly of a human genome. Wenger et al. (2019). Nature Biotechnology. doi:10.1038/s41587-019-0217-9
Abstract
The DNA sequencing technologies in use today produce either highly accurate short reads or less-accurate long reads. We report the optimization of circular consensus sequencing (CCS) to improve the accuracy of single-molecule real-time (SMRT) sequencing (PacBio) and generate highly accurate (99.8%) long high-fidelity (HiFi) reads with an average length of 13.5 kilobases (kb). We applied our approach to sequence the well-characterized human HG002/NA24385 genome and obtained precision and recall rates of at least 99.91% for single-nucleotide variants (SNVs), 95.98% for insertions and deletions 15 megabases (Mb) and concordance of 99.997%, substantially outperforming assembly with less-accurate long reads.
Deamer et al. (2016), M. Jain et al. (2018) — nanopore sequencing and ultra-long reads.
Three decades of nanopore sequencing. Deamer et al. (2016). Nature Biotechnology. doi:10.1038/nbt.3423
Nanopore sequencing and assembly of a human genome with ultra-long reads. Jain et al. (2018). Nature Biotechnology. doi:10.1038/nbt.4060
Abstract
We report the sequencing and assembly of a reference genome for the human GM12878 Utah/Ceph cell line using the MinION (Oxford Nanopore Technologies) nanopore sequencer. 91.2 Gb of sequence data, representing ∼30× theoretical coverage, were produced. Reference-based alignment enabled detection of large structural variants and epigenetic modifications. De novo assembly of nanopore reads alone yielded a contiguous assembly (NG50 ∼3 Mb). We developed a protocol to generate ultra-long reads (N50 > 100 kb, read lengths up to 882 kb). Incorporating an additional 5× coverage of these ultra-long reads more than doubled the assembly contiguity (NG50 ∼6.4 Mb). The final assembled genome was 2,867 million bases in size, covering 85.8% of the reference. Assembly accuracy, after incorporating complementary short-read sequencing data, exceeded 99.8%. Ultra-long reads enabled assembly and phasing of the 4-Mb major histocompatibility complex (MHC) locus in its entirety, measurement of telomere repeat length, and closure of gaps in the reference human genome assembly GRCh38.
Staden (1979), Lander and Waterman (1988) — the shotgun strategy and its mathematics.
A strategy of DNA sequencing employing computer programs. Staden (1979). Nucleic Acids Research. doi:10.1093/nar/6.7.2601
Abstract
With modern fast sequencing techniques and suitable computer programs it is now possible to sequence whole genomes without the need of restriction maps. This paper describes computer programs that can be used to order both sequence gel readings and clones. A method of coding for uncertainties in gel readings is described. These programs are available on request.
Genomic mapping by fingerprinting random clones: A mathematical analysis. Lander & Waterman (1988). Genomics. doi:10.1016/0888-7543(88)90007-9
Abstract
Results from physical mapping projects have recently been reported for the genomes of Escherichia coli, Saccharomyces cerevisiae, and Caenorhabditis elegans, and similar projects are currently being planned for other organisms. In such projects, the physical map is assembled by first "fingerprinting" a large number of clones chosen at random from a recombinant library and then inferring overlaps between clones with sufficiently similar fingerprints. Although the basic approach is the same, there are many possible choices for the fingerprint used to characterize the clones and the rules for declaring overlap. In this paper, we derive simple formulas showing how the progress of a physical mapping project is affected by the nature of the fingerprinting scheme. Using these formulas, we discuss the analytic considerations involved in selecting an appropriate fingerprinting scheme for a particular project.
B.5.2 Genome assembly
Nagarajan and Pop (2013) — sequence assembly demystified.
Sequence assembly demystified. Nagarajan & Pop (2013). Nature Reviews Genetics. doi:10.1038/nrg3367
Abstract
Advances in sequencing technologies and increased access to sequencing services have led to renewed interest in sequence and genome assembly. Concurrently, new applications for sequencing have emerged, including gene expression analysis, discovery of genomic variants and metagenomics, and each of these has different needs and challenges in terms of assembly. We survey the theoretical foundations that underlie modern assembly and highlight the options and practical trade-offs that need to be considered, focusing on how individual features address the needs of specific applications. We also review key software and the interplay between experimental design and efficacy of assembly.
Anson and Myers (1999), Myers et al. (2000) — whole-genome shotgun assembly.
Algorithms for whole genome shotgun sequencing. Anson & Myers (1999). Proceedings of the Third Annual International Conference on Computational Molecular Biology (RECOMB '99).
A Whole-Genome Assembly of Drosophila. Myers et al. (2000). Science. doi:10.1126/science.287.5461.2196
Abstract
We report on the quality of a whole-genome assembly of Drosophila melanogaster and the nature of the computer algorithms that accomplished it. Three independent external data sources essentially agree with and support the assembly's sequence and ordering of contigs across the euchromatic portion of the genome. In addition, there are isolated contigs that we believe represent nonrepetitive pockets within the heterochromatin of the centromeres. Comparison with a previously sequenced 2.9- megabase region indicates that sequencing accuracy within nonrepetitive segments is greater than 99. 99% without manual curation. As such, this initial reconstruction of the Drosophila sequence should be of substantial value to the scientific community.
★ Myers (2005) — the fragment assembly string graph.
The fragment assembly string graph. Myers (2005). Bioinformatics. doi:10.1093/bioinformatics/bti1114
Abstract
We present a concept and formalism, the string graph, which represents all that is inferable about a DNA sequence from a collection of shotgun sequencing reads collected from it. We give time and space efficient algorithms for constructing a string graph given the collection of overlaps between the reads and, in particular, present a novel linear expected time algorithm for transitive reduction in this context. The result demonstrates that the decomposition of reads into kmers employed in the de Bruijn graph approach described earlier is not essential, and exposes its close connection to the unitig approach we developed at Celera. This paper is a preliminary piece giving the basic algorithm and results that demonstrate the efficiency and scalability of the method. These ideas are being used to build a next-generation whole genome assembler called BOA (Berkeley Open Assembler) that will easily scale to mammalian genomes.
Idury and Waterman (1995), Pevzner et al. (2001) — the Eulerian path approach.
A New Algorithm for DNA Sequence Assembly. Idury & Waterman (1995). Journal of Computational Biology. doi:10.1089/cmb.1995.2.291
Abstract
Since the advent of rapid DNA sequencing methods in 1976, scientists have had the problem of inferring DNA sequences from sequenced fragments. Shotgun sequencing is a well-established biological and computational method used in practice. Many conventional algorithms for shotgun sequencing are based on the notion of pairwise fragment overlap. While shotgun sequencing infers a DNA sequence given the sequences of overlapping fragments, a recent and complementary method, called sequencing by hybridization (SBH), infers a DNA sequence given the set of oligomers that represents all subwords of some fixed length, k. In this paper, we propose a new computer algorithm for DNA sequence assembly that combines in a novel way the techniques of both shotgun and SBH methods. Based on our preliminary investigations, the algorithm promises to be very fast and practical for DNA sequence assembly.
An Eulerian path approach to DNA fragment assembly. Pevzner et al. (2001). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.171285098
Abstract
For the last 20 years, fragment assembly in DNA sequencing followed the "overlap-layout-consensus" paradigm that is used in all currently available assembly tools. Although this approach proved useful in assembling clones, it faces difficulties in genomic shotgun assembly. We abandon the classical "overlap-layout-consensus" approach in favor of a new euler algorithm that, for the first time, resolves the 20-year-old "repeat problem" in fragment assembly. Our main result is the reduction of the fragment assembly to a variation of the classical Eulerian path problem that allows one to generate accurate solutions of large-scale sequencing problems. euler, in contrast to the celera assembler, does not mask such repeats but uses them instead as a powerful fragment assembly tool.
Zerbino and Birney (2008), Simpson et al. (2009), Bankevich et al. (2012) — Velvet, ABySS and SPAdes.
Velvet: Algorithms for de novo short read assembly using de Bruijn graphs. Zerbino & Birney (2008). Genome Research. doi:10.1101/gr.074492.107
Abstract
We have developed a new set of algorithms, collectively called "Velvet," to manipulate de Bruijn graphs for genomic sequence assembly. A de Bruijn graph is a compact representation based on short words (k-mers) that is ideal for high coverage, very short read (25-50 bp) data sets. Applying Velvet to very short reads and paired-ends information only, one can produce contigs of significant length, up to 50-kb N50 length in simulations of prokaryotic data and 3-kb N50 on simulated mammalian BACs. When applied to real Solexa data sets without read pairs, Velvet generated contigs of approximately 8 kb in a prokaryote and 2 kb in a mammalian BAC, in close agreement with our simulated results without read-pair information. Velvet represents a new approach to assembly that can leverage very short reads in combination with read pairs to produce useful assemblies.
ABySS: A parallel assembler for short read sequence data. Simpson et al. (2009). Genome Research. doi:10.1101/gr.089532.108
Abstract
Widespread adoption of massively parallel deoxyribonucleic acid (DNA) sequencing instruments has prompted the recent development of de novo short read assembly algorithms. A common shortcoming of the available tools is their inability to efficiently assemble vast amounts of data generated from large-scale sequencing projects, such as the sequencing of individual human genomes to catalog natural genetic variation. To address this limitation, we developed ABySS (Assembly By Short Sequences), a parallelized sequence assembler. As a demonstration of the capability of our software, we assembled 3.5 billion paired-end reads from the genome of an African male publicly released by Illumina, Inc. Approximately 2.76 million contigs > or =100 base pairs (bp) in length were created with an N50 size of 1499 bp, representing 68% of the reference human genome. Analysis of these contigs identified polymorphic and novel sequences not present in the human reference assembly, which were validated by alignment to alternate human assemblies and to other primate genomes.
SPAdes: A New Genome Assembly Algorithm and Its Applications to Single-Cell Sequencing. Bankevich et al. (2012). Journal of Computational Biology. doi:10.1089/cmb.2012.0021
Abstract
The lion's share of bacteria in various environments cannot be cloned in the laboratory and thus cannot be sequenced using existing technologies. A major goal of single-cell genomics is to complement gene-centric metagenomic data with whole-genome assemblies of uncultivated organisms. Assembly of single-cell data is challenging because of highly non-uniform read coverage as well as elevated levels of sequencing errors and chimeric reads. We describe SPAdes, a new assembler for both single-cell and standard (multicell) assembly, and demonstrate that it improves on the recently released E+V-SC assembler (specialized for single-cell data) and on popular assemblers Velvet and SoapDeNovo (for multicell data). SPAdes generates single-cell assemblies, providing information about genomes of uncultivatable bacteria that vastly exceeds what may be obtained via traditional metagenomics studies. SPAdes is available online ( http://bioinf.spbau.ru/spades ). It is distributed as open source software.
Chin et al. (2013), Koren et al. (2017), Kolmogorov et al. (2019) — long-read assembly (HGAP, Canu, Flye).
Nonhybrid, finished microbial genome assemblies from long-read SMRT sequencing data. Chin et al. (2013). Nature Methods. doi:10.1038/nmeth.2474
Abstract
We present a hierarchical genome-assembly process (HGAP) for high-quality de novo microbial genome assemblies using only a single, long-insert shotgun DNA library in conjunction with Single Molecule, Real-Time (SMRT) DNA sequencing. Our method uses the longest reads as seeds to recruit all other reads for construction of highly accurate preassembled reads through a directed acyclic graph-based consensus procedure, which we follow with assembly using off-the-shelf long-read assemblers. In contrast to hybrid approaches, HGAP does not require highly accurate raw reads for error correction. We demonstrate efficient genome assembly for several microorganisms using as few as three SMRT Cell zero-mode waveguide arrays of sequencing and for BACs using just one SMRT Cell. Long repeat regions can be successfully resolved with this workflow. We also describe a consensus algorithm that incorporates SMRT sequencing primary quality values to produce de novo genome sequence exceeding 99.999% accuracy.
Canu: scalable and accurate long-read assembly via adaptive k -mer weighting and repeat separation. Koren et al. (2017). Genome Research. doi:10.1101/gr.215087.116
Abstract
Long-read single-molecule sequencing has revolutionized de novo genome assembly and enabled the automated reconstruction of reference-quality genomes. However, given the relatively high error rates of such technologies, efficient and accurate assembly of large repeats and closely related haplotypes remains challenging. We address these issues with Canu, a successor of Celera Assembler that is specifically designed for noisy single-molecule sequences. Canu introduces support for nanopore sequencing, halves depth-of-coverage requirements, and improves assembly continuity while simultaneously reducing runtime by an order of magnitude on large genomes versus Celera Assembler 8.2. These advances result from new overlapping and assembly algorithms, including an adaptive overlapping strategy based on tf-idf weighted MinHash and a sparse assembly graph construction that avoids collapsing diverged repeats and haplotypes. We demonstrate that Canu can reliably assemble complete microbial genomes and near-complete eukaryotic chromosomes using either Pacific Biosciences (PacBio) or Oxford Nanopore technologies and achieves a contig NG50 of >21 Mbp on both human and Drosophila melanogaster PacBio data sets. For assembly structures that cannot be linearly represented, Canu provides graph-based assembly outputs in graphical fragment assembly (GFA) format for analysis or integration with complementary phasing and scaffolding techniques. The combination of such highly resolved assembly graphs with long-range scaffolding information promises the complete and automated assembly of complex genomes.
Assembly of long, error-prone reads using repeat graphs. Kolmogorov et al. (2019). Nature Biotechnology. doi:10.1038/s41587-019-0072-8
Abstract
Accurate genome assembly is hampered by repetitive regions. Although long single molecule sequencing reads are better able to resolve genomic repeats than short-read data, most long-read assembly algorithms do not provide the repeat characterization necessary for producing optimal assemblies. Here, we present Flye, a long-read assembly algorithm that generates arbitrary paths in an unknown repeat graph, called disjointigs, and constructs an accurate repeat graph from these error-riddled disjointigs. We benchmark Flye against five state-of-the-art assemblers and show that it generates better or comparable assemblies, while being an order of magnitude faster. Flye nearly doubled the contiguity of the human genome assembly (as measured by the NGA50 assembly quality metric) compared with existing assemblers.
Cheng et al. (2021) — hifiasm: haplotype-resolved assembly.
Haplotype-resolved de novo assembly using phased assembly graphs with hifiasm. Cheng et al. (2021). Nature Methods. doi:10.1038/s41592-020-01056-5
Abstract
Haplotype-resolved de novo assembly is the ultimate solution to the study of sequence variations in a genome. However, existing algorithms either collapse heterozygous alleles into one consensus copy or fail to cleanly separate the haplotypes to produce high-quality phased assemblies. Here we describe hifiasm, a de novo assembler that takes advantage of long high-fidelity sequence reads to faithfully represent the haplotype information in a phased assembly graph. Unlike other graph-based assemblers that only aim to maintain the contiguity of one haplotype, hifiasm strives to preserve the contiguity of all haplotypes. This feature enables the development of a graph trio binning algorithm that greatly advances over standard trio binning. On three human and five nonhuman datasets, including California redwood with a ~30-Gb hexaploid genome, we show that hifiasm frequently delivers better assemblies than existing tools and consistently outperforms others on haplotype-resolved assembly.
Burton et al. (2013) — chromosome-scale scaffolding with Hi-C.
Chromosome-scale scaffolding of de novo genome assemblies based on chromatin interactions. Burton et al. (2013). Nature Biotechnology. doi:10.1038/nbt.2727
Abstract
Genomes assembled de novo from short reads are highly fragmented relative to the finished chromosomes of Homo sapiens and key model organisms generated by the Human Genome Project. To address this problem, we need scalable, cost-effective methods to obtain assemblies with chromosome-scale contiguity. Here we show that genome-wide chromatin interaction data sets, such as those generated by Hi-C, are a rich source of long-range information for assigning, ordering and orienting genomic sequences to chromosomes, including across centromeres. To exploit this finding, we developed an algorithm that uses Hi-C data for ultra-long-range scaffolding of de novo genome assemblies. We demonstrate the approach by combining shotgun fragment and short jump mate-pair sequences with Hi-C data to generate chromosome-scale de novo assemblies of the human, mouse and Drosophila genomes, achieving--for the human genome--98% accuracy in assigning scaffolds to chromosome groups and 99% accuracy in ordering and orienting scaffolds within chromosome groups. Hi-C data can also be used to validate chromosomal translocations in cancer genomes.
Bradnam et al. (2013) — Assemblathon 2.
Assemblathon 2: evaluating de novo methods of genome assembly in three vertebrate species. Bradnam et al. (2013). Gigascience. doi:10.1186/2047-217x-2-10
Abstract
Background The process of generating raw genome sequence data continues to become cheaper, faster, and more accurate. However, assembly of such data into high-quality, finished genome sequences remains challenging. Many genome assembly tools are available, but they differ greatly in terms of their performance (speed, scalability, hardware requirements, acceptance of newer read technologies) and in their final output (composition of assembled sequence). More importantly, it remains largely unclear how to best assess the quality of assembled genome sequences. The Assemblathon competitions are intended to assess current state-of-the-art methods in genome assembly. Results In Assemblathon 2, we provided a variety of sequence data to be assembled for three vertebrate species (a bird, a fish, and snake). This resulted in a total of 43 submitted assemblies from 21 participating teams. We evaluated these assemblies using a combination of optical map data, Fosmid sequences, and several statistical methods. From over 100 different metrics, we chose ten key measures by which to assess the overall quality of the assemblies. Conclusions Many current genome assemblers produced useful assemblies, containing a significant representation of their genes and overall genome structure. However, the high degree of variability between the entries suggests that there is still much room for improvement in the field of genome assembly and that approaches which work well in assembling the genome of one species may not necessarily work well for another.
Simão et al. (2015), Rhie et al. (2020) — BUSCO and Merqury.
BUSCO: assessing genome assembly and annotation completeness with single-copy orthologs. Simão et al. (2015). Bioinformatics. doi:10.1093/bioinformatics/btv351
Abstract
Motivation Genomics has revolutionized biological research, but quality assessment of the resulting assembled sequences is complicated and remains mostly limited to technical measures like N50. Results We propose a measure for quantitative assessment of genome assembly and annotation completeness based on evolutionarily informed expectations of gene content. We implemented the assessment procedure in open-source software, with sets of Benchmarking Universal Single-Copy Orthologs, named BUSCO. Availability and implementation Software implemented in Python and datasets available for download from http://busco.ezlab.org. Contact evgeny.zdobnov@unige.ch Supplementary information Supplementary data are available at Bioinformatics online.
Merqury: reference-free quality, completeness, and phasing assessment for genome assemblies. Rhie et al. (2020). Genome Biology. doi:10.1186/s13059-020-02134-9
Abstract
Recent long-read assemblies often exceed the quality and completeness of available reference genomes, making validation challenging. Here we present Merqury, a novel tool for reference-free assembly evaluation based on efficient k-mer set operations. By comparing k-mers in a de novo assembly to those found in unassembled high-accuracy reads, Merqury estimates base-level accuracy and completeness. For trios, Merqury can also evaluate haplotype-specific accuracy, completeness, phase block continuity, and switch errors. Multiple visualizations, such as k-mer spectrum plots, can be generated for evaluation. We demonstrate on both human and plant genomes that Merqury is a fast and robust method for assembly validation.
Nurk et al. (2022) — the complete human genome (T2T).
The complete sequence of a human genome. Nurk et al. (2022). Science. doi:10.1126/science.abj6987
Abstract
Since its initial release in 2000, the human reference genome has covered only the euchromatic fraction of the genome, leaving important heterochromatic regions unfinished. Addressing the remaining 8% of the genome, the Telomere-to-Telomere (T2T) Consortium presents a complete 3.055 billion-base pair sequence of a human genome, T2T-CHM13, that includes gapless assemblies for all chromosomes except Y, corrects errors in the prior references, and introduces nearly 200 million base pairs of sequence containing 1956 gene predictions, 99 of which are predicted to be protein coding. The completed regions include all centromeric satellite arrays, recent segmental duplications, and the short arms of all five acrocentric chromosomes, unlocking these complex regions of the genome to variational and functional studies.
Rhie et al. (2021) — reference-quality genomes for all vertebrates.
Towards complete and error-free genome assemblies of all vertebrate species. Rhie et al. (2021). Nature. doi:10.1038/s41586-021-03451-0
Abstract
High-quality and complete reference genome assemblies are fundamental for the application of genomics to biology, disease, and biodiversity conservation. However, such assemblies are available for only a few non-microbial species 1-4 . To address this issue, the international Genome 10K (G10K) consortium 5,6 has worked over a five-year period to evaluate and develop cost-effective methods for assembling highly accurate and nearly complete reference genomes. Here we present lessons learned from generating assemblies for 16 species that represent six major vertebrate lineages. We confirm that long-read sequencing technologies are essential for maximizing genome quality, and that unresolved complex repeats and haplotype heterozygosity are major sources of assembly error when not handled correctly. Our assemblies correct substantial errors, add missing sequence in some of the best historical reference genomes, and reveal biological discoveries. These include the identification of many false gene duplications, increases in gene sizes, chromosome rearrangements that are specific to lineages, a repeated independent chromosome breakpoint in bat genomes, and a canonical GC-rich pattern in protein-coding genes and their regulatory regions. Adopting these lessons, we have embarked on the Vertebrate Genomes Project (VGP), an international effort to generate high-quality, complete reference genomes for all of the roughly 70,000 extant vertebrate species and to help to enable a new era of discovery across the life sciences.
Liao et al. (2023) — a draft human pangenome reference.
A draft human pangenome reference. Liao et al. (2023). Nature. doi:10.1038/s41586-023-05896-x
Abstract
Here the Human Pangenome Reference Consortium presents a first draft of the human pangenome reference. The pangenome contains 47 phased, diploid assemblies from a cohort of genetically diverse individuals 1 . These assemblies cover more than 99% of the expected sequence in each genome and are more than 99% accurate at the structural and base pair levels. Based on alignments of the assemblies, we generate a draft pangenome that captures known variants and haplotypes and reveals new alleles at structurally complex loci. We also add 119 million base pairs of euchromatic polymorphic sequences and 1,115 gene duplications relative to the existing reference GRCh38. Roughly 90 million of the additional base pairs are derived from structural variation. Using our draft pangenome to analyse short-read data reduced small variant discovery errors by 34% and increased the number of structural variants detected per haplotype by 104% compared with GRCh38-based workflows, which enabled the typing of the vast majority of structural variant alleles per sample.
B.5.3 Read mapping and variant calling
Li and Durbin (2009) — BWA.
Fast and accurate short read alignment with Burrows–Wheeler transform. Li & Durbin (2009). Bioinformatics. doi:10.1093/bioinformatics/btp324
Abstract
Motivation The enormous amount of short reads generated by the new DNA sequencing technologies call for the development of fast and accurate read alignment programs. A first generation of hash table-based methods has been developed, including MAQ, which is accurate, feature rich and fast enough to align short reads from a single individual. However, MAQ does not support gapped alignment for single-end reads, which makes it unsuitable for alignment of longer reads where indels may occur frequently. The speed of MAQ is also a concern when the alignment is scaled up to the resequencing of hundreds of individuals. Results We implemented Burrows-Wheeler Alignment tool (BWA), a new read alignment package that is based on backward search with Burrows-Wheeler Transform (BWT), to efficiently align short sequencing reads against a large reference sequence such as the human genome, allowing mismatches and gaps. BWA supports both base space reads, e.g. from Illumina sequencing machines, and color space reads from AB SOLiD machines. Evaluations on both simulated and real data suggest that BWA is approximately 10-20x faster than MAQ, while achieving similar accuracy. In addition, BWA outputs alignment in the new standard SAM (Sequence Alignment/Map) format. Variant calling and other downstream analyses after the alignment can be achieved with the open source SAMtools software package. Availability http://maq.sourceforge.net.
Li (2018) — minimap2.
Minimap2: pairwise alignment for nucleotide sequences. Li (2018). Bioinformatics. doi:10.1093/bioinformatics/bty191
Abstract
Motivation Recent advances in sequencing technologies promise ultra-long reads of ∼100 kb in average, full-length mRNA or cDNA reads in high throughput and genomic contigs over 100 Mb in length. Existing alignment programs are unable or inefficient to process such data at scale, which presses for the development of new alignment algorithms. Results Minimap2 is a general-purpose alignment program to map DNA or long mRNA sequences against a large reference database. It works with accurate short reads of ≥100 bp in length, ≥1 kb genomic reads at error rate ∼15%, full-length noisy Direct RNA or cDNA reads and assembly contigs or closely related full chromosomes of hundreds of megabases in length. Minimap2 does split-read alignment, employs concave gap cost for long insertions and deletions and introduces new heuristics to reduce spurious alignments. It is 3-4 times as fast as mainstream short-read mappers at comparable accuracy, and is ≥30 times faster than long-read genomic or cDNA mappers at higher accuracy, surpassing most aligners specialized in one type of alignment. Availability and implementation https://github.com/lh3/minimap2. Supplementary information Supplementary data are available at Bioinformatics online.
Li et al. (2009), Danecek et al. (2021) — the SAM format, SAMtools and BCFtools.
The Sequence Alignment/Map format and SAMtools. Li et al. (2009). Bioinformatics. doi:10.1093/bioinformatics/btp352
Abstract
The Sequence Alignment/Map (SAM) format is a generic alignment format for storing read alignments against reference sequences, supporting short and long reads (up to 128 Mbp) produced by different sequencing platforms. It is flexible in style, compact in size, efficient in random access and is the format in which alignments from the 1000 Genomes Project are released. SAMtools implements various utilities for post-processing alignments in the SAM format, such as indexing, variant caller and alignment viewer, and thus provides universal tools for processing read alignments. Availability http://samtools.sourceforge.net.
Twelve years of SAMtools and BCFtools. Danecek et al. (2021). GigaScience. doi:10.1093/gigascience/giab008
Abstract
Background SAMtools and BCFtools are widely used programs for processing and analysing high-throughput sequencing data. They include tools for file format conversion and manipulation, sorting, querying, statistics, variant calling, and effect analysis amongst other methods. Findings The first version appeared online 12 years ago and has been maintained and further developed ever since, with many new features and improvements added over the years. The SAMtools and BCFtools packages represent a unique collection of tools that have been used in numerous other software projects and countless genomic pipelines. Conclusion Both SAMtools and BCFtools are freely available on GitHub under the permissive MIT licence, free for both non-commercial and commercial use. Both packages have been installed >1 million times via Bioconda. The source code and documentation are available from https://www.htslib.org.
Garrison et al. (2018) — graph genomes and mapping with vg.
Variation graph toolkit improves read mapping by representing genetic variation in the reference. Garrison et al. (2018). Nature Biotechnology. doi:10.1038/nbt.4227
Abstract
Reference genomes guide our interpretation of DNA sequence data. However, conventional linear references represent only one version of each locus, ignoring variation in the population. Poor representation of an individual's genome sequence impacts read mapping and introduces bias. Variation graphs are bidirected DNA sequence graphs that compactly represent genetic variation across a population, including large-scale structural variation such as inversions and duplications. Previous graph genome software implementations have been limited by scalability or topological constraints. Here we present vg, a toolkit of computational methods for creating, manipulating, and using these structures as references at the scale of the human genome. vg provides an efficient approach to mapping reads onto arbitrary variation graphs using generalized compressed suffix arrays, with improved accuracy over alignment to a linear reference, and effectively removing reference bias. These capabilities make using variation graphs as references for DNA sequencing practical at a gigabase scale, or at the topological complexity of de novo assemblies.
Li et al. (2008) — mapping quality and variant calling (MAQ).
Mapping short DNA sequencing reads and calling variants using mapping quality scores. Li et al. (2008). Genome Research. doi:10.1101/gr.078212.108
Abstract
New sequencing technologies promise a new era in the use of DNA sequence. However, some of these technologies produce very short reads, typically of a few tens of base pairs, and to use these reads effectively requires new algorithms and software. In particular, there is a major issue in efficiently aligning short reads to a reference genome and handling ambiguity or lack of accuracy in this alignment. Here we introduce the concept of mapping quality, a measure of the confidence that a read actually comes from the position it is aligned to by the mapping algorithm. We describe the software MAQ that can build assemblies by mapping shotgun short reads to a reference genome, using quality scores to derive genotype calls of the consensus sequence of a diploid genome, e.g., from a human sample. MAQ makes full use of mate-pair information and estimates the error probability of each read alignment. Error probabilities are also derived for the final genotype calls, using a Bayesian statistical model that incorporates the mapping qualities, error probabilities from the raw sequence quality scores, sampling of the two haplotypes, and an empirical model for correlated errors at a site. Both read mapping and genotype calling are evaluated on simulated data and real data. MAQ is accurate, efficient, versatile, and user-friendly. It is freely available at http://maq.sourceforge.net.
Marth et al. (1999), Li (2011) — probabilistic SNP discovery and calling.
A general approach to single-nucleotide polymorphism discovery. Marth et al. (1999). Nature Genetics.
A statistical framework for SNP calling, mutation discovery, association mapping and population genetical parameter estimation from sequencing data. Li (2011). Bioinformatics. doi:10.1093/bioinformatics/btr509
Abstract
Motivation Most existing methods for DNA sequence analysis rely on accurate sequences or genotypes. However, in applications of the next-generation sequencing (NGS), accurate genotypes may not be easily obtained (e.g. multi-sample low-coverage sequencing or somatic mutation discovery). These applications press for the development of new methods for analyzing sequence data with uncertainty. Results We present a statistical framework for calling SNPs, discovering somatic mutations, inferring population genetical parameters and performing association tests directly based on sequencing data without explicit genotyping or linkage-based imputation. On real data, we demonstrate that our method achieves comparable accuracy to alternative methods for estimating site allele count, for inferring allele frequency spectrum and for association mapping. We also highlight the necessity of using symmetric datasets for finding somatic mutations and confirm that for discovering rare events, mismapping is frequently the leading source of errors. Availability http://samtools.sourceforge.net. Contact hengli@broadinstitute.org.
McKenna et al. (2010), DePristo et al. (2011) — GATK.
The Genome Analysis Toolkit: A MapReduce framework for analyzing next-generation DNA sequencing data. McKenna et al. (2010). Genome Research. doi:10.1101/gr.107524.110
Abstract
Next-generation DNA sequencing (NGS) projects, such as the 1000 Genomes Project, are already revolutionizing our understanding of genetic variation among individuals. However, the massive data sets generated by NGS--the 1000 Genome pilot alone includes nearly five terabases--make writing feature-rich, efficient, and robust analysis tools difficult for even computationally sophisticated individuals. Indeed, many professionals are limited in the scope and the ease with which they can answer scientific questions by the complexity of accessing and manipulating the data produced by these machines. Here, we discuss our Genome Analysis Toolkit (GATK), a structured programming framework designed to ease the development of efficient and robust analysis tools for next-generation DNA sequencers using the functional programming philosophy of MapReduce. The GATK provides a small but rich set of data access patterns that encompass the majority of analysis tool needs. Separating specific analysis calculations from common data management infrastructure enables us to optimize the GATK framework for correctness, stability, and CPU and memory efficiency and to enable distributed and shared memory parallelization. We highlight the capabilities of the GATK by describing the implementation and application of robust, scale-tolerant tools like coverage calculators and single nucleotide polymorphism (SNP) calling. We conclude that the GATK programming framework enables developers and analysts to quickly and easily write efficient and robust NGS tools, many of which have already been incorporated into large-scale sequencing projects like the 1000 Genomes Project and The Cancer Genome Atlas.
A framework for variation discovery and genotyping using next-generation DNA sequencing data. DePristo et al. (2011). Nature Genetics. doi:10.1038/ng.806
Abstract
Recent advances in sequencing technology make it possible to comprehensively catalog genetic variation in population samples, creating a foundation for understanding human disease, ancestry and evolution. The amounts of raw data produced are prodigious, and many computational steps are required to translate this output into high-quality variant calls. We present a unified analytic framework to discover and genotype variation among multiple samples simultaneously that achieves sensitive and specific results across five sequencing technologies and three distinct, canonical experimental designs. Our process includes (i) initial read mapping; (ii) local realignment around indels; (iii) base quality score recalibration; (iv) SNP discovery and genotyping to find all potential variants; and (v) machine learning to separate true segregating variation from machine artifacts common to next-generation sequencing technologies. We here discuss the application of these tools, instantiated in the Genome Analysis Toolkit, to deep whole-genome, whole-exome capture and multi-sample low-pass (∼4×) 1000 Genomes Project datasets.
Poplin et al. (2018) — DeepVariant.
A universal SNP and small-indel variant caller using deep neural networks. Poplin et al. (2018). Nature Biotechnology. doi:10.1038/nbt.4235
Abstract
Despite rapid advances in sequencing technologies, accurately calling genetic variants present in an individual genome from billions of short, errorful sequence reads remains challenging. Here we show that a deep convolutional neural network can call genetic variation in aligned next-generation sequencing read data by learning statistical relationships between images of read pileups around putative variant and true genotype calls. The approach, called DeepVariant, outperforms existing state-of-the-art tools. The learned model generalizes across genome builds and mammalian species, allowing nonhuman sequencing projects to benefit from the wealth of human ground-truth data. We further show that DeepVariant can learn to call variants in a variety of sequencing technologies and experimental designs, including deep whole genomes from 10X Genomics and Ion Ampliseq exomes, highlighting the benefits of using more automated and generalizable techniques for variant calling.
Zook et al. (2019) — Genome in a Bottle benchmarks.
An open resource for accurately benchmarking small variant and reference calls. Zook et al. (2019). Nature Biotechnology. doi:10.1038/s41587-019-0074-6
Abstract
Benchmark small variant calls are required for developing, optimizing and assessing the performance of sequencing and bioinformatics methods. Here, as part of the Genome in a Bottle (GIAB) Consortium, we apply a reproducible, cloud-based pipeline to integrate multiple short- and linked-read sequencing datasets and provide benchmark calls for human genomes. We generate benchmark calls for one previously analyzed GIAB sample, as well as six genomes from the Personal Genome Project. These new genomes have broad, open consent, making this a 'first of its kind' resource that is available to the community for multiple downstream applications. We produce 17% more benchmark single nucleotide variations, 176% more indels and 12% larger benchmark regions than previously published GIAB benchmarks. We demonstrate that this benchmark reliably identifies errors in existing callsets and highlight challenges in interpreting performance metrics when using benchmarks that are not perfect or comprehensive. Finally, we identify strengths and weaknesses of callsets by stratifying performance according to variant type and genome context.
The 1000 Genomes Project Consortium (2015) — the 1000 Genomes Project.
A global reference for human genetic variation. The 1000 Genomes Project Consortium (2015). Nature. doi:10.1038/nature15393
Abstract
The 1000 Genomes Project set out to provide a comprehensive description of common human genetic variation by applying whole-genome sequencing to a diverse set of individuals from multiple populations. Here we report completion of the project, having reconstructed the genomes of 2,504 individuals from 26 populations using a combination of low-coverage whole-genome sequencing, deep exome sequencing, and dense microarray genotyping. We characterized a broad spectrum of genetic variation, in total over 88 million variants (84.7 million single nucleotide polymorphisms (SNPs), 3.6 million short insertions/deletions (indels), and 60,000 structural variants), all phased onto high-quality haplotypes. This resource includes >99% of SNP variants with a frequency of >1% for a variety of ancestries. We describe the distribution of genetic variation across the global sample, and discuss the implications for common disease studies.
Mahmoud et al. (2019), Sedlazeck et al. (2018) — structural variant calling.
Structural variant calling: the long and the short of it. Mahmoud et al. (2019). Genome Biology. doi:10.1186/s13059-019-1828-7
Abstract
Recent research into structural variants (SVs) has established their importance to medicine and molecular biology, elucidating their role in various diseases, regulation of gene expression, ethnic diversity, and large-scale chromosome evolution-giving rise to the differences within populations and among species. Nevertheless, characterizing SVs and determining the optimal approach for a given experimental design remains a computational and scientific challenge. Multiple approaches have emerged to target various SV classes, zygosities, and size ranges. Here, we review these approaches with respect to their ability to infer SVs across the full spectrum of large, complex variations and present computational methods for each approach.
Accurate detection of complex structural variations using single-molecule sequencing. Sedlazeck et al. (2018). Nature Methods. doi:10.1038/s41592-018-0001-7
Abstract
Structural variations are the greatest source of genetic variation, but they remain poorly understood because of technological limitations. Single-molecule long-read sequencing has the potential to dramatically advance the field, although high error rates are a challenge with existing methods. Addressing this need, we introduce open-source methods for long-read alignment (NGMLR; https://github.com/philres/ngmlr ) and structural variant identification (Sniffles; https://github.com/fritzsedlazeck/Sniffles ) that provide unprecedented sensitivity and precision for variant detection, even in repeat-rich regions and for complex nested events that can have substantial effects on human health. In several long-read datasets, including healthy and cancerous human genomes, we discovered thousands of novel variants and categorized systematic errors in short-read approaches. NGMLR and Sniffles can automatically filter false events and operate on low-coverage data, thereby reducing the high costs that have hindered the application of long reads in clinical and research settings.
B.5.4 Gene prediction
Fickett (1982) — recognising coding regions.
Recognition of protein coding regions in DNA sequences. Fickett (1982). Nucleic Acids Research. doi:10.1093/nar/10.17.5303
Abstract
We give a test for protein coding regions which is based on simple and universal differences between protein-coding and noncoding DNA. The test is simple enough to use without a computer and is completely objective. The test has been thoroughly proven on 400,000 bases of sequence data: it misclassifies 5% of the regions tested and gives an answer of "No Opinion" one fifth of the time. We predict some new coding and noncoding regions in published sequences.
Borodovsky and McIninch (1993) — GeneMark.
GENMARK: Parallel gene recognition for both DNA strands. Borodovsky & McIninch (1993). Computers & Chemistry. doi:10.1016/0097-8485(93)85004-v
★ Lukashin and Borodovsky (1998) — GeneMark.hmm.
GeneMark.hmm: new solutions for gene finding. Lukashin & Borodovsky (1998). Nucleic Acids Research. doi:10.1093/nar/26.4.1107
Abstract
The number of completely sequenced bacterial genomes has been growing fast. There are computer methods available for finding genes but yet there is a need for more accurate algorithms. The GeneMark. hmm algorithm presented here was designed to improve the gene prediction quality in terms of finding exact gene boundaries. The idea was to embed the GeneMark models into naturally derived hidden Markov model framework with gene boundaries modeled as transitions between hidden states. We also used the specially derived ribosome binding site pattern to refine predictions of translation initiation codons. The algorithm was evaluated on several test sets including 10 complete bacterial genomes. It was shown that the new algorithm is significantly more accurate than GeneMark in exact gene prediction. Interestingly, the high gene finding accuracy was observed even in the case when Markov models of order zero, one and two were used. We present the analysis of false positive and false negative predictions with the caution that these categories are not precisely defined if the public database annotation is used as a control.
Delcher (1999), Hyatt et al. (2010) — Glimmer and Prodigal.
Improved microbial gene identification with GLIMMER. Delcher (1999). Nucleic Acids Research. doi:10.1093/nar/27.23.4636
Abstract
The GLIMMER system for microbial gene identification finds approximately 97-98% of all genes in a genome when compared with published annotation. This paper reports on two new results: (i) significant technical improvements to GLIMMER that improve its accuracy still further, and (ii) a comprehensive evaluation that demonstrates that the accuracy of the system is likely to be higher than previously recognized. A significant proportion of the genes missed by the system appear to be hypothetical proteins whose existence is only supported by the predictions of other programs. When the analysis is restricted to genes that have significant homology to genes in other organisms, GLIMMER misses <1% of known genes.
Prodigal: prokaryotic gene recognition and translation initiation site identification. Hyatt et al. (2010). BMC Bioinformatics. doi:10.1186/1471-2105-11-119
Abstract
Background The quality of automated gene prediction in microbial organisms has improved steadily over the past decade, but there is still room for improvement. Increasing the number of correct identifications, both of genes and of the translation initiation sites for each gene, and reducing the overall number of false positives, are all desirable goals. Results With our years of experience in manually curating genomes for the Joint Genome Institute, we developed a new gene prediction algorithm called Prodigal (PROkaryotic DYnamic programming Gene-finding ALgorithm). With Prodigal, we focused specifically on the three goals of improved gene structure prediction, improved translation initiation site recognition, and reduced false positives. We compared the results of Prodigal to existing gene-finding methods to demonstrate that it met each of these objectives. Conclusion We built a fast, lightweight, open source gene prediction program called Prodigal http://compbio.ornl.gov/prodigal/. Prodigal achieved good results compared to existing methods, and we believe it will be a valuable asset to automated microbial annotation pipelines.
Guigó et al. (1992) — GeneID: prediction of gene structure.
Prediction of gene structure. Guigó et al. (1992). Journal of Molecular Biology. doi:10.1016/0022-2836(92)90130-c
Abstract
We have developed a hierarchical rule base system for identifying genes in DNA sequences. Atomic sites (such as initiation codons, stop codons, acceptor sites and donor sites) are identified by a number of different methods and evaluated by a set of filters and rules chosen to maximize sensitivity; these are combined into higher-order gene elements (such as exons), evaluated, filtered and combined as equivalence classes into probable genes, which are evaluated and ranked. The system has been tested on an extensive collection of vertebrate genes smaller than 15,000 bases. Results obtained show that, on average, 88% of the predicted coding region for a transcription unit is actually coding, and 80% of the actual coding is correctly predicted. This will, in most applications, be sufficient for a search against protein sequence databases for the identification of probable gene function. In addition, the system provides a general test platform for both gene atomic site identification and the rules for their evaluation and assembly.
Burge and Karlin (1997) — GENSCAN.
Prediction of complete gene structures in human genomic DNA. Burge & Karlin (1997). Journal of Molecular Biology. doi:10.1006/jmbi.1997.0951
Abstract
We introduce a general probabilistic model of the gene structure of human genomic sequences which incorporates descriptions of the basic transcriptional, translational and splicing signals, as well as length distributions and compositional features of exons, introns and intergenic regions. Distinct sets of model parameters are derived to account for the many substantial differences in gene density and structure observed in distinct C + G compositional regions of the human genome. In addition, new models of the donor and acceptor splice signals are described which capture potentially important dependencies between signal positions. The model is applied to the problem of gene identification in a computer program, GENSCAN, which identifies complete exon/intron structures of genes in genomic DNA. Novel features of the program include the capacity to predict multiple genes in a sequence, to deal with partial as well as complete genes, and to predict consistent sets of genes occurring on either or both DNA strands. GENSCAN is shown to have substantially higher accuracy than existing methods when tested on standardized sets of human and vertebrate genes, with 75 to 80% of exons identified exactly. The program is also capable of indicating fairly accurately the reliability of each predicted exon. Consistently high levels of accuracy are observed for sequences of differing C + G content and for distinct groups of vertebrates.
Burset and Guigó (1996) — evaluation of gene prediction programs.
Evaluation of Gene Structure Prediction Programs. Burset & Guigó (1996). Genomics. doi:10.1006/geno.1996.0298
Abstract
We evaluate a number of computer programs designed to predict the structure of protein coding genes in genomic DNA sequences. Computational gene identification is set to play an increasingly important role in the development of the genome projects, as emphasis turns from mapping to large-scale sequencing. The evaluation presented here serves both to assess the current status of the problem and to identify the most promising approaches to ensure further progress. The programs analyzed were uniformly tested on a large set of vertebrate sequences with simple gene structure, and several measures of predictive accuracy were computed at the nucleotide, exon, and protein product levels. The results indicated that the predictive accuracy of the programs analyzed was lower than originally found. The accuracy was even lower when considering only those sequences that had recently been entered and that did not show any similarity to previously entered sequences. This indicates that the programs are overly dependent on the particularities of the examples they learn from. For most of the programs, accuracy in this test set ranged from 0.60 to 0.70 as measured by the Correlation Coefficient (where 1.0 corresponds to a perfect prediction and 0.0 is the value expected for a random prediction), and the average percentage of exons exactly identified was less than 50%. Only those programs including protein sequence database searches showed substantially greater accuracy. The accuracy of the programs was severely affected by relatively high rates of sequence errors. Since the set on which the programs were tested included only relatively short sequences with simple gene structure, the accuracy of the programs is likely to be even lower when used for large uncharacterized genomic sequences with complex structure. While in such cases, programs currently available may still be of great use in pinpointing the regions likely to contain exons, they are far from being powerful enough to elucidate its genomic structure completely.
Stanke and Waack (2003), Korf (2004) — AUGUSTUS and SNAP.
Gene prediction with a hidden Markov model and a new intron submodel. Stanke & Waack (2003). Bioinformatics. doi:10.1093/bioinformatics/btg1080
Abstract
Motivation The problem of finding the genes in eukaryotic DNA sequences by computational methods is still not satisfactorily solved. Gene finding programs have achieved relatively high accuracy on short genomic sequences but do not perform well on longer sequences with an unknown number of genes in them. Here existing programs tend to predict many false exons. Results We have developed a new program, AUGUSTUS, for the ab initio prediction of protein coding genes in eukaryotic genomes. The program is based on a Hidden Markov Model and integrates a number of known methods and submodels. It employs a new way of modeling intron lengths. We use a new donor splice site model, a new model for a short region directly upstream of the donor splice site model that takes the reading frame into account and apply a method that allows better GC-content dependent parameter estimation. AUGUSTUS predicts on longer sequences far more human and drosophila genes accurately than the ab initio gene prediction programs we compared it with, while at the same time being more specific. Availability A web interface for AUGUSTUS and the executable program are located at http://augustus.gobics.de.
Gene finding in novel genomes. Korf (2004). BMC Bioinformatics. doi:10.1186/1471-2105-5-59
Abstract
Background Computational gene prediction continues to be an important problem, especially for genomes with little experimental data. Results I introduce the SNAP gene finder which has been designed to be easily adaptable to a variety of genomes. In novel genomes without an appropriate gene finder, I demonstrate that employing a foreign gene finder can produce highly inaccurate results, and that the most compatible parameters may not come from the nearest phylogenetic neighbor. I find that foreign gene finders are more usefully employed to bootstrap parameter estimation and that the resulting parameters can be highly accurate. Conclusion Since gene prediction is sensitive to species-specific parameters, every genome needs a dedicated gene finder.
Holt and Yandell (2011), Brůna et al. (2021) — MAKER2 and BRAKER2.
MAKER2: an annotation pipeline and genome-database management tool for second-generation genome projects. Holt & Yandell (2011). BMC Bioinformatics. doi:10.1186/1471-2105-12-491
Abstract
Background Second-generation sequencing technologies are precipitating major shifts with regards to what kinds of genomes are being sequenced and how they are annotated. While the first generation of genome projects focused on well-studied model organisms, many of today's projects involve exotic organisms whose genomes are largely terra incognita. This complicates their annotation, because unlike first-generation projects, there are no pre-existing 'gold-standard' gene-models with which to train gene-finders. Improvements in genome assembly and the wide availability of mRNA-seq data are also creating opportunities to update and re-annotate previously published genome annotations. Today's genome projects are thus in need of new genome annotation tools that can meet the challenges and opportunities presented by second-generation sequencing technologies. Results We present MAKER2, a genome annotation and data management tool designed for second-generation genome projects. MAKER2 is a multi-threaded, parallelized application that can process second-generation datasets of virtually any size. We show that MAKER2 can produce accurate annotations for novel genomes where training-data are limited, of low quality or even non-existent. MAKER2 also provides an easy means to use mRNA-seq data to improve annotation quality; and it can use these data to update legacy annotations, significantly improving their quality. We also show that MAKER2 can evaluate the quality of genome annotations, and identify and prioritize problematic annotations for manual review. Conclusions MAKER2 is the first annotation engine specifically designed for second-generation genome projects. MAKER2 scales to datasets of any size, requires little in the way of training data, and can use mRNA-seq data to improve annotation quality. It can also update and manage legacy genome annotation datasets.
BRAKER2: automatic eukaryotic genome annotation with GeneMark-EP+ and AUGUSTUS supported by a protein database. Brůna et al. (2021). NAR Genomics and Bioinformatics. doi:10.1093/nargab/lqaa108
Abstract
The task of eukaryotic genome annotation remains challenging. Only a few genomes could serve as standards of annotation achieved through a tremendous investment of human curation efforts. Still, the correctness of all alternative isoforms, even in the best-annotated genomes, could be a good subject for further investigation. The new BRAKER2 pipeline generates and integrates external protein support into the iterative process of training and gene prediction by GeneMark-EP+ and AUGUSTUS. BRAKER2 continues the line started by BRAKER1 where self-training GeneMark-ET and AUGUSTUS made gene predictions supported by transcriptomic data. Among the challenges addressed by the new pipeline was a generation of reliable hints to protein-coding exon boundaries from likely homologous but evolutionarily distant proteins. In comparison with other pipelines for eukaryotic genome annotation, BRAKER2 is fully automatic. It is favorably compared under equal conditions with other pipelines, e.g. MAKER2, in terms of accuracy and performance. Development of BRAKER2 should facilitate solving the task of harmonization of annotation of protein-coding genes in genomes of different eukaryotic species. However, we fully understand that several more innovations are needed in transcriptomic and proteomic technologies as well as in algorithmic development to reach the goal of highly accurate annotation of eukaryotic genomes.
Tatusova et al. (2016) — NCBI prokaryotic genome annotation pipeline.
NCBI prokaryotic genome annotation pipeline. Tatusova et al. (2016). Nucleic Acids Research. doi:10.1093/nar/gkw569
Abstract
Recent technological advances have opened unprecedented opportunities for large-scale sequencing and analysis of populations of pathogenic species in disease outbreaks, as well as for large-scale diversity studies aimed at expanding our knowledge across the whole domain of prokaryotes. To meet the challenge of timely interpretation of structure, function and meaning of this vast genetic information, a comprehensive approach to automatic genome annotation is critically needed. In collaboration with Georgia Tech, NCBI has developed a new approach to genome annotation that combines alignment based methods with methods of predicting protein-coding and RNA genes and other functional elements directly from sequence. A new gene finding tool, GeneMarkS+, uses the combined evidence of protein and RNA placement by homology as an initial map of annotation to generate and modify ab initio gene predictions across the whole genome. Thus, the new NCBI's Prokaryotic Genome Annotation Pipeline (PGAP) relies more on sequence similarity when confident comparative data are available, while it relies more on statistical predictions in the absence of external evidence. The pipeline provides a framework for generation and analysis of annotation on the full breadth of prokaryotic taxonomy. For additional information on PGAP see https://www.ncbi.nlm.nih.gov/genome/annotation_prok/ and the NCBI Handbook, https://www.ncbi.nlm.nih.gov/books/NBK174280/.
Stiehler et al. (2020) — Helixer: deep-learning gene prediction.
Helixer: cross-species gene annotation of large eukaryotic genomes using deep learning. Stiehler et al. (2020). Bioinformatics. doi:10.1093/bioinformatics/btaa1044
Abstract
Motivation Current state-of-the-art tools for the de novo annotation of genes in eukaryotic genomes have to be specifically fitted for each species and still often produce annotations that can be improved much further. The fundamental algorithmic architecture for these tools has remained largely unchanged for about two decades, limiting learning capabilities. Here, we set out to improve the cross-species annotation of genes from DNA sequence alone with the help of deep learning. The goal is to eliminate the dependency on a closely related gene model while also improving the predictive quality in general with a fundamentally new architecture. Results We present Helixer, a framework for the development and usage of a cross-species deep learning model that improves significantly on performance and generalizability when compared to more traditional methods. We evaluate our approach by building a single vertebrate model for the base-wise annotation of 186 animal genomes and a separate land plant model for 51 plant genomes. Our predictions are shown to be much less sensitive to the length of the genome than those of a current state-of-the-art tool. We also present two novel post-processing techniques that each worked to further strengthen our annotations and show in-depth results of an RNA-Seq based comparison of our predictions. Our method does not yet produce comprehensive gene models but rather outputs base pair wise probabilities. Availability and implementation The source code of this work is available at https://github.com/weberlab-hhu/Helixer under the GNU General Public License v3.0. The trained models are available at https://doi.org/10.5281/zenodo.3974409. Supplementary information Supplementary data are available at Bioinformatics online.
B.5.5 Metagenomics
Pace (1997), Handelsman et al. (1998) — the molecular view of microbial diversity, and the metagenome.
A Molecular View of Microbial Diversity and the Biosphere. Pace (1997). Science. doi:10.1126/science.276.5313.734
Abstract
Over three decades of molecular-phylogenetic studies, researchers have compiled an increasingly robust map of evolutionary diversification showing that the main diversity of life is microbial, distributed among three primary relatedness groups or domains: Archaea, Bacteria, and Eucarya. The general properties of representatives of the three domains indicate that the earliest life was based on inorganic nutrition and that photosynthesis and use of organic compounds for carbon and energy metabolism came comparatively later. The application of molecular-phylogenetic methods to study natural microbial ecosystems without the traditional requirement for cultivation has resulted in the discovery of many unexpected evolutionary lineages; members of some of these lineages are only distantly related to known organisms but are sufficiently abundant that they are likely to have impact on the chemistry of the biosphere.
Molecular biological access to the chemistry of unknown soil microbes: a new frontier for natural products. Handelsman et al. (1998). Chemistry & Biology. doi:10.1016/s1074-5521(98)90108-9
Abstract
Cultured soil microorganisms have provided a rich source of natural-product chemistry. Because only a tiny fraction of soil microbes from soil are readily cultured, soil might be the greatest untapped resource for novel chemistry. The concept of cloning the metagenome to access the collective genomes and the biosynthetic machinery of soil microflora is explored here.
Tyson et al. (2004), Venter et al. (2004) — the first shotgun metagenomes.
Community structure and metabolism through reconstruction of microbial genomes from the environment. Tyson et al. (2004). Nature. doi:10.1038/nature02340
Abstract
Microbial communities are vital in the functioning of all ecosystems; however, most microorganisms are uncultivated, and their roles in natural systems are unclear. Here, using random shotgun sequencing of DNA from a natural acidophilic biofilm, we report reconstruction of near-complete genomes of Leptospirillum group II and Ferroplasma type II, and partial recovery of three other genomes. This was possible because the biofilm was dominated by a small number of species populations and the frequency of genomic rearrangements and gene insertions or deletions was relatively low. Because each sequence read came from a different individual, we could determine that single-nucleotide polymorphisms are the predominant form of heterogeneity at the strain level. The Leptospirillum group II genome had remarkably few nucleotide polymorphisms, despite the existence of low-abundance variants. The Ferroplasma type II genome seems to be a composite from three ancestral strains that have undergone homologous recombination to form a large population of mosaic genomes. Analysis of the gene complement for each organism revealed the pathways for carbon and nitrogen fixation and energy generation, and provided insights into survival strategies in an extreme environment.
Environmental Genome Shotgun Sequencing of the Sargasso Sea. Venter et al. (2004). Science. doi:10.1126/science.1093857
Abstract
We have applied "whole-genome shotgun sequencing" to microbial populations collected en masse on tangential flow and impact filters from seawater samples collected from the Sargasso Sea near Bermuda. A total of 1.045 billion base pairs of nonredundant sequence was generated, annotated, and analyzed to elucidate the gene content, diversity, and relative abundance of the organisms within these environmental samples. These data are estimated to derive from at least 1800 genomic species based on sequence relatedness, including 148 previously unknown bacterial phylotypes. We have identified over 1.2 million previously unknown genes represented in these samples, including more than 782 new rhodopsin-like photoreceptors. Variation in species present and stoichiometry suggests substantial oceanic microbial diversity.
Qin et al. (2010), The Human Microbiome Project Consortium (2012), Sunagawa et al. (2015) — the gut, the human microbiome and the ocean.
A human gut microbial gene catalogue established by metagenomic sequencing. Qin et al. (2010). Nature. doi:10.1038/nature08821
Abstract
To understand the impact of gut microbes on human health and well-being it is crucial to assess their genetic potential. Here we describe the Illumina-based metagenomic sequencing, assembly and characterization of 3.3 million non-redundant microbial genes, derived from 576.7 gigabases of sequence, from faecal samples of 124 European individuals. The gene set, approximately 150 times larger than the human gene complement, contains an overwhelming majority of the prevalent (more frequent) microbial genes of the cohort and probably includes a large proportion of the prevalent human intestinal microbial genes. The genes are largely shared among individuals of the cohort. Over 99% of the genes are bacterial, indicating that the entire cohort harbours between 1,000 and 1,150 prevalent bacterial species and each individual at least 160 such species, which are also largely shared. We define and describe the minimal gut metagenome and the minimal gut bacterial genome in terms of functions present in all individuals and most bacteria, respectively.
Structure, function and diversity of the healthy human microbiome. The Human Microbiome Project Consortium (2012). Nature. doi:10.1038/nature11234
Abstract
Studies of the human microbiome have revealed that even healthy individuals differ remarkably in the microbes that occupy habitats such as the gut, skin and vagina. Much of this diversity remains unexplained, although diet, environment, host genetics and early microbial exposure have all been implicated. Accordingly, to characterize the ecology of human-associated microbial communities, the Human Microbiome Project has analysed the largest cohort and set of distinct, clinically relevant body habitats so far. We found the diversity and abundance of each habitat's signature microbes to vary widely even among healthy subjects, with strong niche specialization both within and among individuals. The project encountered an estimated 81-99% of the genera, enzyme families and community configurations occupied by the healthy Western microbiome. Metagenomic carriage of metabolic pathways was stable among individuals despite variation in community structure, and ethnic/racial background proved to be one of the strongest associations of both pathways and microbes with clinical metadata. These results thus delineate the range of structural and functional configurations normal in the microbial communities of a healthy population, enabling future characterization of the epidemiology, ecology and translational applications of the human microbiome.
Structure and function of the global ocean microbiome. Sunagawa et al. (2015). Science. doi:10.1126/science.1261359
Abstract
Microbes are dominant drivers of biogeochemical processes, yet drawing a global picture of functional diversity, microbial community structure, and their ecological determinants remains a grand challenge. We analyzed 7.2 terabases of metagenomic data from 243 Tara Oceans samples from 68 locations in epipelagic and mesopelagic waters across the globe to generate an ocean microbial reference gene catalog with >40 million nonredundant, mostly novel sequences from viruses, prokaryotes, and picoeukaryotes. Using 139 prokaryote-enriched samples, containing >35,000 species, we show vertical stratification with epipelagic community composition mostly driven by temperature rather than other environmental factors or geography. We identify ocean microbial core functionality and reveal that >73% of its abundance is shared with the human gut microbiome despite the physicochemical differences between these two ecosystems.
Callahan et al. (2016), Bolyen et al. (2019) — ASVs and QIIME 2.
DADA2: High-resolution sample inference from Illumina amplicon data. Callahan et al. (2016). Nature Methods. doi:10.1038/nmeth.3869
Abstract
We present the open-source software package DADA2 for modeling and correcting Illumina-sequenced amplicon errors (https://github.com/benjjneb/dada2). DADA2 infers sample sequences exactly and resolves differences of as little as 1 nucleotide. In several mock communities, DADA2 identified more real variants and output fewer spurious sequences than other methods. We applied DADA2 to vaginal samples from a cohort of pregnant women, revealing a diversity of previously undetected Lactobacillus crispatus variants.
Reproducible, interactive, scalable and extensible microbiome data science using QIIME 2. Bolyen et al. (2019). Nature Biotechnology. doi:10.1038/s41587-019-0209-9
Wood and Salzberg (2014), Wood et al. (2019), Blanco-Míguez et al. (2023) — Kraken and MetaPhlAn.
Kraken: ultrafast metagenomic sequence classification using exact alignments. Wood & Salzberg (2014). Genome Biology. doi:10.1186/gb-2014-15-3-r46
Abstract
Kraken is an ultrafast and highly accurate program for assigning taxonomic labels to metagenomic DNA sequences. Previous programs designed for this task have been relatively slow and computationally expensive, forcing researchers to use faster abundance estimation programs, which only classify small subsets of metagenomic data. Using exact alignment of k-mers, Kraken achieves classification accuracy comparable to the fastest BLAST program. In its fastest mode, Kraken classifies 100 base pair reads at a rate of over 4.1 million reads per minute, 909 times faster than Megablast and 11 times faster than the abundance estimation program MetaPhlAn. Kraken is available at http://ccb.jhu.edu/software/kraken/.
Improved metagenomic analysis with Kraken 2. Wood et al. (2019). Genome Biology. doi:10.1186/s13059-019-1891-0
Abstract
Although Kraken's k-mer-based approach provides a fast taxonomic classification of metagenomic sequence data, its large memory requirements can be limiting for some applications. Kraken 2 improves upon Kraken 1 by reducing memory usage by 85%, allowing greater amounts of reference genomic data to be used, while maintaining high accuracy and increasing speed fivefold. Kraken 2 also introduces a translated search mode, providing increased sensitivity in viral metagenomics analysis.
Extending and improving metagenomic taxonomic profiling with uncharacterized species using MetaPhlAn 4. Blanco-Míguez et al. (2023). Nature Biotechnology. doi:10.1038/s41587-023-01688-w
Abstract
Metagenomic assembly enables new organism discovery from microbial communities, but it can only capture few abundant organisms from most metagenomes. Here we present MetaPhlAn 4, which integrates information from metagenome assemblies and microbial isolate genomes for more comprehensive metagenomic taxonomic profiling. From a curated collection of 1.01 M prokaryotic reference and metagenome-assembled genomes, we define unique marker genes for 26,970 species-level genome bins, 4,992 of them taxonomically unidentified at the species level. MetaPhlAn 4 explains ~20% more reads in most international human gut microbiomes and >40% in less-characterized environments such as the rumen microbiome and proves more accurate than available alternatives on synthetic evaluations while also reliably quantifying organisms with no cultured isolates. Application of the method to >24,500 metagenomes highlights previously undetected species to be strong biomarkers for host conditions and lifestyles in human and mouse microbiomes and shows that even previously uncharacterized species can be genetically profiled at the resolution of single microbial strains.
Nurk et al. (2017), Li et al. (2015) — metagenome assembly.
metaSPAdes: a new versatile metagenomic assembler. Nurk et al. (2017). Genome Research. doi:10.1101/gr.213959.116
Abstract
While metagenomics has emerged as a technology of choice for analyzing bacterial populations, the assembly of metagenomic data remains challenging, thus stifling biological discoveries. Moreover, recent studies revealed that complex bacterial populations may be composed from dozens of related strains, thus further amplifying the challenge of metagenomic assembly. metaSPAdes addresses various challenges of metagenomic assembly by capitalizing on computational ideas that proved to be useful in assemblies of single cells and highly polymorphic diploid genomes. We benchmark metaSPAdes against other state-of-the-art metagenome assemblers and demonstrate that it results in high-quality assemblies across diverse data sets.
MEGAHIT: an ultra-fast single-node solution for large and complex metagenomics assembly via succinct de Bruijn graph. Li et al. (2015). Bioinformatics. doi:10.1093/bioinformatics/btv033
Abstract
MEGAHIT is a NGS de novo assembler for assembling large and complex metagenomics data in a time- and cost-efficient manner. It finished assembling a soil metagenomics dataset with 252 Gbps in 44.1 and 99.6 h on a single computing node with and without a graphics processing unit, respectively. MEGAHIT assembles the data as a whole, i.e. no pre-processing like partitioning and normalization was needed. When compared with previous methods on assembling the soil data, MEGAHIT generated a three-time larger assembly, with longer contig N50 and average contig length; furthermore, 55.8% of the reads were aligned to the assembly, giving a fourfold improvement.
Kang et al. (2019), Parks et al. (2015), Bowers et al. (2017) — binning, MAG quality and standards.
MetaBAT 2: an adaptive binning algorithm for robust and efficient genome reconstruction from metagenome assemblies. Kang et al. (2019). PeerJ. doi:10.7717/peerj.7359
Abstract
We previously reported on MetaBAT, an automated metagenome binning software tool to reconstruct single genomes from microbial communities for subsequent analyses of uncultivated microbial species. MetaBAT has become one of the most popular binning tools largely due to its computational efficiency and ease of use, especially in binning experiments with a large number of samples and a large assembly. MetaBAT requires users to choose parameters to fine-tune its sensitivity and specificity. If those parameters are not chosen properly, binning accuracy can suffer, especially on assemblies of poor quality. Here, we developed MetaBAT 2 to overcome this problem. MetaBAT 2 uses a new adaptive binning algorithm to eliminate manual parameter tuning. We also performed extensive software engineering optimization to increase both computational and memory efficiency. Comparing MetaBAT 2 to alternative software tools on over 100 real world metagenome assemblies shows superior accuracy and computing speed. Binning a typical metagenome assembly takes only a few minutes on a single commodity workstation. We therefore recommend the community adopts MetaBAT 2 for their metagenome binning experiments. MetaBAT 2 is open source software and available at https://bitbucket.org/berkeleylab/metabat.
CheckM: assessing the quality of microbial genomes recovered from isolates, single cells, and metagenomes. Parks et al. (2015). Genome Research. doi:10.1101/gr.186072.114
Abstract
Large-scale recovery of genomes from isolates, single cells, and metagenomic data has been made possible by advances in computational methods and substantial reductions in sequencing costs. Although this increasing breadth of draft genomes is providing key information regarding the evolutionary and functional diversity of microbial life, it has become impractical to finish all available reference genomes. Making robust biological inferences from draft genomes requires accurate estimates of their completeness and contamination. Current methods for assessing genome quality are ad hoc and generally make use of a limited number of "marker" genes conserved across all bacterial or archaeal genomes. Here we introduce CheckM, an automated method for assessing the quality of a genome using a broader set of marker genes specific to the position of a genome within a reference genome tree and information about the collocation of these genes. We demonstrate the effectiveness of CheckM using synthetic data and a wide range of isolate-, single-cell-, and metagenome-derived genomes. CheckM is shown to provide accurate estimates of genome completeness and contamination and to outperform existing approaches. Using CheckM, we identify a diverse range of errors currently impacting publicly available isolate genomes and demonstrate that genomes obtained from single cells and metagenomic data vary substantially in quality. In order to facilitate the use of draft genomes, we propose an objective measure of genome quality that can be used to select genomes suitable for specific gene- and genome-centric analyses of microbial communities.
Minimum information about a single amplified genome (MISAG) and a metagenome-assembled genome (MIMAG) of bacteria and archaea. Bowers et al. (2017). Nature Biotechnology. doi:10.1038/nbt.3893
Abstract
We present two standards developed by the Genomic Standards Consortium (GSC) for reporting bacterial and archaeal genome sequences. Both are extensions of the Minimum Information about Any (x) Sequence (MIxS). The standards are the Minimum Information about a Single Amplified Genome (MISAG) and the Minimum Information about a Metagenome-Assembled Genome (MIMAG), including, but not limited to, assembly quality, and estimates of genome completeness and contamination. These standards can be used in combination with other GSC checklists, including the Minimum Information about a Genome Sequence (MIGS), Minimum Information about a Metagenomic Sequence (MIMS), and Minimum Information about a Marker Gene Sequence (MIMARKS). Community-wide adoption of MISAG and MIMAG will facilitate more robust comparative genomic analyses of bacterial and archaeal diversity.
Almeida et al. (2019), Pasolli et al. (2019), Nayfach et al. (2020) — large catalogues of MAGs.
A new genomic blueprint of the human gut microbiota. Almeida et al. (2019). Nature. doi:10.1038/s41586-019-0965-1
Abstract
The composition of the human gut microbiota is linked to health and disease, but knowledge of individual microbial species is needed to decipher their biological roles. Despite extensive culturing and sequencing efforts, the complete bacterial repertoire of the human gut microbiota remains undefined. Here we identify 1,952 uncultured candidate bacterial species by reconstructing 92,143 metagenome-assembled genomes from 11,850 human gut microbiomes. These uncultured genomes substantially expand the known species repertoire of the collective human gut microbiota, with a 281% increase in phylogenetic diversity. Although the newly identified species are less prevalent in well-studied populations compared to reference isolate genomes, they improve classification of understudied African and South American samples by more than 200%. These candidate species encode hundreds of newly identified biosynthetic gene clusters and possess a distinctive functional capacity that might explain their elusive nature. Our work expands the known diversity of uncultured gut bacteria, which provides unprecedented resolution for taxonomic and functional characterization of the intestinal microbiota.
Extensive Unexplored Human Microbiome Diversity Revealed by Over 150,000 Genomes from Metagenomes Spanning Age, Geography, and Lifestyle. Pasolli et al. (2019). Cell. doi:10.1016/j.cell.2019.01.001
Abstract
The body-wide human microbiome plays a role in health, but its full diversity remains uncharacterized, particularly outside of the gut and in international populations. We leveraged 9,428 metagenomes to reconstruct 154,723 microbial genomes (45% of high quality) spanning body sites, ages, countries, and lifestyles. We recapitulated 4,930 species-level genome bins (SGBs), 77% without genomes in public repositories (unknown SGBs [uSGBs]). uSGBs are prevalent (in 93% of well-assembled samples), expand underrepresented phyla, and are enriched in non-Westernized populations (40% of the total SGBs). We annotated 2.85 M genes in SGBs, many associated with conditions including infant development (94,000) or Westernization (106,000). SGBs and uSGBs permit deeper microbiome analyses and increase the average mappability of metagenomic reads from 67.76% to 87.51% in the gut (median 94.26%) and 65.14% to 82.34% in the mouth. We thus identify thousands of microbial genomes from yet-to-be-named species, expand the pangenomes of human-associated microbes, and allow better exploitation of metagenomic technologies.
A genomic catalog of Earth’s microbiomes. Nayfach et al. (2020). Nature Biotechnology. doi:10.1038/s41587-020-0718-6
Abstract
The reconstruction of bacterial and archaeal genomes from shotgun metagenomes has enabled insights into the ecology and evolution of environmental and host-associated microbiomes. Here we applied this approach to >10,000 metagenomes collected from diverse habitats covering all of Earth's continents and oceans, including metagenomes from human and animal hosts, engineered environments, and natural and agricultural soils, to capture extant microbial, metabolic and functional potential. This comprehensive catalog includes 52,515 metagenome-assembled genomes representing 12,556 novel candidate species-level operational taxonomic units spanning 135 phyla. The catalog expands the known phylogenetic diversity of bacteria and archaea by 44% and is broadly available for streamlined comparative analyses, interactive exploration, metabolic modeling and bulk download. We demonstrate the utility of this collection for understanding secondary-metabolite biosynthetic potential and for resolving thousands of new host linkages to uncultivated viruses. This resource underscores the value of genome-centric approaches for revealing genomic properties of uncultivated microorganisms that affect ecosystem processes.
Sczyrba et al. (2017), Meyer et al. (2022) — CAMI benchmarks.
Critical Assessment of Metagenome Interpretation—a benchmark of metagenomics software. Sczyrba et al. (2017). Nature Methods. doi:10.1038/nmeth.4458
Abstract
Methods for assembly, taxonomic profiling and binning are key to interpreting metagenome data, but a lack of consensus about benchmarking complicates performance assessment. The Critical Assessment of Metagenome Interpretation (CAMI) challenge has engaged the global developer community to benchmark their programs on highly complex and realistic data sets, generated from ∼700 newly sequenced microorganisms and ∼600 novel viruses and plasmids and representing common experimental setups. Assembly and genome binning programs performed well for species represented by individual genomes but were substantially affected by the presence of related strains. Taxonomic profiling and binning programs were proficient at high taxonomic ranks, with a notable performance decrease below family level. Parameter settings markedly affected performance, underscoring their importance for program reproducibility. The CAMI results highlight current challenges but also provide a roadmap for software selection to answer specific research questions.
Critical Assessment of Metagenome Interpretation: the second round of challenges. Meyer et al. (2022). Nature Methods. doi:10.1038/s41592-022-01431-4
Abstract
Evaluating metagenomic software is key for optimizing metagenome interpretation and focus of the Initiative for the Critical Assessment of Metagenome Interpretation (CAMI). The CAMI II challenge engaged the community to assess methods on realistic and complex datasets with long- and short-read sequences, created computationally from around 1,700 new and known genomes, as well as 600 new plasmids and viruses. Here we analyze 5,002 results by 76 program versions. Substantial improvements were seen in assembly, some due to long-read data. Related strains still were challenging for assembly and genome recovery through binning, as was assembly quality for the latter. Profilers markedly matured, with taxon profilers and binners excelling at higher bacterial ranks, but underperforming for viruses and Archaea. Clinical pathogen detection results revealed a need to improve reproducibility. Runtime and memory usage analyses identified efficient programs, including top performers with other metrics. The results identify challenges and guide researchers in selecting methods for analyses.
B.6 Genomes
Further papers for Chapter 6.
In the HTML version, click Abstract to show the abstract.
B.6.1 Genome projects
Fleischmann et al. (1995), Fraser et al. (1995) — the first bacterial genomes.
Whole-Genome Random Sequencing and Assembly of Haemophilus influenzae Rd. Fleischmann et al. (1995). Science. doi:10.1126/science.7542800
Abstract
An approach for genome analysis based on sequencing and assembly of unselected pieces of DNA from the whole chromosome has been applied to obtain the complete nucleotide sequence (1,830,137 base pairs) of the genome from the bacterium Haemophilus influenzae Rd. This approach eliminates the need for initial mapping efforts and is therefore applicable to the vast array of microbial species for which genome maps are unavailable. The H. influenzae Rd genome sequence (Genome Sequence DataBase accession number L42023) represents the only complete genome sequence from a free-living organism.
The Minimal Gene Complement of Mycoplasma genitalium. Fraser et al. (1995). Science. doi:10.1126/science.270.5235.397
Abstract
The complete nucleotide sequence (580,070 base pairs) of the Mycoplasma genitalium genome, the smallest known genome of any free-living organism, has been determined by whole-genome random sequencing and assembly. A total of only 470 predicted coding regions were identified that include genes required for DNA replication, transcription and translation, DNA repair, cellular transport, and energy metabolism. Comparison of this genome to that of Haemophilus influenzae suggests that differences in genome content are reflected as profound differences in physiology and metabolic capacity between these two organisms.
Goffeau et al. (1996) — yeast: life with 6,000 genes.
Life with 6000 Genes. Goffeau et al. (1996). Science. doi:10.1126/science.274.5287.546
Abstract
The genome of the yeast Saccharomyces cerevisiae has been completely sequenced through a worldwide collaboration. The sequence of 12,068 kilobases defines 5885 potential protein-encoding genes, approximately 140 genes specifying ribosomal RNA, 40 genes for small nuclear RNA molecules, and 275 transfer RNA genes. In addition, the complete sequence provides information about the higher order organization of yeast's 16 chromosomes and allows some insight into their evolutionary history. The genome shows a considerable amount of apparent genetic redundancy, and one of the major problems to be tackled during the next stage of the yeast genome project is to elucidate the biological functions of all of these genes.
Blattner et al. (1997) — Escherichia coli.
The Complete Genome Sequence of Escherichia coli K-12. Blattner et al. (1997). Science. doi:10.1126/science.277.5331.1453
Abstract
The 4,639,221-base pair sequence of Escherichia coli K-12 is presented. Of 4288 protein-coding genes annotated, 38 percent have no attributed function. Comparison with five other sequenced microbes reveals ubiquitous as well as narrowly distributed gene families; many families of similar genes within E. coli are also evident. The largest family of paralogous proteins contains 80 ABC transporters. The genome as a whole is strikingly organized with respect to the local direction of replication; guanines, oligonucleotides possibly related to replication and recombination, and most genes are so oriented. The genome also contains insertion sequence (IS) elements, phage remnants, and many other patches of unusual composition indicating genome plasticity through horizontal transfer.
The C. elegans Sequencing Consortium (1998), Adams et al. (2000), The Arabidopsis Genome Initiative (2000) — the first animal and plant genomes.
Genome Sequence of the Nematode C. elegans: A Platform for Investigating Biology. The C. elegans Sequencing Consortium (1998). Science. doi:10.1126/science.282.5396.2012
Abstract
The 97-megabase genomic sequence of the nematode Caenorhabditis elegans reveals over 19,000 genes. More than 40 percent of the predicted protein products find significant matches in other organisms. There is a variety of repeated sequences, both local and dispersed. The distinctive distribution of some repeats and highly conserved genes provides evidence for a regional organization of the chromosomes.
The Genome Sequence of Drosophila melanogaster. Adams et al. (2000). Science. doi:10.1126/science.287.5461.2185
Abstract
The fly Drosophila melanogaster is one of the most intensively studied organisms in biology and serves as a model system for the investigation of many developmental and cellular processes common to higher eukaryotes, including humans. We have determined the nucleotide sequence of nearly all of the approximately 120-megabase euchromatic portion of the Drosophila genome using a whole-genome shotgun sequencing strategy supported by extensive clone-based sequence and a high-quality bacterial artificial chromosome physical map. Efforts are under way to close the remaining gaps; however, the sequence is of sufficient accuracy and contiguity to be declared substantially complete and to support an initial analysis of genome structure and preliminary gene annotation and interpretation. The genome encodes approximately 13,600 genes, somewhat fewer than the smaller Caenorhabditis elegans genome, but with comparable functional diversity.
Analysis of the genome sequence of the flowering plant Arabidopsis thaliana. The Arabidopsis Genome Initiative (2000). Nature. doi:10.1038/35048692
Abstract
The flowering plant Arabidopsis thaliana is an important model system for identifying genes and determining their functions. Here we report the analysis of the genomic sequence of Arabidopsis. The sequenced regions cover 115.4 megabases of the 125-megabase genome and extend into centromeric regions. The evolution of Arabidopsis involved a whole-genome duplication, followed by subsequent gene loss and extensive local gene duplications, giving rise to a dynamic genome enriched by lateral gene transfer from a cyanobacterial-like ancestor of the plastid. The genome contains 25,498 genes encoding proteins from 11,000 families, similar to the functional diversity of Drosophila and Caenorhabditis elegans--the other sequenced multicellular eukaryotes. Arabidopsis has many families of new proteins but also lacks several common protein families, indicating that the sets of common proteins have undergone differential expansion and contraction in the three multicellular eukaryotes. This is the first complete genome sequence of a plant and provides the foundations for more comprehensive comparison of conserved processes in all eukaryotes, identifying a wide range of plant-specific gene functions and establishing rapid systematic ways to identify genes for crop improvement.
Venter et al. (2001) — the Celera human genome.
The Sequence of the Human Genome. Venter et al. (2001). Science. doi:10.1126/science.1058040
Abstract
A 2.91-billion base pair (bp) consensus sequence of the euchromatic portion of the human genome was generated by the whole-genome shotgun sequencing method. The 14.8-billion bp DNA sequence was generated over 9 months from 27,271,853 high-quality sequence reads (5.11-fold coverage of the genome) from both ends of plasmid clones made from the DNA of five individuals. Two assembly strategies-a whole-genome assembly and a regional chromosome assembly-were used, each combining sequence data from Celera and the publicly funded genome effort. The public data were shredded into 550-bp segments to create a 2.9-fold coverage of those genome regions that had been sequenced, without including biases inherent in the cloning and assembly procedure used by the publicly funded group. This brought the effective coverage in the assemblies to eightfold, reducing the number and size of gaps in the final assembly over what would be obtained with 5.11-fold coverage. The two assembly strategies yielded very similar results that largely agree with independent mapping data. The assemblies effectively cover the euchromatic regions of the human chromosomes. More than 90% of the genome is in scaffold assemblies of 100,000 bp or more, and 25% of the genome is in scaffolds of 10 million bp or larger. Analysis of the genome sequence revealed 26,588 protein-encoding transcripts for which there was strong corroborating evidence and an additional approximately 12,000 computationally derived genes with mouse matches or other weak supporting evidence. Although gene-dense clusters are obvious, almost half the genes are dispersed in low G+C sequence separated by large tracts of apparently noncoding sequence. Only 1.1% of the genome is spanned by exons, whereas 24% is in introns, with 75% of the genome being intergenic DNA. Duplications of segmental blocks, ranging in size up to chromosomal lengths, are abundant throughout the genome and reveal a complex evolutionary history. Comparative genomic analysis indicates vertebrate expansions of genes associated with neuronal function, with tissue-specific developmental regulation, and with the hemostasis and immune systems. DNA sequence comparisons between the consensus sequence and publicly funded genome data provided locations of 2.1 million single-nucleotide polymorphisms (SNPs). A random pair of human haploid genomes differed at a rate of 1 bp per 1250 on average, but there was marked heterogeneity in the level of polymorphism across the genome. Less than 1% of all SNPs resulted in variation in proteins, but the task of determining which SNPs have functional consequences remains an open challenge.
International Human Genome Sequencing Consortium (2004) — finishing the human genome.
Finishing the euchromatic sequence of the human genome. International Human Genome Sequencing Consortium (2004). Nature. doi:10.1038/nature03001
Abstract
The sequence of the human genome encodes the genetic instructions for human physiology, as well as rich information about human evolution. In 2001, the International Human Genome Sequencing Consortium reported a draft sequence of the euchromatic portion of the human genome. Since then, the international collaboration has worked to convert this draft into a genome sequence with high accuracy and nearly complete coverage. Here, we report the result of this finishing process. The current genome sequence (Build 35) contains 2.85 billion nucleotides interrupted by only 341 gaps. It covers approximately 99% of the euchromatic genome and is accurate to an error rate of approximately 1 event per 100,000 bases. Many of the remaining euchromatic gaps are associated with segmental duplications and will require focused work with new methods. The near-complete sequence, the first for a vertebrate, greatly improves the precision of biological analyses of the human genome including studies of gene number, birth and death. Notably, the human genome seems to encode only 20,000-25,000 protein-coding genes. The genome sequence reported here should serve as a firm foundation for biomedical research in the decades ahead.
Mouse Genome Sequencing Consortium (2002) — the mouse genome and the first comparison of two mammals.
Initial sequencing and comparative analysis of the mouse genome. Mouse Genome Sequencing Consortium (2002). Nature. doi:10.1038/nature01262
Abstract
The sequence of the mouse genome is a key informational tool for understanding the contents of the human genome and a key experimental tool for biomedical research. Here, we report the results of an international collaboration to produce a high-quality draft sequence of the mouse genome. We also present an initial comparative analysis of the mouse and human genomes, describing some of the insights that can be gleaned from the two sequences. We discuss topics including the analysis of the evolutionary forces shaping the size, structure and sequence of the genomes; the conservation of large-scale synteny across most of the genomes; the much lower extent of sequence orthology covering less than half of the genomes; the proportions of the genomes under selection; the number of protein-coding genes; the expansion of gene families related to reproduction and immunity; the evolution of proteins; and the identification of intraspecies polymorphism.
Nurk et al. (2022) — the complete human genome.
The complete sequence of a human genome. Nurk et al. (2022). Science. doi:10.1126/science.abj6987
Abstract
Since its initial release in 2000, the human reference genome has covered only the euchromatic fraction of the genome, leaving important heterochromatic regions unfinished. Addressing the remaining 8% of the genome, the Telomere-to-Telomere (T2T) Consortium presents a complete 3.055 billion-base pair sequence of a human genome, T2T-CHM13, that includes gapless assemblies for all chromosomes except Y, corrects errors in the prior references, and introduces nearly 200 million base pairs of sequence containing 1956 gene predictions, 99 of which are predicted to be protein coding. The completed regions include all centromeric satellite arrays, recent segmental duplications, and the short arms of all five acrocentric chromosomes, unlocking these complex regions of the genome to variational and functional studies.
B.6.2 Genome size and organisation
Thomas (1971), Gregory (2001) — the C-value paradox and the C-value enigma.
THE GENETIC ORGANIZATION OF CHROMOSOMES. Thomas (1971). Annual Review of Genetics. doi:10.1146/annurev.ge.05.120171.001321
Coincidence, coevolution, or causation? DNA content, cell size, and the C-value enigma. Gregory (2001). Biological Reviews of the Cambridge Philosophical Society. doi:10.1017/s1464793100005595
Abstract
Variation in DNA content has been largely ignored as a factor in evolution, particularly following the advent of sequence-based approaches to genomic analysis. The significant genome size diversity among organisms (more than 200000-fold among eukaryotes) bears no relationship to organismal complexity and both the origins and reasons for the clearly non-random distribution of this variation remain unclear. Several theories have been proposed to explain this 'C-value enigma' (heretofore known as the 'C-value paradox'), each of which can be described as either a mutation pressure' or 'optimal DNA' theory. Mutation pressure theories consider the large portion of non-coding DNA in eukaryotic genomes as either 'junk' or 'selfish' DNA and are important primarily in considerations of the origin of secondary DNA. Optimal DNA theories differ from mutation pressure theories by emphasizing the strong link between DNA content and cell and nuclear volumes. While mutation pressure theories generally explain this association with cell size as coincidental, the nucleoskeletal theory proposes a coevolutionary interaction between nuclear and cell volume, with DNA content adjusted adaptively following shifts in cell size. Each of these approaches to the C-value enigma is problematic for a variety of reasons and the preponderance of the available evidence instead favours the nucleotypic theory which postulates a causal link between bulk DNA amount and cell volume. Under this view, variation in DNA content is under direct selection via its impacts on cellular and organismal parameters. Until now, no satisfactory mechanism has been presented to explain this nucleotypic effect. However, recent advances in the study of cell cycle regulation suggest a possible 'gene nucleus interaction model' which may account for it. The present article provides a detailed review of the debate surrounding the C-value enigma, the various theories proposed to explain it, and the evidence in favour of a causal connection between DNA content and cell size. In addition, a new model of nucleotypic influence is developed, along with suggestions for further empirical investigation. Finally, some evolutionary implications of genome size diversity are considered, and a broadening of the traditional 'biological hierarchy' is recommended.
Kapusta et al. (2017) — the dynamics of genome size: gains and losses.
Dynamics of genome size evolution in birds and mammals. Kapusta et al. (2017). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1616702114
Abstract
Genome size in mammals and birds shows remarkably little interspecific variation compared with other taxa. However, genome sequencing has revealed that many mammal and bird lineages have experienced differential rates of transposable element (TE) accumulation, which would be predicted to cause substantial variation in genome size between species. Thus, we hypothesize that there has been covariation between the amount of DNA gained by transposition and lost by deletion during mammal and avian evolution, resulting in genome size equilibrium. To test this model, we develop computational methods to quantify the amount of DNA gained by TE expansion and lost by deletion over the last 100 My in the lineages of 10 species of eutherian mammals and 24 species of birds. The results reveal extensive variation in the amount of DNA gained via lineage-specific transposition, but that DNA loss counteracted this expansion to various extents across lineages. Our analysis of the rate and size spectrum of deletion events implies that DNA removal in both mammals and birds has proceeded mostly through large segmental deletions (>10 kb). These findings support a unified "accordion" model of genome size evolution in eukaryotes whereby DNA loss counteracting TE expansion is a major determinant of genome size. Furthermore, we propose that extensive DNA loss, and not necessarily a dearth of TE activity, has been the primary force maintaining the greater genomic compaction of flying birds and bats relative to their flightless relatives.
Fernández et al. (2024) — the largest known genome.
A 160 Gbp fork fern genome shatters size record for eukaryotes. Fernández et al. (2024). iScience. doi:10.1016/j.isci.2024.109889
Abstract
Vascular plants are exceptional among eukaryotes due to their outstanding genome size diversity which ranges ∼2,400-fold, including the largest genome so far recorded in the angiosperm Paris japonica (148.89 Gbp/1C). Despite available data showing that giant genomes are restricted across the Tree of Life, the biological limits to genome size expansion remain to be established. Here, we report the discovery of an even larger eukaryotic genome in Tmesipteris oblanceolata , a New Caledonian fork fern. At 160.45 Gbp/1C, this record-breaking genome challenges current understanding and opens new avenues to explore the evolutionary dynamics of genomic gigantism.
Nakabachi et al. (2006), Giovannoni et al. (2005) — the smallest genomes: endosymbionts and streamlined bacteria.
The 160-Kilobase Genome of the Bacterial Endosymbiont Carsonella. Nakabachi et al. (2006). Science. doi:10.1126/science.1134196
Abstract
Previous studies have suggested that the minimal cellular genome could be as small as 400 kilobases. Here, we report the complete genome sequence of the psyllid symbiont Carsonella ruddii, which consists of a circular chromosome of 159,662 base pairs, averaging 16.5% GC content. It is by far the smallest and most AT-rich bacterial genome yet characterized. The genome has a high coding density (97%) with many overlapping genes and reduced gene length. Genes for translation and amino acid biosynthesis are relatively well represented, but numerous genes considered essential for life are missing, suggesting that Carsonella may have achieved organelle-like status.
Genome Streamlining in a Cosmopolitan Oceanic Bacterium. Giovannoni et al. (2005). Science. doi:10.1126/science.1114057
Abstract
The SAR11 clade consists of very small, heterotrophic marine alpha-proteobacteria that are found throughout the oceans, where they account for about 25% of all microbial cells. Pelagibacter ubique, the first cultured member of this clade, has the smallest genome and encodes the smallest number of predicted open reading frames known for a free-living microorganism. In contrast to parasitic bacteria and archaea with small genomes, P. ubique has complete biosynthetic pathways for all 20 amino acids and all but a few cofactors. P. ubique has no pseudogenes, introns, transposons, extrachromosomal elements, or inteins; few paralogs; and the shortest intergenic spacers yet observed for any cell.
Rocha (2008) — the organisation of the bacterial genome.
The Organization of the Bacterial Genome. Rocha (2008). Annual Review of Genetics. doi:10.1146/annurev.genet.42.110807.091653
Abstract
Many bacterial cellular processes interact intimately with the chromosome. Such interplay is the major driving force of genome structure or organization. Interactions take place at different scales-local for gene expression, global for replication-and lead to the differentiation of the chromosome into organizational units such as operons, replichores, or macrodomains. These processes are intermingled in the cell and create complex higher-level organizational features that are adaptive because they favor the interplay between the processes. The surprising result of selection for genome organization is that gene repertoires change much more quickly than chromosomal structure. Comparative genomics and experimental genomic manipulations are untangling the different cellular and evolutionary mechanisms causing such resilience to change. Since organization results from cellular processes, a better understanding of chromosome organization will help unravel the underlying cellular processes and their diversity.
Lobry (1996) — strand asymmetry and GC skew.
Asymmetric substitution patterns in the two DNA strands of bacteria. Lobry (1996). Molecular Biology and Evolution. doi:10.1093/oxfordjournals.molbev.a025626
Abstract
Analyses of the genomes of three prokaryotes, Escherichia coli, Bacillus subtilis, and Haemophilus influenzae, revealed a new type of genomic compartmentalization of base frequencies. There was a departure from intrastrand equifrequency between A and T or between C and G, showing that the substitution patterns of the two strands of DNA were asymmetric. The positions of the boundaries between these compartments were found to coincide with the origin and terminus of chromosome replication, and there were more A-T and C-G deviations in intergenic regions and third codon positions, suggesting that a mutational bias was responsible for this asymmetry. The strand asymmetry was found to be due to a difference in base compositions of transcripts in the leading and lagging strands. This difference is sufficient to affect codon usage, but it is small compared to the effects of gene expressivity and amino-acid composition.
Jacob and Monod (1961) — the operon.
Genetic regulatory mechanisms in the synthesis of proteins. Jacob & Monod (1961). Journal of Molecular Biology. doi:10.1016/s0022-2836(61)80072-7
Berget et al. (1977), Chow et al. (1977) — the discovery of split genes.
Spliced segments at the 5′ terminus of adenovirus 2 late mRNA. Berget et al. (1977). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.74.8.3171
Abstract
An mRNA fraction coding for hexon polypeptide, the major virion structural protein, was purified by gel electrophoresis from extracts of adenovirus 2-infected cells late in the lytic cycle. The mRNA sequences in this fraction were mapped between 51.7 and 61.3 units on the genome by visualizing RNA-DNA hybrids in the electron microscope. When hybrids of hexon mRNA and single-stranded restriction endonuclease cleavage fragments of viral DNA were visualized in the electron microscope,branched forms were observed in which 160 nucleotides of RNA from the 5' terminus were not hydrogen bonded to the single-stranded DNA. DNA sequences complementary to the RNA sequences in each 5' tail were found by electron microscopy to be located at 17, 20, and 27 units on the same strand as that coding for the body of the hexon mRNA. Thus, four segments of viral RNA may be joined together during the synthesis of mature hexon mRNA. A model is presented for adenovirus late mRNA synthesis that involves multiple splicing during maturation of a larger precursor nuclear RNA.
An amazing sequence arrangement at the 5′ ends of adenovirus 2 messenger RNA. Chow et al. (1977). Cell. doi:10.1016/0092-8674(77)90180-5
Abstract
The 5' terminal sequences of several adenovirus 2 (Ad2) mRNAs, isolated late in infection, are complementary to sequences within the Ad2 genome which are remote from the DNA from which the main coding sequence of each mRNA is transcribed. This has been observed by forming RNA displacement loops (R loops) between Ad2 DNA and unfractionated polysomal RNA from infected cells. The 5' terminal sequences of mRNAs in R loops, variously located between positions 36 and 92, form complex secondary hybrids with single-stranded DNA from restriction endonuclease fragments containing sequences to the left of position 36 on the Ad2 genome. The structures visualized in the electron microscope show that short sequences coded at map positions 16.6, 19.6 and 26.6 on the R strand are joined to form a leader sequence of 150-200 nucleotides at the 5' end of many late mRNAs. A late mRNA which maps to the left of position 16.6 shows a different pattern of second site hybridization. It contains sequences from 4.9-6.0 linked directly to those from 9.6-10.9. These findings imply a new mechanism for the biosynthesis of Ad2 mRNA in mammalian cells.
Greider and Blackburn (1985) — telomerase.
Identification of a specific telomere terminal transferase activity in tetrahymena extracts. Greider & Blackburn (1985). Cell. doi:10.1016/0092-8674(85)90170-9
Abstract
We have found a novel activity in Tetrahymena cell free extracts that adds tandem TTGGGG repeats onto synthetic telomere primers. The single-stranded DNA oligonucleotides (TTGGGG)4 and TGTGTGGGTGTGTGGGTGTGTGGG, consisting of the Tetrahymena and yeast telomeric sequences respectively, each functioned as primers for elongation, while (CCCCAA)4 and two nontelomeric sequence DNA oligomers did not. Efficient synthesis of the TTGGGG repeats depended only on addition of micromolar concentrations of oligomer primer, dGTP, and dTTP to the extract. The activity was sensitive to heat and proteinase K treatment. The repeat addition was independent of both endogenous Tetrahymena DNA and the endogenous alpha-type DNA polymerase; and a greater elongation activity was present during macronuclear development, when a large number of telomeres are formed and replicated, than during vegetative cell growth. We propose that the novel telomere terminal transferase is involved in the addition of telomeric repeats necessary for the replication of chromosome ends in eukaryotes.
Pan et al. (2008), Wang et al. (2008) — alternative splicing in the human transcriptome.
Deep surveying of alternative splicing complexity in the human transcriptome by high-throughput sequencing. Pan et al. (2008). Nature Genetics. doi:10.1038/ng.259
Abstract
We carried out the first analysis of alternative splicing complexity in human tissues using mRNA-Seq data. New splice junctions were detected in approximately 20% of multiexon genes, many of which are tissue specific. By combining mRNA-Seq and EST-cDNA sequence data, we estimate that transcripts from approximately 95% of multiexon genes undergo alternative splicing and that there are approximately 100,000 intermediate- to high-abundance alternative splicing events in major human tissues. From a comparison with quantitative alternative splicing microarray profiling data, we also show that mRNA-Seq data provide reliable measurements for exon inclusion levels.
Alternative isoform regulation in human tissue transcriptomes. Wang et al. (2008). Nature. doi:10.1038/nature07509
Abstract
Through alternative processing of pre-messenger RNAs, individual mammalian genes often produce multiple mRNA and protein isoforms that may have related, distinct or even opposing functions. Here we report an in-depth analysis of 15 diverse human tissue and cell line transcriptomes on the basis of deep sequencing of complementary DNA fragments, yielding a digital inventory of gene and mRNA isoform expression. Analyses in which sequence reads are mapped to exon-exon junctions indicated that 92-94% of human genes undergo alternative splicing, 86% with a minor isoform frequency of 15% or more. Differences in isoform-specific read densities indicated that most alternative splicing and alternative cleavage and polyadenylation events vary between tissues, whereas variation between individuals was approximately twofold to threefold less common. Extreme or 'switch-like' regulation of splicing between tissues was associated with increased sequence conservation in regulatory regions and with generation of full-length open reading frames. Patterns of alternative splicing and alternative cleavage and polyadenylation were strongly correlated across tissues, suggesting coordinated regulation of these processes, and sequence conservation of a subset of known regulatory motifs in both alternative introns and 3' untranslated regions suggested common involvement of specific factors in tissue-level regulation of both splicing and polyadenylation.
Koonin and Wolf (2008) — the dynamic view of prokaryotic genomes.
Genomics of bacteria and archaea: the emerging dynamic view of the prokaryotic world. Koonin & Wolf (2008). Nucleic Acids Research. doi:10.1093/nar/gkn668
Abstract
The first bacterial genome was sequenced in 1995, and the first archaeal genome in 1996. Soon after these breakthroughs, an exponential rate of genome sequencing was established, with a doubling time of approximately 20 months for bacteria and approximately 34 months for archaea. Comparative analysis of the hundreds of sequenced bacterial and dozens of archaeal genomes leads to several generalizations on the principles of genome organization and evolution. A crucial finding that enables functional characterization of the sequenced genomes and evolutionary reconstruction is that the majority of archaeal and bacterial genes have conserved orthologs in other, often, distant organisms. However, comparative genomics also shows that horizontal gene transfer (HGT) is a dominant force of prokaryotic evolution, along with the loss of genetic material resulting in genome contraction. A crucial component of the prokaryotic world is the mobilome, the enormous collection of viruses, plasmids and other selfish elements, which are in constant exchange with more stable chromosomes and serve as HGT vehicles. Thus, the prokaryotic genome space is a tightly connected, although compartmentalized, network, a novel notion that undermines the 'Tree of Life' model of evolution and requires a new conceptual framework and tools for the study of prokaryotic evolution.
B.6.3 Organelles and endosymbiosis
Sagan (1967) — the origin of mitosing cells (endosymbiotic theory).
On the origin of mitosing cells. Sagan (1967). Journal of Theoretical Biology. doi:10.1016/0022-5193(67)90079-3
Abstract
A theory of the origin of eukaryotic cells ("higher" cells which divide by classical mitosis) is presented. By hypothesis, three fundamental organelles: the mitochondria, the photosynthetic plastids and the (9+2) basal bodies of flagella were themselves once free-living (prokaryotic) cells. The evolution of photosynthesis under the anaerobic conditions of the early atmosphere to form anaerobic bacteria, photosynthetic bacteria and eventually blue-green algae (and protoplastids) is described. The subsequent evolution of aerobic metabolism in prokaryotes to form aerobic bacteria (protoflagella and protomitochondria) presumably occurred during the transition to the oxidizing atmosphere. Classical mitosis evolved in protozoan-type cells millions of years after the evolution of photosynthesis. A plausible scheme for the origin of classical mitosis in primitive amoeboflagellates is presented. During the course of the evolution of mitosis, photosynthetic plastids (themselves derived from prokaryotes) were symbiotically acquired by some of these protozoans to form the eukaryotic algae and the green plants. The cytological, biochemical and paleontological evidence for this theory is presented, along with suggestions for further possible experimental verification. The implications of this scheme for the systematics of the lower organisms is discussed.
Anderson et al. (1981) — the human mitochondrial genome.
Sequence and organization of the human mitochondrial genome. Anderson et al. (1981). Nature. doi:10.1038/290457a0
Abstract
The complete sequence of the 16,569-base pair human mitochondrial genome is presented. The genes for the 12S and 16S rRNAs, 22 tRNAs, cytochrome c oxidase subunits I, II and III, ATPase subunit 6, cytochrome b and eight other predicted protein coding genes have been located. The sequence shows extreme economy in that the genes have none or only a few noncoding bases between them, and in many cases the termination codons are not coded in the DNA but are created post-transcriptionally by polyadenylation of the mRNAs.
Lang et al. (1997) — the most bacteria-like mitochondrial genome.
An ancestral mitochondrial DNA resembling a eubacterial genome in miniature. Lang et al. (1997). Nature. doi:10.1038/387493a0
Abstract
Mitochondria, organelles specialized in energy conservation reactions in eukaryotic cells, have evolved from eubacteria-like endosymbionts whose closest known relatives are the rickettsial group of alpha-proteobacteria. Because characterized mitochondrial genomes vary markedly in structure, it has been impossible to infer from them the initial form of the proto-mitochondrial genome. This would require the identification of minimally derived mitochondrial DNAs that better reflect the ancestral state. Here we describe such a primitive mitochondrial genome, in the freshwater protozoon Reclinomonas americana. This protist displays ultrastructural characteristics that ally it with the retortamonads, a protozoan group that lacks mitochondria. R. americana mtDNA (69,034 base pairs) contains the largest collection of genes (97) so far identified in any mtDNA, including genes for 5S ribosomal RNA, the RNA component of RNase P, and at least 18 proteins not previously known to be encoded in mitochondria. Most surprising are four genes specifying a multisubunit, eubacterial-type RNA polymerase. Features of gene content together with eubacterial characteristics of genome organization and expression not found before in mitochondrial genomes indicate that R. americana mtDNA more closely resembles the ancestral proto-mitochondrial genome than any other mtDNA investigated to date.
Timmis et al. (2004) — endosymbiotic gene transfer.
Endosymbiotic gene transfer: organelle genomes forge eukaryotic chromosomes. Timmis et al. (2004). Nature Reviews Genetics. doi:10.1038/nrg1271
Roger et al. (2017) — the origin and diversification of mitochondria.
The Origin and Diversification of Mitochondria. Roger et al. (2017). Current Biology. doi:10.1016/j.cub.2017.09.015
Abstract
Mitochondria are best known for their role in the generation of ATP by aerobic respiration. Yet, research in the past half century has shown that they perform a much larger suite of functions and that these functions can vary substantially among diverse eukaryotic lineages. Despite this diversity, all mitochondria derive from a common ancestral organelle that originated from the integration of an endosymbiotic alphaproteobacterium into a host cell related to Asgard Archaea. The transition from endosymbiotic bacterium to permanent organelle entailed a massive number of evolutionary changes including the origins of hundreds of new genes and a protein import system, insertion of membrane transporters, integration of metabolism and reproduction, genome reduction, endosymbiotic gene transfer, lateral gene transfer and the retargeting of proteins. These changes occurred incrementally as the endosymbiont and the host became integrated. Although many insights into this transition have been gained, controversy persists regarding the nature of the original endosymbiont, its initial interactions with the host and the timing of its integration relative to the origin of other features of eukaryote cells. Since the establishment of the organelle, proteins have been gained, lost, transferred and retargeted as mitochondria have specialized into the spectrum of functional types seen across the eukaryotic tree of life.
B.6.4 Populations
Tettelin et al. (2005), Medini et al. (2005) — the pan-genome concept.
Genome analysis of multiple pathogenic isolates of Streptococcus agalactiae : Implications for the microbial “pan-genome”. Tettelin et al. (2005). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.0506758102
Abstract
The development of efficient and inexpensive genome sequencing methods has revolutionized the study of human bacterial pathogens and improved vaccine design. Unfortunately, the sequence of a single genome does not reflect how genetic variability drives pathogenesis within a bacterial species and also limits genome-wide screens for vaccine candidates or for antimicrobial targets. We have generated the genomic sequence of six strains representing the five major disease-causing serotypes of Streptococcus agalactiae, the main cause of neonatal infection in humans. Analysis of these genomes and those available in databases showed that the S. agalactiae species can be described by a pan-genome consisting of a core genome shared by all isolates, accounting for approximately 80% of any single genome, plus a dispensable genome consisting of partially shared and strain-specific genes. Mathematical extrapolation of the data suggests that the gene reservoir available for inclusion in the S. agalactiae pan-genome is vast and that unique genes will continue to be identified even after sequencing hundreds of genomes.
The microbial pan-genome. Medini et al. (2005). Current Opinion in Genetics & Development. doi:10.1016/j.gde.2005.09.006
Abstract
A decade after the beginning of the genomic era, the question of how genomics can describe a bacterial species has not been fully addressed. Experimental data have shown that in some species new genes are discovered even after sequencing the genomes of several strains. Mathematical modeling predicts that new genes will be discovered even after sequencing hundreds of genomes per species. Therefore, a bacterial species can be described by its pan-genome, which is composed of a "core genome" containing genes present in all strains, and a "dispensable genome" containing genes present in two or more strains and genes unique to single strains. Given that the number of unique genes is vast, the pan-genome of a bacterial species might be orders of magnitude larger than any single genome.
Touchon et al. (2009) — the E. coli pan-genome.
Organised Genome Dynamics in the Escherichia coli Species Results in Highly Diverse Adaptive Paths. Touchon et al. (2009). PLoS Genetics. doi:10.1371/journal.pgen.1000344
Abstract
The Escherichia coli species represents one of the best-studied model organisms, but also encompasses a variety of commensal and pathogenic strains that diversify by high rates of genetic change. We uniformly (re-) annotated the genomes of 20 commensal and pathogenic E. coli strains and one strain of E. fergusonii (the closest E. coli related species), including seven that we sequenced to completion. Within the approximately 18,000 families of orthologous genes, we found approximately 2,000 common to all strains. Although recombination rates are much higher than mutation rates, we show, both theoretically and using phylogenetic inference, that this does not obscure the phylogenetic signal, which places the B2 phylogenetic group and one group D strain at the basal position. Based on this phylogeny, we inferred past evolutionary events of gain and loss of genes, identifying functional classes under opposite selection pressures. We found an important adaptive role for metabolism diversification within group B2 and Shigella strains, but identified few or no extraintestinal virulence-specific genes, which could render difficult the development of a vaccine against extraintestinal infections. Genome flux in E. coli is confined to a small number of conserved positions in the chromosome, which most often are not associated with integrases or tRNA genes. Core genes flanking some of these regions show higher rates of recombination, suggesting that a gene, once acquired by a strain, spreads within the species by homologous recombination at the flanking genes. Finally, the genome's long-scale structure of recombination indicates lower recombination rates, but not higher mutation rates, at the terminus of replication. The ensuing effect of background selection and biased gene conversion may thus explain why this region is A+T-rich and shows high sequence divergence but low sequence polymorphism. Overall, despite a very high gene flow, genes co-exist in an organised genome.
Sela et al. (2016) — a theory of prokaryotic genome evolution.
Theory of prokaryotic genome evolution. Sela et al. (2016). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1614083113
Abstract
Bacteria and archaea typically possess small genomes that are tightly packed with protein-coding genes. The compactness of prokaryotic genomes is commonly perceived as evidence of adaptive genome streamlining caused by strong purifying selection in large microbial populations. In such populations, even the small cost incurred by nonfunctional DNA because of extra energy and time expenditure is thought to be sufficient for this extra genetic material to be eliminated by selection. However, contrary to the predictions of this model, there exists a consistent, positive correlation between the strength of selection at the protein sequence level, measured as the ratio of nonsynonymous to synonymous substitution rates, and microbial genome size. Here, by fitting the genome size distributions in multiple groups of prokaryotes to predictions of mathematical models of population evolution, we show that only models in which acquisition of additional genes is, on average, slightly beneficial yield a good fit to genomic data. These results suggest that the number of genes in prokaryotic genomes reflects the equilibrium between the benefit of additional genes that diminishes as the genome grows and deletion bias (i.e., the rate of deletion of genetic material being slightly greater than the rate of acquisition). Thus, new genes acquired by microbial genomes, on average, appear to be adaptive. The tight spacing of protein-coding genes likely results from a combination of the deletion bias and purifying selection that efficiently eliminates nonfunctional, noncoding sequences.
Lynch et al. (2016) — genetic drift, selection and the evolution of the mutation rate.
Genetic drift, selection and the evolution of the mutation rate. Lynch et al. (2016). Nature Reviews Genetics. doi:10.1038/nrg.2016.104
Abstract
As one of the few cellular traits that can be quantified across the tree of life, DNA-replication fidelity provides an excellent platform for understanding fundamental evolutionary processes. Furthermore, because mutation is the ultimate source of all genetic variation, clarifying why mutation rates vary is crucial for understanding all areas of biology. A potentially revealing hypothesis for mutation-rate evolution is that natural selection primarily operates to improve replication fidelity, with the ultimate limits to what can be achieved set by the power of random genetic drift. This drift-barrier hypothesis is consistent with comparative measures of mutation rates, provides a simple explanation for the existence of error-prone polymerases and yields a formal counter-argument to the view that selection fine-tunes gene-specific mutation rates.
Cann et al. (1987), Hublin et al. (2017) — mitochondrial Eve and the age of Homo sapiens.
Mitochondrial DNA and human evolution. Cann et al. (1987). Nature. doi:10.1038/325031a0
Abstract
Mitochondrial DNAs from 147 people, drawn from five geographic populations have been analysed by restriction mapping. All these mitochondrial DNAs stem from one woman who is postulated to have lived about 200,000 years ago, probably in Africa. All the populations examined except the African population have multiple origins, implying that each area was colonised repeatedly.
New fossils from Jebel Irhoud, Morocco and the pan-African origin of Homo sapiens. Hublin et al. (2017). Nature. doi:10.1038/nature22336
Abstract
Fossil evidence points to an African origin of Homo sapiens from a group called either H. heidelbergensis or H. rhodesiensis. However, the exact place and time of emergence of H. sapiens remain obscure because the fossil record is scarce and the chronological age of many key specimens remains uncertain. In particular, it is unclear whether the present day 'modern' morphology rapidly emerged approximately 200 thousand years ago (ka) among earlier representatives of H. sapiens or evolved gradually over the last 400 thousand years. Here we report newly discovered human fossils from Jebel Irhoud, Morocco, and interpret the affinities of the hominins from this site with other archaic and recent human groups. We identified a mosaic of features including facial, mandibular and dental morphology that aligns the Jebel Irhoud material with early or recent anatomically modern humans and more primitive neurocranial and endocranial morphology. In combination with an age of 315 ± 34 thousand years (as determined by thermoluminescence dating), this evidence makes Jebel Irhoud the oldest and richest African Middle Stone Age hominin site that documents early stages of the H. sapiens clade in which key features of modern morphology were established. Furthermore, it shows that the evolutionary processes behind the emergence of H. sapiens involved the whole African continent.
Krings et al. (1997) — the first Neanderthal DNA.
Neandertal DNA Sequences and the Origin of Modern Humans. Krings et al. (1997). Cell. doi:10.1016/s0092-8674(00)80310-4
Abstract
DNA was extracted from the Neandertal-type specimen found in 1856 in western Germany. By sequencing clones from short overlapping PCR products, a hitherto unknown mitochondrial (mt) DNA sequence was determined. Multiple controls indicate that this sequence is endogenous to the fossil. Sequence comparisons with human mtDNA sequences, as well as phylogenetic analyses, show that the Neandertal sequence falls outside the variation of modern humans. Furthermore, the age of the common ancestor of the Neandertal and modern human mtDNAs is estimated to be four times greater than that of the common ancestor of human mtDNAs. This suggests that Neandertals went extinct without contributing mtDNA to modern humans.
Reich et al. (2010), Meyer et al. (2012) — the Denisovans.
Genetic history of an archaic hominin group from Denisova Cave in Siberia. Reich et al. (2010). Nature. doi:10.1038/nature09710
Abstract
Using DNA extracted from a finger bone found in Denisova Cave in southern Siberia, we have sequenced the genome of an archaic hominin to about 1.9-fold coverage. This individual is from a group that shares a common origin with Neanderthals. This population was not involved in the putative gene flow from Neanderthals into Eurasians; however, the data suggest that it contributed 4-6% of its genetic material to the genomes of present-day Melanesians. We designate this hominin population 'Denisovans' and suggest that it may have been widespread in Asia during the Late Pleistocene epoch. A tooth found in Denisova Cave carries a mitochondrial genome highly similar to that of the finger bone. This tooth shares no derived morphological features with Neanderthals or modern humans, further indicating that Denisovans have an evolutionary history distinct from Neanderthals and modern humans.
A High-Coverage Genome Sequence from an Archaic Denisovan Individual. Meyer et al. (2012). Science. doi:10.1126/science.1224344
Abstract
We present a DNA library preparation method that has allowed us to reconstruct a high-coverage (30×) genome sequence of a Denisovan, an extinct relative of Neandertals. The quality of this genome allows a direct estimation of Denisovan heterozygosity indicating that genetic diversity in these archaic hominins was extremely low. It also allows tentative dating of the specimen on the basis of "missing evolution" in its genome, detailed measurements of Denisovan and Neandertal admixture into present-day human populations, and the generation of a near-complete catalog of genetic changes that swept to high frequency in modern humans since their divergence from Denisovans.
Prüfer et al. (2013) — a high-coverage Neanderthal genome.
The complete genome sequence of a Neanderthal from the Altai Mountains. Prüfer et al. (2013). Nature. doi:10.1038/nature12886
Abstract
We present a high-quality genome sequence of a Neanderthal woman from Siberia. We show that her parents were related at the level of half-siblings and that mating among close relatives was common among her recent ancestors. We also sequenced the genome of a Neanderthal from the Caucasus to low coverage. An analysis of the relationships and population history of available archaic genomes and 25 present-day human genomes shows that several gene flow events occurred among Neanderthals, Denisovans and early modern humans, possibly including gene flow into Denisovans from an unknown archaic group. Thus, interbreeding, albeit of low magnitude, occurred among many hominin groups in the Late Pleistocene. In addition, the high-quality Neanderthal genome allows us to establish a definitive list of substitutions that became fixed in modern humans after their separation from the ancestors of Neanderthals and Denisovans.
Slon et al. (2018) — a Neanderthal–Denisovan hybrid.
The genome of the offspring of a Neanderthal mother and a Denisovan father. Slon et al. (2018). Nature. doi:10.1038/s41586-018-0455-x
Abstract
Neanderthals and Denisovans are extinct groups of hominins that separated from each other more than 390,000 years ago 1,2 . Here we present the genome of 'Denisova 11', a bone fragment from Denisova Cave (Russia) 3 and show that it comes from an individual who had a Neanderthal mother and a Denisovan father. The father, whose genome bears traces of Neanderthal ancestry, came from a population related to a later Denisovan found in the cave 4-6 . The mother came from a population more closely related to Neanderthals who lived later in Europe 2,7 than to an earlier Neanderthal found in Denisova Cave 8 , suggesting that migrations of Neanderthals between eastern and western Eurasia occurred sometime after 120,000 years ago. The finding of a first-generation Neanderthal-Denisovan offspring among the small number of archaic specimens sequenced to date suggests that mixing between Late Pleistocene hominin groups was common when they met.
Sankararaman et al. (2014) — the landscape of Neanderthal ancestry in present-day humans.
The genomic landscape of Neanderthal ancestry in present-day humans. Sankararaman et al. (2014). Nature. doi:10.1038/nature12961
Abstract
Genomic studies have shown that Neanderthals interbred with modern humans, and that non-Africans today are the products of this mixture. The antiquity of Neanderthal gene flow into modern humans means that genomic regions that derive from Neanderthals in any one human today are usually less than a hundred kilobases in size. However, Neanderthal haplotypes are also distinctive enough that several studies have been able to detect Neanderthal ancestry at specific loci. We systematically infer Neanderthal haplotypes in the genomes of 1,004 present-day humans. Regions that harbour a high frequency of Neanderthal alleles are enriched for genes affecting keratin filaments, suggesting that Neanderthal alleles may have helped modern humans to adapt to non-African environments. We identify multiple Neanderthal-derived alleles that confer risk for disease, suggesting that Neanderthal alleles continue to shape human biology. An unexpected finding is that regions with reduced Neanderthal ancestry are enriched in genes, implying selection to remove genetic material derived from Neanderthals. Genes that are more highly expressed in testes than in any other tissue are especially reduced in Neanderthal ancestry, and there is an approximately fivefold reduction of Neanderthal ancestry on the X chromosome, which is known from studies of diverse species to be especially dense in male hybrid sterility genes. These results suggest that part of the explanation for genomic regions of reduced Neanderthal ancestry is Neanderthal alleles that caused decreased fertility in males when moved to a modern human genetic background.
Huerta-Sánchez et al. (2014) — adaptive introgression in Tibetans.
Altitude adaptation in Tibetans caused by introgression of Denisovan-like DNA. Huerta-Sánchez et al. (2014). Nature. doi:10.1038/nature13408
Abstract
As modern humans migrated out of Africa, they encountered many new environmental conditions, including greater temperature extremes, different pathogens and higher altitudes. These diverse environments are likely to have acted as agents of natural selection and to have led to local adaptations. One of the most celebrated examples in humans is the adaptation of Tibetans to the hypoxic environment of the high-altitude Tibetan plateau. A hypoxia pathway gene, EPAS1, was previously identified as having the most extreme signature of positive selection in Tibetans, and was shown to be associated with differences in haemoglobin concentration at high altitude. Re-sequencing the region around EPAS1 in 40 Tibetan and 40 Han individuals, we find that this gene has a highly unusual haplotype structure that can only be convincingly explained by introgression of DNA from Denisovan or Denisovan-related individuals into humans. Scanning a larger set of worldwide populations, we find that the selected haplotype is only found in Denisovans and in Tibetans, and at very low frequency among Han Chinese. Furthermore, the length of the haplotype, and the fact that it is not found in any other populations, makes it unlikely that the haplotype sharing between Tibetans and Denisovans was caused by incomplete ancestral lineage sorting rather than introgression. Our findings illustrate that admixture with other hominin species has provided genetic variation that helped humans to adapt to new environments.
The 1000 Genomes Project Consortium (2015) — the 1000 Genomes Project.
A global reference for human genetic variation. The 1000 Genomes Project Consortium (2015). Nature. doi:10.1038/nature15393
Abstract
The 1000 Genomes Project set out to provide a comprehensive description of common human genetic variation by applying whole-genome sequencing to a diverse set of individuals from multiple populations. Here we report completion of the project, having reconstructed the genomes of 2,504 individuals from 26 populations using a combination of low-coverage whole-genome sequencing, deep exome sequencing, and dense microarray genotyping. We characterized a broad spectrum of genetic variation, in total over 88 million variants (84.7 million single nucleotide polymorphisms (SNPs), 3.6 million short insertions/deletions (indels), and 60,000 structural variants), all phased onto high-quality haplotypes. This resource includes >99% of SNP variants with a frequency of >1% for a variety of ancestries. We describe the distribution of genetic variation across the global sample, and discuss the implications for common disease studies.
B.6.5 Repeats and the 3D genome
McClintock (1950) — the discovery of transposable elements.
The origin and behavior of mutable loci in maize. McClintock (1950). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.36.6.344
Wells and Feschotte (2020) — a field guide to transposable elements.
A Field Guide to Eukaryotic Transposable Elements. Wells & Feschotte (2020). Annual Review of Genetics. doi:10.1146/annurev-genet-040620-022145
Abstract
Transposable elements (TEs) are mobile DNA sequences that propagate within genomes. Through diverse invasion strategies, TEs have come to occupy a substantial fraction of nearly all eukaryotic genomes, and they represent a major source of genetic variation and novelty. Here we review the defining features of each major group of eukaryotic TEs and explore their evolutionary origins and relationships. We discuss how the unique biology of different TEs influences their propagation and distribution within and across genomes. Environmental and genetic factors acting at the level of the host species further modulate the activity, diversification, and fate of TEs, producing the dramatic variation in TE content observed across eukaryotes. We argue that cataloging TE diversity and dissecting the idiosyncratic behavior of individual elements are crucial to expanding our comprehension of their impact on the biology of genomes and the evolution of species.
Storer et al. (2021) — Dfam.
The Dfam community resource of transposable element families, sequence models, and genome annotations. Storer et al. (2021). Mobile DNA. doi:10.1186/s13100-020-00230-y
Abstract
Dfam is an open access database of repetitive DNA families, sequence models, and genome annotations. The 3.0-3.3 releases of Dfam ( https://dfam.org ) represent an evolution from a proof-of-principle collection of transposable element families in model organisms into a community resource for a broad range of species, and for both curated and uncurated datasets. In addition, releases since Dfam 3.0 provide auxiliary consensus sequence models, transposable element protein alignments, and a formalized classification system to support the growing diversity of organisms represented in the resource. The latest release includes 266,740 new de novo generated transposable element families from 336 species contributed by the EBI. This expansion demonstrates the utility of many of Dfam's new features and provides insight into the long term challenges ahead for improving de novo generated transposable element datasets.
Bailey et al. (2002) — segmental duplications in the human genome.
Recent Segmental Duplications in the Human Genome. Bailey et al. (2002). Science. doi:10.1126/science.1072047
Abstract
Primate-specific segmental duplications are considered important in human disease and evolution. The inability to distinguish between allelic and duplication sequence overlap has hampered their characterization as well as assembly and annotation of our genome. We developed a method whereby each public sequence is analyzed at the clone level for overrepresentation within a whole-genome shotgun sequence. This test has the ability to detect duplications larger than 15 kilobases irrespective of copy number, location, or high sequence similarity. We mapped 169 large regions flanked by highly similar duplications. Twenty-four of these hot spots of genomic instability have been associated with genetic disease. Our analysis indicates a highly nonrandom chromosomal and genic distribution of recent segmental duplications, with a likely role in expanding protein diversity.
Altemose et al. (2022) — complete maps of human centromeres.
Complete genomic and epigenetic maps of human centromeres. Altemose et al. (2022). Science. doi:10.1126/science.abl4178
Abstract
Existing human genome assemblies have almost entirely excluded repetitive sequences within and near centromeres, limiting our understanding of their organization, evolution, and functions, which include facilitating proper chromosome segregation. Now, a complete, telomere-to-telomere human genome assembly (T2T-CHM13) has enabled us to comprehensively characterize pericentromeric and centromeric repeats, which constitute 6.2% of the genome (189.9 megabases). Detailed maps of these regions revealed multimegabase structural rearrangements, including in active centromeric repeat arrays. Analysis of centromere-associated sequences uncovered a strong relationship between the position of the centromere and the evolution of the surrounding DNA through layered repeat expansions. Furthermore, comparisons of chromosome X centromeres across a diverse panel of individuals illuminated high degrees of structural, epigenetic, and sequence variation in these complex and rapidly evolving regions.
Dekker et al. (2002) — chromosome conformation capture.
Capturing Chromosome Conformation. Dekker et al. (2002). Science. doi:10.1126/science.1067799
Abstract
We describe an approach to detect the frequency of interaction between any two genomic loci. Generation of a matrix of interaction frequencies between sites on the same or different chromosomes reveals their relative spatial disposition and provides information about the physical properties of the chromatin fiber. This methodology can be applied to the spatial organization of entire genomes in organisms from bacteria to human. Using the yeast Saccharomyces cerevisiae, we could confirm known qualitative features of chromosome organization within the nucleus and dynamic changes in that organization during meiosis. We also analyzed yeast chromosome III at the G1 stage of the cell cycle. We found that chromatin is highly flexible throughout. Furthermore, functionally distinct AT- and GC-rich domains were found to exhibit different conformations, and a population-average 3D model of chromosome III could be determined. Chromosome III emerges as a contorted ring.
Lieberman-Aiden et al. (2009) — Hi-C.
Comprehensive Mapping of Long-Range Interactions Reveals Folding Principles of the Human Genome. Lieberman-Aiden et al. (2009). Science. doi:10.1126/science.1181369
Abstract
We describe Hi-C, a method that probes the three-dimensional architecture of whole genomes by coupling proximity-based ligation with massively parallel sequencing. We constructed spatial proximity maps of the human genome with Hi-C at a resolution of 1 megabase. These maps confirm the presence of chromosome territories and the spatial proximity of small, gene-rich chromosomes. We identified an additional level of genome organization that is characterized by the spatial segregation of open and closed chromatin to form two genome-wide compartments. At the megabase scale, the chromatin conformation is consistent with a fractal globule, a knot-free, polymer conformation that enables maximally dense packing while preserving the ability to easily fold and unfold any genomic locus. The fractal globule is distinct from the more commonly used globular equilibrium model. Our results demonstrate the power of Hi-C to map the dynamic conformations of whole genomes.
Dixon et al. (2012), Nora et al. (2012) — topologically associating domains.
Topological domains in mammalian genomes identified by analysis of chromatin interactions. Dixon et al. (2012). Nature. doi:10.1038/nature11082
Abstract
The spatial organization of the genome is intimately linked to its biological function, yet our understanding of higher order genomic structure is coarse, fragmented and incomplete. In the nucleus of eukaryotic cells, interphase chromosomes occupy distinct chromosome territories, and numerous models have been proposed for how chromosomes fold within chromosome territories. These models, however, provide only few mechanistic details about the relationship between higher order chromatin structure and genome function. Recent advances in genomic technologies have led to rapid advances in the study of three-dimensional genome organization. In particular, Hi-C has been introduced as a method for identifying higher order chromatin interactions genome wide. Here we investigate the three-dimensional organization of the human and mouse genomes in embryonic stem cells and terminally differentiated cell types at unprecedented resolution. We identify large, megabase-sized local chromatin interaction domains, which we term 'topological domains', as a pervasive structural feature of the genome organization. These domains correlate with regions of the genome that constrain the spread of heterochromatin. The domains are stable across different cell types and highly conserved across species, indicating that topological domains are an inherent property of mammalian genomes. Finally, we find that the boundaries of topological domains are enriched for the insulator binding protein CTCF, housekeeping genes, transfer RNAs and short interspersed element (SINE) retrotransposons, indicating that these factors may have a role in establishing the topological domain structure of the genome.
Spatial partitioning of the regulatory landscape of the X-inactivation centre. Nora et al. (2012). Nature. doi:10.1038/nature11049
Abstract
In eukaryotes transcriptional regulation often involves multiple long-range elements and is influenced by the genomic environment. A prime example of this concerns the mouse X-inactivation centre (Xic), which orchestrates the initiation of X-chromosome inactivation (XCI) by controlling the expression of the non-protein-coding Xist transcript. The extent of Xic sequences required for the proper regulation of Xist remains unknown. Here we use chromosome conformation capture carbon-copy (5C) and super-resolution microscopy to analyse the spatial organization of a 4.5-megabases (Mb) region including Xist. We discover a series of discrete 200-kilobase to 1 Mb topologically associating domains (TADs), present both before and after cell differentiation and on the active and inactive X. TADs align with, but do not rely on, several domain-wide features of the epigenome, such as H3K27me3 or H3K9me2 blocks and lamina-associated domains. TADs also align with coordinately regulated gene clusters. Disruption of a TAD boundary causes ectopic chromosomal contacts and long-range transcriptional misregulation. The Xist/Tsix sense/antisense unit illustrates how TADs enable the spatial segregation of oppositely regulated chromosomal neighbourhoods, with the respective promoters of Xist and Tsix lying in adjacent TADs, each containing their known positive regulators. We identify a novel distal regulatory region of Tsix within its TAD, which produces a long intervening RNA, Linx. In addition to uncovering a new principle of cis-regulatory architecture of mammalian chromosomes, our study sets the stage for the full genetic dissection of the X-inactivation centre.
Rao et al. (2014), Fudenberg et al. (2016) — loops and loop extrusion.
A 3D Map of the Human Genome at Kilobase Resolution Reveals Principles of Chromatin Looping. Rao et al. (2014). Cell. doi:10.1016/j.cell.2014.11.021
Abstract
We use in situ Hi-C to probe the 3D architecture of genomes, constructing haploid and diploid maps of nine cell types. The densest, in human lymphoblastoid cells, contains 4.9 billion contacts, achieving 1 kb resolution. We find that genomes are partitioned into contact domains (median length, 185 kb), which are associated with distinct patterns of histone marks and segregate into six subcompartments. We identify ∼10,000 loops. These loops frequently link promoters and enhancers, correlate with gene activation, and show conservation across cell types and species. Loop anchors typically occur at domain boundaries and bind CTCF. CTCF sites at loop anchors occur predominantly (>90%) in a convergent orientation, with the asymmetric motifs "facing" one another. The inactive X chromosome splits into two massive domains and contains large loops anchored at CTCF-binding repeats.
Formation of Chromosomal Domains by Loop Extrusion. Fudenberg et al. (2016). Cell Reports. doi:10.1016/j.celrep.2016.04.085
Abstract
Topologically associating domains (TADs) are fundamental structural and functional building blocks of human interphase chromosomes, yet the mechanisms of TAD formation remain unclear. Here, we propose that loop extrusion underlies TAD formation. In this process, cis-acting loop-extruding factors, likely cohesins, form progressively larger loops but stall at TAD boundaries due to interactions with boundary proteins, including CTCF. Using polymer simulations, we show that this model produces TADs and finer-scale features of Hi-C data. Each TAD emerges from multiple loops dynamically formed through extrusion, contrary to typical illustrations of single static loops. Loop extrusion both explains diverse experimental observations-including the preferential orientation of CTCF motifs, enrichments of architectural proteins at TAD boundaries, and boundary deletion experiments-and makes specific predictions for the depletion of CTCF versus cohesin. Finally, loop extrusion has potentially far-ranging consequences for processes such as enhancer-promoter interactions, orientation-specific chromosomal looping, and compaction of mitotic chromosomes.
B.6.6 Non-coding RNA and the function debate
Lee et al. (1993), Bartel (2018) — microRNAs.
The C. elegans heterochronic gene lin-4 encodes small RNAs with antisense complementarity to lin-14. Lee et al. (1993). Cell. doi:10.1016/0092-8674(93)90529-y
Abstract
lin-4 is essential for the normal temporal control of diverse postembryonic developmental events in C. elegans. lin-4 acts by negatively regulating the level of LIN-14 protein, creating a temporal decrease in LIN-14 protein starting in the first larval stage (L1). We have cloned the C. elegans lin-4 locus by chromosomal walking and transformation rescue. We used the C. elegans clone to isolate the gene from three other Caenorhabditis species; all four Caenorhabditis clones functionally rescue the lin-4 null allele of C. elegans. Comparison of the lin-4 genomic sequence from these four species and site-directed mutagenesis of potential open reading frames indicated that lin-4 does not encode a protein. Two small lin-4 transcripts of approximately 22 and 61 nt were identified in C. elegans and found to contain sequences complementary to a repeated sequence element in the 3' untranslated region (UTR) of lin-14 mRNA, suggesting that lin-4 regulates lin-14 translation via an antisense RNA-RNA interaction.
Metazoan MicroRNAs. Bartel (2018). Cell. doi:10.1016/j.cell.2018.03.006
Abstract
MicroRNAs (miRNAs) are ∼22 nt RNAs that direct posttranscriptional repression of mRNA targets in diverse eukaryotic lineages. In humans and other mammals, these small RNAs help sculpt the expression of most mRNAs. This article reviews advances in our understanding of the defining features of metazoan miRNAs and their biogenesis, genomics, and evolution. It then reviews how metazoan miRNAs are regulated, how they recognize and cause repression of their targets, and the biological functions of this repression, with a compilation of knockout phenotypes that shows that important biological functions have been identified for most of the broadly conserved miRNAs of mammals.
Mattick et al. (2023) — long non-coding RNAs.
Long non-coding RNAs: definitions, functions, challenges and recommendations. Mattick et al. (2023). Nature Reviews Molecular Cell Biology. doi:10.1038/s41580-022-00566-8
Abstract
Genes specifying long non-coding RNAs (lncRNAs) occupy a large fraction of the genomes of complex organisms. The term 'lncRNAs' encompasses RNA polymerase I (Pol I), Pol II and Pol III transcribed RNAs, and RNAs from processed introns. The various functions of lncRNAs and their many isoforms and interleaved relationships with other genes make lncRNA classification and annotation difficult. Most lncRNAs evolve more rapidly than protein-coding sequences, are cell type specific and regulate many aspects of cell differentiation and development and other physiological processes. Many lncRNAs associate with chromatin-modifying complexes, are transcribed from enhancers and nucleate phase separation of nuclear condensates and domains, indicating an intimate link between lncRNA expression and the spatial control of gene expression during development. lncRNAs also have important roles in the cytoplasm and beyond, including in the regulation of translation, metabolism and signalling. lncRNAs often have a modular structure and are rich in repeats, which are increasingly being shown to be relevant to their function. In this Consensus Statement, we address the definition and nomenclature of lncRNAs and their conservation, expression, phenotypic visibility, structure and functions. We also discuss research challenges and provide recommendations to advance the understanding of the roles of lncRNAs in development, cell biology and disease.
The ENCODE Project Consortium (2012), The ENCODE Project Consortium (2020) — ENCODE.
An integrated encyclopedia of DNA elements in the human genome. The ENCODE Project Consortium (2012). Nature. doi:10.1038/nature11247
Abstract
The human genome encodes the blueprint of life, but the function of the vast majority of its nearly three billion bases is unknown. The Encyclopedia of DNA Elements (ENCODE) project has systematically mapped regions of transcription, transcription factor association, chromatin structure and histone modification. These data enabled us to assign biochemical functions for 80% of the genome, in particular outside of the well-studied protein-coding regions. Many discovered candidate regulatory elements are physically associated with one another and with expressed genes, providing new insights into the mechanisms of gene regulation. The newly identified elements also show a statistical correspondence to sequence variants linked to human disease, and can thereby guide interpretation of this variation. Overall, the project provides new insights into the organization and regulation of our genes and genome, and is an expansive resource of functional annotations for biomedical research.
Expanded encyclopaedias of DNA elements in the human and mouse genomes. The ENCODE Project Consortium (2020). Nature. doi:10.1038/s41586-020-2493-4
Abstract
The human and mouse genomes contain instructions that specify RNAs and proteins and govern the timing, magnitude, and cellular context of their production. To better delineate these elements, phase III of the Encyclopedia of DNA Elements (ENCODE) Project has expanded analysis of the cell and tissue repertoires of RNA transcription, chromatin structure and modification, DNA methylation, chromatin looping, and occupancy by transcription factors and RNA-binding proteins. Here we summarize these efforts, which have produced 5,992 new experimental datasets, including systematic determinations across mouse fetal development. All data are available through the ENCODE data portal (https://www.encodeproject.org), including phase II ENCODE 1 and Roadmap Epigenomics 2 data. We have developed a registry of 926,535 human and 339,815 mouse candidate cis-regulatory elements, covering 7.9 and 3.4% of their respective genomes, by integrating selected datatypes associated with gene regulation, and constructed a web-based server (SCREEN; http://screen.encodeproject.org) to provide flexible, user-defined access to this resource. Collectively, the ENCODE data and registry provide an expansive resource for the scientific community to build a better understanding of the organization and function of the human and mouse genomes.
Graur et al. (2013), Doolittle (2013) — critiques of ENCODE.
On the Immortality of Television Sets: “Function” in the Human Genome According to the Evolution-Free Gospel of ENCODE. Graur et al. (2013). Genome Biology and Evolution. doi:10.1093/gbe/evt028
Abstract
A recent slew of ENCyclopedia Of DNA Elements (ENCODE) Consortium publications, specifically the article signed by all Consortium members, put forward the idea that more than 80% of the human genome is functional. This claim flies in the face of current estimates according to which the fraction of the genome that is evolutionarily conserved through purifying selection is less than 10%. Thus, according to the ENCODE Consortium, a biological function can be maintained indefinitely without selection, which implies that at least 80 - 10 = 70% of the genome is perfectly invulnerable to deleterious mutations, either because no mutation can ever occur in these "functional" regions or because no mutation in these regions can ever be deleterious. This absurd conclusion was reached through various means, chiefly by employing the seldom used "causal role" definition of biological function and then applying it inconsistently to different biochemical properties, by committing a logical fallacy known as "affirming the consequent," by failing to appreciate the crucial difference between "junk DNA" and "garbage DNA," by using analytical methods that yield biased errors and inflate estimates of functionality, by favoring statistical sensitivity over specificity, and by emphasizing statistical significance rather than the magnitude of the effect. Here, we detail the many logical and methodological transgressions involved in assigning functionality to almost every nucleotide in the human genome. The ENCODE results were predicted by one of its authors to necessitate the rewriting of textbooks. We agree, many textbooks dealing with marketing, mass-media hype, and public relations may well have to be rewritten.
Is junk DNA bunk? A critique of ENCODE. Doolittle (2013). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1221376110
Abstract
Do data from the Encyclopedia Of DNA Elements (ENCODE) project render the notion of junk DNA obsolete? Here, I review older arguments for junk grounded in the C-value paradox and propose a thought experiment to challenge ENCODE's ontology. Specifically, what would we expect for the number of functional elements (as ENCODE defines them) in genomes much larger than our own genome? If the number were to stay more or less constant, it would seem sensible to consider the rest of the DNA of larger genomes to be junk or, at least, assign it a different sort of role (structural rather than informational). If, however, the number of functional elements were to rise significantly with C-value then, (i) organisms with genomes larger than our genome are more complex phenotypically than we are, (ii) ENCODE's definition of functional element identifies many sites that would not be considered functional or phenotype-determining by standard uses in biology, or (iii) the same phenotypic functions are often determined in a more diffuse fashion in larger-genomed organisms. Good cases can be made for propositions ii and iii. A larger theoretical framework, embracing informational and structural roles for DNA, neutral as well as adaptive causes of complexity, and selection as a multilevel phenomenon, is needed.
Kellis et al. (2014) — defining functional DNA elements.
Defining functional DNA elements in the human genome. Kellis et al. (2014). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1318948111
Abstract
With the completion of the human genome sequence, attention turned to identifying and annotating its functional DNA elements. As a complement to genetic and comparative genomics approaches, the Encyclopedia of DNA Elements Project was launched to contribute maps of RNA transcripts, transcriptional regulator binding sites, and chromatin states in many cell types. The resulting genome-wide data reveal sites of biochemical activity with high positional resolution and cell type specificity that facilitate studies of gene regulation and interpretation of noncoding variants associated with human disease. However, the biochemically active regions cover a much larger fraction of the genome than do evolutionarily conserved regions, raising the question of whether nonconserved but biochemically active regions are truly functional. Here, we review the strengths and limitations of biochemical, evolutionary, and genetic approaches for defining functional DNA segments, potential sources for the observed differences in estimated genomic coverage, and the biological implications of these discrepancies. We also analyze the relationship between signal intensity, genomic coverage, and evolutionary conservation. Our results reinforce the principle that each approach provides complementary information and that we need to use combinations of all three to elucidate genome function in human biology and disease.
Rands et al. (2014), Christmas et al. (2023) — the constrained fraction of the human genome.
8.2% of the Human Genome Is Constrained: Variation in Rates of Turnover across Functional Element Classes in the Human Lineage. Rands et al. (2014). PLoS Genetics. doi:10.1371/journal.pgen.1004525
Abstract
Ten years on from the finishing of the human reference genome sequence, it remains unclear what fraction of the human genome confers function, where this sequence resides, and how much is shared with other mammalian species. When addressing these questions, functional sequence has often been equated with pan-mammalian conserved sequence. However, functional elements that are short-lived, including those contributing to species-specific biology, will not leave a footprint of long-lasting negative selection. Here, we address these issues by identifying and characterising sequence that has been constrained with respect to insertions and deletions for pairs of eutherian genomes over a range of divergences. Within noncoding sequence, we find increasing amounts of mutually constrained sequence as species pairs become more closely related, indicating that noncoding constrained sequence turns over rapidly. We estimate that half of present-day noncoding constrained sequence has been gained or lost in approximately the last 130 million years (half-life in units of divergence time, d1/2 = 0.25-0.31). While enriched with ENCODE biochemical annotations, much of the short-lived constrained sequences we identify are not detected by models optimized for wider pan-mammalian conservation. Constrained DNase 1 hypersensitivity sites, promoters and untranslated regions have been more evolutionarily stable than long noncoding RNA loci which have turned over especially rapidly. By contrast, protein coding sequence has been highly stable, with an estimated half-life of over a billion years (d1/2 = 2.1-5.0). From extrapolations we estimate that 8.2% (7.1-9.2%) of the human genome is presently subject to negative selection and thus is likely to be functional, while only 2.2% has maintained constraint in both human and mouse since these species diverged. These results reveal that the evolutionary history of the human genome has been highly dynamic, particularly for its noncoding yet biologically functional fraction.
Evolutionary constraint and innovation across hundreds of placental mammals. Christmas et al. (2023). Science. doi:10.1126/science.abn3943
Abstract
Zoonomia is the largest comparative genomics resource for mammals produced to date. By aligning genomes for 240 species, we identify bases that, when mutated, are likely to affect fitness and alter disease risk. At least 332 million bases (~10.7%) in the human genome are unusually conserved across species (evolutionarily constrained) relative to neutrally evolving repeats, and 4552 ultraconserved elements are nearly perfectly conserved. Of 101 million significantly constrained single bases, 80% are outside protein-coding exons and half have no functional annotations in the Encyclopedia of DNA Elements (ENCODE) resource. Changes in genes and regulatory elements are associated with exceptional mammalian traits, such as hibernation, that could inform therapeutic development. Earth's vast and imperiled biodiversity offers distinctive power for identifying genetic variants that affect genome function and organismal phenotypes.
Koonin (2011) — are there laws of genome evolution?
Are There Laws of Genome Evolution?. Koonin (2011). PLoS Computational Biology. doi:10.1371/journal.pcbi.1002173
Abstract
Research in quantitative evolutionary genomics and systems biology led to the discovery of several universal regularities connecting genomic and molecular phenomic variables. These universals include the log-normal distribution of the evolutionary rates of orthologous genes; the power law-like distributions of paralogous family size and node degree in various biological networks; the negative correlation between a gene's sequence evolution rate and expression level; and differential scaling of functional classes of genes with genome size. The universals of genome evolution can be accounted for by simple mathematical models similar to those used in statistical physics, such as the birth-death-innovation model. These models do not explicitly incorporate selection; therefore, the observed universal regularities do not appear to be shaped by selection but rather are emergent properties of gene ensembles. Although a complete physical theory of evolutionary biology is inconceivable, the universals of genome evolution might qualify as "laws of evolutionary genomics" in the same sense "law" is understood in modern physics.
B.7 Expression analysis
Further papers for Chapter 7.
Papers marked ★ were mandatory in earlier reading lists. In the HTML version, click Abstract to show the abstract.
B.7.1 Microarrays
Velculescu et al. (1995) — serial analysis of gene expression (SAGE).
Serial Analysis of Gene Expression. Velculescu et al. (1995). Science. doi:10.1126/science.270.5235.484
Abstract
The characteristics of an organism are determined by the genes expressed within it. A method was developed, called serial analysis of gene expression (SAGE), that allows the quantitative and simultaneous analysis of a large number of transcripts. To demonstrate this strategy, short diagnostic sequence tags were isolated from pancreas, concatenated, and cloned. Manual sequencing of 1000 tags revealed a gene expression pattern characteristic of pancreatic function. New pancreatic transcripts corresponding to novel tags were identified. SAGE should provide a broadly applicable means for the quantitative cataloging and comparison of expressed genes in a variety of normal, developmental, and disease states.
Fodor et al. (1991), Lockhart et al. (1996) — Affymetrix oligonucleotide arrays.
Light-Directed, Spatially Addressable Parallel Chemical Synthesis. Fodor et al. (1991). Science. doi:10.1126/science.1990438
Abstract
Solid-phase chemistry, photolabile protecting groups, and photolithography have been combined to achieve light-directed, spatially addressable parallel chemical synthesis to yield a highly diverse set of chemical products. Binary masking, one of many possible combinatorial synthesis strategies, yields 2n compounds in n chemical steps. An array of 1024 peptides was synthesized in ten steps, and its interaction with a monoclonal antibody was assayed by epifluorescence microscopy. High-density arrays formed by light-directed synthesis are potentially rich sources of chemical diversity for discovering new ligands that bind to biological receptors and for elucidating principles governing molecular interactions. The generality of this approach is illustrated by the light-directed synthesis of a dinucleotide. Spatially directed synthesis of complex compounds could also be used for microfabrication of devices.
Expression monitoring by hybridization to high-density oligonucleotide arrays. Lockhart et al. (1996). Nature Biotechnology. doi:10.1038/nbt1296-1675
Abstract
The human genome encodes approximately 100,000 different genes, and at least partial sequence information for nearly all will be available soon. Sequence information alone, however, is insufficient for a full understanding of gene function, expression, regulation, and splice-site variation. Because cellular processes are governed by the repertoire of expressed genes, and the levels and timing of expression, it is important to have experimental tools for the direct monitoring of large numbers of mRNAs in parallel. We have developed an approach that is based on hybridization to small, high-density arrays containing tens of thousands of synthetic oligonucleotides. The arrays are designed based on sequence information alone and are synthesized in situ using a combination of photolithography and oligonucleotide chemistry. RNAs present at a frequency of 1:300,000 are unambiguously detected, and detection is quantitative over more than three orders of magnitude. This approach provides a way to use directly the growing body of sequence information for highly parallel experimental investigations. Because of the combinatorial nature of the chemistry and the ability to synthesize small arrays containing hundreds of thousands of specifically chosen oligonucleotides, the method is readily scalable to the simultaneous monitoring of tens of thousands of genes.
Schena et al. (1995) — spotted cDNA microarrays.
Quantitative Monitoring of Gene Expression Patterns with a Complementary DNA Microarray. Schena et al. (1995). Science. doi:10.1126/science.270.5235.467
Abstract
A high-capacity system was developed to monitor the expression of many genes in parallel. Microarrays prepared by high-speed robotic printing of complementary DNAs on glass were used for quantitative expression measurements of the corresponding genes. Because of the small format and high density of the arrays, hybridization volumes of 2 microliters could be used that enabled detection of rare transcripts in probe mixtures derived from 2 micrograms of total cellular messenger RNA. Differential expression measurements of 45 Arabidopsis genes were made by means of simultaneous, two-color fluorescence hybridization.
Gunderson et al. (2004), Kuhn et al. (2004) — Illumina BeadArrays.
Decoding Randomly Ordered DNA Arrays. Gunderson et al. (2004). Genome Research. doi:10.1101/gr.2255804
Abstract
We have developed a simple and efficient algorithm to identify each member of a large collection of DNA-linked objects through the use of hybridization, and have applied it to the manufacture of randomly assembled arrays of beads in wells. Once the algorithm has been used to determine the identity of each bead, the microarray can be used in a wide variety of applications, including single nucleotide polymorphism genotyping and gene expression profiling. The algorithm requires only a few labels and several sequential hybridizations to identify thousands of different DNA sequences with great accuracy. We have decoded tens of thousands of arrays, each with 1520 sequences represented at approximately 30-fold redundancy by up to approximately 50,000 beads, with a median error rate of <1 x 10(-4) per bead. The approach makes use of error checking codes and provides, for the first time, a direct functional quality control of every element of each array that is manufactured. The algorithm can be applied to any spatially fixed collection of objects or molecules that are associated with specific DNA sequences.
A novel, high-performance random array platform for quantitative gene expression profiling. Kuhn et al. (2004). Genome Research. doi:10.1101/gr.2739104
Abstract
We have developed a new microarray technology for quantitative gene-expression profiling on the basis of randomly assembled arrays of beads. Each bead carries a gene-specific probe sequence. There are multiple copies of each sequence-specific bead in an array, which contributes to measurement precision and reliability. We optimized the system for specific and sensitive analysis of mammalian RNA, and using RNA controls of defined concentration, obtained the following estimates of system performance: specificity of 1:250,000 in mammalian poly(A(+)) mRNA; limit of detection 0.13 pM; dynamic range 3.2 logs; and sufficient precision to detect 1.3-fold differences with 95% confidence within the dynamic range. Measurements of expression differences between human brain and liver were validated by concordance with quantitative real-time PCR (R(2) = 0.98 for log-transformed ratios, and slope of the best-fit line = 1.04, for 20 genes). Quantitative performance was further verified using a mouse B- and T-cell model system. We found published reports of B- or T-cell-specific expression for 42 of 59 genes that showed the greatest differential expression between B- and T-cells in our system. All of the literature observations were concordant with our results. Our experiments were carried out on a 96-array matrix system that requires only 100 ng of input RNA and uses standard microtiter plates to process samples in parallel. Our technology has advantages for analyzing multiple samples, is scalable to all known genes in a genome, and is flexible, allowing the use of standard or custom probes in an array.
Quackenbush (2001) — computational analysis of microarray data.
Computational analysis of microarray data. Quackenbush (2001). Nature Reviews Genetics. doi:10.1038/35076576
Abstract
Microarray experiments are providing unprecedented quantities of genome-wide data on gene-expression patterns. Although this technique has been enthusiastically developed and applied in many biological contexts, the management and analysis of the millions of data points that result from these experiments has received less attention. Sophisticated computational tools are available, but the methods that are used to analyse the data can have a profound influence on the interpretation of the results. A basic understanding of these computational tools is therefore required for optimal experimental design and meaningful data analysis.
Bolstad et al. (2003), Irizarry (2003) — quantile normalisation and RMA.
A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. Bolstad et al. (2003). Bioinformatics. doi:10.1093/bioinformatics/19.2.185
Abstract
Motivation When running experiments that involve multiple high density oligonucleotide arrays, it is important to remove sources of variation between arrays of non-biological origin. Normalization is a process for reducing this variation. It is common to see non-linear relations between arrays and the standard normalization provided by Affymetrix does not perform well in these situations. Results We present three methods of performing normalization at the probe intensity level. These methods are called complete data methods because they make use of data from all arrays in an experiment to form the normalizing relation. These algorithms are compared to two methods that make use of a baseline array: a one number scaling based algorithm and a method that uses a non-linear normalizing relation by comparing the variability and bias of an expression measure. Two publicly available datasets are used to carry out the comparisons. The simplest and quickest complete data method is found to perform favorably. Availability Software implementing all three of the complete data normalization methods is available as part of the R package Affy, which is a part of the Bioconductor project http://www.bioconductor.org. Supplementary information Additional figures may be found at http://www.stat.berkeley.edu/~bolstad/normalize/index.html
Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Irizarry (2003). Biostatistics. doi:10.1093/biostatistics/4.2.249
Abstract
In this paper we report exploratory analyses of high-density oligonucleotide array data from the Affymetrix GeneChip system with the objective of improving upon currently used measures of gene expression. Our analyses make use of three data sets: a small experimental study consisting of five MGU74A mouse GeneChip arrays, part of the data from an extensive spike-in study conducted by Gene Logic and Wyeth's Genetics Institute involving 95 HG-U95A human GeneChip arrays; and part of a dilution study conducted by Gene Logic involving 75 HG-U95A GeneChip arrays. We display some familiar features of the perfect match and mismatch probe (PM and MM) values of these data, and examine the variance-mean relationship with probe-level data from probes believed to be defective, and so delivering noise only. We explain why we need to normalize the arrays to one another using probe level intensities. We then examine the behavior of the PM and MM using spike-in data and assess three commonly used summary measures: Affymetrix's (i) average difference (AvDiff) and (ii) MAS 5.0 signal, and (iii) the Li and Wong multiplicative model-based expression index (MBEI). The exploratory data analyses of the probe level data motivate a new summary measure that is a robust multi-array average (RMA) of background-adjusted, normalized, and log-transformed PM values. We evaluate the four expression summary measures using the dilution study data, assessing their behavior in terms of bias, variance and (for MBEI and RMA) model fit. Finally, we evaluate the algorithms in terms of their ability to detect known levels of differential expression using the spike-in data. We conclude that there is no obvious downside to using RMA and attaching a standard error (SE) to this quantity using a linear model which removes probe-specific affinities.
Smyth (2004) — limma: linear models and empirical Bayes.
Linear Models and Empirical Bayes Methods for Assessing Differential Expression in Microarray Experiments. Smyth (2004). Statistical Applications in Genetics and Molecular Biology. doi:10.2202/1544-6115.1027
Abstract
The problem of identifying differentially expressed genes in designed microarray experiments is considered. Lonnstedt and Speed (2002) derived an expression for the posterior odds of differential expression in a replicated two-color experiment using a simple hierarchical parametric model. The purpose of this paper is to develop the hierarchical model of Lonnstedt and Speed (2002) into a practical approach for general microarray experiments with arbitrary numbers of treatments and RNA samples. The model is reset in the context of general linear models with arbitrary coefficients and contrasts of interest. The approach applies equally well to both single channel and two color microarray experiments. Consistent, closed form estimators are derived for the hyperparameters in the model. The estimators proposed have robust behavior even for small numbers of arrays and allow for incomplete data arising from spot filtering or spot quality weights. The posterior odds statistic is reformulated in terms of a moderated t-statistic in which posterior residual standard deviations are used in place of ordinary standard deviations. The empirical Bayes approach is equivalent to shrinkage of the estimated sample variances towards a pooled estimate, resulting in far more stable inference when the number of arrays is small. The use of moderated t-statistics has the advantage over the posterior odds that the number of hyperparameters which need to estimated is reduced; in particular, knowledge of the non-null prior for the fold changes are not required. The moderated t-statistic is shown to follow a t-distribution with augmented degrees of freedom. The moderated t inferential approach extends to accommodate tests of composite null hypotheses through the use of moderated F-statistics. The performance of the methods is demonstrated in a simulation study. Results are presented for two publicly available data sets.
Eisen et al. (1998) — cluster analysis and display of genome-wide expression.
Cluster analysis and display of genome-wide expression patterns. Eisen et al. (1998). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.95.25.14863
Abstract
A system of cluster analysis for genome-wide expression data from DNA microarray hybridization is described that uses standard statistical algorithms to arrange genes according to similarity in pattern of gene expression. The output is displayed graphically, conveying the clustering and the underlying expression data simultaneously in a form intuitive for biologists. We have found in the budding yeast Saccharomyces cerevisiae that clustering gene expression data groups together efficiently genes of known similar function, and we find a similar tendency in human data. Thus patterns seen in genome-wide expression experiments can be interpreted as indications of the status of cellular processes. Also, coexpression of genes of known function with poorly characterized or novel genes may provide a simple means of gaining leads to the functions of many genes for which information is not available currently.
Golub et al. (1999), Perou et al. (2000), Veer et al. (2002) — classifying cancers by expression.
Molecular Classification of Cancer: Class Discovery and Class Prediction by Gene Expression Monitoring. Golub et al. (1999). Science. doi:10.1126/science.286.5439.531
Abstract
Although cancer classification has improved over the past 30 years, there has been no general approach for identifying new cancer classes (class discovery) or for assigning tumors to known classes (class prediction). Here, a generic approach to cancer classification based on gene expression monitoring by DNA microarrays is described and applied to human acute leukemias as a test case. A class discovery procedure automatically discovered the distinction between acute myeloid leukemia (AML) and acute lymphoblastic leukemia (ALL) without previous knowledge of these classes. An automatically derived class predictor was able to determine the class of new leukemia cases. The results demonstrate the feasibility of cancer classification based solely on gene expression monitoring and suggest a general strategy for discovering and predicting cancer classes for other types of cancer, independent of previous biological knowledge.
Molecular portraits of human breast tumours. Perou et al. (2000). Nature. doi:10.1038/35021093
Abstract
Human breast tumours are diverse in their natural history and in their responsiveness to treatments. Variation in transcriptional programs accounts for much of the biological diversity of human cells and tumours. In each cell, signal transduction and regulatory systems transduce information from the cell's identity to its environmental status, thereby controlling the level of expression of every gene in the genome. Here we have characterized variation in gene expression patterns in a set of 65 surgical specimens of human breast tumours from 42 different individuals, using complementary DNA microarrays representing 8,102 human genes. These patterns provided a distinctive molecular portrait of each tumour. Twenty of the tumours were sampled twice, before and after a 16-week course of doxorubicin chemotherapy, and two tumours were paired with a lymph node metastasis from the same patient. Gene expression patterns in two tumour samples from the same individual were almost always more similar to each other than either was to any other sample. Sets of co-expressed genes were identified for which variation in messenger RNA levels could be related to specific features of physiological variation. The tumours could be classified into subtypes distinguished by pervasive differences in their gene expression patterns.
Gene expression profiling predicts clinical outcome of breast cancer. van ’t Veer et al. (2002). Nature. doi:10.1038/415530a
Abstract
Breast cancer patients with the same stage of disease can have markedly different treatment responses and overall outcome. The strongest predictors for metastases (for example, lymph node status and histological grade) fail to classify accurately breast tumours according to their clinical behaviour. Chemotherapy or hormonal therapy reduces the risk of distant metastases by approximately one-third; however, 70-80% of patients receiving this treatment would have survived without it. None of the signatures of breast cancer gene expression reported to date allow for patient-tailored therapy strategies. Here we used DNA microarray analysis on primary breast tumours of 117 young patients, and applied supervised classification to identify a gene expression signature strongly predictive of a short interval to distant metastases ('poor prognosis' signature) in patients without tumour cells in local lymph nodes at diagnosis (lymph node negative). In addition, we established a signature that identifies tumours of BRCA1 carriers. The poor prognosis signature consists of genes regulating cell cycle, invasion, metastasis and angiogenesis. This gene expression profile will outperform all currently used clinical parameters in predicting disease outcome. Our findings provide a strategy to select patients who would benefit from adjuvant therapy.
Su et al. (2004) — a gene atlas of mouse and human tissues.
A gene atlas of the mouse and human protein-encoding transcriptomes. Su et al. (2004). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.0400782101
Abstract
The tissue-specific pattern of mRNA expression can indicate important clues about gene function. High-density oligonucleotide arrays offer the opportunity to examine patterns of gene expression on a genome scale. Toward this end, we have designed custom arrays that interrogate the expression of the vast majority of protein-encoding human and mouse genes and have used them to profile a panel of 79 human and 61 mouse tissues. The resulting data set provides the expression patterns for thousands of predicted genes, as well as known and poorly characterized genes, from mice and humans. We have explored this data set for global trends in gene expression, evaluated commonly used lines of evidence in gene prediction methodologies, and investigated patterns indicative of chromosomal organization of transcription. We describe hundreds of regions of correlated transcription and show that some are subject to both tissue and parental allele-specific expression, suggesting a link between spatial expression and imprinting.
Brazma et al. (2001), MAQC Consortium (2006) — standards and reproducibility.
Minimum information about a microarray experiment (MIAME)—toward standards for microarray data. Brazma et al. (2001). Nature Genetics. doi:10.1038/ng1201-365
Abstract
Microarray analysis has become a widely used tool for the generation of gene expression data on a genomic scale. Although many significant results have been derived from microarray studies, one limitation has been the lack of standards for presenting and exchanging such data. Here we present a proposal, the Minimum Information About a Microarray Experiment (MIAME), that describes the minimum information required to ensure that microarray data can be easily interpreted and that results derived from its analysis can be independently verified. The ultimate goal of this work is to establish a standard for recording and reporting microarray-based gene expression data, which will in turn facilitate the establishment of databases and public repositories and enable the development of data analysis tools. With respect to MIAME, we concentrate on defining the content and structure of the necessary information rather than the technical format for capturing it.
The MicroArray Quality Control (MAQC) project shows inter- and intraplatform reproducibility of gene expression measurements. MAQC Consortium (2006). Nature Biotechnology. doi:10.1038/nbt1239
Abstract
Over the last decade, the introduction of microarray technology has had a profound impact on gene expression research. The publication of studies with dissimilar or altogether contradictory results, obtained using different microarray platforms to analyze identical RNA samples, has raised concerns about the reliability of this technology. The MicroArray Quality Control (MAQC) project was initiated to address these concerns, as well as other performance and data analysis issues. Expression data on four titration pools from two distinct reference RNA samples were generated at multiple test sites using a variety of microarray-based and alternative technology platforms. Here we describe the experimental design and probe mapping efforts behind the MAQC project. We show intraplatform consistency across test sites as well as a high level of interplatform concordance in terms of genes identified as differentially expressed. This study provides a resource that represents an important first step toward establishing a framework for the use of microarrays in clinical and regulatory settings.
B.7.2 RNA-seq
Mortazavi et al. (2008), Nagalakshmi et al. (2008) — the first RNA-seq studies.
Mapping and quantifying mammalian transcriptomes by RNA-Seq. Mortazavi et al. (2008). Nature Methods. doi:10.1038/nmeth.1226
Abstract
We have mapped and quantified mouse transcriptomes by deeply sequencing them and recording how frequently each gene is represented in the sequence sample (RNA-Seq). This provides a digital measure of the presence and prevalence of transcripts from known and previously unknown genes. We report reference measurements composed of 41-52 million mapped 25-base-pair reads for poly(A)-selected RNA from adult mouse brain, liver and skeletal muscle tissues. We used RNA standards to quantify transcript prevalence and to test the linear range of transcript detection, which spanned five orders of magnitude. Although >90% of uniquely mapped reads fell within known exons, the remaining data suggest new and revised gene models, including changed or additional promoters, exons and 3' untranscribed regions, as well as new candidate microRNA precursors. RNA splice events, which are not readily measured by standard gene expression microarray or serial analysis of gene expression methods, were detected directly by mapping splice-crossing sequence reads. We observed 1.45 x 10(5) distinct splices, and alternative splices were prominent, with 3,500 different genes expressing one or more alternate internal splices.
The Transcriptional Landscape of the Yeast Genome Defined by RNA Sequencing. Nagalakshmi et al. (2008). Science. doi:10.1126/science.1158441
Abstract
The identification of untranslated regions, introns, and coding regions within an organism remains challenging. We developed a quantitative sequencing-based method called RNA-Seq for mapping transcribed regions, in which complementary DNA fragments are subjected to high-throughput sequencing and mapped to the genome. We applied RNA-Seq to generate a high-resolution transcriptome map of the yeast genome and demonstrated that most (74.5%) of the nonrepetitive sequence of the yeast genome is transcribed. We confirmed many known and predicted introns and demonstrated that others are not actively used. Alternative initiation codons and upstream open reading frames also were identified for many yeast genes. We also found unexpected 3'-end heterogeneity and the presence of many overlapping genes. These results indicate that the yeast transcriptome is more complex than previously appreciated.
Marioni et al. (2008) — RNA-seq compared with arrays.
RNA-seq: An assessment of technical reproducibility and comparison with gene expression arrays. Marioni et al. (2008). Genome Research. doi:10.1101/gr.079558.108
Abstract
Ultra-high-throughput sequencing is emerging as an attractive alternative to microarrays for genotyping, analysis of methylation patterns, and identification of transcription factor binding sites. Here, we describe an application of the Illumina sequencing (formerly Solexa sequencing) platform to study mRNA expression levels. Our goals were to estimate technical variance associated with Illumina sequencing in this context and to compare its ability to identify differentially expressed genes with existing array technologies. To do so, we estimated gene expression differences between liver and kidney RNA samples using multiple sequencing replicates, and compared the sequencing data to results obtained from Affymetrix arrays using the same RNA samples. We find that the Illumina sequencing data are highly replicable, with relatively little technical variation, and thus, for many purposes, it may suffice to sequence each mRNA sample only once (i.e., using one lane). The information in a single lane of Illumina sequencing data appears comparable to that in a single array in enabling identification of differentially expressed genes, while allowing for additional analyses such as detection of low-expressed genes, alternative splice variants, and novel transcripts. Based on our observations, we propose an empirical protocol and a statistical framework for the analysis of gene expression using ultra-high-throughput sequencing technology.
Stark et al. (2019) — RNA sequencing: the teenage years.
RNA sequencing: the teenage years. Stark et al. (2019). Nature Reviews Genetics. doi:10.1038/s41576-019-0150-2
Abstract
Over the past decade, RNA sequencing (RNA-seq) has become an indispensable tool for transcriptome-wide analysis of differential gene expression and differential splicing of mRNAs. However, as next-generation sequencing technologies have developed, so too has RNA-seq. Now, RNA-seq methods are available for studying many different aspects of RNA biology, including single-cell gene expression, translation (the translatome) and RNA structure (the structurome). Exciting new applications are being explored, such as spatial transcriptomics (spatialomics). Together with new long-read and direct RNA-seq technologies and better computational tools for data analysis, innovations in RNA-seq are contributing to a fuller understanding of RNA biology, from questions such as when and where transcription occurs to the folding and intermolecular interactions that govern RNA function.
Trapnell et al. (2009), Dobin et al. (2012), Kim et al. (2019) — spliced aligners (TopHat, STAR, HISAT2).
TopHat: discovering splice junctions with RNA-Seq. Trapnell et al. (2009). Bioinformatics. doi:10.1093/bioinformatics/btp120
Abstract
Motivation A new protocol for sequencing the messenger RNA in a cell, known as RNA-Seq, generates millions of short sequence fragments in a single run. These fragments, or 'reads', can be used to measure levels of gene expression and to identify novel splice variants of genes. However, current software for aligning RNA-Seq data to a genome relies on known splice junctions and cannot identify novel ones. TopHat is an efficient read-mapping algorithm designed to align reads from an RNA-Seq experiment to a reference genome without relying on known splice sites. Results We mapped the RNA-Seq reads from a recent mammalian RNA-Seq experiment and recovered more than 72% of the splice junctions reported by the annotation-based software from that study, along with nearly 20,000 previously unreported junctions. The TopHat pipeline is much faster than previous systems, mapping nearly 2.2 million reads per CPU hour, which is sufficient to process an entire RNA-Seq experiment in less than a day on a standard desktop computer. We describe several challenges unique to ab initio splice site discovery from RNA-Seq reads that will require further algorithm development. Availability TopHat is free, open-source software available from http://tophat.cbcb.umd.edu. Supplementary information Supplementary data are available at Bioinformatics online.
STAR: ultrafast universal RNA-seq aligner. Dobin et al. (2012). Bioinformatics. doi:10.1093/bioinformatics/bts635
Abstract
Motivation Accurate alignment of high-throughput RNA-seq data is a challenging and yet unsolved problem because of the non-contiguous transcript structure, relatively short read lengths and constantly increasing throughput of the sequencing technologies. Currently available RNA-seq aligners suffer from high mapping error rates, low mapping speed, read length limitation and mapping biases. Results To align our large (>80 billon reads) ENCODE Transcriptome RNA-seq dataset, we developed the Spliced Transcripts Alignment to a Reference (STAR) software based on a previously undescribed RNA-seq alignment algorithm that uses sequential maximum mappable seed search in uncompressed suffix arrays followed by seed clustering and stitching procedure. STAR outperforms other aligners by a factor of >50 in mapping speed, aligning to the human genome 550 million 2 × 76 bp paired-end reads per hour on a modest 12-core server, while at the same time improving alignment sensitivity and precision. In addition to unbiased de novo detection of canonical junctions, STAR can discover non-canonical splices and chimeric (fusion) transcripts, and is also capable of mapping full-length RNA sequences. Using Roche 454 sequencing of reverse transcription polymerase chain reaction amplicons, we experimentally validated 1960 novel intergenic splice junctions with an 80-90% success rate, corroborating the high precision of the STAR mapping strategy. Availability and implementation STAR is implemented as a standalone C++ code. STAR is free open source software distributed under GPLv3 license and can be downloaded from http://code.google.com/p/rna-star/.
Graph-based genome alignment and genotyping with HISAT2 and HISAT-genotype. Kim et al. (2019). Nature Biotechnology. doi:10.1038/s41587-019-0201-4
Abstract
The human reference genome represents only a small number of individuals, which limits its usefulness for genotyping. We present a method named HISAT2 (hierarchical indexing for spliced alignment of transcripts 2) that can align both DNA and RNA sequences using a graph Ferragina Manzini index. We use HISAT2 to represent and search an expanded model of the human reference genome in which over 14.5 million genomic variants in combination with haplotypes are incorporated into the data structure used for searching and alignment. We benchmark HISAT2 using simulated and real datasets to demonstrate that our strategy of representing a population of genomes, together with a fast, memory-efficient search algorithm, provides more detailed and accurate variant analyses than other methods. We apply HISAT2 for HLA typing and DNA fingerprinting; both applications form part of the HISAT-genotype software that enables analysis of haplotype-resolved genes or genomic regions. HISAT-genotype outperforms other computational methods and matches or exceeds the performance of laboratory-based assays.
Trapnell et al. (2010), Pertea et al. (2015), Grabherr et al. (2011) — transcript assembly (Cufflinks, StringTie, Trinity).
Transcript assembly and quantification by RNA-Seq reveals unannotated transcripts and isoform switching during cell differentiation. Trapnell et al. (2010). Nature Biotechnology. doi:10.1038/nbt.1621
Abstract
High-throughput mRNA sequencing (RNA-Seq) promises simultaneous transcript discovery and abundance estimation. However, this would require algorithms that are not restricted by prior gene annotations and that account for alternative transcription and splicing. Here we introduce such algorithms in an open-source software program called Cufflinks. To test Cufflinks, we sequenced and analyzed >430 million paired 75-bp RNA-Seq reads from a mouse myoblast cell line over a differentiation time series. We detected 13,692 known transcripts and 3,724 previously unannotated ones, 62% of which are supported by independent expression data or by homologous genes in other species. Over the time series, 330 genes showed complete switches in the dominant transcription start site (TSS) or splice isoform, and we observed more subtle shifts in 1,304 other genes. These results suggest that Cufflinks can illuminate the substantial regulatory flexibility and complexity in even this well-studied model of muscle development and that it can improve transcriptome-based genome annotation.
StringTie enables improved reconstruction of a transcriptome from RNA-seq reads. Pertea et al. (2015). Nature Biotechnology. doi:10.1038/nbt.3122
Abstract
Methods used to sequence the transcriptome often produce more than 200 million short sequences. We introduce StringTie, a computational method that applies a network flow algorithm originally developed in optimization theory, together with optional de novo assembly, to assemble these complex data sets into transcripts. When used to analyze both simulated and real data sets, StringTie produces more complete and accurate reconstructions of genes and better estimates of expression levels, compared with other leading transcript assembly programs including Cufflinks, IsoLasso, Scripture and Traph. For example, on 90 million reads from human blood, StringTie correctly assembled 10,990 transcripts, whereas the next best assembly was of 7,187 transcripts by Cufflinks, which is a 53% increase in transcripts assembled. On a simulated data set, StringTie correctly assembled 7,559 transcripts, which is 20% more than the 6,310 assembled by Cufflinks. As well as producing a more complete transcriptome assembly, StringTie runs faster on all data sets tested to date compared with other assembly software, including Cufflinks.
Full-length transcriptome assembly from RNA-Seq data without a reference genome. Grabherr et al. (2011). Nature Biotechnology. doi:10.1038/nbt.1883
Abstract
Massively parallel sequencing of cDNA has enabled deep and efficient probing of transcriptomes. Current approaches for transcript reconstruction from such data often rely on aligning reads to a reference genome, and are thus unsuitable for samples with a partial or missing reference genome. Here we present the Trinity method for de novo assembly of full-length transcripts and evaluate it on samples from fission yeast, mouse and whitefly, whose reference genome is not yet available. By efficiently constructing and analyzing sets of de Bruijn graphs, Trinity fully reconstructs a large fraction of transcripts, including alternatively spliced isoforms and transcripts from recently duplicated genes. Compared with other de novo transcriptome assemblers, Trinity recovers more full-length transcripts across a broad range of expression levels, with a sensitivity similar to methods that rely on genome alignments. Our approach provides a unified solution for transcriptome reconstruction in any sample, especially in the absence of a reference genome.
Li and Dewey (2011) — RSEM and the EM algorithm.
RSEM: accurate transcript quantification from RNA-Seq data with or without a reference genome. Li & Dewey (2011). BMC Bioinformatics. doi:10.1186/1471-2105-12-323
Abstract
Background RNA-Seq is revolutionizing the way transcript abundances are measured. A key challenge in transcript quantification from RNA-Seq data is the handling of reads that map to multiple genes or isoforms. This issue is particularly important for quantification with de novo transcriptome assemblies in the absence of sequenced genomes, as it is difficult to determine which transcripts are isoforms of the same gene. A second significant issue is the design of RNA-Seq experiments, in terms of the number of reads, read length, and whether reads come from one or both ends of cDNA fragments. Results We present RSEM, an user-friendly software package for quantifying gene and isoform abundances from single-end or paired-end RNA-Seq data. RSEM outputs abundance estimates, 95% credibility intervals, and visualization files and can also simulate RNA-Seq data. In contrast to other existing tools, the software does not require a reference genome. Thus, in combination with a de novo transcriptome assembler, RSEM enables accurate transcript quantification for species without sequenced genomes. On simulated and real data sets, RSEM has superior or comparable performance to quantification methods that rely on a reference genome. Taking advantage of RSEM's ability to effectively use ambiguously-mapping reads, we show that accurate gene-level abundance estimates are best obtained with large numbers of short single-end reads. On the other hand, estimates of the relative frequencies of isoforms within single genes may be improved through the use of paired-end reads, depending on the number of possible splice forms for each gene. Conclusions RSEM is an accurate and user-friendly software tool for quantifying transcript abundances from RNA-Seq data. As it does not rely on the existence of a reference genome, it is particularly useful for quantification with de novo transcriptome assemblies. In addition, RSEM has enabled valuable guidance for cost-efficient design of quantification experiments with RNA-Seq, which is currently relatively expensive.
N. L. Bray et al. (2016), Patro et al. (2017) — kallisto and Salmon.
Near-optimal probabilistic RNA-seq quantification. Bray et al. (2016). Nature Biotechnology. doi:10.1038/nbt.3519
Abstract
We present kallisto, an RNA-seq quantification program that is two orders of magnitude faster than previous approaches and achieves similar accuracy. Kallisto pseudoaligns reads to a reference, producing a list of transcripts that are compatible with each read while avoiding alignment of individual bases. We use kallisto to analyze 30 million unaligned paired-end RNA-seq reads in <10 min on a standard laptop computer. This removes a major computational bottleneck in RNA-seq analysis.
Salmon provides fast and bias-aware quantification of transcript expression. Patro et al. (2017). Nature Methods. doi:10.1038/nmeth.4197
Abstract
We introduce Salmon, a lightweight method for quantifying transcript abundance from RNA-seq reads. Salmon combines a new dual-phase parallel inference algorithm and feature-rich bias models with an ultra-fast read mapping procedure. It is the first transcriptome-wide quantifier to correct for fragment GC-content bias, which, as we demonstrate here, substantially improves the accuracy of abundance estimates and the sensitivity of subsequent differential expression analysis.
Wagner et al. (2012) — why TPM rather than RPKM.
Measurement of mRNA abundance using RNA-seq data: RPKM measure is inconsistent among samples. Wagner et al. (2012). Theory in Biosciences. doi:10.1007/s12064-012-0162-3
Abstract
Measures of RNA abundance are important for many areas of biology and often obtained from high-throughput RNA sequencing methods such as Illumina sequence data. These measures need to be normalized to remove technical biases inherent in the sequencing approach, most notably the length of the RNA species and the sequencing depth of a sample. These biases are corrected in the widely used reads per kilobase per million reads (RPKM) measure. Here, we argue that the intended meaning of RPKM is a measure of relative molar RNA concentration (rmc) and show that for each set of transcripts the average rmc is a constant, namely the inverse of the number of transcripts mapped. Further, we show that RPKM does not respect this invariance property and thus cannot be an accurate measure of rmc. We propose a slight modification of RPKM that eliminates this inconsistency and call it TPM for transcripts per million. TPM respects the average invariance and eliminates statistical biases inherent in the RPKM measure.
Schurch et al. (2016) — how many replicates are needed.
How many biological replicates are needed in an RNA-seq experiment and which differential expression tool should you use?. Schurch et al. (2016). RNA. doi:10.1261/rna.053959.115
Abstract
RNA-seq is now the technology of choice for genome-wide differential gene expression experiments, but it is not clear how many biological replicates are needed to ensure valid biological interpretation of the results or which statistical tools are best for analyzing the data. An RNA-seq experiment with 48 biological replicates in each of two conditions was performed to answer these questions and provide guidelines for experimental design. With three biological replicates, nine of the 11 tools evaluated found only 20%-40% of the significantly differentially expressed (SDE) genes identified with the full set of 42 clean replicates. This rises to >85% for the subset of SDE genes changing in expression by more than fourfold. To achieve >85% for all SDE genes regardless of fold change requires more than 20 biological replicates. The same nine tools successfully control their false discovery rate at ≲5% for all numbers of replicates, while the remaining two tools fail to control their FDR adequately, particularly for low numbers of replicates. For future RNA-seq experiments, these results suggest that at least six biological replicates should be used, rising to at least 12 when it is important to identify SDE genes for all fold changes. If fewer than 12 replicates are used, a superior combination of true positive and false positive performances makes edgeR and DESeq2 the leading tools. For higher replicate numbers, minimizing false positives is more important and DESeq marginally outperforms the other tools.
Leek et al. (2010) — batch effects.
Tackling the widespread and critical impact of batch effects in high-throughput data. Leek et al. (2010). Nature Reviews Genetics. doi:10.1038/nrg2825
Abstract
High-throughput technologies are widely used, for example to assay genetic variants, gene and protein expression, and epigenetic modifications. One often overlooked complication with such studies is batch effects, which occur because measurements are affected by laboratory conditions, reagent lots and personnel differences. This becomes a major problem when batch effects are correlated with an outcome of interest and lead to incorrect conclusions. Using both published studies and our own analyses, we argue that batch effects (as well as other technical and biological artefacts) are widespread and critical to address. We review experimental and computational approaches for doing so.
B.7.3 Statistics and interpretation
Oshlack et al. (2010) — from RNA-seq reads to differential expression.
From RNA-seq reads to differential expression results. Oshlack et al. (2010). Genome Biology. doi:10.1186/gb-2010-11-12-220
Abstract
Many methods and tools are available for preprocessing high-throughput RNA sequencing data and detecting differential expression.
Robinson et al. (2009), Anders and Huber (2010) — edgeR and DESeq.
edgeR : a Bioconductor package for differential expression analysis of digital gene expression data. Robinson et al. (2009). Bioinformatics. doi:10.1093/bioinformatics/btp616
Abstract
It is expected that emerging digital gene expression (DGE) technologies will overtake microarray technologies in the near future for many functional genomics applications. One of the fundamental data analysis tasks, especially for gene expression studies, involves determining whether there is evidence that counts for a transcript or exon are significantly different across experimental conditions. edgeR is a Bioconductor software package for examining differential expression of replicated count data. An overdispersed Poisson model is used to account for both biological and technical variability. Empirical Bayes methods are used to moderate the degree of overdispersion across transcripts, improving the reliability of inference. The methodology can be used even with the most minimal levels of replication, provided at least one phenotype or experimental condition is replicated. The software may have other applications beyond sequencing data, such as proteome peptide count data. Availability The package is freely available under the LGPL licence from the Bioconductor web site (http://bioconductor.org).
Differential expression analysis for sequence count data. Anders & Huber (2010). Genome Biology. doi:10.1186/gb-2010-11-10-r106
Abstract
High-throughput sequencing assays such as RNA-Seq, ChIP-Seq or barcode counting provide quantitative readouts in the form of count data. To infer differential signal in such data correctly and with good statistical power, estimation of data variability throughout the dynamic range and a suitable error model are required. We propose a method based on the negative binomial distribution, with variance and mean linked by local regression and present an implementation, DESeq, as an R/Bioconductor package.
Law et al. (2014) — voom.
voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Law et al. (2014). Genome Biology. doi:10.1186/gb-2014-15-2-r29
Abstract
New normal linear modeling strategies are presented for analyzing read counts from RNA-seq experiments. The voom method estimates the mean-variance relationship of the log-counts, generates a precision weight for each observation and enters these into the limma empirical Bayes analysis pipeline. This opens access for RNA-seq analysts to a large body of methodology developed for microarrays. Simulation studies show that voom performs as well or better than count-based RNA-seq methods even when the data are generated according to the assumptions of the earlier methods. Two case studies illustrate the use of linear modeling and gene set testing methods.
Storey and Tibshirani (2003), Efron and Tibshirani (2002) — q-values and local FDR.
Statistical significance for genomewide studies. Storey & Tibshirani (2003). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1530509100
Abstract
With the increase in genomewide experiments and the sequencing of multiple genomes, the analysis of large data sets has become commonplace in biology. It is often the case that thousands of features in a genomewide data set are tested against some null hypothesis, where a number of features are expected to be significant. Here we propose an approach to measuring statistical significance in these genomewide studies based on the concept of the false discovery rate. This approach offers a sensible balance between the number of true and false positives that is automatically calibrated and easily interpreted. In doing so, a measure of statistical significance called the q value is associated with each tested feature. The q value is similar to the well known p value, except it is a measure of significance in terms of the false discovery rate rather than the false positive rate. Our approach avoids a flood of false positive results, while offering a more liberal criterion than what has been used in genome scans for linkage.
Empirical bayes methods and false discovery rates for microarrays. Efron & Tibshirani (2002). Genetic Epidemiology. doi:10.1002/gepi.1124
Abstract
In a classic two-sample problem, one might use Wilcoxon's statistic to test for a difference between treatment and control subjects. The analogous microarray experiment yields thousands of Wilcoxon statistics, one for each gene on the array, and confronts the statistician with a difficult simultaneous inference situation. We will discuss two inferential approaches to this problem: an empirical Bayes method that requires very little a priori Bayesian modeling, and the frequentist method of "false discovery rates" proposed by Benjamini and Hochberg in 1995. It turns out that the two methods are closely related and can be used together to produce sensible simultaneous inferences.
Krzywinski and Altman (2013), Nuzzo (2014), Wasserstein and Lazar (2016) — the importance of being uncertain; p-values and their misuse.
Significance, P values and t-tests. Krzywinski & Altman (2013). Nature Methods. doi:10.1038/nmeth.2698
Scientific method: Statistical errors. Nuzzo (2014). Nature. doi:10.1038/506150a
The ASA Statement on p -Values: Context, Process, and Purpose. Wasserstein & Lazar (2016). The American Statistician. doi:10.1080/00031305.2016.1154108
Noble (2009) — how does multiple testing correction work?
How does multiple testing correction work?. Noble (2009). Nature Biotechnology. doi:10.1038/nbt1209-1135
Holm (1979) — Holm’s step-down procedure.
A simple sequentially rejective multiple test procedure. Holm (1979). Scandinavian Journal of Statistics.
Ioannidis (2005) — why most published research findings are false.
Why Most Published Research Findings Are False. Ioannidis (2005). PLoS Medicine. doi:10.1371/journal.pmed.0020124
Abstract
There is increasing concern that most current published research findings are false. The probability that a research claim is true may depend on study power and bias, the number of other studies on the same question, and, importantly, the ratio of true to no relationships among the relationships probed in each scientific field. In this framework, a research finding is less likely to be true when the studies conducted in a field are smaller; when effect sizes are smaller; when there is a greater number and lesser preselection of tested relationships; where there is greater flexibility in designs, definitions, outcomes, and analytical modes; when there is greater financial and other interest and prejudice; and when more teams are involved in a scientific field in chase of statistical significance. Simulations show that for most study designs and settings, it is more likely for a research claim to be false than true. Moreover, for many current scientific fields, claimed research findings may often be simply accurate measures of the prevailing bias. In this essay, I discuss the implications of these problems for the conduct and interpretation of research.
Lever et al. (2017) — principal component analysis.
Principal component analysis. Lever et al. (2017). Nature Methods. doi:10.1038/nmeth.4346
Maaten and Hinton (2008), McInnes et al. (2018), Becht et al. (2018) — t-SNE and UMAP.
Visualizing Data using t-SNE. van der Maaten & Hinton (2008). Journal of Machine Learning Research. https://www.jmlr.org/papers/v9/vandermaaten08a.html
UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. McInnes et al. (2018). https://arxiv.org/abs/1802.03426
Dimensionality reduction for visualizing single-cell data using UMAP. Becht et al. (2018). Nature Biotechnology. doi:10.1038/nbt.4314
Abstract
Advances in single-cell technologies have enabled high-resolution dissection of tissue composition. Several tools for dimensionality reduction are available to analyze the large number of parameters generated in single-cell studies. Recently, a nonlinear dimensionality-reduction technique, uniform manifold approximation and projection (UMAP), was developed for the analysis of any type of high-dimensional data. Here we apply it to biological data, using three well-characterized mass cytometry and single-cell RNA sequencing datasets. Comparing the performance of UMAP with five other tools, we find that UMAP provides the fastest run times, highest reproducibility and the most meaningful organization of cell clusters. The work highlights the use of UMAP for improved visualization and interpretation of single-cell data.
Chari and Pachter (2023) — the specious art of single-cell genomics.
The specious art of single-cell genomics. Chari & Pachter (2023). PLOS Computational Biology. doi:10.1371/journal.pcbi.1011288
Abstract
Dimensionality reduction is standard practice for filtering noise and identifying relevant features in large-scale data analyses. In biology, single-cell genomics studies typically begin with reduction to 2 or 3 dimensions to produce "all-in-one" visuals of the data that are amenable to the human eye, and these are subsequently used for qualitative and quantitative exploratory analysis. However, there is little theoretical support for this practice, and we show that extreme dimension reduction, from hundreds or thousands of dimensions to 2, inevitably induces significant distortion of high-dimensional datasets. We therefore examine the practical implications of low-dimensional embedding of single-cell data and find that extensive distortions and inconsistent practices make such embeddings counter-productive for exploratory, biological analyses. In lieu of this, we discuss alternative approaches for conducting targeted embedding and feature exploration to enable hypothesis-driven biological discovery.
Subramanian et al. (2005) — gene set enrichment analysis.
Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Subramanian et al. (2005). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.0506580102
Abstract
Although genomewide RNA expression analysis has become a routine tool in biomedical research, extracting biological insight from such information remains a major challenge. Here, we describe a powerful analytical method called Gene Set Enrichment Analysis (GSEA) for interpreting gene expression data. The method derives its power by focusing on gene sets, that is, groups of genes that share common biological function, chromosomal location, or regulation. We demonstrate how GSEA yields insights into several cancer-related data sets, including leukemia and lung cancer. Notably, where single-gene analysis finds little similarity between two independent studies of patient survival in lung cancer, GSEA reveals many biological pathways in common. The GSEA method is embodied in a freely available software package, together with an initial database of 1,325 biologically defined gene sets.
Huang et al. (2008), Liberzon et al. (2015) — DAVID and MSigDB.
Systematic and integrative analysis of large gene lists using DAVID bioinformatics resources. Huang et al. (2008). Nature Protocols. doi:10.1038/nprot.2008.211
Abstract
DAVID bioinformatics resources consists of an integrated biological knowledgebase and analytic tools aimed at systematically extracting biological meaning from large gene/protein lists. This protocol explains how to use DAVID, a high-throughput and integrated data-mining environment, to analyze gene lists derived from high-throughput genomic experiments. The procedure first requires uploading a gene list containing any number of common gene identifiers followed by analysis using one or more text and pathway-mining tools such as gene functional classification, functional annotation chart or clustering and functional annotation table. By following this protocol, investigators are able to gain an in-depth understanding of the biological themes in lists of genes that are enriched in genome-scale studies.
The Molecular Signatures Database Hallmark Gene Set Collection. Liberzon et al. (2015). Cell Systems. doi:10.1016/j.cels.2015.12.004
Abstract
The Molecular Signatures Database (MSigDB) is one of the most widely used and comprehensive databases of gene sets for performing gene set enrichment analysis. Since its creation, MSigDB has grown beyond its roots in metabolic disease and cancer to include >10,000 gene sets. These better represent a wider range of biological processes and diseases, but the utility of the database is reduced by increased redundancy across, and heterogeneity within, gene sets. To address this challenge, here we use a combination of automated approaches and expert curation to develop a collection of "hallmark" gene sets as part of MSigDB. Each hallmark in this collection consists of a "refined" gene set, derived from multiple "founder" sets, that conveys a specific biological state or process and displays coherent expression. The hallmarks effectively summarize most of the relevant information of the original founder sets and, by reducing both variation and redundancy, provide more refined and concise inputs for gene set enrichment analysis.
Langfelder and Horvath (2008) — WGCNA co-expression networks.
WGCNA: an R package for weighted correlation network analysis. Langfelder & Horvath (2008). BMC Bioinformatics. doi:10.1186/1471-2105-9-559
Abstract
Background Correlation networks are increasingly being used in bioinformatics applications. For example, weighted gene co-expression network analysis is a systems biology method for describing the correlation patterns among genes across microarray samples. Weighted correlation network analysis (WGCNA) can be used for finding clusters (modules) of highly correlated genes, for summarizing such clusters using the module eigengene or an intramodular hub gene, for relating modules to one another and to external sample traits (using eigengene network methodology), and for calculating module membership measures. Correlation networks facilitate network based gene screening methods that can be used to identify candidate biomarkers or therapeutic targets. These methods have been successfully applied in various biological contexts, e.g. cancer, mouse genetics, yeast genetics, and analysis of brain imaging data. While parts of the correlation network methodology have been described in separate publications, there is a need to provide a user-friendly, comprehensive, and consistent software implementation and an accompanying tutorial. Results The WGCNA R software package is a comprehensive collection of R functions for performing various aspects of weighted correlation network analysis. The package includes functions for network construction, module detection, gene selection, calculations of topological properties, data simulation, visualization, and interfacing with external software. Along with the R package we also present R software tutorials. While the methods development was motivated by gene expression data, the underlying data mining approach can be applied to a variety of different settings. Conclusion The WGCNA package provides R functions for weighted correlation network analysis, e.g. co-expression network analysis of gene expression data. The R package along with its source code and additional material are freely available at http://www.genetics.ucla.edu/labs/horvath/CoexpressionNetwork/Rpackages/WGCNA.
The GTEx Consortium (2020) — GTEx: genetic effects on expression across tissues.
The GTEx Consortium atlas of genetic regulatory effects across human tissues. The GTEx Consortium (2020). Science. doi:10.1126/science.aaz1776
Abstract
The Genotype-Tissue Expression (GTEx) project was established to characterize genetic effects on the transcriptome across human tissues and to link these regulatory mechanisms to trait and disease associations. Here, we present analyses of the version 8 data, examining 15,201 RNA-sequencing samples from 49 tissues of 838 postmortem donors. We comprehensively characterize genetic associations for gene expression and splicing in cis and trans, showing that regulatory associations are found for almost all genes, and describe the underlying molecular mechanisms and their contribution to allelic heterogeneity and pleiotropy of complex traits. Leveraging the large diversity of tissues, we provide insights into the tissue specificity of genetic effects and show that cell type composition is a key factor in understanding gene regulatory mechanisms in human tissues.
Schwanhäusser et al. (2011), Liu et al. (2016) — how mRNA levels relate to protein levels.
Global quantification of mammalian gene expression control. Schwanhäusser et al. (2011). Nature. doi:10.1038/nature10098
Abstract
Gene expression is a multistep process that involves the transcription, translation and turnover of messenger RNAs and proteins. Although it is one of the most fundamental processes of life, the entire cascade has never been quantified on a genome-wide scale. Here we simultaneously measured absolute mRNA and protein abundance and turnover by parallel metabolic pulse labelling for more than 5,000 genes in mammalian cells. Whereas mRNA and protein levels correlated better than previously thought, corresponding half-lives showed no correlation. Using a quantitative model we have obtained the first genome-scale prediction of synthesis rates of mRNAs and proteins. We find that the cellular abundance of proteins is predominantly controlled at the level of translation. Genes with similar combinations of mRNA and protein stability shared functional properties, indicating that half-lives evolved under energetic and dynamic constraints. Quantitative information about all stages of gene expression provides a rich resource and helps to provide a greater understanding of the underlying design principles.
On the Dependency of Cellular Protein Levels on mRNA Abundance. Liu et al. (2016). Cell. doi:10.1016/j.cell.2016.03.014
Abstract
The question of how genomic information is expressed to determine phenotypes is of central importance for basic and translational life science research and has been studied by transcriptomic and proteomic profiling. Here, we review the relationship between protein and mRNA levels under various scenarios, such as steady state, long-term state changes, and short-term adaptation, demonstrating the complexity of gene expression regulation, especially during dynamic transitions. The spatial and temporal variations of mRNAs, as well as the local availability of resources for protein biosynthesis, strongly influence the relationship between protein levels and their coding transcripts. We further discuss the buffering of mRNA fluctuations at the level of protein concentrations. We conclude that transcript levels by themselves are not sufficient to predict protein levels in many scenarios and to thus explain genotype-phenotype relationships and that high-quality data quantifying different levels of gene expression are indispensable for the complete understanding of biological processes.
Uhlén et al. (2015) — the Human Protein Atlas.
Tissue-based map of the human proteome. Uhlén et al. (2015). Science. doi:10.1126/science.1260419
Abstract
Resolving the molecular details of proteome variation in the different tissues and organs of the human body will greatly increase our knowledge of human biology and disease. Here, we present a map of the human tissue proteome based on an integrated omics approach that involves quantitative transcriptomics at the tissue and organ level, combined with tissue microarray-based immunohistochemistry, to achieve spatial localization of proteins down to the single-cell level. Our tissue-based analysis detected more than 90% of the putative protein-coding genes. We used this approach to explore the human secretome, the membrane proteome, the druggable proteome, the cancer proteome, and the metabolic functions in 32 different tissues and organs. All the data are integrated in an interactive Web-based database that allows exploration of individual proteins, as well as navigation of global expression patterns, in all major tissues and organs in the human body.
B.7.4 Single-cell and spatial genomics
Heumos et al. (2023) — best practices for single-cell analysis across modalities.
Best practices for single-cell analysis across modalities. Heumos et al. (2023). Nature Reviews Genetics. doi:10.1038/s41576-023-00586-w
Abstract
Recent advances in single-cell technologies have enabled high-throughput molecular profiling of cells across modalities and locations. Single-cell transcriptomics data can now be complemented by chromatin accessibility, surface protein expression, adaptive immune receptor repertoire profiling and spatial information. The increasing availability of single-cell data across modalities has motivated the development of novel computational methods to help analysts derive biological insights. As the field grows, it becomes increasingly difficult to navigate the vast landscape of tools and analysis steps. Here, we summarize independent benchmarking studies of unimodal and multimodal single-cell analysis across modalities to suggest comprehensive best-practice workflows for the most common analysis steps. Where independent benchmarks are not available, we review and contrast popular methods. Our article serves as an entry point for novices in the field of single-cell (multi-)omic analysis and guides advanced users to the most recent best practices.
Tang et al. (2009) — the first single-cell transcriptome.
mRNA-Seq whole-transcriptome analysis of a single cell. Tang et al. (2009). Nature Methods. doi:10.1038/nmeth.1315
Abstract
Next-generation sequencing technology is a powerful tool for transcriptome analysis. However, under certain conditions, only a small amount of material is available, which requires more sensitive techniques that can preferably be used at the single-cell level. Here we describe a single-cell digital gene expression profiling assay. Using our mRNA-Seq assay with only a single mouse blastomere, we detected the expression of 75% (5,270) more genes than microarray techniques and identified 1,753 previously unknown splice junctions called by at least 5 reads. Moreover, 8-19% of the genes with multiple known transcript isoforms expressed at least two isoforms in the same blastomere or oocyte, which unambiguously demonstrated the complexity of the transcript variants at whole-genome scale in individual cells. Finally, for Dicer1(-/-) and Ago2(-/-) (Eif2c2(-/-)) oocytes, we found that 1,696 and 1,553 genes, respectively, were abnormally upregulated compared to wild-type controls, with 619 genes in common.
Svensson et al. (2018) — exponential scaling of single-cell RNA-seq.
Exponential scaling of single-cell RNA-seq in the past decade. Svensson et al. (2018). Nature Protocols. doi:10.1038/nprot.2017.149
Abstract
Measurement of the transcriptomes of single cells has been feasible for only a few years, but it has become an extremely popular assay. While many types of analysis can be carried out and various questions can be answered by single-cell RNA-seq, a central focus is the ability to survey the diversity of cell types in a sample. Unbiased and reproducible cataloging of gene expression patterns in distinct cell types requires large numbers of cells. Technological developments and protocol improvements have fueled consistent and exponential increases in the number of cells that can be studied in single-cell RNA-seq analyses. In this Perspective, we highlight the key technological developments that have enabled this growth in the data obtained from single-cell RNA-seq experiments.
Picelli et al. (2013), Macosko et al. (2015), Klein et al. (2015), Zheng et al. (2017) — Smart-seq2 and droplet methods.
Smart-seq2 for sensitive full-length transcriptome profiling in single cells. Picelli et al. (2013). Nature Methods. doi:10.1038/nmeth.2639
Abstract
Single-cell gene expression analyses hold promise for characterizing cellular heterogeneity, but current methods compromise on either the coverage, the sensitivity or the throughput. Here, we introduce Smart-seq2 with improved reverse transcription, template switching and preamplification to increase both yield and length of cDNA libraries generated from individual cells. Smart-seq2 transcriptome libraries have improved detection, coverage, bias and accuracy compared to Smart-seq libraries and are generated with off-the-shelf reagents at lower cost.
Highly Parallel Genome-wide Expression Profiling of Individual Cells Using Nanoliter Droplets. Macosko et al. (2015). Cell. doi:10.1016/j.cell.2015.05.002
Abstract
Cells, the basic units of biological structure and function, vary broadly in type and state. Single-cell genomics can characterize cell identity and function, but limitations of ease and scale have prevented its broad application. Here we describe Drop-seq, a strategy for quickly profiling thousands of individual cells by separating them into nanoliter-sized aqueous droplets, associating a different barcode with each cell's RNAs, and sequencing them all together. Drop-seq analyzes mRNA transcripts from thousands of individual cells simultaneously while remembering transcripts' cell of origin. We analyzed transcriptomes from 44,808 mouse retinal cells and identified 39 transcriptionally distinct cell populations, creating a molecular atlas of gene expression for known retinal cell classes and novel candidate cell subtypes. Drop-seq will accelerate biological discovery by enabling routine transcriptional profiling at single-cell resolution. VIDEO ABSTRACT.
Droplet Barcoding for Single-Cell Transcriptomics Applied to Embryonic Stem Cells. Klein et al. (2015). Cell. doi:10.1016/j.cell.2015.04.044
Abstract
It has long been the dream of biologists to map gene expression at the single-cell level. With such data one might track heterogeneous cell sub-populations, and infer regulatory relationships between genes and pathways. Recently, RNA sequencing has achieved single-cell resolution. What is limiting is an effective way to routinely isolate and process large numbers of individual cells for quantitative in-depth sequencing. We have developed a high-throughput droplet-microfluidic approach for barcoding the RNA from thousands of individual cells for subsequent analysis by next-generation sequencing. The method shows a surprisingly low noise profile and is readily adaptable to other sequencing-based assays. We analyzed mouse embryonic stem cells, revealing in detail the population structure and the heterogeneous onset of differentiation after leukemia inhibitory factor (LIF) withdrawal. The reproducibility of these high-throughput single-cell data allowed us to deconstruct cell populations and infer gene expression relationships. VIDEO ABSTRACT.
Massively parallel digital transcriptional profiling of single cells. Zheng et al. (2017). Nature Communications. doi:10.1038/ncomms14049
Abstract
Characterizing the transcriptome of individual cells is fundamental to understanding complex biological systems. We describe a droplet-based system that enables 3' mRNA counting of tens of thousands of single cells per sample. Cell encapsulation, of up to 8 samples at a time, takes place in ∼6 min, with ∼50% cell capture efficiency. To demonstrate the system's technical performance, we collected transcriptome data from ∼250k single cells across 29 samples. We validated the sensitivity of the system and its ability to detect rare populations using cell lines and synthetic RNAs. We profiled 68k peripheral blood mononuclear cells to demonstrate the system's ability to characterize large immune populations. Finally, we used sequence variation in the transcriptome data to determine host and donor chimerism at single-cell resolution from bone marrow mononuclear cells isolated from transplant patients.
Islam et al. (2013) — unique molecular identifiers.
Quantitative single-cell RNA-seq with unique molecular identifiers. Islam et al. (2013). Nature Methods. doi:10.1038/nmeth.2772
Abstract
Single-cell RNA sequencing (RNA-seq) is a powerful tool to reveal cellular heterogeneity, discover new cell types and characterize tumor microevolution. However, losses in cDNA synthesis and bias in cDNA amplification lead to severe quantitative errors. We show that molecular labels--random sequences that label individual molecules--can nearly eliminate amplification noise, and that microfluidic sample preparation and optimized reagents produce a fivefold improvement in mRNA capture efficiency.
Svensson (2020) — droplet scRNA-seq is not zero-inflated.
Droplet scRNA-seq is not zero-inflated. Svensson (2020). Nature Biotechnology. doi:10.1038/s41587-019-0379-5
Lun et al. (2016), Hafemeister and Satija (2019) — normalisation of single-cell data.
Pooling across cells to normalize single-cell RNA sequencing data with many zero counts. Lun et al. (2016). Genome Biology. doi:10.1186/s13059-016-0947-7
Abstract
Normalization of single-cell RNA sequencing data is necessary to eliminate cell-specific biases prior to downstream analyses. However, this is not straightforward for noisy single-cell data where many counts are zero. We present a novel approach where expression values are summed across pools of cells, and the summed values are used for normalization. Pool-based size factors are then deconvolved to yield cell-based factors. Our deconvolution approach outperforms existing methods for accurate normalization of cell-specific biases in simulated data. Similar behavior is observed in real data, where deconvolution improves the relevance of results of downstream analyses.
Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression. Hafemeister & Satija (2019). Genome Biology. doi:10.1186/s13059-019-1874-1
Abstract
Single-cell RNA-seq (scRNA-seq) data exhibits significant cell-to-cell variation due to technical factors, including the number of molecules detected in each cell, which can confound biological heterogeneity with technical effects. To address this, we present a modeling framework for the normalization and variance stabilization of molecular count data from scRNA-seq experiments. We propose that the Pearson residuals from "regularized negative binomial regression," where cellular sequencing depth is utilized as a covariate in a generalized linear model, successfully remove the influence of technical characteristics from downstream analyses while preserving biological heterogeneity. Importantly, we show that an unconstrained negative binomial model may overfit scRNA-seq data, and overcome this by pooling information across genes with similar abundances to obtain stable parameter estimates. Our procedure omits the need for heuristic steps including pseudocount addition or log-transformation and improves common downstream analytical tasks such as variable gene selection, dimensional reduction, and differential expression. Our approach can be applied to any UMI-based scRNA-seq dataset and is freely available as part of the R package sctransform, with a direct interface to our single-cell toolkit Seurat.
Traag et al. (2019) — the Leiden algorithm.
From Louvain to Leiden: guaranteeing well-connected communities. Traag et al. (2019). Scientific Reports. doi:10.1038/s41598-019-41695-z
Abstract
Community detection is often used to understand the structure of large and complex networks. One of the most popular algorithms for uncovering community structure is the so-called Louvain algorithm. We show that this algorithm has a major defect that largely went unnoticed until now: the Louvain algorithm may yield arbitrarily badly connected communities. In the worst case, communities may even be disconnected, especially when running the algorithm iteratively. In our experimental analysis, we observe that up to 25% of the communities are badly connected and up to 16% are disconnected. To address this problem, we introduce the Leiden algorithm. We prove that the Leiden algorithm yields communities that are guaranteed to be connected. In addition, we prove that, when the Leiden algorithm is applied iteratively, it converges to a partition in which all subsets of all communities are locally optimally assigned. Furthermore, by relying on a fast local move approach, the Leiden algorithm runs faster than the Louvain algorithm. We demonstrate the performance of the Leiden algorithm for several benchmark and real-world networks. We find that the Leiden algorithm is faster than the Louvain algorithm and uncovers better partitions, in addition to providing explicit guarantees.
Hao et al. (2021), Wolf et al. (2018) — Seurat and Scanpy.
Integrated analysis of multimodal single-cell data. Hao et al. (2021). Cell. doi:10.1016/j.cell.2021.04.048
Abstract
The simultaneous measurement of multiple modalities represents an exciting frontier for single-cell genomics and necessitates computational methods that can define cellular states based on multimodal data. Here, we introduce "weighted-nearest neighbor" analysis, an unsupervised framework to learn the relative utility of each data type in each cell, enabling an integrative analysis of multiple modalities. We apply our procedure to a CITE-seq dataset of 211,000 human peripheral blood mononuclear cells (PBMCs) with panels extending to 228 antibodies to construct a multimodal reference atlas of the circulating immune system. Multimodal analysis substantially improves our ability to resolve cell states, allowing us to identify and validate previously unreported lymphoid subpopulations. Moreover, we demonstrate how to leverage this reference to rapidly map new datasets and to interpret immune responses to vaccination and coronavirus disease 2019 (COVID-19). Our approach represents a broadly applicable strategy to analyze single-cell multimodal datasets and to look beyond the transcriptome toward a unified and multimodal definition of cellular identity.
SCANPY: large-scale single-cell gene expression data analysis. Wolf et al. (2018). Genome Biology. doi:10.1186/s13059-017-1382-0
Abstract
SCANPY is a scalable toolkit for analyzing single-cell gene expression data. It includes methods for preprocessing, visualization, clustering, pseudotime and trajectory inference, differential expression testing, and simulation of gene regulatory networks. Its Python-based implementation efficiently deals with data sets of more than one million cells ( https://github.com/theislab/Scanpy ). Along with SCANPY, we present ANNDATA, a generic class for handling annotated data matrices ( https://github.com/theislab/anndata ).
Korsunsky et al. (2019), Lopez et al. (2018) — data integration (Harmony, scVI).
Fast, sensitive and accurate integration of single-cell data with Harmony. Korsunsky et al. (2019). Nature Methods. doi:10.1038/s41592-019-0619-0
Abstract
The emerging diversity of single-cell RNA-seq datasets allows for the full transcriptional characterization of cell types across a wide variety of biological and clinical conditions. However, it is challenging to analyze them together, particularly when datasets are assayed with different technologies, because biological and technical differences are interspersed. We present Harmony (https://github.com/immunogenomics/harmony), an algorithm that projects cells into a shared embedding in which cells group by cell type rather than dataset-specific conditions. Harmony simultaneously accounts for multiple experimental and biological factors. In six analyses, we demonstrate the superior performance of Harmony to previously published algorithms while requiring fewer computational resources. Harmony enables the integration of ~10 6 cells on a personal computer. We apply Harmony to peripheral blood mononuclear cells from datasets with large experimental differences, five studies of pancreatic islet cells, mouse embryogenesis datasets and the integration of scRNA-seq with spatial transcriptomics data.
Deep generative modeling for single-cell transcriptomics. Lopez et al. (2018). Nature Methods. doi:10.1038/s41592-018-0229-2
Abstract
Single-cell transcriptome measurements can reveal unexplored biological diversity, but they suffer from technical noise and bias that must be modeled to account for the resulting uncertainty in downstream analyses. Here we introduce single-cell variational inference (scVI), a ready-to-use scalable framework for the probabilistic representation and analysis of gene expression in single cells ( https://github.com/YosefLab/scVI ). scVI uses stochastic optimization and deep neural networks to aggregate information across similar cells and genes and to approximate the distributions that underlie observed expression values, while accounting for batch effects and limited sensitivity. We used scVI for a range of fundamental analysis tasks including batch correction, visualization, clustering, and differential expression, and achieved high accuracy for each task.
Trapnell et al. (2014), La Manno et al. (2018) — pseudotime and RNA velocity.
The dynamics and regulators of cell fate decisions are revealed by pseudotemporal ordering of single cells. Trapnell et al. (2014). Nature Biotechnology. doi:10.1038/nbt.2859
Abstract
Defining the transcriptional dynamics of a temporal process such as cell differentiation is challenging owing to the high variability in gene expression between individual cells. Time-series gene expression analyses of bulk cells have difficulty distinguishing early and late phases of a transcriptional cascade or identifying rare subpopulations of cells, and single-cell proteomic methods rely on a priori knowledge of key distinguishing markers. Here we describe Monocle, an unsupervised algorithm that increases the temporal resolution of transcriptome dynamics using single-cell RNA-Seq data collected at multiple time points. Applied to the differentiation of primary human myoblasts, Monocle revealed switch-like changes in expression of key regulatory factors, sequential waves of gene regulation, and expression of regulators that were not known to act in differentiation. We validated some of these predicted regulators in a loss-of function screen. Monocle can in principle be used to recover single-cell gene expression kinetics from a wide array of cellular processes, including differentiation, proliferation and oncogenic transformation.
RNA velocity of single cells. La Manno et al. (2018). Nature. doi:10.1038/s41586-018-0414-6
Abstract
RNA abundance is a powerful indicator of the state of individual cells. Single-cell RNA sequencing can reveal RNA abundance with high quantitative accuracy, sensitivity and throughput 1 . However, this approach captures only a static snapshot at a point in time, posing a challenge for the analysis of time-resolved phenomena such as embryogenesis or tissue regeneration. Here we show that RNA velocity-the time derivative of the gene expression state-can be directly estimated by distinguishing between unspliced and spliced mRNAs in common single-cell RNA sequencing protocols. RNA velocity is a high-dimensional vector that predicts the future state of individual cells on a timescale of hours. We validate its accuracy in the neural crest lineage, demonstrate its use on multiple published datasets and technical platforms, reveal the branching lineage tree of the developing mouse hippocampus, and examine the kinetics of transcription in human embryonic brain. We expect RNA velocity to greatly aid the analysis of developmental lineages and cellular dynamics, particularly in humans.
Ståhl et al. (2016), Rodriques et al. (2019), Chen et al. (2015), Marx (2021) — spatial transcriptomics.
Visualization and analysis of gene expression in tissue sections by spatial transcriptomics. Ståhl et al. (2016). Science. doi:10.1126/science.aaf2403
Abstract
Analysis of the pattern of proteins or messengerRNAs (mRNAs) in histological tissue sections is a cornerstone in biomedical research and diagnostics. This typically involves the visualization of a few proteins or expressed genes at a time. We have devised a strategy, which we call "spatial transcriptomics," that allows visualization and quantitative analysis of the transcriptome with spatial resolution in individual tissue sections. By positioning histological sections on arrayed reverse transcription primers with unique positional barcodes, we demonstrate high-quality RNA-sequencing data with maintained two-dimensional positional information from the mouse brain and human breast cancer. Spatial transcriptomics provides quantitative gene expression data and visualization of the distribution of mRNAs within tissue sections and enables novel types of bioinformatics analyses, valuable in research and diagnostics.
Slide-seq: A scalable technology for measuring genome-wide expression at high spatial resolution. Rodriques et al. (2019). Science. doi:10.1126/science.aaw1219
Abstract
Spatial positions of cells in tissues strongly influence function, yet a high-throughput, genome-wide readout of gene expression with cellular resolution is lacking. We developed Slide-seq, a method for transferring RNA from tissue sections onto a surface covered in DNA-barcoded beads with known positions, allowing the locations of the RNA to be inferred by sequencing. Using Slide-seq, we localized cell types identified by single-cell RNA sequencing datasets within the cerebellum and hippocampus, characterized spatial gene expression patterns in the Purkinje layer of mouse cerebellum, and defined the temporal evolution of cell type-specific responses in a mouse model of traumatic brain injury. These studies highlight how Slide-seq provides a scalable method for obtaining spatially resolved gene expression data at resolutions comparable to the sizes of individual cells.
Spatially resolved, highly multiplexed RNA profiling in single cells. Chen et al. (2015). Science. doi:10.1126/science.aaa6090
Abstract
Knowledge of the expression profile and spatial landscape of the transcriptome in individual cells is essential for understanding the rich repertoire of cellular behaviors. Here, we report multiplexed error-robust fluorescence in situ hybridization (MERFISH), a single-molecule imaging approach that allows the copy numbers and spatial localizations of thousands of RNA species to be determined in single cells. Using error-robust encoding schemes to combat single-molecule labeling and detection errors, we demonstrated the imaging of 100 to 1000 distinct RNA species in hundreds of individual cells. Correlation analysis of the ~10(4) to 10(6) pairs of genes allowed us to constrain gene regulatory networks, predict novel functions for many unannotated genes, and identify distinct spatial distribution patterns of RNAs that correlate with properties of the encoded proteins.
Method of the Year: spatially resolved transcriptomics. Marx (2021). Nature Methods. doi:10.1038/s41592-020-01033-y
Theodoris et al. (2023), Cui et al. (2024) — single-cell foundation models.
Transfer learning enables predictions in network biology. Theodoris et al. (2023). Nature. doi:10.1038/s41586-023-06139-9
Abstract
Mapping gene networks requires large amounts of transcriptomic data to learn the connections between genes, which impedes discoveries in settings with limited data, including rare diseases and diseases affecting clinically inaccessible tissues. Recently, transfer learning has revolutionized fields such as natural language understanding 1,2 and computer vision 3 by leveraging deep learning models pretrained on large-scale general datasets that can then be fine-tuned towards a vast array of downstream tasks with limited task-specific data. Here, we developed a context-aware, attention-based deep learning model, Geneformer, pretrained on a large-scale corpus of about 30 million single-cell transcriptomes to enable context-specific predictions in settings with limited data in network biology. During pretraining, Geneformer gained a fundamental understanding of network dynamics, encoding network hierarchy in the attention weights of the model in a completely self-supervised manner. Fine-tuning towards a diverse panel of downstream tasks relevant to chromatin and network dynamics using limited task-specific data demonstrated that Geneformer consistently boosted predictive accuracy. Applied to disease modelling with limited patient data, Geneformer identified candidate therapeutic targets for cardiomyopathy. Overall, Geneformer represents a pretrained deep learning model from which fine-tuning towards a broad range of downstream applications can be pursued to accelerate discovery of key network regulators and candidate therapeutic targets.
scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Cui et al. (2024). Nature Methods. doi:10.1038/s41592-024-02201-0
Abstract
Generative pretrained models have achieved remarkable success in various domains such as language and computer vision. Specifically, the combination of large-scale diverse datasets and pretrained transformers has emerged as a promising approach for developing foundation models. Drawing parallels between language and cellular biology (in which texts comprise words; similarly, cells are defined by genes), our study probes the applicability of foundation models to advance cellular biology and genetic research. Using burgeoning single-cell sequencing data, we have constructed a foundation model for single-cell biology, scGPT, based on a generative pretrained transformer across a repository of over 33 million cells. Our findings illustrate that scGPT effectively distills critical biological insights concerning genes and cells. Through further adaptation of transfer learning, scGPT can be optimized to achieve superior performance across diverse downstream applications. This includes tasks such as cell type annotation, multi-batch integration, multi-omic integration, perturbation response prediction and gene network inference.
Kedzierska et al. (2025), Ahlmann-Eltze et al. (2025) — evaluating foundation models against simple baselines.
Zero-shot evaluation reveals limitations of single-cell foundation models. Kedzierska et al. (2025). Genome Biology. doi:10.1186/s13059-025-03574-x
Abstract
Foundation models such as scGPT and Geneformer have not been rigorously evaluated in a setting where they are used without any further training (i.e., zero-shot). Understanding the performance of models in zero-shot settings is critical to applications that exclude the ability to fine-tune, such as discovery settings where labels are unknown. Our evaluation of the zero-shot performance of Geneformer and scGPT suggests that, in some cases, these models may face reliability challenges and could be outperformed by simpler methods. Our findings underscore the importance of zero-shot evaluations in development and deployment of foundation models in single-cell research.
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Ahlmann-Eltze et al. (2025). Nature Methods. doi:10.1038/s41592-025-02772-6
Abstract
Recent research in deep-learning-based foundation models promises to learn representations of single-cell data that enable prediction of the effects of genetic perturbations. Here we compared five foundation models and two other deep learning models against deliberately simple baselines for predicting transcriptome changes after single or double perturbations. None outperformed the baselines, which highlights the importance of critical benchmarking in directing and evaluating method development.
Regev et al. (2017), The Tabula Muris Consortium (2018), The Tabula Sapiens Consortium (2022) — cell atlases.
The Human Cell Atlas. Regev et al. (2017). eLife. doi:10.7554/elife.27041
Abstract
The recent advent of methods for high-throughput single-cell molecular profiling has catalyzed a growing sense in the scientific community that the time is ripe to complete the 150-year-old effort to identify all cell types in the human body. The Human Cell Atlas Project is an international collaborative effort that aims to define all human cell types in terms of distinctive molecular profiles (such as gene expression profiles) and to connect this information with classical cellular descriptions (such as location and morphology). An open comprehensive reference map of the molecular state of cells in healthy human tissues would propel the systematic study of physiological states, developmental trajectories, regulatory circuitry and interactions of cells, and also provide a framework for understanding cellular dysregulation in human disease. Here we describe the idea, its potential utility, early proofs-of-concept, and some design considerations for the Human Cell Atlas, including a commitment to open data, code, and community.
Single-cell transcriptomics of 20 mouse organs creates a Tabula Muris. The Tabula Muris Consortium (2018). Nature. doi:10.1038/s41586-018-0590-4
Abstract
Here we present a compendium of single-cell transcriptomic data from the model organism Mus musculus that comprises more than 100,000 cells from 20 organs and tissues. These data represent a new resource for cell biology, reveal gene expression in poorly characterized cell populations and enable the direct and controlled comparison of gene expression in cell types that are shared between tissues, such as T lymphocytes and endothelial cells from different anatomical locations. Two distinct technical approaches were used for most organs: one approach, microfluidic droplet-based 3'-end counting, enabled the survey of thousands of cells at relatively low coverage, whereas the other, full-length transcript analysis based on fluorescence-activated cell sorting, enabled the characterization of cell types with high sensitivity and coverage. The cumulative data provide the foundation for an atlas of transcriptomic cell biology.
The Tabula Sapiens: A multiple-organ, single-cell transcriptomic atlas of humans. The Tabula Sapiens Consortium (2022). Science. doi:10.1126/science.abl4896
Abstract
Molecular characterization of cell types using single-cell transcriptome sequencing is revolutionizing cell biology and enabling new insights into the physiology of human organs. We created a human reference atlas comprising nearly 500,000 cells from 24 different tissues and organs, many from the same donor. This atlas enabled molecular characterization of more than 400 cell types, their distribution across tissues, and tissue-specific variation in gene expression. Using multiple tissues from a single donor enabled identification of the clonal distribution of T cells between tissues, identification of the tissue-specific mutation rate in B cells, and analysis of the cell cycle state and proliferative potential of shared cell types across tissues. Cell type-specific RNA splicing was discovered and analyzed across tissues within an individual.
Zeng (2022) — what is a cell type?
What is a cell type and how to define it?. Zeng (2022). Cell. doi:10.1016/j.cell.2022.06.031
Abstract
Cell types are the basic functional units of an organism. Cell types exhibit diverse phenotypic properties at multiple levels, making them challenging to define, categorize, and understand. This review provides an overview of the basic principles of cell types rooted in evolution and development and discusses approaches to characterize and classify cell types and investigate how they contribute to the organism's function, using the mammalian brain as a primary example. I propose a roadmap toward a conceptual framework and knowledge base of cell types that will enable a deeper understanding of the dynamic changes of cellular function under healthy and diseased conditions.
B.7.5 Cancer evolution
Nowell (1976) — the clonal evolution of tumour cell populations.
The Clonal Evolution of Tumor Cell Populations: Acquired genetic lability permits stepwise selection of variant sublines and underlies tumor progression.. Nowell (1976). Science. doi:10.1126/science.959840
Abstract
It is proposed that most neoplasms arise from a single cell of origin, and tumor progression results from acquired genetic variability within the original clone allowing sequential selection of more aggressive sublines. Tumor cell populations are apparently more genetically unstable than normal cells, perhaps from activation of specific gene loci in the neoplasm, continued presence of carcinogen, or even nutritional deficiencies within the tumor. The acquired genetic insta0ility and associated selection process, most readily recognized cytogenetically, results in advanced human malignancies being highly individual karyotypically and biologically. Hence, each patient's cancer may require individual specific therapy, and even this may be thwarted by emergence of a genetically variant subline resistant to the treatment. More research should be directed toward understanding and controlling the evolutionary process in tumors before it reaches the late stage usually seen in clinical cancer.
Gerlinger et al. (2012), McGranahan and Swanton (2017) — intratumour heterogeneity and branched evolution.
Intratumor Heterogeneity and Branched Evolution Revealed by Multiregion Sequencing. Gerlinger et al. (2012). New England Journal of Medicine. doi:10.1056/nejmoa1113205
Abstract
Background Intratumor heterogeneity may foster tumor evolution and adaptation and hinder personalized-medicine strategies that depend on results from single tumor-biopsy samples. Methods To examine intratumor heterogeneity, we performed exome sequencing, chromosome aberration analysis, and ploidy profiling on multiple spatially separated samples obtained from primary renal carcinomas and associated metastatic sites. We characterized the consequences of intratumor heterogeneity using immunohistochemical analysis, mutation functional analysis, and profiling of messenger RNA expression. Results Phylogenetic reconstruction revealed branched evolutionary tumor growth, with 63 to 69% of all somatic mutations not detectable across every tumor region. Intratumor heterogeneity was observed for a mutation within an autoinhibitory domain of the mammalian target of rapamycin (mTOR) kinase, correlating with S6 and 4EBP phosphorylation in vivo and constitutive activation of mTOR kinase activity in vitro. Mutational intratumor heterogeneity was seen for multiple tumor-suppressor genes converging on loss of function; SETD2, PTEN, and KDM5C underwent multiple distinct and spatially separated inactivating mutations within a single tumor, suggesting convergent phenotypic evolution. Gene-expression signatures of good and poor prognosis were detected in different regions of the same tumor. Allelic composition and ploidy profiling analysis revealed extensive intratumor heterogeneity, with 26 of 30 tumor samples from four tumors harboring divergent allelic-imbalance profiles and with ploidy heterogeneity in two of four tumors. Conclusions Intratumor heterogeneity can lead to underestimation of the tumor genomics landscape portrayed from single tumor-biopsy samples and may present major challenges to personalized-medicine and biomarker development. Intratumor heterogeneity, associated with heterogeneous protein function, may foster tumor adaptation and therapeutic failure through Darwinian selection. (Funded by the Medical Research Council and others.).
Clonal Heterogeneity and Tumor Evolution: Past, Present, and the Future. McGranahan & Swanton (2017). Cell. doi:10.1016/j.cell.2017.01.018
Abstract
Intratumor heterogeneity, which fosters tumor evolution, is a key challenge in cancer medicine. Here, we review data and technologies that have revealed intra-tumor heterogeneity across cancer types and the dynamics, constraints, and contingencies inherent to tumor evolution. We emphasize the importance of macro-evolutionary leaps, often involving large-scale chromosomal alterations, in driving tumor evolution and metastasis and consider the role of the tumor microenvironment in engendering heterogeneity and drug resistance. We suggest that bold approaches to drug development, harnessing the adaptive properties of the immune-microenvironment while limiting those of the tumor, combined with advances in clinical trial-design, will improve patient outcome.
Schwartz and Schäffer (2017) — tumour phylogenetics.
The evolution of tumour phylogenetics: principles and practice. Schwartz & Schäffer (2017). Nature Reviews Genetics. doi:10.1038/nrg.2016.170
Abstract
Rapid advances in high-throughput sequencing and a growing realization of the importance of evolutionary theory to cancer genomics have led to a proliferation of phylogenetic studies of tumour progression. These studies have yielded not only new insights but also a plethora of experimental approaches, sometimes reaching conflicting or poorly supported conclusions. Here, we consider this body of work in light of the key computational principles underpinning phylogenetic inference, with the goal of providing practical guidance on the design and analysis of scientifically rigorous tumour phylogeny studies. We survey the range of methods and tools available to the researcher, their key applications, and the various unsolved problems, closing with a perspective on the prospects and broader implications of this field.
Gawad et al. (2016) — single-cell genome sequencing.
Single-cell genome sequencing: current state of the science. Gawad et al. (2016). Nature Reviews Genetics. doi:10.1038/nrg.2015.16
Abstract
The field of single-cell genomics is advancing rapidly and is generating many new insights into complex biological systems, ranging from the diversity of microbial ecosystems to the genomics of human cancer. In this Review, we provide an overview of the current state of the field of single-cell genome sequencing. First, we focus on the technical challenges of making measurements that start from a single molecule of DNA, and then explore how some of these recent methodological advancements have enabled the discovery of unexpected new biology. Areas highlighted include the application of single-cell genomics to interrogate microbial dark matter and to evaluate the pathogenic roles of genetic mosaicism in multicellular organisms, with a focus on cancer. We then attempt to predict advances we expect to see in the next few years.
Navin et al. (2011), Minussi et al. (2021) — tumour evolution from single-cell DNA sequencing.
Tumour evolution inferred by single-cell sequencing. Navin et al. (2011). Nature. doi:10.1038/nature09807
Abstract
Genomic analysis provides insights into the role of copy number variation in disease, but most methods are not designed to resolve mixed populations of cells. In tumours, where genetic heterogeneity is common, very important information may be lost that would be useful for reconstructing evolutionary history. Here we show that with flow-sorted nuclei, whole genome amplification and next generation sequencing we can accurately quantify genomic copy number within an individual nucleus. We apply single-nucleus sequencing to investigate tumour population structure and evolution in two human breast cancer cases. Analysis of 100 single cells from a polygenomic tumour revealed three distinct clonal subpopulations that probably represent sequential clonal expansions. Additional analysis of 100 single cells from a monogenomic primary tumour and its liver metastasis indicated that a single clonal expansion formed the primary tumour and seeded the metastasis. In both primary tumours, we also identified an unexpectedly abundant subpopulation of genetically diverse 'pseudodiploid' cells that do not travel to the metastatic site. In contrast to gradual models of tumour progression, our data indicate that tumours grow by punctuated clonal expansions with few persistent intermediates.
Breast tumours maintain a reservoir of subclonal diversity during expansion. Minussi et al. (2021). Nature. doi:10.1038/s41586-021-03357-x
Abstract
Our knowledge of copy number evolution during the expansion of primary breast tumours is limited 1,2 . Here, to investigate this process, we developed a single-cell, single-molecule DNA-sequencing method and performed copy number analysis of 16,178 single cells from 8 human triple-negative breast cancers and 4 cell lines. The results show that breast tumours and cell lines comprise a large milieu of subclones (7-22) that are organized into a few (3-5) major superclones. Evolutionary analysis suggests that after clonal TP53 mutations, multiple loss-of-heterozygosity events and genome doubling, there was a period of transient genomic instability followed by ongoing copy number evolution during the primary tumour expansion. By subcloning single daughter cells in culture, we show that tumour cells rediversify their genomes and do not retain isogenic properties. These data show that triple-negative breast cancers continue to evolve chromosome aberrations and maintain a reservoir of subclonal diversity during primary tumour growth.
Patel et al. (2014), Tirosh et al. (2016) — single-cell RNA-seq of tumours.
Single-cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma. Patel et al. (2014). Science. doi:10.1126/science.1254257
Abstract
Human cancers are complex ecosystems composed of cells with distinct phenotypes, genotypes, and epigenetic states, but current models do not adequately reflect tumor composition in patients. We used single-cell RNA sequencing (RNA-seq) to profile 430 cells from five primary glioblastomas, which we found to be inherently variable in their expression of diverse transcriptional programs related to oncogenic signaling, proliferation, complement/immune response, and hypoxia. We also observed a continuum of stemness-related expression states that enabled us to identify putative regulators of stemness in vivo. Finally, we show that established glioblastoma subtype classifiers are variably expressed across individual cells within a tumor and demonstrate the potential prognostic implications of such intratumoral heterogeneity. Thus, we reveal previously unappreciated heterogeneity in diverse regulatory programs central to glioblastoma biology, prognosis, and therapy.
Dissecting the multicellular ecosystem of metastatic melanoma by single-cell RNA-seq. Tirosh et al. (2016). Science. doi:10.1126/science.aad0501
Abstract
To explore the distinct genotypic and phenotypic states of melanoma tumors, we applied single-cell RNA sequencing (RNA-seq) to 4645 single cells isolated from 19 patients, profiling malignant, immune, stromal, and endothelial cells. Malignant cells within the same tumor displayed transcriptional heterogeneity associated with the cell cycle, spatial context, and a drug-resistance program. In particular, all tumors harbored malignant cells from two distinct transcriptional cell states, such that tumors characterized by high levels of the MITF transcription factor also contained cells with low MITF and elevated levels of the AXL kinase. Single-cell analyses suggested distinct tumor microenvironmental patterns, including cell-to-cell interactions. Analysis of tumor-infiltrating T cells revealed exhaustion programs, their connection to T cell activation and clonal expansion, and their variability across patients. Overall, we begin to unravel the cellular ecosystem of tumors and how single-cell genomics offers insights with implications for both targeted and immune therapies.
Gao et al. (2021) — copy number from single-cell transcriptomes.
Delineating copy number and clonal substructure in human tumors from single-cell transcriptomes. Gao et al. (2021). Nature Biotechnology. doi:10.1038/s41587-020-00795-2
Abstract
Single-cell transcriptomic analysis is widely used to study human tumors. However, it remains challenging to distinguish normal cell types in the tumor microenvironment from malignant cells and to resolve clonal substructure within the tumor. To address these challenges, we developed an integrative Bayesian segmentation approach called copy number karyotyping of aneuploid tumors (CopyKAT) to estimate genomic copy number profiles at an average genomic resolution of 5 Mb from read depth in high-throughput single-cell RNA sequencing (scRNA-seq) data. We applied CopyKAT to analyze 46,501 single cells from 21 tumors, including triple-negative breast cancer, pancreatic ductal adenocarcinoma, anaplastic thyroid cancer, invasive ductal carcinoma and glioblastoma, to accurately (98%) distinguish cancer cells from normal cell types. In three breast tumors, CopyKAT resolved clonal subpopulations that differed in the expression of cancer genes, such as KRAS, and signatures, including epithelial-to-mesenchymal transition, DNA repair, apoptosis and hypoxia. These data show that CopyKAT can aid in the analysis of scRNA-seq data in a variety of solid human tumors.
Lawson et al. (2018) — tumour heterogeneity and metastasis at single-cell resolution.
Tumour heterogeneity and metastasis at single-cell resolution. Lawson et al. (2018). Nature Cell Biology. doi:10.1038/s41556-018-0236-7
Abstract
Tumours comprise a heterogeneous collection of cells with distinct genetic and phenotypic properties that can differentially promote progression, metastasis and drug resistance. Emerging single-cell technologies provide a new opportunity to profile individual cells within tumours and investigate what roles they play in these processes. This Review discusses key technological considerations for single-cell studies in cancer, new findings using single-cell technologies and critical open questions for future applications.
B.8 Proteomics
Further papers for Chapter 8.
In the HTML version, click Abstract to show the abstract.
B.8.1 Reviews
Aebersold and Mann (2003) — mass spectrometry-based proteomics.
Mass spectrometry-based proteomics. Aebersold & Mann (2003). Nature. doi:10.1038/nature01511
Abstract
Recent successes illustrate the role of mass spectrometry-based proteomics as an indispensable tool for molecular and cellular biology and for the emerging field of systems biology. These include the study of protein-protein interactions via affinity-based isolations on a small and proteome-wide scale, the mapping of numerous organelles, the concurrent description of the malaria parasite genome and proteome, and the generation of quantitative protein profiles from diverse species. The ability of mass spectrometry to identify and, increasingly, to precisely quantify thousands of proteins from complex samples can be expected to impact broadly on biology and medicine.
Nesvizhskii (2010) — computational methods and error-rate estimation for peptide and protein identification.
A survey of computational methods and error rate estimation procedures for peptide and protein identification in shotgun proteomics. Nesvizhskii (2010). Journal of Proteomics. doi:10.1016/j.jprot.2010.08.009
Abstract
This manuscript provides a comprehensive review of the peptide and protein identification process using tandem mass spectrometry (MS/MS) data generated in shotgun proteomic experiments. The commonly used methods for assigning peptide sequences to MS/MS spectra are critically discussed and compared, from basic strategies to advanced multi-stage approaches. A particular attention is paid to the problem of false-positive identifications. Existing statistical approaches for assessing the significance of peptide to spectrum matches are surveyed, ranging from single-spectrum approaches such as expectation values to global error rate estimation procedures such as false discovery rates and posterior probabilities. The importance of using auxiliary discriminant information (mass accuracy, peptide separation coordinates, digestion properties, and etc.) is discussed, and advanced computational approaches for joint modeling of multiple sources of information are presented. This review also includes a detailed analysis of the issues affecting the interpretation of data at the protein level, including the amplification of error rates when going from peptide to protein level, and the ambiguities in inferring the identifies of sample proteins in the presence of shared peptides. Commonly used methods for computing protein-level confidence scores are discussed in detail. The review concludes with a discussion of several outstanding computational issues.
Toby et al. (2016) — top-down proteomics and proteoforms.
Progress in Top-Down Proteomics and the Analysis of Proteoforms. Toby et al. (2016). Annual Review of Analytical Chemistry. doi:10.1146/annurev-anchem-071015-041550
Abstract
From a molecular perspective, enactors of function in biology are intact proteins that can be variably modified at the genetic, transcriptional, or post-translational level. Over the past 30 years, mass spectrometry (MS) has become a powerful method for the analysis of proteomes. Prevailing bottom-up proteomics operates at the level of the peptide, leading to issues with protein inference, connectivity, and incomplete sequence/modification information. Top-down proteomics (TDP), alternatively, applies MS at the proteoform level to analyze intact proteins with diverse sources of intramolecular complexity preserved during analysis. Fortunately, advances in prefractionation workflows, MS instrumentation, and dissociation methods for whole-protein ions have helped TDP emerge as an accessible and potentially disruptive modality with increasingly translational value. In this review, we discuss technical and conceptual advances in TDP, along with the growing power of proteoform-resolved measurements in clinical and translational research.
B.8.2 History and instruments
O’Farrell (1975) — two-dimensional gel electrophoresis.
High resolution two-dimensional electrophoresis of proteins.. O’Farrell (1975). Journal of Biological Chemistry. doi:10.1016/s0021-9258(19)41496-8
Abstract
A technique has been developed for the separation of proteins by two-dimensional polyacrylamide gel electrophoresis. Due to its resolution and sensitivity, this technique is a powerful tool for the analysis and detection of proteins from complex biological sources. Proteins are separated according to isoelectric point by isoelectric focusing in the first dimension, and according to molecular weight by sodium dodecyl sulfate electrophoresis in the second dimension. Since these two parameters are unrelated, it is possible to obtain an almost uniform distribution of protein spots across a two-diminsional gel. This technique has resolved 1100 different components from Escherichia coli and should be capable of resolving a maximum of 5000 proteins. A protein containing as little as one disintegration per min of either 14C or 35S can be detected by autoradiography. A protein which constitutes 10 minus 4 to 10 minus 5% of the total protein can be detected and quantified by autoradiography. The reproducibility of the separation is sufficient to permit each spot on one separation to be matched with a spot on a different separation. This technique provides a method for estimation (at the described sensitivities) of the number of proteins made by any biological system. This system can resolve proteins differing in a single charge and consequently can be used in the analysis of in vivo modifications resulting in a change in charge. Proteins whose charge is changed by missense mutations can be identified. A detailed description of the methods as well as the characteristics of this system are presented.
Fenn et al. (1989), Karas and Hillenkamp (1988), Tanaka et al. (1988) — electrospray and MALDI.
Electrospray Ionization for Mass Spectrometry of Large Biomolecules. Fenn et al. (1989). Science. doi:10.1126/science.2675315
Abstract
Electrospray ionization has recently emerged as a powerful technique for producing intact ions in vacuo from large and complex species in solution. To an extent greater than has previously been possible with the more familiar "soft" ionization methods, this technique makes the power and elegance of mass spectrometric analysis applicable to the large and fragile polar molecules that play such vital roles in biological systems. The distinguishing features of electrospray spectra for large molecules are coherent sequences of peaks whose component ions are multiply charged, the ions of each peak differing by one charge from those of adjacent neighbors in the sequence. Spectra have been obtained for biopolymers including oligonucleotides and proteins, the latter having molecular weights up to 130,000, with as yet no evidence of an upper limit.
Laser desorption ionization of proteins with molecular masses exceeding 10,000 daltons. Karas & Hillenkamp (1988). Analytical Chemistry. doi:10.1021/ac00171a028
Protein and polymer analyses up to m/z 100 000 by laser ionization time-of-flight mass spectrometry. Tanaka et al. (1988). Rapid Communications in Mass Spectrometry. doi:10.1002/rcm.1290020802
Henzel et al. (1993) — peptide mass fingerprinting.
Identifying proteins from two-dimensional gels by molecular mass searching of peptide fragments in protein sequence databases.. Henzel et al. (1993). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.90.11.5011
Abstract
A rapid method for the identification of known proteins separated by two-dimensional gel electrophoresis is described in which molecular masses of peptide fragments are used to search a protein sequence database. The peptides are generated by in situ reduction, alkylation, and tryptic digestion of proteins electroblotted from two-dimensional gels. Masses are determined at the subpicomole level by matrix-assisted laser desorption/ionization mass spectrometry of the unfractionated digest. A computer program has been developed that searches the protein sequence database for multiple peptides of individual proteins that match the measured masses. To ensure that the most recent database updates are included, a theoretical digest of the entire database is generated each time the program is executed. This method facilitates simultaneous processing of a large number of two-dimensional gel spots. The method was applied to a two-dimensional gel of a crude Escherichia coli extract that was electroblotted onto poly(vinylidene difluoride) membrane. Ten randomly chosen spots were analyzed. With as few as three peptide masses, each protein was uniquely identified from over 91,000 protein sequences. All identifications were verified by concurrent N-terminal sequencing of identical spots from a second blot. One of the spots contained an N-terminally blocked protein that required enzymatic cleavage, peptide separation, and Edman degradation for confirmation of its identity.
Makarov (2000), Hu et al. (2005) — the Orbitrap.
Electrostatic Axially Harmonic Orbital Trapping: A High-Performance Technique of Mass Analysis. Makarov (2000). Analytical Chemistry. doi:10.1021/ac991131p
Abstract
This work describes a new type of mass analyzer which employs trapping in an electrostatic field. The potential distribution of the field can be represented as a combination of quadrupole and logarithmic potentials. In the absence of any magnetic or rf fields, ion stability is achieved only due to ions orbiting around an axial electrode. Orbiting ions also perform harmonic oscillations along the electrode with frequency proportional to (m/z)-1/2. These oscillations are detected using image current detection and are transformed into mass spectra using fast FT, similarly to FTICR. Practical aspects of the trap design are presented. High-mass resolution up to 150,000 for ions produced by laser ablation has been demonstrated, along with high-energy acceptance and wide mass range.
The Orbitrap: a new mass spectrometer. Hu et al. (2005). Journal of Mass Spectrometry. doi:10.1002/jms.856
Abstract
Research areas such as proteomics and metabolomics are driving the demand for mass spectrometers that have high performance but modest power requirements, size, and cost. This paper describes such an instrument, the Orbitrap, based on a new type of mass analyzer invented by Makarov. The Orbitrap operates by radially trapping ions about a central spindle electrode. An outer barrel-like electrode is coaxial with the inner spindlelike electrode and mass/charge values are measured from the frequency of harmonic ion oscillations, along the axis of the electric field, undergone by the orbitally trapped ions. This axial frequency is independent of the energy and spatial spread of the ions. Ion frequencies are measured non-destructively by acquisition of time-domain image current transients, with subsequent fast Fourier transforms (FFTs) being used to obtain the mass spectra. In addition to describing the Orbitrap mass analyzer, this paper also describes a complete Orbitrap-based mass spectrometer, equipped with an electrospray ionization source (ESI). Ions are transferred from the ESI source through three stages of differential pumping using RF guide quadrupoles. The third quadrupole, pressurized to less than 10(-3) Torr with collision gas, acts as an ion accumulator; ion/neutral collisions slow the ions and cause them to pool in an axial potential well at the end of the quadrupole. Ion bunches are injected from this pool into the Orbitrap analyzer for mass analysis. The ion injection process is described in a simplified way, including a description of electrodynamic squeezing, field compensation for the effects of the ion injection slit, and criteria for orbital stability. Features of the Orbitrap at its present stage of development include high mass resolution (up to 150,000), large space charge capacity, high mass accuracy (2-5 ppm), a mass/charge range of at least 6000, and dynamic range greater than 10(3). Applications based on electrospray ionization are described, including characterization of transition-metal complexes, oligosaccharides, peptides, and proteins. Use is also made of the high-resolution capabilities of the Orbitrap to confirm the presence of metaclusters of serine octamers in ESI mass spectra and to perform H/D exchange experiments on these ions in the storage quadrupole.
Meier et al. (2018) — PASEF on a trapped ion mobility instrument.
Online Parallel Accumulation–Serial Fragmentation (PASEF) with a Novel Trapped Ion Mobility Mass Spectrometer. Meier et al. (2018). Molecular & Cellular Proteomics. doi:10.1074/mcp.tir118.000900
Abstract
In bottom-up proteomics, peptides are separated by liquid chromatography with elution peak widths in the range of seconds, whereas mass spectra are acquired in about 100 microseconds with time-of-flight (TOF) instruments. This allows adding ion mobility as a third dimension of separation. Among several formats, trapped ion mobility spectrometry (TIMS) is attractive because of its small size, low voltage requirements and high efficiency of ion utilization. We have recently demonstrated a scan mode termed parallel accumulation - serial fragmentation (PASEF), which multiplies the sequencing speed without any loss in sensitivity (Meier et al. , PMID: 26538118). Here we introduce the timsTOF Pro instrument, which optimally implements online PASEF. It features an orthogonal ion path into the ion mobility device, limiting the amount of debris entering the instrument and making it very robust in daily operation. We investigate different precursor selection schemes for shotgun proteomics to optimally allocate in excess of 100 fragmentation events per second. More than 600,000 fragmentation spectra in standard 120 min LC runs are achievable, which can be used for near exhaustive precursor selection in complex mixtures or accumulating the signal of weak precursors. In 120 min single runs of HeLa digest, MaxQuant identified more than 6,000 proteins without matching to a library and with high quantitative reproducibility (R > 0.97). Online PASEF achieves a remarkable sensitivity with more than 2,500 proteins identified in 30 min runs of only 10 ng HeLa digest. We also show that highly reproducible collisional cross sections can be acquired on a large scale (R > 0.99). PASEF on the timsTOF Pro is a valuable addition to the technological toolbox in proteomics, with a number of unique operating modes that are only beginning to be explored.
B.8.3 Acquisition and quantification
Gillet et al. (2012) — SWATH-MS.
Targeted Data Extraction of the MS/MS Spectra Generated by Data-independent Acquisition: A New Concept for Consistent and Accurate Proteome Analysis. Gillet et al. (2012). Molecular & Cellular Proteomics. doi:10.1074/mcp.o111.016717
Abstract
Most proteomic studies use liquid chromatography coupled to tandem mass spectrometry to identify and quantify the peptides generated by the proteolysis of a biological sample. However, with the current methods it remains challenging to rapidly, consistently, reproducibly, accurately, and sensitively detect and quantify large fractions of proteomes across multiple samples. Here we present a new strategy that systematically queries sample sets for the presence and quantity of essentially any protein of interest. It consists of using the information available in fragment ion spectral libraries to mine the complete fragment ion maps generated using a data-independent acquisition method. For this study, the data were acquired on a fast, high resolution quadrupole-quadrupole time-of-flight (TOF) instrument by repeatedly cycling through 32 consecutive 25-Da precursor isolation windows (swaths). This SWATH MS acquisition setup generates, in a single sample injection, time-resolved fragment ion spectra for all the analytes detectable within the 400-1200 m/z precursor range and the user-defined retention time window. We show that suitable combinations of fragment ions extracted from these data sets are sufficiently specific to confidently identify query peptides over a dynamic range of 4 orders of magnitude, even if the precursors of the queried peptides are not detectable in the survey scans. We also show that queried peptides are quantified with a consistency and accuracy comparable with that of selected reaction monitoring, the gold standard proteomic quantification method. Moreover, targeted data extraction enables ad libitum quantification refinement and dynamic extension of protein probing by iterative re-mining of the once-and-forever acquired data sets. This combination of unbiased, broad range precursor ion fragmentation and targeted data extraction alleviates most constraints of present proteomic methods and should be equally applicable to the comprehensive analysis of other classes of analytes, beyond proteomics.
Lange et al. (2008), Peterson et al. (2012) — targeted proteomics (SRM and PRM).
Selected reaction monitoring for quantitative proteomics: a tutorial. Lange et al. (2008). Molecular Systems Biology. doi:10.1038/msb.2008.61
Abstract
Systems biology relies on data sets in which the same group of proteins is consistently identified and precisely quantified across multiple samples, a requirement that is only partially achieved by current proteomics approaches. Selected reaction monitoring (SRM)-also called multiple reaction monitoring-is emerging as a technology that ideally complements the discovery capabilities of shotgun strategies by its unique potential for reliable quantification of analytes of low abundance in complex mixtures. In an SRM experiment, a predefined precursor ion and one of its fragments are selected by the two mass filters of a triple quadrupole instrument and monitored over time for precise quantification. A series of transitions (precursor/fragment ion pairs) in combination with the retention time of the targeted peptide can constitute a definitive assay. Typically, a large number of peptides are quantified during a single LC-MS experiment. This tutorial explains the application of SRM for quantitative proteomics, including the selection of proteotypic peptides and the optimization and validation of transitions. Furthermore, normalization and various factors affecting sensitivity and accuracy are discussed.
Parallel Reaction Monitoring for High Resolution and High Mass Accuracy Quantitative, Targeted Proteomics. Peterson et al. (2012). Molecular & Cellular Proteomics. doi:10.1074/mcp.o112.020131
Abstract
Selected reaction monitoring on a triple quadrupole mass spectrometer is currently experiencing a renaissance within the proteomics community for its, as yet, unparalleled ability to characterize and quantify a set of proteins reproducibly, completely, and with high sensitivity. Given the immense benefit that high resolution and accurate mass instruments have brought to the discovery proteomics field, we wondered if highly accurate mass measurement capabilities could be leveraged to provide benefits in the targeted proteomics domain as well. Here, we propose a new targeted proteomics paradigm centered on the use of next generation, quadrupole-equipped high resolution and accurate mass instruments: parallel reaction monitoring (PRM). In PRM, the third quadrupole of a triple quadrupole is substituted with a high resolution and accurate mass mass analyzer to permit the parallel detection of all target product ions in one, concerted high resolution mass analysis. We detail the analytical performance of the PRM method, using a quadrupole-equipped bench-top Orbitrap MS, and draw a performance comparison to selected reaction monitoring in terms of run-to-run reproducibility, dynamic range, and measurement accuracy. In addition to requiring minimal upfront method development and facilitating automated data analysis, PRM yielded quantitative data over a wider dynamic range than selected reaction monitoring in the presence of a yeast background matrix because of PRM's high selectivity in the mass-to-charge domain. With achievable linearity over the quantifiable dynamic range found to be statistically equal between the two methods, our investigation suggests that PRM will be a promising new addition to the quantitative proteomics toolbox.
Ong et al. (2002) — SILAC.
Stable Isotope Labeling by Amino Acids in Cell Culture, SILAC, as a Simple and Accurate Approach to Expression Proteomics. Ong et al. (2002). Molecular & Cellular Proteomics. doi:10.1074/mcp.m200025-mcp200
Abstract
Quantitative proteomics has traditionally been performed by two-dimensional gel electrophoresis, but recently, mass spectrometric methods based on stable isotope quantitation have shown great promise for the simultaneous and automated identification and quantitation of complex protein mixtures. Here we describe a method, termed SILAC, for stable isotope labeling by amino acids in cell culture, for the in vivo incorporation of specific amino acids into all mammalian proteins. Mammalian cell lines are grown in media lacking a standard essential amino acid but supplemented with a non-radioactive, isotopically labeled form of that amino acid, in this case deuterated leucine (Leu-d3). We find that growth of cells maintained in these media is no different from growth in normal media as evidenced by cell morphology, doubling time, and ability to differentiate. Complete incorporation of Leu-d3 occurred after five doublings in the cell lines and proteins studied. Protein populations from experimental and control samples are mixed directly after harvesting, and mass spectrometric identification is straightforward as every leucine-containing peptide incorporates either all normal leucine or all Leu-d3. We have applied this technique to the relative quantitation of changes in protein expression during the process of muscle cell differentiation. Proteins that were found to be up-regulated during this process include glyceraldehyde-3-phosphate dehydrogenase, fibronectin, and pyruvate kinase M2. SILAC is a simple, inexpensive, and accurate procedure that can be used as a quantitative proteomic approach in any cell culture system.
Ross et al. (2004), Thompson et al. (2003) — iTRAQ and TMT.
Multiplexed Protein Quantitation in Saccharomyces cerevisiae Using Amine-reactive Isobaric Tagging Reagents. Ross et al. (2004). Molecular & Cellular Proteomics. doi:10.1074/mcp.m400129-mcp200
Abstract
We describe here a multiplexed protein quantitation strategy that provides relative and absolute measurements of proteins in complex mixtures. At the core of this methodology is a multiplexed set of isobaric reagents that yield amine-derivatized peptides. The derivatized peptides are indistinguishable in MS, but exhibit intense low-mass MS/MS signature ions that support quantitation. In this study, we have examined the global protein expression of a wild-type yeast strain and the isogenic upf1Delta and xrn1Delta mutant strains that are defective in the nonsense-mediated mRNA decay and the general 5' to 3' decay pathways, respectively. We also demonstrate the use of 4-fold multiplexing to enable relative protein measurements simultaneously with determination of absolute levels of a target protein using synthetic isobaric peptide standards. We find that inactivation of Upf1p and Xrn1p causes common as well as unique effects on protein expression.
Tandem Mass Tags: A Novel Quantification Strategy for Comparative Analysis of Complex Protein Mixtures by MS/MS. Thompson et al. (2003). Analytical Chemistry. doi:10.1021/ac0262560
Abstract
A novel MS/MS-based analysis strategy using isotopomer labels, referred to as "tandem mass tags" (TMTs), for the accurate quantification of peptides and proteins is described. The new tags are designed to ensure that identical peptides labeled with different TMTs exactly comigrate in all separations. The tags require novel methods of quantification analysis using tandem mass spectrometry. The new tags and analysis methods allow peptides from different samples to be identified by their relative abundance with greater ease and accuracy than other methods. The new TMTs permit simultaneous determination of both the identity and relative abundances of peptide pairs using a collision induced dissociation (CID)-based analysis method. Relative abundance measurements made in the MS/MS mode using the new tags are accurate and sensitive. Compared to MS-mode measurements, a very high signal-to-noise ratio is achieved with MS/MS based detection. The new tags should be applicable to a wide variety of peptide isolation methods.
Cox et al. (2014) — MaxLFQ label-free quantification.
Accurate Proteome-wide Label-free Quantification by Delayed Normalization and Maximal Peptide Ratio Extraction, Termed MaxLFQ. Cox et al. (2014). Molecular & Cellular Proteomics. doi:10.1074/mcp.m113.031591
Abstract
Protein quantification without isotopic labels has been a long-standing interest in the proteomics field. However, accurate and robust proteome-wide quantification with label-free approaches remains a challenge. We developed a new intensity determination and normalization procedure called MaxLFQ that is fully compatible with any peptide or protein separation prior to LC-MS analysis. Protein abundance profiles are assembled using the maximum possible information from MS signals, given that the presence of quantifiable peptides varies from sample to sample. For a benchmark dataset with two proteomes mixed at known ratios, we accurately detected the mixing ratio over the entire protein expression range, with greater precision for abundant proteins. The significance of individual label-free quantifications was obtained via a t test approach. For a second benchmark dataset, we accurately quantify fold changes over several orders of magnitude, a task that is challenging with label-based methods. MaxLFQ is a generic label-free quantification technology that is readily applicable to many biological questions; it is compatible with standard statistical analysis workflows, and it has been validated in many and diverse biological projects. Our algorithms can handle very large experiments of 500+ samples in a manageable computing time. It is implemented in the freely available MaxQuant computational proteomics platform and works completely seamlessly at the click of a button.
Gerber et al. (2003) — absolute quantification with heavy peptides.
Absolute quantification of proteins and phosphoproteins from cell lysates by tandem MS. Gerber et al. (2003). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.0832254100
Abstract
A need exists for technologies that permit the direct quantification of differences in protein and posttranslationally modified protein expression levels. Here we present a strategy for the absolute quantification (termed AQUA) of proteins and their modification states. Peptides are synthesized with incorporated stable isotopes as ideal internal standards to mimic native peptides formed by proteolysis. These synthetic peptides can also be prepared with covalent modifications (e.g., phosphorylation, methylation, acetylation, etc.) that are chemically identical to naturally occurring posttranslational modifications. Such AQUA internal standard peptides are then used to precisely and quantitatively measure the absolute levels of proteins and posttranslationally modified proteins after proteolysis by using a selected reaction monitoring analysis in a tandem mass spectrometer. In the present work, the AQUA strategy was used to (i) quantify low abundance yeast proteins involved in gene silencing, (ii) quantitatively determine the cell cycle-dependent phosphorylation of Ser-1126 of human separase protein, and (iii) identify kinases capable of phosphorylating Ser-1501 of separase in an in vitro kinase assay. The methods described here represent focused, alternative approaches for studying the dynamically changing proteome.
B.8.4 Identification algorithms
Eng et al. (1994) — SEQUEST.
An approach to correlate tandem mass spectral data of peptides with amino acid sequences in a protein database. Eng et al. (1994). Journal of the American Society for Mass Spectrometry. doi:10.1016/1044-0305(94)80016-2
Abstract
A method to correlate the uninterpreted tandem mass spectra of peptides produced under low energy (10–50 eV) collision conditions with amino acid sequences in the Genpept database has been developed. In this method the protein database is searched to identify linear amino acid sequences within a mass tolerance of +/- 1 mu of the precursor ion molecular weight A cross-correlation function is then used to provide a measurement of similarity between the mass-to-charge ratios for the fragment ions predicted from amino acid sequences obtained from the database and the fragment ions observed in the tandem mass spectrum. In general, a difference greater than 0.1 between the normalized cross-correlation functions of the first- and second-ranked search results indicates a successful match between sequence and spectrum. Searches of species-specific protein databases with tandem mass spectra acquired from peptides obtained from the enzymatically digested total proteins of E. coli and S. cerevisiae cells allowed matching of the spectra to amino acid sequences within proteins of these organisms. The approach described in this manuscript provides a convenient method to interpret tandem mass spectra with known sequences in a protein database.
Perkins et al. (1999) — Mascot.
Probability-based protein identification by searching sequence databases using mass spectrometry data. Perkins et al. (1999). Electrophoresis. doi:10.1002/(sici)1522-2683(19991201)20:18<3551::aid-elps3551>3.0.co;2-2
Abstract
Several algorithms have been described in the literature for protein identification by searching a sequence database using mass spectrometry data. In some approaches, the experimental data are peptide molecular weights from the digestion of a protein by an enzyme. Other approaches use tandem mass spectrometry (MS/MS) data from one or more peptides. Still others combine mass data with amino acid sequence data. We present results from a new computer program, Mascot, which integrates all three types of search. The scoring algorithm is probability based, which has a number of advantages: (i) A simple rule can be used to judge whether a result is significant or not. This is particularly useful in guarding against false positives. (ii) Scores can be compared with those from other types of search, such as sequence homology. (iii) Search parameters can be readily optimised by iteration. The strengths and limitations of probability-based scoring are discussed, particularly in the context of high throughput, fully automated protein identification.
Craig and Beavis (2004), Cox et al. (2011), Kim and Pevzner (2014) — X!Tandem, Andromeda and MS-GF+.
TANDEM: matching proteins with tandem mass spectra. Craig & Beavis (2004). Bioinformatics. doi:10.1093/bioinformatics/bth092
Abstract
Tandem mass spectra obtained from fragmenting peptide ions contain some peptide sequence specific information, but often there is not enough information to sequence the original peptide completely. Several proprietary software applications have been developed to attempt to match the spectra with a list of protein sequences that may contain the sequence of the peptide. The application TANDEM was written to provide the proteomics research community with a set of components that can be used to test new methods and algorithms for performing this type of sequence-to-data matching. Availability The source code and binaries for this software are available at http://www.proteome.ca/opensource.html, for Windows, Linux and Macintosh OSX. The source code is made available under the Artistic License, from the authors.
Andromeda: A Peptide Search Engine Integrated into the MaxQuant Environment. Cox et al. (2011). Journal of Proteome Research. doi:10.1021/pr101065j
Abstract
A key step in mass spectrometry (MS)-based proteomics is the identification of peptides in sequence databases by their fragmentation spectra. Here we describe Andromeda, a novel peptide search engine using a probabilistic scoring model. On proteome data, Andromeda performs as well as Mascot, a widely used commercial search engine, as judged by sensitivity and specificity analysis based on target decoy searches. Furthermore, it can handle data with arbitrarily high fragment mass accuracy, is able to assign and score complex patterns of post-translational modifications, such as highly phosphorylated peptides, and accommodates extremely large databases. The algorithms of Andromeda are provided. Andromeda can function independently or as an integrated search engine of the widely used MaxQuant computational proteomics platform and both are freely available at www.maxquant.org. The combination enables analysis of large data sets in a simple analysis workflow on a desktop computer. For searching individual spectra Andromeda is also accessible via a web server. We demonstrate the flexibility of the system by implementing the capability to identify cofragmented peptides, significantly improving the total number of identified peptides.
MS-GF+ makes progress towards a universal database search tool for proteomics. Kim & Pevzner (2014). Nature Communications. doi:10.1038/ncomms6277
Abstract
Mass spectrometry (MS) instruments and experimental protocols are rapidly advancing, but the software tools to analyse tandem mass spectra are lagging behind. We present a database search tool MS-GF+ that is sensitive (it identifies more peptides than most other database search tools) and universal (it works well for diverse types of spectra, different configurations of MS instruments and different experimental protocols). We benchmark MS-GF+ using diverse spectral data sets: (i) spectra of varying fragmentation methods; (ii) spectra of multiple enzyme digests; (iii) spectra of phosphorylated peptides; and (iv) spectra of peptides with unusual fragmentation propensities produced by a novel alpha-lytic protease. For all these data sets, MS-GF+ significantly increases the number of identified peptides compared with commonly used methods for peptide identifications. We emphasize that although MS-GF+ is not specifically designed for any particular experimental set-up, it improves on the performance of tools specifically designed for these applications (for example, specialized tools for phosphoproteomics).
Kong et al. (2017) — MSFragger and open search.
MSFragger: ultrafast and comprehensive peptide identification in mass spectrometry–based proteomics. Kong et al. (2017). Nature Methods. doi:10.1038/nmeth.4256
Abstract
There is a need to better understand and handle the 'dark matter' of proteomics-the vast diversity of post-translational and chemical modifications that are unaccounted in a typical mass spectrometry-based analysis and thus remain unidentified. We present a fragment-ion indexing method, and its implementation in peptide identification tool MSFragger, that enables a more than 100-fold improvement in speed over most existing proteome database search tools. Using several large proteomic data sets, we demonstrate how MSFragger empowers the open database search concept for comprehensive identification of peptides and all their modified forms, uncovering dramatic differences in modification rates across experimental samples and conditions. We further illustrate its utility using protein-RNA cross-linked peptide data and using affinity purification experiments where we observe, on average, a 300% increase in the number of identified spectra for enriched proteins. We also discuss the benefits of open searching for improved false discovery rate estimation in proteomics.
Lam et al. (2007) — spectral library search.
Development and validation of a spectral library searching method for peptide identification from MS/MS. Lam et al. (2007). PROTEOMICS. doi:10.1002/pmic.200600625
Abstract
A notable inefficiency of shotgun proteomics experiments is the repeated rediscovery of the same identifiable peptides by sequence database searching methods, which often are time-consuming and error-prone. A more precise and efficient method, in which previously observed and identified peptide MS/MS spectra are catalogued and condensed into searchable spectral libraries to allow new identifications by spectral matching, is seen as a promising alternative. To that end, an open-source, functionally complete, high-throughput and readily extensible MS/MS spectral searching tool, SpectraST, was developed. A high-quality spectral library was constructed by combining the high-confidence identifications of millions of spectra taken from various data repositories and searched using four sequence search engines. The resulting library consists of over 30,000 spectra for Saccharomyces cerevisiae. Using this library, SpectraST vastly outperforms the sequence search engine SEQUEST in terms of speed and the ability to discriminate good and bad hits. A unique advantage of SpectraST is its full integration into the popular Trans Proteomic Pipeline suite of software, which facilitates user adoption and provides important functionalities such as peptide and protein probability assignment, quantification, and data visualization. This method of spectral library searching is especially suited for targeted proteomics applications, offering superior performance to traditional sequence searching.
Dančík et al. (1999), Ma et al. (2003) — de novo sequencing and PEAKS.
De Novo Peptide Sequencing via Tandem Mass Spectrometry. Dančík et al. (1999). Journal of Computational Biology. doi:10.1089/106652799318300
Abstract
Peptide sequencing via tandem mass spectrometry (MS/MS) is one of the most powerful tools in proteomics for identifying proteins. Because complete genome sequences are accumulating rapidly, the recent trend in interpretation of MS/MS spectra has been database search. However, de novo MS/MS spectral interpretation remains an open problem typically involving manual interpretation by expert mass spectrometrists. We have developed a new algorithm, SHERENGA, for de novo interpretation that automatically learns fragment ion types and intensity thresholds from a collection of test spectra generated from any type of mass spectrometer. The test data are used to construct optimal path scoring in the graph representations of MS/MS spectra. A ranked list of high scoring paths corresponds to potential peptide sequences. SHERENGA is most useful for interpreting sequences of peptides resulting from unknown proteins and for validating the results of database search algorithms in fully automated, high-throughput peptide sequencing.
PEAKS: powerful software for peptide de novo sequencing by tandem mass spectrometry. Ma et al. (2003). Rapid Communications in Mass Spectrometry. doi:10.1002/rcm.1196
Abstract
A number of different approaches have been described to identify proteins from tandem mass spectrometry (MS/MS) data. The most common approaches rely on the available databases to match experimental MS/MS data. These methods suffer from several drawbacks and cannot be used for the identification of proteins from unknown genomes. In this communication, we describe a new de novo sequencing software package, PEAKS, to extract amino acid sequence information without the use of databases. PEAKS uses a new model and a new algorithm to efficiently compute the best peptide sequences whose fragment ions can best interpret the peaks in the MS/MS spectrum. The output of the software gives amino acid sequences with confidence scores for the entire sequences, as well as an additional novel positional scoring scheme for portions of the sequences. The performance of PEAKS is compared with Lutefisk, a well-known de novo sequencing software, using quadrupole-time-of-flight (Q-TOF) data obtained for several tryptic peptides from standard proteins.
Tran et al. (2017), Eloff et al. (2025) — DeepNovo and InstaNovo.
De novo peptide sequencing by deep learning. Tran et al. (2017). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1705691114
Abstract
De novo peptide sequencing from tandem MS data is the key technology in proteomics for the characterization of proteins, especially for new sequences, such as mAbs. In this study, we propose a deep neural network model, DeepNovo, for de novo peptide sequencing. DeepNovo architecture combines recent advances in convolutional neural networks and recurrent neural networks to learn features of tandem mass spectra, fragment ions, and sequence patterns of peptides. The networks are further integrated with local dynamic programming to solve the complex optimization task of de novo sequencing. We evaluated the method on a wide variety of species and found that DeepNovo considerably outperformed state of the art methods, achieving 7.7-22.9% higher accuracy at the amino acid level and 38.1-64.0% higher accuracy at the peptide level. We further used DeepNovo to automatically reconstruct the complete sequences of antibody light and heavy chains of mouse, achieving 97.5-100% coverage and 97.2-99.5% accuracy, without assisting databases. Moreover, DeepNovo is retrainable to adapt to any sources of data and provides a complete end-to-end training and prediction solution to the de novo sequencing problem. Not only does our study extend the deep learning revolution to a new field, but it also shows an innovative approach in solving optimization problems by using deep learning and dynamic programming.
InstaNovo enables diffusion-powered de novo peptide sequencing in large-scale proteomics experiments. Eloff et al. (2025). Nature Machine Intelligence. doi:10.1038/s42256-025-01019-5
Abstract
Mass spectrometry-based proteomics focuses on identifying the peptide that generates a tandem mass spectrum. Traditional methods rely on protein databases but are often limited or inapplicable in certain contexts. De novo peptide sequencing, which assigns peptide sequences to spectra without prior information, is valuable for diverse biological applications; however, owing to a lack of accuracy, it remains challenging to apply. Here we introduce InstaNovo, a transformer model that translates fragment ion peaks into peptide sequences. We demonstrate that InstaNovo outperforms state-of-the-art methods and showcase its utility in several applications. We also introduce InstaNovo+, a diffusion model that improves performance through iterative refinement of predicted sequences. Using these models, we achieve improved therapeutic sequencing coverage, discover novel peptides and detect unreported organisms in diverse datasets, thereby expanding the scope and detection rate of proteomics searches. Our models unlock opportunities across domains such as direct protein sequencing, immunopeptidomics and exploration of the dark proteome.
Gessulat et al. (2019), Gabriels et al. (2019), Zeng et al. (2022) — Prosit, MS²PIP and AlphaPeptDeep.
Prosit: proteome-wide prediction of peptide tandem mass spectra by deep learning. Gessulat et al. (2019). Nature Methods. doi:10.1038/s41592-019-0426-7
Abstract
In mass-spectrometry-based proteomics, the identification and quantification of peptides and proteins heavily rely on sequence database searching or spectral library matching. The lack of accurate predictive models for fragment ion intensities impairs the realization of the full potential of these approaches. Here, we extended the ProteomeTools synthetic peptide library to 550,000 tryptic peptides and 21 million high-quality tandem mass spectra. We trained a deep neural network, termed Prosit, resulting in chromatographic retention time and fragment ion intensity predictions that exceed the quality of the experimental data. Integrating Prosit into database search pipelines led to more identifications at >10× lower false discovery rates. We show the general applicability of Prosit by predicting spectra for proteases other than trypsin, generating spectral libraries for data-independent acquisition and improving the analysis of metaproteomes. Prosit is integrated into ProteomicsDB, allowing search result re-scoring and custom spectral library generation for any organism on the basis of peptide sequence alone.
Updated MS²PIP web server delivers fast and accurate MS² peak intensity prediction for multiple fragmentation methods, instruments and labeling techniques. Gabriels et al. (2019). Nucleic Acids Research. doi:10.1093/nar/gkz299
Abstract
MS²PIP is a data-driven tool that accurately predicts peak intensities for a given peptide's fragmentation mass spectrum. Since the release of the MS²PIP web server in 2015, we have brought significant updates to both the tool and the web server. In addition to the original models for CID and HCD fragmentation, we have added specialized models for the TripleTOF 5600+ mass spectrometer, for TMT-labeled peptides, for iTRAQ-labeled peptides, and for iTRAQ-labeled phosphopeptides. Because the fragmentation pattern is heavily altered in each of these cases, these additional models greatly improve the prediction accuracy for their corresponding data types. We have also substantially reduced the computational resources required to run MS²PIP, and have completely rebuilt the web server, which now allows predictions of up to 100 000 peptide sequences in a single request. The MS²PIP web server is freely available at https://iomics.ugent.be/ms2pip/.
AlphaPeptDeep: a modular deep learning framework to predict peptide properties for proteomics. Zeng et al. (2022). Nature Communications. doi:10.1038/s41467-022-34904-3
Abstract
Machine learning and in particular deep learning (DL) are increasingly important in mass spectrometry (MS)-based proteomics. Recent DL models can predict the retention time, ion mobility and fragment intensities of a peptide just from the amino acid sequence with good accuracy. However, DL is a very rapidly developing field with new neural network architectures frequently appearing, which are challenging to incorporate for proteomics researchers. Here we introduce AlphaPeptDeep, a modular Python framework built on the PyTorch DL library that learns and predicts the properties of peptides ( https://github.com/MannLabs/alphapeptdeep ). It features a model shop that enables non-specialists to create models in just a few lines of code. AlphaPeptDeep represents post-translational modifications in a generic manner, even if only the chemical composition is known. Extensive use of transfer learning obviates the need for large data sets to refine models for particular experimental conditions. The AlphaPeptDeep models for predicting retention time, collisional cross sections and fragment intensities are at least on par with existing tools. Additional sequence-based properties can also be predicted by AlphaPeptDeep, as demonstrated with a HLA peptide prediction model to improve HLA peptide identification for data-independent acquisition ( https://github.com/MannLabs/PeptDeep-HLA ).
Röst et al. (2014), Demichev et al. (2019) — OpenSWATH and DIA-NN.
OpenSWATH enables automated, targeted analysis of data-independent acquisition MS data. Röst et al. (2014). Nature Biotechnology. doi:10.1038/nbt.2841
DIA-NN: neural networks and interference correction enable deep proteome coverage in high throughput. Demichev et al. (2019). Nature Methods. doi:10.1038/s41592-019-0638-x
Abstract
We present an easy-to-use integrated software suite, DIA-NN, that exploits deep neural networks and new quantification and signal correction strategies for the processing of data-independent acquisition (DIA) proteomics experiments. DIA-NN improves the identification and quantification performance in conventional DIA proteomic applications, and is particularly beneficial for high-throughput applications, as it is fast and enables deep and confident proteome coverage when used in combination with fast chromatographic methods.
B.8.5 Statistics
Keller et al. (2002) — PeptideProphet.
Empirical Statistical Model To Estimate the Accuracy of Peptide Identifications Made by MS/MS and Database Search. Keller et al. (2002). Analytical Chemistry. doi:10.1021/ac025747h
Abstract
We present a statistical model to estimate the accuracy of peptide assignments to tandem mass (MS/MS) spectra made by database search applications such as SEQUEST. Employing the expectation maximization algorithm, the analysis learns to distinguish correct from incorrect database search results, computing probabilities that peptide assignments to spectra are correct based upon database search scores and the number of tryptic termini of peptides. Using SEQUEST search results for spectra generated from a sample of known protein components, we demonstrate that the computed probabilities are accurate and have high power to discriminate between correctly and incorrectly assigned peptides. This analysis makes it possible to filter large volumes of MS/MS database search results with predictable false identification error rates and can serve as a common standard by which the results of different research groups are compared.
Käll et al. (2007) — Percolator.
Semi-supervised learning for peptide identification from shotgun proteomics datasets. Käll et al. (2007). Nature Methods. doi:10.1038/nmeth1113
Abstract
Shotgun proteomics uses liquid chromatography-tandem mass spectrometry to identify proteins in complex biological samples. We describe an algorithm, called Percolator, for improving the rate of confident peptide identifications from a collection of tandem mass spectra. Percolator uses semi-supervised machine learning to discriminate between correct and decoy spectrum identifications, correctly assigning peptides to 17% more spectra from a tryptic Saccharomyces cerevisiae dataset, and up to 77% more spectra from non-tryptic digests, relative to a fully supervised approach.
Nesvizhskii et al. (2003) — ProteinProphet and protein inference.
A Statistical Model for Identifying Proteins by Tandem Mass Spectrometry. Nesvizhskii et al. (2003). Analytical Chemistry. doi:10.1021/ac0341261
Abstract
A statistical model is presented for computing probabilities that proteins are present in a sample on the basis of peptides assigned to tandem mass (MS/MS) spectra acquired from a proteolytic digest of the sample. Peptides that correspond to more than a single protein in the sequence database are apportioned among all corresponding proteins, and a minimal protein list sufficient to account for the observed peptide assignments is derived using the expectation-maximization algorithm. Using peptide assignments to spectra generated from a sample of 18 purified proteins, as well as complex H. influenzae and Halobacterium samples, the model is shown to produce probabilities that are accurate and have high power to discriminate correct from incorrect protein identifications. This method allows filtering of large-scale proteomics data sets with predictable sensitivity and false positive identification error rates. Fast, consistent, and transparent, it provides a standard for publishing large-scale protein identification data sets in the literature and for comparing the results obtained from different experiments.
Savitski et al. (2015) — protein-level FDR.
A Scalable Approach for Protein False Discovery Rate Estimation in Large Proteomic Data Sets. Savitski et al. (2015). Molecular & Cellular Proteomics. doi:10.1074/mcp.m114.046995
Abstract
Calculating the number of confidently identified proteins and estimating false discovery rate (FDR) is a challenge when analyzing very large proteomic data sets such as entire human proteomes. Biological and technical heterogeneity in proteomic experiments further add to the challenge and there are strong differences in opinion regarding the conceptual validity of a protein FDR and no consensus regarding the methodology for protein FDR determination. There are also limitations inherent to the widely used classic target-decoy strategy that particularly show when analyzing very large data sets and that lead to a strong over-representation of decoy identifications. In this study, we investigated the merits of the classic, as well as a novel target-decoy-based protein FDR estimation approach, taking advantage of a heterogeneous data collection comprised of ∼19,000 LC-MS/MS runs deposited in ProteomicsDB (https://www.proteomicsdb.org). The "picked" protein FDR approach treats target and decoy sequences of the same protein as a pair rather than as individual entities and chooses either the target or the decoy sequence depending on which receives the highest score. We investigated the performance of this approach in combination with q-value based peptide scoring to normalize sample-, instrument-, and search engine-specific differences. The "picked" target-decoy strategy performed best when protein scoring was based on the best peptide q-value for each protein yielding a stable number of true positive protein identifications over a wide range of q-value thresholds. We show that this simple and unbiased strategy eliminates a conceptual issue in the commonly used "classic" protein FDR approach that causes overprediction of false-positive protein identification in large data sets. The approach scales from small to very large data sets without losing performance, consistently increases the number of true-positive protein identifications and is readily implemented in proteomics analysis software.
Choi et al. (2014) — MSstats.
MSstats: an R package for statistical analysis of quantitative mass spectrometry-based proteomic experiments. Choi et al. (2014). Bioinformatics. doi:10.1093/bioinformatics/btu305
Abstract
Unlabelled MSstats is an R package for statistical relative quantification of proteins and peptides in mass spectrometry-based proteomics. Version 2.0 of MSstats supports label-free and label-based experimental workflows and data-dependent, targeted and data-independent spectral acquisition. It takes as input identified and quantified spectral peaks, and outputs a list of differentially abundant peptides or proteins, or summaries of peptide or protein relative abundance. MSstats relies on a flexible family of linear mixed models. Availability and implementation The code, the documentation and example datasets are available open-source at www.msstats.org under the Artistic-2.0 license. The package can be downloaded from www.msstats.org or from Bioconductor www.bioconductor.org and used in an R command line workflow. The package can also be accessed as an external tool in Skyline (Broudy et al., 2014) and used via graphical user interface.
B.8.6 Beyond abundance
Olsen et al. (2006), Beausoleil et al. (2006) — phosphoproteomics and site localisation.
Global, In Vivo, and Site-Specific Phosphorylation Dynamics in Signaling Networks. Olsen et al. (2006). Cell. doi:10.1016/j.cell.2006.09.026
Abstract
Cell signaling mechanisms often transmit information via posttranslational protein modifications, most importantly reversible protein phosphorylation. Here we develop and apply a general mass spectrometric technology for identification and quantitation of phosphorylation sites as a function of stimulus, time, and subcellular location. We have detected 6,600 phosphorylation sites on 2,244 proteins and have determined their temporal dynamics after stimulating HeLa cells with epidermal growth factor (EGF) and recorded them in the Phosida database. Fourteen percent of phosphorylation sites are modulated at least 2-fold by EGF, and these were classified by their temporal profiles. Surprisingly, a majority of proteins contain multiple phosphorylation sites showing different kinetics, suggesting that they serve as platforms for integrating signals. In addition to protein kinase cascades, the targets of reversible phosphorylation include ubiquitin ligases, guanine nucleotide exchange factors, and at least 46 different transcriptional regulators. The dynamic phosphoproteome provides a missing link in a global, integrative view of cellular regulation.
A probability-based approach for high-throughput protein phosphorylation analysis and site localization. Beausoleil et al. (2006). Nature Biotechnology. doi:10.1038/nbt1240
Abstract
Data analysis and interpretation remain major logistical challenges when attempting to identify large numbers of protein phosphorylation sites by nanoscale reverse-phase liquid chromatography/tandem mass spectrometry (LC-MS/MS) (Supplementary Figure 1 online). In this report we address challenges that are often only addressable by laborious manual validation, including data set error, data set sensitivity and phosphorylation site localization. We provide a large-scale phosphorylation data set with a measured error rate as determined by the target-decoy approach, we demonstrate an approach to maximize data set sensitivity by efficiently distracting incorrect peptide spectral matches (PSMs), and we present a probability-based score, the Ascore, that measures the probability of correct phosphorylation site localization based on the presence and intensity of site-determining ions in MS/MS spectra. We applied our methods in a fully automated fashion to nocodazole-arrested HeLa cell lysate where we identified 1,761 nonredundant phosphorylation sites from 491 proteins with a peptide false-positive rate of 1.3%.
Gavin et al. (2002), Ho et al. (2002), Huttlin et al. (2021) — AP-MS maps of yeast and human complexes.
Functional organization of the yeast proteome by systematic analysis of protein complexes. Gavin et al. (2002). Nature. doi:10.1038/415141a
Abstract
Most cellular processes are carried out by multiprotein complexes. The identification and analysis of their components provides insight into how the ensemble of expressed proteins (proteome) is organized into functional units. We used tandem-affinity purification (TAP) and mass spectrometry in a large-scale approach to characterize multiprotein complexes in Saccharomyces cerevisiae. We processed 1,739 genes, including 1,143 human orthologues of relevance to human biology, and purified 589 protein assemblies. Bioinformatic analysis of these assemblies defined 232 distinct multiprotein complexes and proposed new cellular roles for 344 proteins, including 231 proteins with no previous functional annotation. Comparison of yeast and human complexes showed that conservation across species extends from single proteins to their molecular environment. Our analysis provides an outline of the eukaryotic proteome as a network of protein complexes at a level of organization beyond binary interactions. This higher-order map contains fundamental biological information and offers the context for a more reasoned and informed approach to drug discovery.
Systematic identification of protein complexes in Saccharomyces cerevisiae by mass spectrometry. Ho et al. (2002). Nature. doi:10.1038/415180a
Abstract
The recent abundance of genome sequence data has brought an urgent need for systematic proteomics to decipher the encoded protein networks that dictate cellular function. To date, generation of large-scale protein-protein interaction maps has relied on the yeast two-hybrid system, which detects binary interactions through activation of reporter gene expression. With the advent of ultrasensitive mass spectrometric protein identification methods, it is feasible to identify directly protein complexes on a proteome-wide scale. Here we report, using the budding yeast Saccharomyces cerevisiae as a test case, an example of this approach, which we term high-throughput mass spectrometric protein complex identification (HMS-PCI). Beginning with 10% of predicted yeast proteins as baits, we detected 3,617 associated proteins covering 25% of the yeast proteome. Numerous protein complexes were identified, including many new interactions in various signalling pathways and in the DNA damage response. Comparison of the HMS-PCI data set with interactions reported in the literature revealed an average threefold higher success rate in detection of known complexes compared with large-scale two-hybrid studies. Given the high degree of connectivity observed in this study, even partial HMS-PCI coverage of complex proteomes, including that of humans, should allow comprehensive identification of cellular networks.
Dual proteome-scale networks reveal cell-specific remodeling of the human interactome. Huttlin et al. (2021). Cell. doi:10.1016/j.cell.2021.04.011
Abstract
Thousands of interactions assemble proteins into modules that impart spatial and functional organization to the cellular proteome. Through affinity-purification mass spectrometry, we have created two proteome-scale, cell-line-specific interaction networks. The first, BioPlex 3.0, results from affinity purification of 10,128 human proteins-half the proteome-in 293T cells and includes 118,162 interactions among 14,586 proteins. The second results from 5,522 immunoprecipitations in HCT116 cells. These networks model the interactome whose structure encodes protein function, localization, and complex membership. Comparison across cell lines validates thousands of interactions and reveals extensive customization. Whereas shared interactions reside in core complexes and involve essential proteins, cell-specific interactions link these complexes, "rewiring" subnetworks within each cell's interactome. Interactions covary among proteins of shared function as the proteome remodels to produce each cell's phenotype. Viewable interactively online through BioPlexExplorer, these networks define principles of proteome organization and enable unknown protein characterization.
Choi et al. (2010), Mellacheruvu et al. (2013) — scoring AP-MS data.
SAINT: probabilistic scoring of affinity purification–mass spectrometry data. Choi et al. (2010). Nature Methods. doi:10.1038/nmeth.1541
Abstract
We present 'significance analysis of interactome' (SAINT), a computational tool that assigns confidence scores to protein-protein interaction data generated using affinity purification-mass spectrometry (AP-MS). The method uses label-free quantitative data and constructs separate distributions for true and false interactions to derive the probability of a bona fide protein-protein interaction. We show that SAINT is applicable to data of different scales and protein connectivity and allows transparent analysis of AP-MS data.
The CRAPome: a contaminant repository for affinity purification–mass spectrometry data. Mellacheruvu et al. (2013). Nature Methods. doi:10.1038/nmeth.2557
Abstract
Affinity purification coupled with mass spectrometry (AP-MS) is a widely used approach for the identification of protein-protein interactions. However, for any given protein of interest, determining which of the identified polypeptides represent bona fide interactors versus those that are background contaminants (for example, proteins that interact with the solid-phase support, affinity reagent or epitope tag) is a challenging task. The standard approach is to identify nonspecific interactions using one or more negative-control purifications, but many small-scale AP-MS studies do not capture a complete, accurate background protein set when available controls are limited. Fortunately, negative controls are largely bait independent. Hence, aggregating negative controls from multiple AP-MS studies can increase coverage and improve the characterization of background associated with a given experimental protocol. Here we present the contaminant repository for affinity purification (the CRAPome) and describe its use for scoring protein-protein interactions. The repository (currently available for Homo sapiens and Saccharomyces cerevisiae) and computational tools are freely accessible at http://www.crapome.org/.
Roux et al. (2012) — BioID proximity labelling.
A promiscuous biotin ligase fusion protein identifies proximal and interacting proteins in mammalian cells. Roux et al. (2012). Journal of Cell Biology. doi:10.1083/jcb.201112098
Abstract
We have developed a new technique for proximity-dependent labeling of proteins in eukaryotic cells. Named BioID for proximity-dependent biotin identification, this approach is based on fusion of a promiscuous Escherichia coli biotin protein ligase to a targeting protein. BioID features proximity-dependent biotinylation of proteins that are near-neighbors of the fusion protein. Biotinylated proteins may be isolated by affinity capture and identified by mass spectrometry. We apply BioID to lamin-A (LaA), a well-characterized intermediate filament protein that is a constituent of the nuclear lamina, an important structural element of the nuclear envelope (NE). We identify multiple proteins that associate with and/or are proximate to LaA in vivo. The most abundant of these include known interactors of LaA that are localized to the NE, as well as a new NE-associated protein named SLAP75. Our results suggest BioID is a useful and generally applicable method to screen for both interacting and neighboring proteins in their native cellular environment.
Havugimana et al. (2012), Wan et al. (2015) — co-fractionation MS.
A Census of Human Soluble Protein Complexes. Havugimana et al. (2012). Cell. doi:10.1016/j.cell.2012.08.011
Abstract
Cellular processes often depend on stable physical associations between proteins. Despite recent progress, knowledge of the composition of human protein complexes remains limited. To close this gap, we applied an integrative global proteomic profiling approach, based on chromatographic separation of cultured human cell extracts into more than one thousand biochemical fractions that were subsequently analyzed by quantitative tandem mass spectrometry, to systematically identify a network of 13,993 high-confidence physical interactions among 3,006 stably associated soluble human proteins. Most of the 622 putative protein complexes we report are linked to core biological processes and encompass both candidate disease genes and unannotated proteins to inform on mechanism. Strikingly, whereas larger multiprotein assemblies tend to be more extensively annotated and evolutionarily conserved, human protein complexes with five or fewer subunits are far more likely to be functionally unannotated or restricted to vertebrates, suggesting more recent functional innovations.
Panorama of ancient metazoan macromolecular complexes. Wan et al. (2015). Nature. doi:10.1038/nature14877
Abstract
Macromolecular complexes are essential to conserved biological processes, but their prevalence across animals is unclear. By combining extensive biochemical fractionation with quantitative mass spectrometry, here we directly examined the composition of soluble multiprotein complexes among diverse metazoan models. Using an integrative approach, we generated a draft conservation map consisting of more than one million putative high-confidence co-complex interactions for species with fully sequenced genomes that encompasses functional modules present broadly across all extant animals. Clustering reveals a spectrum of conservation, ranging from ancient eukaryotic assemblies that have probably served cellular housekeeping roles for at least one billion years, ancestral complexes that have accrued contemporary components, and rarer metazoan innovations linked to multicellularity. We validated these projections by independent co-fractionation experiments in evolutionarily distant species, affinity purification and functional analyses. The comprehensiveness, centrality and modularity of these reconstructed interactomes reflect their fundamental mechanistic importance and adaptive value to animal cell systems.
Drew et al. (2021), Fischer et al. (2025) — hu.MAP 2.0 and 3.0.
hu.MAP 2.0: integration of over 15,000 proteomic experiments builds a global compendium of human multiprotein assemblies. Drew et al. (2021). Molecular Systems Biology. doi:10.15252/msb.202010016
Abstract
A general principle of biology is the self-assembly of proteins into functional complexes. Characterizing their composition is, therefore, required for our understanding of cellular functions. Unfortunately, we lack knowledge of the comprehensive set of identities of protein complexes in human cells. To address this gap, we developed a machine learning framework to identify protein complexes in over 15,000 mass spectrometry experiments which resulted in the identification of nearly 7,000 physical assemblies. We show our resource, hu.MAP 2.0, is more accurate and comprehensive than previous state of the art high-throughput protein complex resources and gives rise to many new hypotheses, including for 274 completely uncharacterized proteins. Further, we identify 253 promiscuous proteins that participate in multiple complexes pointing to possible moonlighting roles. We have made hu.MAP 2.0 easily searchable in a web interface (http://humap2.proteincomplexes.org/), which will be a valuable resource for researchers across a broad range of interests including systems biology, structural biology, and molecular explanations of disease.
hu.MAP3.0: atlas of human protein complexes by integration of >25,000 proteomic experiments. Fischer et al. (2025). Molecular Systems Biology. doi:10.1038/s44320-025-00121-5
Abstract
Macromolecular protein complexes carry out most cellular functions. Unfortunately, we lack the subunit composition for many human protein complexes. To address this gap we integrated >25,000 mass spectrometry experiments using a machine learning approach to identify >15,000 human protein complexes. We show our map of protein complexes is highly accurate and more comprehensive than previous maps, placing nearly 70% of human proteins into their physical contexts. We globally characterize our complexes using mass spectrometry based protein covariation data (ProteomeHD.2) and identify covarying complexes suggesting common functional associations. hu.MAP3.0 generates testable functional hypotheses for 472 uncharacterized proteins which we support using AlphaFold modeling. Additionally, we use AlphaFold modeling to identify 5871 mutually exclusive proteins in hu.MAP3.0 complexes suggesting complexes serve different functional roles depending on their subunit composition. We identify expression as the primary way cells and organisms relieve the conflict of mutually exclusive subunits. Finally, we import our complexes to EMBL-EBI's Complex Portal ( https://www.ebi.ac.uk/complexportal/home ) and provide complexes through our hu.MAP3.0 web interface ( https://humap3.proteincomplexes.org/ ). We expect our resource to be highly impactful to the broader research community.
O’Reilly and Rappsilber (2018) — cross-linking MS.
Cross-linking mass spectrometry: methods and applications in structural, molecular and systems biology. O’Reilly & Rappsilber (2018). Nature Structural & Molecular Biology. doi:10.1038/s41594-018-0147-0
Abstract
Over the past decade, cross-linking mass spectrometry (CLMS) has developed into a robust and flexible tool that provides medium-resolution structural information. CLMS data provide a measure of the proximity of amino acid residues and thus offer information on the folds of proteins and the topology of their complexes. Here, we highlight notable successes of this technique as well as common pipelines. Novel CLMS applications, such as in-cell cross-linking, probing conformational changes and tertiary-structure determination, are now beginning to make contributions to molecular biology and the emerging fields of structural systems biology and interactomics.
Schopper et al. (2017), Savitski et al. (2014) — LiP-MS and thermal proteome profiling.
Measuring protein structural changes on a proteome-wide scale using limited proteolysis-coupled mass spectrometry. Schopper et al. (2017). Nature Protocols. doi:10.1038/nprot.2017.100
Abstract
Protein structural changes induced by external perturbations or internal cues can profoundly influence protein activity and thus modulate cellular physiology. A number of biophysical approaches are available to probe protein structural changes, but these are not applicable to a whole proteome in a biological extract. Limited proteolysis-coupled mass spectrometry (LiP-MS) is a recently developed proteomics approach that enables the identification of protein structural changes directly in their complex biological context on a proteome-wide scale. After perturbations of interest, proteome extracts are subjected to a double-protease digestion step with a nonspecific protease applied under native conditions, followed by complete digestion with the sequence-specific protease trypsin under denaturing conditions. This sequential treatment generates structure-specific peptides amenable to bottom-up MS analysis. Next, a proteomics workflow involving shotgun or targeted MS and label-free quantification is applied to measure structure-dependent proteolytic patterns directly in the proteome extract. Possible applications of LiP-MS include discovery of perturbation-induced protein structural alterations, identification of drug targets, detection of disease-associated protein structural states, and analysis of protein aggregates directly in biological samples. The approach also enables identification of the specific protein regions involved in the structural transition or affected by the binding event. Sample preparation takes approximately 2 d, followed by one to several days of MS and data analysis time, depending on the number of samples analyzed. Scientists with basic biochemistry training can implement the sample preparation steps. MS measurement and data analysis require a background in proteomics.
Tracking cancer drugs in living cells by thermal profiling of the proteome. Savitski et al. (2014). Science. doi:10.1126/science.1255784
Abstract
The thermal stability of proteins can be used to assess ligand binding in living cells. We have generalized this concept by determining the thermal profiles of more than 7000 proteins in human cells by means of mass spectrometry. Monitoring the effects of small-molecule ligands on the profiles delineated more than 50 targets for the kinase inhibitor staurosporine. We identified the heme biosynthesis enzyme ferrochelatase as a target of kinase inhibitors and suggest that its inhibition causes the phototoxicity observed with vemurafenib and alectinib. Thermal shifts were also observed for downstream effectors of drug treatment. In live cells, dasatinib induced shifts in BCR-ABL pathway proteins, including CRK/CRKL. Thermal proteome profiling provides an unbiased measure of drug-target engagement and facilitates identification of markers for drug efficacy and toxicity.
Christoforou et al. (2016), Thul et al. (2017) — spatial proteomics of organelles and cells.
A draft map of the mouse pluripotent stem cell spatial proteome. Christoforou et al. (2016). Nature Communications. doi:10.1038/ncomms9992
Abstract
Knowledge of the subcellular distribution of proteins is vital for understanding cellular mechanisms. Capturing the subcellular proteome in a single experiment has proven challenging, with studies focusing on specific compartments or assigning proteins to subcellular niches with low resolution and/or accuracy. Here we introduce hyperLOPIT, a method that couples extensive fractionation, quantitative high-resolution accurate mass spectrometry with multivariate data analysis. We apply hyperLOPIT to a pluripotent stem cell population whose subcellular proteome has not been extensively studied. We provide localization data on over 5,000 proteins with unprecedented spatial resolution to reveal the organization of organelles, sub-organellar compartments, protein complexes, functional networks and steady-state dynamics of proteins and unexpected subcellular locations. The method paves the way for characterizing the impact of post-transcriptional and post-translational modification on protein location and studies involving proteome-level locational changes on cellular perturbation. An interactive open-source resource is presented that enables exploration of these data.
A subcellular map of the human proteome. Thul et al. (2017). Science. doi:10.1126/science.aal3321
Abstract
Resolving the spatial distribution of the human proteome at a subcellular level can greatly increase our understanding of human biology and disease. Here we present a comprehensive image-based map of subcellular protein distribution, the Cell Atlas, built by integrating transcriptomics and antibody-based immunofluorescence microscopy with validation by mass spectrometry. Mapping the in situ localization of 12,003 human proteins at a single-cell level to 30 subcellular structures enabled the definition of the proteomes of 13 major organelles. Exploration of the proteomes revealed single-cell variations in abundance or spatial distribution and localization of about half of the proteins to multiple compartments. This subcellular map can be used to refine existing protein-protein interaction networks and provides an important resource to deconvolute the highly complex architecture of the human cell.
Caprioli et al. (1997), Mund et al. (2022), Budnik et al. (2018) — MS imaging, deep visual proteomics and single-cell proteomics.
Molecular Imaging of Biological Samples: Localization of Peptides and Proteins Using MALDI-TOF MS. Caprioli et al. (1997). Analytical Chemistry. doi:10.1021/ac970888i
Abstract
Matrix-assisted laser desorption/ionization mass spectrometry (MALDI MS) has been used to generate ion images of samples in one or more mass-to-charge (m/z) values, providing the capability of mapping specific molecules to two-dimensional coordinates of the original sample. The high sensitivity of the technique (low-femtomole to attomole levels for proteins and peptides) allows the study of organized biochemical processes occurring in, for example, mammalian tissue sections. The mass spectrometer is used to determine the molecular weights of the molecular in the surface layers of the tissue. Molecules desorbed from the sample typically are singly protonated, giving an ion at (M + H)+, where M is the molecular mass. The procedure involves coating the tissue section, or a blotted imprint of the section, with a thin layer of energy-absorbing matrix and then analyzing the sample to produce an ordered array of mass spectra, each containing nominal m/z values typically covering a range of over 50,000 Da. Images can be displayed in individual m/z values as a selected ion image, which would localize individual compounds in the tissue, or as summed ion images. MALDI ion images of tissue sections can be obtained directly from tissue slices following preparative steps, and this is demonstrated for the mapping of insulin contained in an islet in a section of rat pancreas, hormone peptides in a small area of a section of rat pituitary, and a small protein bound to the membrane of human mucosa cells. Alternatively, imprints of the tissue can be analyzed by blotting the tissue sections on specially prepared targets containing an adsorbent material, e.g., C-18 coated resin beads. Peptides and small proteins bind to the C-18 and create a positive imprint of the tissue which can then be imaged by the mass spectrometer. This is demonstrated for the MALDI ion image analysis of regions of rat splenic pancreas and for an area of rat pituitary traversing the anterior, intermediate, and posterior regions where localized peptides were mapped. In a single spectrum from the anterior/intermediate lobe of a rat pituitary print, over 50 ions corresponding to the peptides present in this tissue were observed as well as precursors, isoforms, and metabolic fragments.
Deep Visual Proteomics defines single-cell identity and heterogeneity. Mund et al. (2022). Nature Biotechnology. doi:10.1038/s41587-022-01302-5
Abstract
Despite the availabilty of imaging-based and mass-spectrometry-based methods for spatial proteomics, a key challenge remains connecting images with single-cell-resolution protein abundance measurements. Here, we introduce Deep Visual Proteomics (DVP), which combines artificial-intelligence-driven image analysis of cellular phenotypes with automated single-cell or single-nucleus laser microdissection and ultra-high-sensitivity mass spectrometry. DVP links protein abundance to complex cellular or subcellular phenotypes while preserving spatial context. By individually excising nuclei from cell culture, we classified distinct cell states with proteomic profiles defined by known and uncharacterized proteins. In an archived primary melanoma tissue, DVP identified spatially resolved proteome changes as normal melanocytes transition to fully invasive melanoma, revealing pathways that change in a spatial manner as cancer progresses, such as mRNA splicing dysregulation in metastatic vertical growth that coincides with reduced interferon signaling and antigen presentation. The ability of DVP to retain precise spatial proteomic information in the tissue context has implications for the molecular profiling of clinical samples.
SCoPE-MS: mass spectrometry of single mammalian cells quantifies proteome heterogeneity during cell differentiation. Budnik et al. (2018). Genome Biology. doi:10.1186/s13059-018-1547-5
Abstract
Some exciting biological questions require quantifying thousands of proteins in single cells. To achieve this goal, we develop Single Cell ProtEomics by Mass Spectrometry (SCoPE-MS) and validate its ability to identify distinct human cancer cell types based on their proteomes. We use SCoPE-MS to quantify over a thousand proteins in differentiating mouse embryonic stem cells. The single-cell proteomes enable us to deconstruct cell populations and infer protein abundance relationships. Comparison between single-cell proteomes and transcriptomes indicates coordinated mRNA and protein covariation, yet many genes exhibit functionally concerted and distinct regulatory patterns at the mRNA and the protein level.
B.8.7 Non-MS proteomics and resources
Thul et al. (2017), Uhlen et al. (2017), Karlsson et al. (2021) — the Human Protein Atlas.
A subcellular map of the human proteome. Thul et al. (2017). Science. doi:10.1126/science.aal3321
Abstract
Resolving the spatial distribution of the human proteome at a subcellular level can greatly increase our understanding of human biology and disease. Here we present a comprehensive image-based map of subcellular protein distribution, the Cell Atlas, built by integrating transcriptomics and antibody-based immunofluorescence microscopy with validation by mass spectrometry. Mapping the in situ localization of 12,003 human proteins at a single-cell level to 30 subcellular structures enabled the definition of the proteomes of 13 major organelles. Exploration of the proteomes revealed single-cell variations in abundance or spatial distribution and localization of about half of the proteins to multiple compartments. This subcellular map can be used to refine existing protein-protein interaction networks and provides an important resource to deconvolute the highly complex architecture of the human cell.
A pathology atlas of the human cancer transcriptome. Uhlen et al. (2017). Science. doi:10.1126/science.aan2507
Abstract
Cancer is one of the leading causes of death, and there is great interest in understanding the underlying molecular mechanisms involved in the pathogenesis and progression of individual tumors. We used systems-level approaches to analyze the genome-wide transcriptome of the protein-coding genes of 17 major cancer types with respect to clinical outcome. A general pattern emerged: Shorter patient survival was associated with up-regulation of genes involved in cell growth and with down-regulation of genes involved in cellular differentiation. Using genome-scale metabolic models, we show that cancer patients have widespread metabolic heterogeneity, highlighting the need for precise and personalized medicine for cancer treatment. All data are presented in an interactive open-access database (www.proteinatlas.org/pathology) to allow genome-wide exploration of the impact of individual proteins on clinical outcomes.
A single–cell type transcriptomics map of human tissues. Karlsson et al. (2021). Science Advances. doi:10.1126/sciadv.abh2169
Abstract
Advances in molecular profiling have opened up the possibility to map the expression of genes in cells, tissues, and organs in the human body. Here, we combined single-cell transcriptomics analysis with spatial antibody-based protein profiling to create a high-resolution single-cell type map of human tissues. An open access atlas has been launched to allow researchers to explore the expression of human protein-coding genes in 192 individual cell type clusters. An expression specificity classification was performed to determine the number of genes elevated in each cell type, allowing comparisons with bulk transcriptomics data. The analysis highlights distinct expression clusters corresponding to cell types sharing similar functions, both within the same organs and between organs.
Uhlen et al. (2016) — validation of antibodies.
A proposal for validation of antibodies. Uhlen et al. (2016). Nature Methods. doi:10.1038/nmeth.3995
Assarsson et al. (2014), Gold et al. (2010) — PEA and SomaScan.
Homogenous 96-Plex PEA Immunoassay Exhibiting High Sensitivity, Specificity, and Excellent Scalability. Assarsson et al. (2014). PLoS ONE. doi:10.1371/journal.pone.0095192
Abstract
Medical research is developing an ever greater need for comprehensive high-quality data generation to realize the promises of personalized health care based on molecular biomarkers. The nucleic acid proximity-based methods proximity ligation and proximity extension assays have, with their dual reporters, shown potential to relieve the shortcomings of antibodies and their inherent cross-reactivity in multiplex protein quantification applications. The aim of the present study was to develop a robust 96-plex immunoassay based on the proximity extension assay (PEA) for improved high throughput detection of protein biomarkers. This was enabled by: (1) a modified design leading to a reduced number of pipetting steps compared to the existing PEA protocol, as well as improved intra-assay precision; (2) a new enzymatic system that uses a hyper-thermostabile enzyme, Pwo, for uniting the two probes allowing for room temperature addition of all reagents and improved the sensitivity; (3) introduction of an inter-plate control and a new normalization procedure leading to improved inter-assay precision (reproducibility). The multiplex proximity extension assay was found to perform well in complex samples, such as serum and plasma, and also in xenografted mice and resuspended dried blood spots, consuming only 1 µL sample per test. All-in-all, the development of the current multiplex technique is a step toward robust high throughput protein marker discovery and research.
Aptamer-Based Multiplexed Proteomic Technology for Biomarker Discovery. Gold et al. (2010). PLoS ONE. doi:10.1371/journal.pone.0015004
Abstract
Background The interrogation of proteomes ("proteomics") in a highly multiplexed and efficient manner remains a coveted and challenging goal in biology and medicine. Methodology/principal findings We present a new aptamer-based proteomic technology for biomarker discovery capable of simultaneously measuring thousands of proteins from small sample volumes (15 µL of serum or plasma). Our current assay measures 813 proteins with low limits of detection (1 pM median), 7 logs of overall dynamic range (~100 fM-1 µM), and 5% median coefficient of variation. This technology is enabled by a new generation of aptamers that contain chemically modified nucleotides, which greatly expand the physicochemical diversity of the large randomized nucleic acid libraries from which the aptamers are selected. Proteins in complex matrices such as plasma are measured with a process that transforms a signature of protein concentrations into a corresponding signature of DNA aptamer concentrations, which is quantified on a DNA microarray. Our assay takes advantage of the dual nature of aptamers as both folded protein-binding entities with defined shapes and unique nucleotide sequences recognizable by specific hybridization probes. To demonstrate the utility of our proteomics biomarker discovery technology, we applied it to a clinical study of chronic kidney disease (CKD). We identified two well known CKD biomarkers as well as an additional 58 potential CKD biomarkers. These results demonstrate the potential utility of our technology to rapidly discover unique protein signatures characteristic of various disease states. Conclusions/significance We describe a versatile and powerful tool that allows large-scale comparison of proteome profiles among discrete populations. This unbiased and highly multiplexed search engine will enable the discovery of novel biomarkers in a manner that is unencumbered by our incomplete knowledge of biology, thereby helping to advance the next generation of evidence-based medicine.
Sun et al. (2023) — plasma proteomics in the UK Biobank.
Plasma proteomic associations with genetics and health in the UK Biobank. Sun et al. (2023). Nature. doi:10.1038/s41586-023-06592-6
Abstract
The Pharma Proteomics Project is a precompetitive biopharmaceutical consortium characterizing the plasma proteomic profiles of 54,219 UK Biobank participants. Here we provide a detailed summary of this initiative, including technical and biological validations, insights into proteomic disease signatures, and prediction modelling for various demographic and health indicators. We present comprehensive protein quantitative trait locus (pQTL) mapping of 2,923 proteins that identifies 14,287 primary genetic associations, of which 81% are previously undescribed, alongside ancestry-specific pQTL mapping in non-European individuals. The study provides an updated characterization of the genetic architecture of the plasma proteome, contextualized with projected pQTL discovery rates as sample sizes and proteomic assay coverages increase over time. We offer extensive insights into trans pQTLs across multiple biological domains, highlight genetic influences on ligand-receptor interactions and pathway perturbations across a diverse collection of cytokines and complement networks, and illustrate long-range epistatic effects of ABO blood group and FUT2 secretor status on proteins with gastrointestinal tissue-enriched expression. We demonstrate the utility of these data for drug discovery by extending the genetic proxied effects of protein targets, such as PCSK9, on additional endpoints, and disentangle specific genes and proteins perturbed at loci associated with COVID-19 susceptibility. This public-private partnership provides the scientific community with an open-access proteomics resource of considerable breadth and depth to help to elucidate the biological mechanisms underlying proteo-genomic discoveries and accelerate the development of biomarkers, predictive models and therapeutics 1 .
Goltsev et al. (2018), Giesen et al. (2014) — CODEX and imaging mass cytometry.
Deep Profiling of Mouse Splenic Architecture with CODEX Multiplexed Imaging. Goltsev et al. (2018). Cell. doi:10.1016/j.cell.2018.07.010
Abstract
A highly multiplexed cytometric imaging approach, termed co-detection by indexing (CODEX), is used here to create multiplexed datasets of normal and lupus (MRL/lpr) murine spleens. CODEX iteratively visualizes antibody binding events using DNA barcodes, fluorescent dNTP analogs, and an in situ polymerization-based indexing procedure. An algorithmic pipeline for single-cell antigen quantification in tightly packed tissues was developed and used to overlay well-known morphological features with de novo characterization of lymphoid tissue architecture at a single-cell and cellular neighborhood levels. We observed an unexpected, profound impact of the cellular neighborhood on the expression of protein receptors on immune cells. By comparing normal murine spleen to spleens from animals with systemic autoimmune disease (MRL/lpr), extensive and previously uncharacterized splenic cell-interaction dynamics in the healthy versus diseased state was observed. The fidelity of multiplexed spatial cytometry demonstrated here allows for quantitative systemic characterization of tissue architecture in normal and clinically aberrant samples.
Highly multiplexed imaging of tumor tissues with subcellular resolution by mass cytometry. Giesen et al. (2014). Nature Methods. doi:10.1038/nmeth.2869
Abstract
Mass cytometry enables high-dimensional, single-cell analysis of cell type and state. In mass cytometry, rare earth metals are used as reporters on antibodies. Analysis of metal abundances using the mass cytometer allows determination of marker expression in individual cells. Mass cytometry has previously been applied only to cell suspensions. To gain spatial information, we have coupled immunohistochemical and immunocytochemical methods with high-resolution laser ablation to CyTOF mass cytometry. This approach enables the simultaneous imaging of 32 proteins and protein modifications at subcellular resolution; with the availability of additional isotopes, measurement of over 100 markers will be possible. We applied imaging mass cytometry to human breast cancer samples, allowing delineation of cell subpopulations and cell-cell interactions and highlighting tumor heterogeneity. Imaging mass cytometry complements existing imaging approaches. It will enable basic studies of tissue heterogeneity and function and support the transition of medicine toward individualized molecularly targeted diagnosis and therapies.
Alfaro et al. (2021) — single-molecule protein sequencing.
The emerging landscape of single-molecule protein sequencing technologies. Alfaro et al. (2021). Nature Methods. doi:10.1038/s41592-021-01143-1
Abstract
Single-cell profiling methods have had a profound impact on the understanding of cellular heterogeneity. While genomes and transcriptomes can be explored at the single-cell level, single-cell profiling of proteomes is not yet established. Here we describe new single-molecule protein sequencing and identification technologies alongside innovations in mass spectrometry that will eventually enable broad sequence coverage in single-cell profiling. These technologies will in turn facilitate biological discovery and open new avenues for ultrasensitive disease diagnostics.
Perez-Riverol et al. (2024), Vizcaíno et al. (2014), Deutsch et al. (2022), Desiere (2006) — PRIDE, ProteomeXchange and PeptideAtlas.
The PRIDE database at 20 years: 2025 update. Perez-Riverol et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae1011
Abstract
The PRoteomics IDEntifications (PRIDE) database (https://www.ebi.ac.uk/pride/) is the world's leading mass spectrometry (MS)-based proteomics data repository and one of the founding members of the ProteomeXchange consortium. This manuscript summarizes the developments in PRIDE resources and related tools for the last three years. The number of submitted datasets to PRIDE Archive (the archival component of PRIDE) has reached on average around 534 datasets per month. This has been possible thanks to continuous improvements in infrastructure such as a new file transfer protocol for very large datasets (Globus), a new data resubmission pipeline and an automatic dataset validation process. Additionally, we will highlight novel activities such as the availability of the PRIDE chatbot (based on the use of open-source Large Language Models), and our work to improve support for MS crosslinking datasets. Furthermore, we will describe how we have increased our efforts to reuse, reanalyze and disseminate high-quality proteomics data into added-value resources such as UniProt, Ensembl and Expression Atlas.
ProteomeXchange provides globally coordinated proteomics data submission and dissemination. Vizcaíno et al. (2014). Nature Biotechnology. doi:10.1038/nbt.2839
The ProteomeXchange consortium at 10 years: 2023 update. Deutsch et al. (2022). Nucleic Acids Research. doi:10.1093/nar/gkac1040
Abstract
Mass spectrometry (MS) is by far the most used experimental approach in high-throughput proteomics. The ProteomeXchange (PX) consortium of proteomics resources (http://www.proteomexchange.org) was originally set up to standardize data submission and dissemination of public MS proteomics data. It is now 10 years since the initial data workflow was implemented. In this manuscript, we describe the main developments in PX since the previous update manuscript in Nucleic Acids Research was published in 2020. The six members of the Consortium are PRIDE, PeptideAtlas (including PASSEL), MassIVE, jPOST, iProX and Panorama Public. We report the current data submission statistics, showcasing that the number of datasets submitted to PX resources has continued to increase every year. As of June 2022, more than 34 233 datasets had been submitted to PX resources, and from those, 20 062 (58.6%) just in the last three years. We also report the development of the Universal Spectrum Identifiers and the improvements in capturing the experimental metadata annotations. In parallel, we highlight that data re-use activities of public datasets continue to increase, enabling connections between PX resources and other popular bioinformatics resources, novel research and also new data resources. Finally, we summarise the current state-of-the-art in data management practices for sensitive human (clinical) proteomics data.
The PeptideAtlas project. Desiere (2006). Nucleic Acids Research. doi:10.1093/nar/gkj040
Abstract
The completion of the sequencing of the human genome and the concurrent, rapid development of high-throughput proteomic methods have resulted in an increasing need for automated approaches to archive proteomic data in a repository that enables the exchange of data among researchers and also accurate integration with genomic data. PeptideAtlas (http://www.peptideatlas.org/) addresses these needs by identifying peptides by tandem mass spectrometry (MS/MS), statistically validating those identifications and then mapping identified sequences to the genomes of eukaryotic organisms. A meaningful comparison of data across different experiments generated by different groups using different types of instruments is enabled by the implementation of a uniform analytic process. This uniform statistical validation ensures a consistent and high-quality set of peptide and protein identifications. The raw data from many diverse proteomic experiments are made available in the associated PeptideAtlas repository in several formats. Here we present a summary of our process and details about the Human, Drosophila and Yeast PeptideAtlas builds.
Martens et al. (2011), Taylor et al. (2007) — mzML and MIAPE.
mzML—a Community Standard for Mass Spectrometry Data. Martens et al. (2011). Molecular & Cellular Proteomics. doi:10.1074/mcp.r110.000133
Abstract
Mass spectrometry is a fundamental tool for discovery and analysis in the life sciences. With the rapid advances in mass spectrometry technology and methods, it has become imperative to provide a standard output format for mass spectrometry data that will facilitate data sharing and analysis. Initially, the efforts to develop a standard format for mass spectrometry data resulted in multiple formats, each designed with a different underlying philosophy. To resolve the issues associated with having multiple formats, vendors, researchers, and software developers convened under the banner of the HUPO PSI to develop a single standard. The new data format incorporated many of the desirable technical attributes from the previous data formats, while adding a number of improvements, including features such as a controlled vocabulary with validation tools to ensure consistent usage of the format, improved support for selected reaction monitoring data, and immediately available implementations to facilitate rapid adoption by the community. The resulting standard data format, mzML, is a well tested open-source format for mass spectrometer output files that can be readily utilized by the community and easily adapted for incremental advances in mass spectrometry technology.
The minimum information about a proteomics experiment (MIAPE). Taylor et al. (2007). Nature Biotechnology. doi:10.1038/nbt1329
Abstract
Both the generation and the analysis of proteomics data are now widespread, and high-throughput approaches are commonplace. Protocols continue to increase in complexity as methods and technologies evolve and diversify. To encourage the standardized collection, integration, storage and dissemination of proteomics data, the Human Proteome Organization's Proteomics Standards Initiative develops guidance modules for reporting the use of techniques such as gel electrophoresis and mass spectrometry. This paper describes the processes and principles underpinning the development of these modules; discusses the ramifications for various interest groups such as experimentalists, funders, publishers and the private sector; addresses the issue of overlap with other reporting guidelines; and highlights the criticality of appropriate tools and resources in enabling 'MIAPE-compliant' reporting.
Kim et al. (2014), Wilhelm et al. (2014), Ezkurdia et al. (2014) — draft maps of the human proteome and their reanalysis.
A draft map of the human proteome. Kim et al. (2014). Nature. doi:10.1038/nature13302
Abstract
The availability of human genome sequence has transformed biomedical research over the past decade. However, an equivalent map for the human proteome with direct measurements of proteins and peptides does not exist yet. Here we present a draft map of the human proteome using high-resolution Fourier-transform mass spectrometry. In-depth proteomic profiling of 30 histologically normal human samples, including 17 adult tissues, 7 fetal tissues and 6 purified primary haematopoietic cells, resulted in identification of proteins encoded by 17,294 genes accounting for approximately 84% of the total annotated protein-coding genes in humans. A unique and comprehensive strategy for proteogenomic analysis enabled us to discover a number of novel protein-coding regions, which includes translated pseudogenes, non-coding RNAs and upstream open reading frames. This large human proteome catalogue (available as an interactive web-based resource at http://www.humanproteomemap.org) will complement available human genome and transcriptome data to accelerate biomedical research in health and disease.
Mass-spectrometry-based draft of the human proteome. Wilhelm et al. (2014). Nature. doi:10.1038/nature13319
Abstract
Proteomes are characterized by large protein-abundance differences, cell-type- and time-dependent expression patterns and post-translational modifications, all of which carry biological information that is not accessible by genomics or transcriptomics. Here we present a mass-spectrometry-based draft of the human proteome and a public, high-performance, in-memory database for real-time analysis of terabytes of big data, called ProteomicsDB. The information assembled from human tissues, cell lines and body fluids enabled estimation of the size of the protein-coding genome, and identified organ-specific proteins and a large number of translated lincRNAs (long intergenic non-coding RNAs). Analysis of messenger RNA and protein-expression profiles of human tissues revealed conserved control of protein abundance, and integration of drug-sensitivity data enabled the identification of proteins predicting resistance or sensitivity. The proteome profiles also hold considerable promise for analysing the composition and stoichiometry of protein complexes. ProteomicsDB thus enables navigation of proteomes, provides biological insight and fosters the development of proteomic technology.
Analyzing the First Drafts of the Human Proteome. Ezkurdia et al. (2014). Journal of Proteome Research. doi:10.1021/pr500572z
Abstract
This letter analyzes two large-scale proteomics studies published in the same issue of Nature. At the time of the release, both studies were portrayed as draft maps of the human proteome and great advances in the field. As with the initial publication of the human genome, these papers have broad appeal and will no doubt lead to a great deal of further analysis by the scientific community. However, we were intrigued by the number of protein-coding genes detected by the two studies, numbers that far exceeded what has been reported for the multinational Human Proteome Project effort. We carried out a simple quality test on the data using the olfactory receptor family. A high-quality proteomics experiment that does not specifically analyze nasal tissues should not expect to detect many peptides for olfactory receptors. Neither of the studies carried out experiments on nasal tissues, yet we found peptide evidence for more than 100 olfactory receptors in the two studies. These results suggest that the two studies are substantially overestimating the number of protein coding genes they identify. We conclude that the experimental data from these two studies should be used with caution.
Legrain et al. (2011), Omenn et al. (2024) — the Human Proteome Project.
The Human Proteome Project: Current State and Future Direction. Legrain et al. (2011). Molecular & Cellular Proteomics. doi:10.1074/mcp.m111.009993
Abstract
After the successful completion of the Human Genome Project, the Human Proteome Organization has recently officially launched a global Human Proteome Project (HPP), which is designed to map the entire human protein set. Given the lack of protein-level evidence for about 30% of the estimated 20,300 protein-coding genes, a systematic global effort will be necessary to achieve this goal with respect to protein abundance, distribution, subcellular localization, interaction with other biomolecules, and functions at specific time points. As a general experimental strategy, HPP research groups will use the three working pillars for HPP: mass spectrometry, antibody capture, and bioinformatics tools and knowledge bases. The HPP participants will take advantage of the output and cross-analyses from the ongoing Human Proteome Organization initiatives and a chromosome-centric protein mapping strategy, termed C-HPP, with which many national teams are currently engaged. In addition, numerous biologically driven and disease-oriented projects will be stimulated and facilitated by the HPP. Timely planning with proper governance of HPP will deliver a protein parts list, reagents, and tools for protein studies and analyses, and a stronger basis for personalized medicine. The Human Proteome Organization urges each national research funding agency and the scientific community at large to identify their preferred pathways to participate in aspects of this highly promising project in a HPP consortium of funders and investigators.
The 2023 Report on the Proteome from the HUPO Human Proteome Project. Omenn et al. (2024). Journal of Proteome Research. doi:10.1021/acs.jproteome.3c00591
Abstract
Since 2010, the Human Proteome Project (HPP), the flagship initiative of the Human Proteome Organization (HUPO), has pursued two goals: (1) to credibly identify the protein parts list and (2) to make proteomics an integral part of multiomics studies of human health and disease. The HPP relies on international collaboration, data sharing, standardized reanalysis of MS data sets by PeptideAtlas and MassIVE-KB using HPP Guidelines for quality assurance, integration and curation of MS and non-MS protein data by neXtProt, plus extensive use of antibody profiling carried out by the Human Protein Atlas. According to the neXtProt release 2023-04-18, protein expression has now been credibly detected (PE1) for 18,397 of the 19,778 neXtProt predicted proteins coded in the human genome (93%). Of these PE1 proteins, 17,453 were detected with mass spectrometry (MS) in accordance with HPP Guidelines and 944 by a variety of non-MS methods. The number of neXtProt PE2, PE3, and PE4 missing proteins now stands at 1381. Achieving the unambiguous identification of 93% of predicted proteins encoded from across all chromosomes represents remarkable experimental progress on the Human Proteome parts list. Meanwhile, there are several categories of predicted proteins that have proved resistant to detection regardless of protein-based methods used. Additionally there are some PE1-4 proteins that probably should be reclassified to PE5, specifically 21 LINC entries and ∼30 HERV entries; these are being addressed in the present year. Applying proteomics in a wide array of biological and clinical studies ensures integration with other omics platforms as reported by the Biology and Disease-driven HPP teams and the antibody and pathology resource pillars. Current progress has positioned the HPP to transition to its Grand Challenge Project focused on determining the primary function(s) of every protein itself and in networks and pathways within the context of human health and disease.
Ideker et al. (2001) — integrated genomic and proteomic analysis of a perturbed network.
Integrated Genomic and Proteomic Analyses of a Systematically Perturbed Metabolic Network. Ideker et al. (2001). Science. doi:10.1126/science.292.5518.929
Abstract
We demonstrate an integrated approach to build, test, and refine a model of a cellular pathway, in which perturbations to critical pathway components are analyzed using DNA microarrays, quantitative proteomics, and databases of known physical interactions. Using this approach, we identify 997 messenger RNAs responding to 20 systematic perturbations of the yeast galactose-utilization pathway, provide evidence that approximately 15 of 289 detected proteins are regulated posttranscriptionally, and identify explicit physical interactions governing the cellular response to each perturbation. We refine the model through further iterations of perturbation and global measurements, suggesting hypotheses about the regulation of galactose utilization and physical interactions between this and a variety of other metabolic pathways.
B.9 Protein structure
Further papers for Chapter 9.
Papers marked ★ were mandatory in earlier reading lists. In the HTML version, click Abstract to show the abstract.
B.9.1 CASP and evaluation
Moult et al. (1995) — the first CASP.
A large-scale experiment to assess protein structure prediction methods. Moult et al. (1995). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.340230303
Kryshtafovych et al. (2021), Kryshtafovych et al. (2023), Yuan et al. (2025) — CASP14, CASP15 and CASP16.
Critical assessment of methods of protein structure prediction (CASP)—Round XIV. Kryshtafovych et al. (2021). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.26237
Abstract
Critical assessment of structure prediction (CASP) is a community experiment to advance methods of computing three-dimensional protein structure from amino acid sequence. Core components are rigorous blind testing of methods and evaluation of the results by independent assessors. In the most recent experiment (CASP14), deep-learning methods from one research group consistently delivered computed structures rivaling the corresponding experimental ones in accuracy. In this sense, the results represent a solution to the classical protein-folding problem, at least for single proteins. The models have already been shown to be capable of providing solutions for problematic crystal structures, and there are broad implications for the rest of structural biology. Other research groups also substantially improved performance. Here, we describe these results and outline some of the many implications. Other related areas of CASP, including modeling of protein complexes, structure refinement, estimation of model accuracy, and prediction of inter-residue contacts and distances, are also described.
Critical assessment of methods of protein structure prediction ( CASP )—Round XV. Kryshtafovych et al. (2023). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.26617
Abstract
Computing protein structure from amino acid sequence information has been a long-standing grand challenge. Critical assessment of structure prediction (CASP) conducts community experiments aimed at advancing solutions to this and related problems. Experiments are conducted every 2 years. The 2020 experiment (CASP14) saw major progress, with the second generation of deep learning methods delivering accuracy comparable with experiment for many single proteins. There is an expectation that these methods will have much wider application in computational structural biology. Here we summarize results from the most recent experiment, CASP15, in 2022, with an emphasis on new deep learning-driven progress. Other papers in this special issue of proteins provide more detailed analysis. For single protein structures, the AlphaFold2 deep learning method is still superior to other approaches, but there are two points of note. First, although AlphaFold2 was the core of all the most successful methods, there was a wide variety of implementation and combination with other methods. Second, using the standard AlphaFold2 protocol and default parameters only produces the highest quality result for about two thirds of the targets, and more extensive sampling is required for the others. The major advance in this CASP is the enormous increase in the accuracy of computed protein complexes, achieved by the use of deep learning methods, although overall these do not fully match the performance for single proteins. Here too, AlphaFold2 based method perform best, and again more extensive sampling than the defaults is often required. Also of note are the encouraging early results on the use of deep learning to compute ensembles of macromolecular structures. Critically for the usability of computed structures, for both single proteins and protein complexes, deep learning derived estimates of both local and global accuracy are of high quality, however the estimates in interface regions are slightly less reliable. CASP15 also included computation of RNA structures for the first time. Here, the classical approaches produced better agreement with experiment than the new deep learning ones, and accuracy is limited. Also, for the first time, CASP included the computation of protein-ligand complexes, an area of special interest for drug design. Here too, classical methods were still superior to deep learning ones. Many new approaches were discussed at the CASP conference, and it is clear methods will continue to advance.
CASP16 Protein Monomer Structure Prediction Assessment. Yuan et al. (2025). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.70031
Abstract
The assessment of monomer targets in the Critical Assessment of Structure Prediction Round 16 (CASP16) underscores that the problem of single-domain protein fold prediction is nearly solved-no target folds were incorrectly predicted across all Evaluation Units. However, challenges remain in accurately modeling truncated sequences, irregular secondary structures, and interaction-induced conformational changes. The release of AlphaFold3 (AF3) during CASP16, and its effective integration by many groups, demonstrated its superiority over AlphaFold2 (AF2), particularly in confidence estimation and model selection. Additional improvements in multiple sequence alignments (MSAs) and fragment-based prediction, that is, selecting the optimal fragment of the full sequence for modeling, also contributed to enhanced prediction accuracy. The top three groups-all from the Yang lab-consistently outperformed others across CASP16 monomer targets, reflecting their robust modeling pipelines and successful adoption of AF3. CASP16 also introduced three new challenges: Phase 0, in which stoichiometry was withheld; Phase 2, which supplied ~8000 MassiveFold models per target to test model selection strategies; and Model 6, which limited predictors to using MSAs provided by the organizers. While we evaluated group performance in these additional challenges, the insights gained were limited due to low participation and caveats in the design of experiments. We suggest improvements for the organization of these challenges and encourage broader engagement from the prediction community. The progress in monomer modeling from CASP15 to CASP16 was subtle, but more groups in CASP16 were able to outperform ColabFold, reflecting the community's improved ability in optimizing AF2 and the growing adoption of AF3. We anticipate that the recent release of the AF3 source code will stimulate future progress through user-driven optimization and innovations in model architecture. Finally, model ranking remains a persistent weakness across most groups, highlighting a critical area for future development.
Jumper, Evans, Pritzel, Green, Figurnov, Ronneberger, Tunyasuvunakool, Bates, Žídek, Potapenko, Bridgland, Meyer, Kohl, Ballard, Cowie, Romera-Paredes, Nikolov, Jain, Adler, Back, Petersen, Reiman, Clancy, Zielinski, Steinegger, Pacholska, Berghammer, Silver, et al. (2021) — AlphaFold2 at CASP14.
Applying and improving AlphaFold at CASP14. Jumper et al. (2021). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.26257
Abstract
We describe the operation and improvement of AlphaFold, the system that was entered by the team AlphaFold2 to the "human" category in the 14th Critical Assessment of Protein Structure Prediction (CASP14). The AlphaFold system entered in CASP14 is entirely different to the one entered in CASP13. It used a novel end-to-end deep neural network trained to produce protein structures from amino acid sequence, multiple sequence alignments, and homologous proteins. In the assessors' ranking by summed z scores (>2.0), AlphaFold scored 244.0 compared to 90.8 by the next best group. The predictions made by AlphaFold had a median domain GDT_TS of 92.4; this is the first time that this level of average accuracy has been achieved during CASP, especially on the more difficult Free Modeling targets, and represents a significant improvement in the state of the art in protein structure prediction. We reported how AlphaFold was run as a human team during CASP14 and improved such that it now achieves an equivalent level of performance without intervention, opening the door to highly accurate large-scale structure prediction.
Zemla (2003), Zhang and Skolnick (2004), Mariani et al. (2013), Basu and Wallner (2016) — GDT, TM-score, lDDT and DockQ.
LGA: a method for finding 3D similarities in protein structures. Zemla (2003). Nucleic Acids Research. doi:10.1093/nar/gkg571
Abstract
We present the LGA (Local-Global Alignment) method, designed to facilitate the comparison of protein structures or fragments of protein structures in sequence dependent and sequence independent modes. The LGA structure alignment program is available as an online service at http://PredictionCenter.llnl.gov/local/lga. Data generated by LGA can be successfully used in a scoring function to rank the level of similarity between two structures and to allow structure classification when many proteins are being analyzed. LGA also allows the clustering of similar fragments of protein structures.
Scoring function for automated assessment of protein structure template quality. Zhang & Skolnick (2004). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.20264
Abstract
We have developed a new scoring function, the template modeling score (TM-score), to assess the quality of protein structure templates and predicted full-length models by extending the approaches used in Global Distance Test (GDT)1 and MaxSub.2 First, a protein size-dependent scale is exploited to eliminate the inherent protein size dependence of the previous scores and appropriately account for random protein structure pairs. Second, rather than setting specific distance cutoffs and calculating only the fractions with errors below the cutoff, all residue pairs in alignment/modeling are evaluated in the proposed score. For comparison of various scoring functions, we have constructed a large-scale benchmark set of structure templates for 1489 small to medium size proteins using the threading program PROSPECTOR_3 and built the full-length models using MODELLER and TASSER. The TM-score of the initial threading alignments, compared to the GDT and MaxSub scoring functions, shows a much stronger correlation to the quality of the final full-length models. The TM-score is further exploited as an assessment of all 'new fold' targets in the recent CASP5 experiment and shows a close coincidence with the results of human-expert visual assessment. These data suggest that the TM-score is a useful complement to the fully automated assessment of protein structure predictions. The executable program of TM-score is freely downloadable at http://bioinformatics.buffalo.edu/TM-score.
lDDT: a local superposition-free score for comparing protein structures and models using distance difference tests. Mariani et al. (2013). Bioinformatics. doi:10.1093/bioinformatics/btt473
Abstract
Motivation The assessment of protein structure prediction techniques requires objective criteria to measure the similarity between a computational model and the experimentally determined reference structure. Conventional similarity measures based on a global superposition of carbon α atoms are strongly influenced by domain motions and do not assess the accuracy of local atomic details in the model. Results The Local Distance Difference Test (lDDT) is a superposition-free score that evaluates local distance differences of all atoms in a model, including validation of stereochemical plausibility. The reference can be a single structure, or an ensemble of equivalent structures. We demonstrate that lDDT is well suited to assess local model quality, even in the presence of domain movements, while maintaining good correlation with global measures. These properties make lDDT a robust tool for the automated assessment of structure prediction servers without manual intervention. Availability and implementation Source code, binaries for Linux and MacOSX, and an interactive web server are available at http://swissmodel.expasy.org/lddt. Contact torsten.schwede@unibas.ch. Supplementary information Supplementary data are available at Bioinformatics online.
DockQ: A Quality Measure for Protein-Protein Docking Models. Basu & Wallner (2016). PLOS ONE. doi:10.1371/journal.pone.0161879
Abstract
The state-of-the-art to assess the structural quality of docking models is currently based on three related yet independent quality measures: Fnat, LRMS, and iRMS as proposed and standardized by CAPRI. These quality measures quantify different aspects of the quality of a particular docking model and need to be viewed together to reveal the true quality, e.g. a model with relatively poor LRMS (>10Å) might still qualify as 'acceptable' with a descent Fnat (>0.50) and iRMS (<3.0Å). This is also the reason why the so called CAPRI criteria for assessing the quality of docking models is defined by applying various ad-hoc cutoffs on these measures to classify a docking model into the four classes: Incorrect, Acceptable, Medium, or High quality. This classification has been useful in CAPRI, but since models are grouped in only four bins it is also rather limiting, making it difficult to rank models, correlate with scoring functions or use it as target function in machine learning algorithms. Here, we present DockQ, a continuous protein-protein docking model quality measure derived by combining Fnat, LRMS, and iRMS to a single score in the range [0, 1] that can be used to assess the quality of protein docking models. By using DockQ on CAPRI models it is possible to almost completely reproduce the original CAPRI classification into Incorrect, Acceptable, Medium and High quality. An average PPV of 94% at 90% Recall demonstrating that there is no need to apply predefined ad-hoc cutoffs to classify docking models. Since DockQ recapitulates the CAPRI classification almost perfectly, it can be viewed as a higher resolution version of the CAPRI classification, making it possible to estimate model quality in a more quantitative way using Z-scores or sum of top ranked models, which has been so valuable for the CASP community. The possibility to directly correlate a quality measure to a scoring function has been crucial for the development of scoring functions for protein structure prediction, and DockQ should be useful in a similar development in the protein docking field. DockQ is available at http://github.com/bjornwallner/DockQ/.
Janin et al. (2003), Haas et al. (2017) — CAPRI and CAMEO.
CAPRI: A Critical Assessment of PRedicted Interactions. Janin et al. (2003). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.10381
Abstract
CAPRI is a communitywide experiment to assess the capacity of protein-docking methods to predict protein-protein interactions. Nineteen groups participated in rounds 1 and 2 of CAPRI and submitted blind structure predictions for seven protein-protein complexes based on the known structure of the component proteins. The predictions were compared to the unpublished X-ray structures of the complexes. We describe here the motivations for launching CAPRI, the rules that we applied to select targets and run the experiment, and some conclusions that can already be drawn. The results stress the need for new scoring functions and for methods handling the conformation changes that were observed in some of the target systems. CAPRI has already been a powerful drive for the community of computational biologists who development docking algorithms. We hope that this issue of Proteins will also be of interest to the community of structural biologists, which we call upon to provide new targets for future rounds of CAPRI, and to all molecular biologists who view protein-protein recognition as an essential process.
Continuous Automated Model EvaluatiOn (CAMEO) complementing the critical assessment of structure prediction in CASP12. Haas et al. (2017). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.25431
Abstract
Every second year, the community experiment "Critical Assessment of Techniques for Structure Prediction" (CASP) is conducting an independent blind assessment of structure prediction methods, providing a framework for comparing the performance of different approaches and discussing the latest developments in the field. Yet, developers of automated computational modeling methods clearly benefit from more frequent evaluations based on larger sets of data. The "Continuous Automated Model EvaluatiOn (CAMEO)" platform complements the CASP experiment by conducting fully automated blind prediction assessments based on the weekly pre-release of sequences of those structures, which are going to be published in the next release of the PDB Protein Data Bank. CAMEO publishes weekly benchmarking results based on models collected during a 4-day prediction window, on average assessing ca. 100 targets during a time frame of 5 weeks. CAMEO benchmarking data is generated consistently for all participating methods at the same point in time, enabling developers to benchmark and cross-validate their method's performance, and directly refer to the benchmarking results in publications. In order to facilitate server development and promote shorter release cycles, CAMEO sends weekly email with submission statistics and low performance warnings. Many participants of CASP have successfully employed CAMEO when preparing their methods for upcoming community experiments. CAMEO offers a variety of scores to allow benchmarking diverse aspects of structure prediction methods. By introducing new scoring schemes, CAMEO facilitates new development in areas of active research, for example, modeling quaternary structure, complexes, or ligand binding sites.
B.9.2 Structure and its determination
Pauling et al. (1951), Ramachandran et al. (1963) — the α-helix and the Ramachandran plot.
The structure of proteins: Two hydrogen-bonded helical configurations of the polypeptide chain. Pauling et al. (1951). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.37.4.205
Stereochemistry of polypeptide chain configurations. Ramachandran et al. (1963). Journal of Molecular Biology. doi:10.1016/s0022-2836(63)80023-6
Kendrew et al. (1958), PERUTZ et al. (1960) — the first protein structures.
A Three-Dimensional Model of the Myoglobin Molecule Obtained by X-Ray Analysis. Kendrew et al. (1958). Nature. doi:10.1038/181662a0
Structure of Hæmoglobin: A Three-Dimensional Fourier Synthesis at 5.5-Å. Resolution, Obtained by X-Ray Analysis. PERUTZ et al. (1960). Nature. doi:10.1038/185416a0
Chothia et al. (1977), Levitt and Chothia (1976), Richardson (1981) — packing, structural classes and the anatomy of proteins.
Structure of proteins: packing of alpha-helices and pleated sheets.. Chothia et al. (1977). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.74.10.4130
Abstract
Simple models are presented that describe the rules for almost all the packing that occurs between and among alpha-helices and pleated sheets. These packing rules, together with the primary and secondary structures, are the major determinants of the three-dimensional structure of proteins.
Structural patterns in globular proteins. Levitt & Chothia (1976). Nature. doi:10.1038/261552a0
Abstract
A simple diagrammatic representation has been used to show the arrangement of alpha helices and beta sheets in 31 globular proteins, which are classified into four clearly separated classes. The observed arrangements are significantly non-random in that pieces of secondary structure adjacent in sequence along the polypeptide chain are also often in contact in three dimensions.
The Anatomy and Taxonomy of Protein Structure. Richardson (1981). Advances in Protein Chemistry Volume 34. doi:10.1016/s0065-3233(08)60520-3
Williamson et al. (1985) — the first NMR structure of a protein.
Solution conformation of proteinase inhibitor IIA from bull seminal plasma by 1H nuclear magnetic resonance and distance geometry. Williamson et al. (1985). Journal of Molecular Biology. doi:10.1016/0022-2836(85)90347-x
Abstract
A determination of the solution conformation of the proteinase inhibitor IIA from bull seminal plasma (BUSI IIA) is described. Two-dimensional nuclear Overhauser enhancement spectroscopy (NOESY) was used to obtain a list of 202 distance constraints between individually assigned hydrogen atoms of the polypeptide chain, to identify the positions of the three disulfide bridges, and to locate the single cis peptide bond. Supplementary geometric constraints were derived from the vicinal spin-spin couplings and the locations of certain hydrogen bonds, as determined by nuclear magnetic resonance (n.m.r.). Using a new distance geometry program (DISGEO) which is capable of computing all-atom structures for proteins the size of BUSI IIA, five conformers were computed from the NOE distance constraints alone, and another five were computed with the supplementary constraints included. Comparison of the different structures computed from the n.m.r. data among themselves and with the crystal structures of two homologous proteins shows that the global features of the conformation of BUSI IIA (i.e. the overall dimensions of the molecule and the threading of the polypeptide chain) were well-defined by the available n.m.r. data. In the Appendix, we describe a preliminary energy refinement of the structure, which showed that the constraints derived from the n.m.r. data are compatible with a low energy spatial structure.
Brünger (1992) — the free R value.
Free R value: a novel statistical quantity for assessing the accuracy of crystal structures. Brünger (1992). Nature. doi:10.1038/355472a0
Abstract
The determination of macromolecular structure by crystallography involves fitting atomic models to the observed diffraction data. The traditional measure of the quality of this fit, and presumably the accuracy of the model, is the R value. Despite stereochemical restraints, it is possible to overfit or 'misfit' the diffraction data: an incorrect model can be refined to fairly good R values as several recent examples have shown. Here I propose a reliable and unbiased indicator of the accuracy of such models. By analogy with the cross-validation method of testing statistical models I define a statistical quantity (R(free) (T) that measures the agreement between observed and computed structure factor amplitudes for a 'test' set of reflections that is omitted in the modelling and refinement process. As examples show, there is a high correlation between R(free) (T) and the accuracy of the atomic model phases. This is useful because experimental phase information is usually inaccurate, incomplete or unavailable. I expect that R(free) (T) will provide a measure of the information content of recently proposed models of thermal motion and disorder, time-averaging and bulk solvent.
Henderson et al. (1990), Kühlbrandt (2014), Scheres (2012) — cryo-EM and the resolution revolution.
Model for the structure of bacteriorhodopsin based on high-resolution electron cryo-microscopy. Henderson et al. (1990). Journal of Molecular Biology. doi:10.1016/s0022-2836(05)80271-2
Abstract
The light-driven proton pump bacteriorhodopsin occurs naturally as two-dimensional crystals. A three-dimensional density map of the structure, at near-atomic resolution, has been obtained by studying the crystals using electron cryo-microscopy to obtain electron diffraction patterns and high-resolution micrographs. New methods were developed for analysing micrographs from tilted specimens, incorporating methods previously developed for untilted specimens that enable large areas to be analysed and corrected for distortions. Data from 72 images, from both tilted and untilted specimens, were analysed to produce the phases of 2700 independent Fourier components of the structure. The amplitudes of these components were accurately measured from 150 diffraction patterns. Together, these data represent about half of the full three-dimensional transform to 3.5 A. The map of the structure has a resolution of 3.5 A in a direction parallel to the membrane plane but lower than this in the perpendicular direction. It shows many features in the density that are resolved from the main density of the seven alpha-helices. We interpret these features as the bulky aromatic side-chains of phenylalanine, tyrosine and tryptophan residues. There is also a very dense feature, which is the beta-ionone ring of the retinal chromophore. Using these bulky side-chains as guide points and taking account of bulges in the helices that indicate smaller side-chains such as leucine, a complete atomic model for bacteriorhodopsin between amino acid residues 8 and 225 has been built. There are 21 amino acid residues, contributed by all seven helices, surrounding the retinal and 26 residues, contributed by five helices, forming the proton pathway or channel. Ten of the amino acid residues in the middle of the proton channel are also part of the retinal binding site. The model also provides a useful basis for consideration of the mechanism of proton pumping and allows a consistent interpretation of a great deal of other experimental data. In particular, the structure suggests that pK changes in the Schiff base must act as the means by which light energy is converted into proton pumping pressure in the channel. Asp96 is on the pathway from the cytoplasm to the Schiff base and Asp85 is on the pathway from the Schiff base to the extracellular surface.
The Resolution Revolution. Kühlbrandt (2014). Science. doi:10.1126/science.1251652
Abstract
Advances in detector technology and image processing are yielding high-resolution electron cryo-microscopy structures of biomolecules. [Also see Report by Amunts et al. ]
A Bayesian View on Cryo-EM Structure Determination. Scheres (2012). Journal of Molecular Biology. doi:10.1016/j.jmb.2011.11.010
Abstract
Three-dimensional (3D) structure determination by single-particle analysis of cryo-electron microscopy (cryo-EM) images requires many parameters to be determined from extremely noisy data. This makes the method prone to overfitting, that is, when structures describe noise rather than signal, in particular near their resolution limit where noise levels are highest. Cryo-EM structures are typically filtered using ad hoc procedures to prevent overfitting, but the tuning of arbitrary parameters may lead to subjectivity in the results. I describe a Bayesian interpretation of cryo-EM structure determination, where smoothness in the reconstructed density is imposed through a Gaussian prior in the Fourier domain. The statistical framework dictates how data and prior knowledge should be combined, so that the optimal 3D linear filter is obtained without the need for arbitrariness and objective resolution estimates may be obtained. Application to experimental data indicates that the statistical approach yields more reliable structures than existing methods and is capable of detecting smaller classes in data sets that contain multiple different structures.
B.9.3 Folding
Levinthal (1966) — Levinthal’s paradox.
Molecular Model-Building by Computer. Levinthal (1966). Scientific American. doi:10.1038/scientificamerican0666-42
Jackson and Fersht (1991) — two-state folding.
Folding of chymotrypsin inhibitor 2. 1. Evidence for a two-state transition. Jackson & Fersht (1991). Biochemistry. doi:10.1021/bi00107a010
Abstract
The reversible folding and unfolding of barley chymotrypsin inhibitor 2 (CI2) appears to be a rare example in which both equilibria and kinetics are described by a two-state model. Equilibrium denaturation by guanidinium chloride and heat is completely reversible, and the data can be fitted to a simple two-state model involving only native and denatured forms. The free energy of folding in the absence of denaturant, delta GH2O, at pH 6.3, is calculated to be 7.03 +/- 0.16 and 7.18 +/- 0.43 kcal mol-1 for guanidinium chloride and thermal denaturation, respectively. Scanning microcalorimetry shows that the ratio of the van't Hoff enthalpy of denaturation to the calorimetric enthalpy of denaturation does not deviate from unity, the value observed for a two-state transition, over the pH range 2.2-3.5. The heat capacity change for denaturation is found to be 0.789 kcal mol-1 K-1. The rate of unfolding of CI2 is first order and increases exponentially with increasing guanidinium chloride concentration. Refolding, however, is complex and involves at least three well-resolved phases. The three phases result from heterogeneity of the unfolded form due to proline isomerization. The fast phase, 77% of the amplitude, corresponds to the refolding of the fraction of the protein that has all its prolines in a native trans conformation. The rate of this major phase decreases exponentially with increasing guanidinium chloride concentration. The unfolding and refolding kinetics can also be fitted to a two-state model.(ABSTRACT TRUNCATED AT 250 WORDS)
Bryngelson et al. (1995), Dill and Chan (1997) — energy landscapes and funnels.
Funnels, pathways, and the energy landscape of protein folding: A synthesis. Bryngelson et al. (1995). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.340210302
Abstract
The understanding, and even the description of protein folding is impeded by the complexity of the process. Much of this complexity can be described and understood by taking a statistical approach to the energetics of protein conformation, that is, to the energy landscape. The statistical energy landscape approach explains when and why unique behaviors, such as specific folding pathways, occur in some proteins and more generally explains the distinction between folding processes common to all sequences and those peculiar to individual sequences. This approach also gives new, quantitative insights into the interpretation of experiments and simulations of protein folding thermodynamics and kinetics. Specifically, the picture provides simple explanations for folding as a two-state first-order phase transition, for the origin of metastable collapsed unfolded states and for the curved Arrhenius plots observed in both laboratory experiments and discrete lattice simulations. The relation of these quantitative ideas to folding pathways, to uniexponential vs. multiexponential behavior in protein folding experiments and to the effect of mutations on folding is also discussed. The success of energy landscape ideas in protein structure prediction is also described. The use of the energy landscape approach for analyzing data is illustrated with a quantitative analysis of some recent simulations, and a qualitative analysis of experiments on the folding of three proteins. The work unifies several previously proposed ideas concerning the mechanism protein folding and delimits the regions of validity of these ideas under different thermodynamic conditions.
From Levinthal to pathways to funnels. Dill & Chan (1997). Nature Structural & Molecular Biology. doi:10.1038/nsb0197-10
Abstract
While the classical view of protein folding kinetics relies on phenomenological models, and regards folding intermediates in a structural way, the new view emphasizes the ensemble nature of protein conformations. Although folding has sometimes been regarded as a linear sequence of events, the new view sees folding as parallel microscopic multi-pathway diffusion-like processes. While the classical view invoked pathways to solve the problem of searching for the needle in the haystack, the pathway idea was then seen as conflicting with Anfinsen's experiments showing that folding is pathway-independent (Levinthal's paradox). In contrast, the new view sees no inherent paradox because it eliminates the pathway idea: folding can funnel to a single stable state by multiple routes in conformational space. The general energy landscape picture provides a conceptual framework for understanding both two-state and multi-state folding kinetics. Better tests of these ideas will come when new experiments become available for measuring not just averages of structural observables, but also correlations among their fluctuations. At that point we hope to learn much more about the real shapes of protein folding landscapes.
Dill and MacCallum (2012) — the protein-folding problem, 50 years on.
The Protein-Folding Problem, 50 Years On. Dill & MacCallum (2012). Science. doi:10.1126/science.1219021
Abstract
The protein-folding problem was first posed about one half-century ago. The term refers to three broad questions: (i) What is the physical code by which an amino acid sequence dictates a protein's native structure? (ii) How can proteins fold so fast? (iii) Can we devise a computer algorithm to predict protein structures from their sequences? We review progress on these problems. In a few cases, computer simulations of the physical forces in chemically detailed models have now achieved the accurate folding of small proteins. We have learned that proteins fold rapidly because random thermal motions cause conformational changes leading energetically downhill toward the native structure, a principle that is captured in funnel-shaped energy landscapes. And thanks in part to the large Protein Data Bank of known structures, predicting protein structures is now far more successful than was thought possible in the early days. What began as three questions of basic science one half-century ago has now grown into the full-fledged research field of protein physical science.
Shaw et al. (2010), Lindorff-Larsen et al. (2011) — folding by molecular dynamics.
Atomic-Level Characterization of the Structural Dynamics of Proteins. Shaw et al. (2010). Science. doi:10.1126/science.1187409
Abstract
Molecular dynamics (MD) simulations are widely used to study protein motions at an atomic level of detail, but they have been limited to time scales shorter than those of many biologically critical conformational changes. We examined two fundamental processes in protein dynamics--protein folding and conformational change within the folded state--by means of extremely long all-atom MD simulations conducted on a special-purpose machine. Equilibrium simulations of a WW protein domain captured multiple folding and unfolding events that consistently follow a well-defined folding pathway; separate simulations of the protein's constituent substructures shed light on possible determinants of this pathway. A 1-millisecond simulation of the folded protein BPTI reveals a small number of structurally distinct conformational states whose reversible interconversion is slower than local relaxations within those states by a factor of more than 1000.
How Fast-Folding Proteins Fold. Lindorff-Larsen et al. (2011). Science. doi:10.1126/science.1208351
Abstract
An outstanding challenge in the field of molecular biology has been to understand the process by which proteins fold into their characteristic three-dimensional structures. Here, we report the results of atomic-level molecular dynamics simulations, over periods ranging between 100 μs and 1 ms, that reveal a set of common principles underlying the folding of 12 structurally diverse proteins. In simulations conducted with a single physics-based energy function, the proteins, representing all three major structural classes, spontaneously and repeatedly fold to their experimentally determined native structures. Early in the folding process, the protein backbone adopts a nativelike topology while certain secondary structure elements and a small number of nonlocal contacts form. In most cases, folding follows a single dominant route in which elements of the native structure appear in an order highly correlated with their propensity to form in the unfolded state.
B.9.4 Comparison and classification
Chothia and Lesk (1986) — structure is more conserved than sequence.
The relation between the divergence of sequence and structure in proteins.. Chothia & Lesk (1986). The EMBO Journal. doi:10.1002/j.1460-2075.1986.tb04288.x
Abstract
Homologous proteins have regions which retain the same general fold and regions where the folds differ. For pairs of distantly related proteins (residue identity approximately 20%), the regions with the same fold may comprise less than half of each molecule. The regions with the same general fold differ in structure by amounts that increase as the amino acid sequences diverge. The root mean square deviation in the positions of the main chain atoms, delta, is related to the fraction of mutated residues, H, by the expression: delta(A) = 0.40 e1.87H.
Holm and Sander (1993), Holm and Sander (1996), Holm (2019) — DALI and the map of the protein universe.
Protein Structure Comparison by Alignment of Distance Matrices. Holm & Sander (1993). Journal of Molecular Biology. doi:10.1006/jmbi.1993.1489
Abstract
With a rapidly growing pool of known tertiary structures, the importance of protein structure comparison parallels that of sequence alignment. We have developed a novel algorithm (DALI) for optimal pairwise alignment of protein structures. The three-dimensional co-ordinates of each protein are used to calculate residue-residue (C alpha-C alpha) distance matrices. The distance matrices are first decomposed into elementary contact patterns, e.g. hexapeptide-hexapeptide submatrices. Then, similar contact patterns in the two matrices are paired and combined into larger consistent sets of pairs. A Monte Carlo procedure is used to optimize a similarity score defined in terms of equivalent intramolecular distances. Several alignments are optimized in parallel, leading to simultaneous detection of the best, second-best and so on solutions. The method allows sequence gaps of any length, reversal of chain direction and free topological connectivity of aligned segments. Sequential connectivity can be imposed as an option. The method is fully automatic and identifies structural resemblances and common structural cores accurately and sensitively, even in the presence of geometrical distortions. An all-against-all alignment of over 200 representative protein structures results in an objective classification of known three-dimensional folds in agreement with visual classifications. Unexpected topological similarities of biological interest have been detected, e.g. between the bacterial toxin colicin A and globins, and between the eukaryotic POU-specific DNA-binding domain and the bacterial lambda repressor.
Mapping the Protein Universe. Holm & Sander (1996). Science. doi:10.1126/science.273.5275.595
Abstract
The comparison of the three-dimensional shapes of protein molecules poses a complex algorithmic problem. Its solution provides biologists with computational tools to organize the rapidly growing set of thousands of known protein shapes, to identify new types of protein architecture, and to discover unexpected evolutionary relations, reaching back billions of years, between protein molecules. Protein shape comparison also improves tools for identifying gene functions in genome databases by defining the essential sequence-structure features of a protein family. Finally, an exhaustive all-on-all shape comparison provides a map of physical attractor regions in the abstract shape space of proteins, with implications for the processes of protein folding and evolution.
DALI and the persistence of protein shape. Holm (2019). Protein Science. doi:10.1002/pro.3749
Abstract
DALI is a popular resource for comparing protein structures. The software is based on distance-matrix alignment. The associated web server provides tools to navigate, integrate and organize some data pushed out by genomics and structural genomics. The server has been running continuously for the past 25 years. Structural biologists routinely use DALI to compare a new structure against previously known protein structures. If significant similarities are discovered, it may indicate a distant homology, that is, that the structures are of shared origin. This may be significant in determining the molecular mechanisms, as these may remain very similar from a distant predecessor to the present day, for example, from the last common ancestor of humans and bacteria. Meta-analysis of independent reference-based evaluations of alignment accuracy and fold discrimination shows DALI at top rank in six out of 12 studies. The web server and standalone software are available from http://ekhidna2.biocenter.helsinki.fi/dali.
Zhang and Skolnick (2005), Kolodny et al. (2005) — TM-align and the evaluation of structural aligners.
TM-align: a protein structure alignment algorithm based on the TM-score. Zhang & Skolnick (2005). Nucleic Acids Research. doi:10.1093/nar/gki524
Abstract
We have developed TM-align, a new algorithm to identify the best structural alignment between protein pairs that combines the TM-score rotation matrix and Dynamic Programming (DP). The algorithm is approximately 4 times faster than CE and 20 times faster than DALI and SAL. On average, the resulting structure alignments have higher accuracy and coverage than those provided by these most often-used methods. TM-align is applied to an all-against-all structure comparison of 10 515 representative protein chains from the Protein Data Bank (PDB) with a sequence identity cutoff <95%: 1996 distinct folds are found when a TM-score threshold of 0.5 is used. We also use TM-align to match the models predicted by TASSER for solved non-homologous proteins in PDB. For both folded and misfolded models, TM-align can almost always find close structural analogs, with an average root mean square deviation, RMSD, of 3 A and 87% alignment coverage. Nevertheless, there exists a significant correlation between the correctness of the predicted structure and the structural similarity of the model to the other proteins in the PDB. This correlation could be used to assist in model selection in blind protein structure predictions. The TM-align program is freely downloadable at http://bioinformatics.buffalo.edu/TM-align.
Comprehensive Evaluation of Protein Structure Alignment Methods: Scoring by Geometric Measures. Kolodny et al. (2005). Journal of Molecular Biology. doi:10.1016/j.jmb.2004.12.032
Abstract
We report the largest and most comprehensive comparison of protein structural alignment methods. Specifically, we evaluate six publicly available structure alignment programs: SSAP, STRUCTAL, DALI, LSQMAN, CE and SSM by aligning all 8,581,970 protein structure pairs in a test set of 2930 protein domains specially selected from CATH v.2.4 to ensure sequence diversity. We consider an alignment good if it matches many residues, and the two substructures are geometrically similar. Even with this definition, evaluating structural alignment methods is not straightforward. At first, we compared the rates of true and false positives using receiver operating characteristic (ROC) curves with the CATH classification taken as a gold standard. This proved unsatisfactory in that the quality of the alignments is not taken into account: sometimes a method that finds less good alignments scores better than a method that finds better alignments. We correct this intrinsic limitation by using four different geometric match measures (SI, MI, SAS, and GSAS) to evaluate the quality of each structural alignment. With this improved analysis we show that there is a wide variation in the performance of different methods; the main reason for this is that it can be difficult to find a good structural alignment between two proteins even when such an alignment exists. We find that STRUCTAL and SSM perform best, followed by LSQMAN and CE. Our focus on the intrinsic quality of each alignment allows us to propose a new method, called "Best-of-All" that combines the best results of all methods. Many commonly used methods miss 10-50% of the good Best-of-All alignments. By putting existing structural alignments into proper perspective, our study allows better comparison of protein structures. By highlighting limitations of existing methods, it will spur the further development of better structural alignment methods. This will have significant biological implications now that structural comparison has come to play a central role in the analysis of experimental work on protein structure, protein function and protein evolution.
van Kempen et al. (2023), Barrio-Hernandez et al. (2023) — Foldseek and the clustering of the AlphaFold Database.
Fast and accurate protein structure search with Foldseek. van Kempen et al. (2023). Nature Biotechnology. doi:10.1038/s41587-023-01773-0
Abstract
As structure prediction methods are generating millions of publicly available protein structures, searching these databases is becoming a bottleneck. Foldseek aligns the structure of a query protein against a database by describing tertiary amino acid interactions within proteins as sequences over a structural alphabet. Foldseek decreases computation times by four to five orders of magnitude with 86%, 88% and 133% of the sensitivities of Dali, TM-align and CE, respectively.
Clustering predicted structures at the scale of the known protein universe. Barrio-Hernandez et al. (2023). Nature. doi:10.1038/s41586-023-06510-w
Abstract
Proteins are key to all cellular processes and their structure is important in understanding their function and evolution. Sequence-based predictions of protein structures have increased in accuracy 1 , and over 214 million predicted structures are available in the AlphaFold database 2 . However, studying protein structures at this scale requires highly efficient methods. Here, we developed a structural-alignment-based clustering algorithm-Foldseek cluster-that can cluster hundreds of millions of structures. Using this method, we have clustered all of the structures in the AlphaFold database, identifying 2.30 million non-singleton structural clusters, of which 31% lack annotations representing probable previously undescribed structures. Clusters without annotation tend to have few representatives covering only 4% of all proteins in the AlphaFold database. Evolutionary analysis suggests that most clusters are ancient in origin but 4% seem to be species specific, representing lower-quality predictions or examples of de novo gene birth. We also show how structural comparisons can be used to predict domain families and their relationships, identifying examples of remote structural similarity. On the basis of these analyses, we identify several examples of human immune-related proteins with putative remote homology in prokaryotic species, illustrating the value of this resource for studying protein function and evolution across the tree of life.
Chothia (1992) — one thousand families.
One thousand families for the molecular biologist. Chothia (1992). Nature. doi:10.1038/357543a0
Murzin et al. (1995), Sillitoe et al. (2020), Cheng et al. (2014), Lau et al. (2024) — SCOP, CATH, ECOD and TED.
SCOP: A structural classification of proteins database for the investigation of sequences and structures. Murzin et al. (1995). Journal of Molecular Biology. doi:10.1016/s0022-2836(05)80134-2
CATH: increased structural coverage of functional space. Sillitoe et al. (2020). Nucleic Acids Research. doi:10.1093/nar/gkaa1079
Abstract
CATH (https://www.cathdb.info) identifies domains in protein structures from wwPDB and classifies these into evolutionary superfamilies, thereby providing structural and functional annotations. There are two levels: CATH-B, a daily snapshot of the latest domain structures and superfamily assignments, and CATH+, with additional derived data, such as predicted sequence domains, and functionally coherent sequence subsets (Functional Families or FunFams). The latest CATH+ release, version 4.3, significantly increases coverage of structural and sequence data, with an addition of 65,351 fully-classified domains structures (+15%), providing 500 238 structural domains, and 151 million predicted sequence domains (+59%) assigned to 5481 superfamilies. The FunFam generation pipeline has been re-engineered to cope with the increased influx of data. Three times more sequences are captured in FunFams, with a concomitant increase in functional purity, information content and structural coverage. FunFam expansion increases the structural annotations provided for experimental GO terms (+59%). We also present CATH-FunVar web-pages displaying variations in protein sequences and their proximity to known or predicted functional sites. We present two case studies (1) putative cancer drivers and (2) SARS-CoV-2 proteins. Finally, we have improved links to and from CATH including SCOP, InterPro, Aquaria and 2DProt.
ECOD: An Evolutionary Classification of Protein Domains. Cheng et al. (2014). PLoS Computational Biology. doi:10.1371/journal.pcbi.1003926
Abstract
Understanding the evolution of a protein, including both close and distant relationships, often reveals insight into its structure and function. Fast and easy access to such up-to-date information facilitates research. We have developed a hierarchical evolutionary classification of all proteins with experimentally determined spatial structures, and presented it as an interactive and updatable online database. ECOD (Evolutionary Classification of protein Domains) is distinct from other structural classifications in that it groups domains primarily by evolutionary relationships (homology), rather than topology (or "fold"). This distinction highlights cases of homology between domains of differing topology to aid in understanding of protein structure evolution. ECOD uniquely emphasizes distantly related homologs that are difficult to detect, and thus catalogs the largest number of evolutionary links among structural domain classifications. Placing distant homologs together underscores the ancestral similarities of these proteins and draws attention to the most important regions of sequence and structure, as well as conserved functional sites. ECOD also recognizes closer sequence-based relationships between protein domains. Currently, approximately 100,000 protein structures are classified in ECOD into 9,000 sequence families clustered into close to 2,000 evolutionary groups. The classification is assisted by an automated pipeline that quickly and consistently classifies weekly releases of PDB structures and allows for continual updates. This synchronization with PDB uniquely distinguishes ECOD among all protein classifications. Finally, we present several case studies of homologous proteins not recorded in other classifications, illustrating the potential of how ECOD can be used to further biological and evolutionary studies.
Exploring structural diversity across the protein universe with The Encyclopedia of Domains. Lau et al. (2024). Science. doi:10.1126/science.adq4946
Abstract
The AlphaFold Protein Structure Database (AFDB) contains more than 214 million predicted protein structures composed of domains, which are independently folding units found in multiple structural and functional contexts. Identifying domains can enable many functional and evolutionary analyses but has remained challenging because of the sheer scale of the data. Using deep learning methods, we have detected and classified every domain in the AFDB, producing The Encyclopedia of Domains. We detected nearly 365 million domains, over 100 million more than can be found by sequence methods, covering more than 1 million taxa. Reassuringly, 77% of the nonredundant domains are similar to known superfamilies, greatly expanding representation of their domain space. We uncovered more than 10,000 new structural interactions between superfamilies and thousands of new folds across the fold space continuum.
Varadi et al. (2023), Bertoni et al. (2025) — the AlphaFold Database.
AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences. Varadi et al. (2023). Nucleic Acids Research. doi:10.1093/nar/gkad1011
Abstract
The AlphaFold Database Protein Structure Database (AlphaFold DB, https://alphafold.ebi.ac.uk) has significantly impacted structural biology by amassing over 214 million predicted protein structures, expanding from the initial 300k structures released in 2021. Enabled by the groundbreaking AlphaFold2 artificial intelligence (AI) system, the predictions archived in AlphaFold DB have been integrated into primary data resources such as PDB, UniProt, Ensembl, InterPro and MobiDB. Our manuscript details subsequent enhancements in data archiving, covering successive releases encompassing model organisms, global health proteomes, Swiss-Prot integration, and a host of curated protein datasets. We detail the data access mechanisms of AlphaFold DB, from direct file access via FTP to advanced queries using Google Cloud Public Datasets and the programmatic access endpoints of the database. We also discuss the improvements and services added since its initial release, including enhancements to the Predicted Aligned Error viewer, customisation options for the 3D viewer, and improvements in the search engine of AlphaFold DB.
AlphaFold Protein Structure Database 2025: a redesigned interface and updated structural coverage. Bertoni et al. (2025). Nucleic Acids Research. doi:10.1093/nar/gkaf1226
Abstract
The AlphaFold Protein Structure Database (AFDB; https://alphafold.ebi.ac.uk), developed by EMBL-EBI and Google DeepMind, provides open access to hundreds of millions of high-accuracy protein structure predictions, transforming research in structural biology and the wider life sciences. Since its launch, AFDB has become a widely used bioinformatics resource, integrated into major databases, visualization platforms, and analysis pipelines. Here, we report the update of the database to align with the UniProt 2025_03 release, along with a comprehensive redesign of the entry page to enhance usability, accessibility, and structural interpretation. The new design integrates annotations directly with an interactive 3D viewer and introduces dedicated domains and summary tabs. Structural coverage has also been updated to include isoforms plus underlying multiple sequence alignments. Data are available through the website, FTP, Google Cloud, and updated APIs. Together, these advances reinforce AFDB as a sustainable resource for exploring protein sequence-structure relationships.
B.9.5 Structure prediction before deep learning
Chou and Fasman (1974a), Chou and Fasman (1974b), Garnier et al. (1978) — early secondary structure prediction.
Conformational parameters for amino acids in helical, β-sheet, and random coil regions calculated from proteins. Chou & Fasman (1974). Biochemistry. doi:10.1021/bi00699a001
Prediction of protein conformation. Chou & Fasman (1974). Biochemistry. doi:10.1021/bi00699a002
Analysis of the accuracy and implications of simple methods for predicting the secondary structure of globular proteins. Garnier et al. (1978). Journal of Molecular Biology. doi:10.1016/0022-2836(78)90297-8
★ Rost and Sander (1993), Jones (1999) — PHD and PSIPRED.
Prediction of Protein Secondary Structure at Better than 70% Accuracy. Rost & Sander (1993). Journal of Molecular Biology. doi:10.1006/jmbi.1993.1413
Abstract
We have trained a two-layered feed-forward neural network on a non-redundant data base of 130 protein chains to predict the secondary structure of water-soluble proteins. A new key aspect is the use of evolutionary information in the form of multiple sequence alignments that are used as input in place of single sequences. The inclusion of protein family information in this form increases the prediction accuracy by six to eight percentage points. A combination of three levels of networks results in an overall three-state accuracy of 70.8% for globular proteins (sustained performance). If four membrane protein chains are included in the evaluation, the overall accuracy drops to 70.2%. The prediction is well balanced between alpha-helix, beta-strand and loop: 65% of the observed strand residues are predicted correctly. The accuracy in predicting the content of three secondary structure types is comparable to that of circular dichroism spectroscopy. The performance accuracy is verified by a sevenfold cross-validation test, and an additional test on 26 recently solved proteins. Of particular practical importance is the definition of a position-specific reliability index. For half of the residues predicted with a high level of reliability the overall accuracy increases to better than 82%. A further strength of the method is the more realistic prediction of segment length. The protein family prediction method is available for testing by academic researchers via an electronic mail server.
Protein secondary structure prediction based on position-specific scoring matrices. Jones (1999). Journal of Molecular Biology. doi:10.1006/jmbi.1999.3091
Abstract
A two-stage neural network has been used to predict protein secondary structure based on the position specific scoring matrices generated by PSI-BLAST. Despite the simplicity and convenience of the approach used, the results are found to be superior to those produced by other methods, including the popular PHD method according to our own benchmarking results and the results from the recent Critical Assessment of Techniques for Protein Structure Prediction experiment (CASP3), where the method was evaluated by stringent blind testing. Using a new testing set based on a set of 187 unique folds, and three-way cross-validation based on structural similarity criteria rather than sequence similarity criteria used previously (no similar folds were present in both the testing and training sets) the method presented here (PSIPRED) achieved an average Q3 score of between 76.5% to 78.3% depending on the precise definition of observed secondary structure used, which is the highest published score for any method to date. Given the success of the method in CASP3, it is reasonable to be confident that the evaluation presented here gives a fair indication of the performance of the method in general.
★ Šali and Blundell (1993) — MODELLER.
Comparative Protein Modelling by Satisfaction of Spatial Restraints. Šali & Blundell (1993). Journal of Molecular Biology. doi:10.1006/jmbi.1993.1626
Abstract
We describe a comparative protein modelling method designed to find the most probable structure for a sequence given its alignment with related structures. The three-dimensional (3D) model is obtained by optimally satisfying spatial restraints derived from the alignment and expressed as probability density functions (pdfs) for the features restrained. For example, the probabilities for main-chain conformations of a modelled residue may be restrained by its residue type, main-chain conformation of an equivalent residue in a related protein, and the local similarity between the two sequences. Several such pdfs are obtained from the correlations between structural features in 17 families of homologous proteins which have been aligned on the basis of their 3D structures. The pdfs restrain C alpha-C alpha distances, main-chain N-O distances, main-chain and side-chain dihedral angles. A smoothing procedure is used in the derivation of these relationships to minimize the problem of a sparse database. The 3D model of a protein is obtained by optimization of the molecular pdf such that the model violates the input restraints as little as possible. The molecular pdf is derived as a combination of pdfs restraining individual spatial features of the whole molecule. The optimization procedure is a variable target function method that applies the conjugate gradients algorithm to positions of all non-hydrogen atoms. The method is automated and is illustrated by the modelling of trypsin from two other serine proteinases.
Bowie et al. (1991), Jones et al. (1992), Fischer and Eisenberg (1996) — 3D profiles, threading and fold recognition.
A Method to Identify Protein Sequences That Fold into a Known Three-Dimensional Structure. Bowie et al. (1991). Science. doi:10.1126/science.1853201
Abstract
The inverse protein folding problem, the problem of finding which amino acid sequences fold into a known three-dimensional (3D) structure, can be effectively attacked by finding sequences that are most compatible with the environments of the residues in the 3D structure. The environments are described by: (i) the area of the residue buried in the protein and inaccessible to solvent; (ii) the fraction of side-chain area that is covered by polar atoms (O and N); and (iii) the local secondary structure. Examples of this 3D profile method are presented for four families of proteins: the globins, cyclic AMP (adenosine 3',5'-monophosphate) receptor-like proteins, the periplasmic binding proteins, and the actins. This method is able to detect the structural similarity of the actins and 70- kilodalton heat shock proteins, even though these protein families share no detectable sequence similarity.
A new approach to protein fold recognition. Jones et al. (1992). Nature. doi:10.1038/358086a0
Abstract
The prediction of protein tertiary structure from sequence using molecular energy calculations has not yet been successful; an alternative strategy of recognizing known motifs or folds in sequences looks more promising. We present here a new approach to fold recognition, whereby sequences are fitted directly onto the backbone coordinates of known protein structures. Our method for protein fold recognition involves automatic modelling of protein structures using a given sequence, and is based on the frameworks of known protein folds. The plausibility of each model, and hence the degree of compatibility between the sequence and the proposed structure, is evaluated by means of a set of empirical potentials derived from proteins of known structure. The novel aspect of our approach is that the matching of sequences to backbone coordinates is performed in full three-dimensional space, incorporating specific pair interactions explicitly.
Protein fold recognition using sequence-derived predictions. Fischer & Eisenberg (1996). Protein Science. doi:10.1002/pro.5560050516
Abstract
In protein fold recognition, one assigns a probe amino acid sequence of unknown structure to one of a library of target 3D structures. Correct assignment depends on effective scoring of the probe sequence for its compatibility with each of the target structures. Here we show that, in addition to the amino acid sequence of the probe, sequence-derived properties of the probe sequence (such as the predicted secondary structure) are useful in fold assignment. The additional measure of compatibility between probe and target is the level of agreement between the predicted secondary structure of the probe and the known secondary structure of the target fold. That is, we recommend a sequence-structure compatibility function that combines previously developed compatibility functions (such as the 3D-1D scores of Bowie et al. [1991] or sequence-sequence replacement tables) with the predicted secondary structure of the probe sequence. The effect on fold assignment of adding predicted secondary structure is evaluated here by using a benchmark set of proteins (Fischer et al., 1996a). The 3D structures of the probe sequences of the benchmark are actually known, but are ignored by our method. The results show that the inclusion of the predicted secondary structure improves fold assignment by about 25%. The results also show that, if the true secondary structure of the probe were known, correct fold assignment would increase by an additional 8-32%. We conclude that incorporating sequence-derived predictions significantly improves assignment of sequences to known 3D folds. Finally, we apply the new method to assign folds to sequences in the SWISSPROT database; six fold assignments are given that are not detectable by standard sequence-sequence comparison methods; for two of these, the fold is known from X-ray crystallography and the fold assignment is correct.
Simons et al. (1997), Jones (2001), ★ Bradley et al. (2005) — fragment assembly with Rosetta and FRAGFOLD.
Assembly of protein tertiary structures from fragments with similar local sequences using simulated annealing and bayesian scoring functions. Simons et al. (1997). Journal of Molecular Biology. doi:10.1006/jmbi.1997.0959
Abstract
We explore the ability of a simple simulated annealing procedure to assemble native-like structures from fragments of unrelated protein structures with similar local sequences using Bayesian scoring functions. Environment and residue pair specific contributions to the scoring functions appear as the first two terms in a series expansion for the residue probability distributions in the protein database; the decoupling of the distance and environment dependencies of the distributions resolves the major problems with current database-derived scoring functions noted by Thomas and Dill. The simulated annealing procedure rapidly and frequently generates native-like structures for small helical proteins and better than random structures for small beta sheet containing proteins. Most of the simulated structures have native-like solvent accessibility and secondary structure patterns, and thus ensembles of these structures provide a particularly challenging set of decoys for evaluating scoring functions. We investigate the effects of multiple sequence information and different types of conformational constraints on the overall performance of the method, and the ability of a variety of recently developed scoring functions to recognize the native-like conformations in the ensembles of simulated structures.
Predicting novel protein folds by using FRAGFOLD. Jones (2001). Proteins: Structure, Function, and Genetics. doi:10.1002/prot.1171
Abstract
The results of applying a fragment-based protein tertiary structure prediction method to the prediction of 8 CASP4 targets are described. The method is based on the assembly of supersecondary structural fragments taken from highly resolved protein structures using a simulated annealing algorithm. Despite the significant degree of success in this case, there is clearly much more developmental work required before predictions with the accuracy of a good homology model, or even a good fold recognition model, can be made with use of this kind of approach.
Toward High-Resolution de Novo Structure Prediction for Small Proteins. Bradley et al. (2005). Science. doi:10.1126/science.1113801
Abstract
The prediction of protein structure from amino acid sequence is a grand challenge of computational molecular biology. By using a combination of improved low- and high-resolution conformational sampling methods, improved atomically detailed potential functions that capture the jigsaw puzzle-like packing of protein cores, and high-performance computing, high-resolution structure prediction (<1.5 angstroms) can be achieved for small protein domains (<85 residues). The primary bottleneck to consistent high-resolution prediction appears to be conformational sampling.
B.9.6 Coevolution and deep learning
Göbel et al. (1994) — correlated mutations.
Correlated mutations and residue contacts in proteins. Göbel et al. (1994). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.340180402
Abstract
The maintenance of protein function and structure constrains the evolution of amino acid sequences. This fact can be exploited to interpret correlated mutations observed in a sequence family as an indication of probable physical contact in three dimensions. Here we present a simple and general method to analyze correlations in mutational behavior between different positions in a multiple sequence alignment. We then use these correlations to predict contact maps for each of 11 protein families and compare the result with the contacts determined by crystallography. For the most strongly correlated residue pairs predicted to be in contact, the prediction accuracy ranges from 37 to 68% and the improvement ratio relative to a random prediction from 1.4 to 5.1. Predicted contact maps can be used as input for the calculation of protein tertiary structure, either from sequence information alone or in combination with experimental information.
Weigt et al. (2009), Morcos et al. (2011), Marks et al. (2011), Jones et al. (2011) — DCA and PSICOV.
Identification of direct residue contacts in protein–protein interaction by message passing. Weigt et al. (2009). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.0805923106
Abstract
Understanding the molecular determinants of specificity in protein-protein interaction is an outstanding challenge of postgenome biology. The availability of large protein databases generated from sequences of hundreds of bacterial genomes enables various statistical approaches to this problem. In this context covariance-based methods have been used to identify correlation between amino acid positions in interacting proteins. However, these methods have an important shortcoming, in that they cannot distinguish between directly and indirectly correlated residues. We developed a method that combines covariance analysis with global inference analysis, adopted from use in statistical physics. Applied to a set of >2,500 representatives of the bacterial two-component signal transduction system, the combination of covariance with global inference successfully and robustly identified residue pairs that are proximal in space without resorting to ad hoc tuning parameters, both for heterointeractions between sensor kinase (SK) and response regulator (RR) proteins and for homointeractions between RR proteins. The spectacular success of this approach illustrates the effectiveness of the global inference approach in identifying direct interaction based on sequence information alone. We expect this method to be applicable soon to interaction surfaces between proteins present in only 1 copy per genome as the number of sequenced genomes continues to expand. Use of this method could significantly increase the potential targets for therapeutic intervention, shed light on the mechanism of protein-protein interaction, and establish the foundation for the accurate prediction of interacting protein partners.
Direct-coupling analysis of residue coevolution captures native contacts across many protein families. Morcos et al. (2011). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1111471108
Abstract
The similarity in the three-dimensional structures of homologous proteins imposes strong constraints on their sequence variability. It has long been suggested that the resulting correlations among amino acid compositions at different sequence positions can be exploited to infer spatial contacts within the tertiary protein structure. Crucial to this inference is the ability to disentangle direct and indirect correlations, as accomplished by the recently introduced direct-coupling analysis (DCA). Here we develop a computationally efficient implementation of DCA, which allows us to evaluate the accuracy of contact prediction by DCA for a large number of protein domains, based purely on sequence information. DCA is shown to yield a large number of correctly predicted contacts, recapitulating the global structure of the contact map for the majority of the protein domains examined. Furthermore, our analysis captures clear signals beyond intradomain residue contacts, arising, e.g., from alternative protein conformations, ligand-mediated residue couplings, and interdomain interactions in protein oligomers. Our findings suggest that contacts predicted by DCA can be used as a reliable guide to facilitate computational predictions of alternative protein conformations, protein complex formation, and even the de novo prediction of protein domain structures, contingent on the existence of a large number of homologous sequences which are being rapidly made available due to advances in genome sequencing.
Protein 3D Structure Computed from Evolutionary Sequence Variation. Marks et al. (2011). PLoS ONE. doi:10.1371/journal.pone.0028766
Abstract
The evolutionary trajectory of a protein through sequence space is constrained by its function. Collections of sequence homologs record the outcomes of millions of evolutionary experiments in which the protein evolves according to these constraints. Deciphering the evolutionary record held in these sequences and exploiting it for predictive and engineering purposes presents a formidable challenge. The potential benefit of solving this challenge is amplified by the advent of inexpensive high-throughput genomic sequencing.In this paper we ask whether we can infer evolutionary constraints from a set of sequence homologs of a protein. The challenge is to distinguish true co-evolution couplings from the noisy set of observed correlations. We address this challenge using a maximum entropy model of the protein sequence, constrained by the statistics of the multiple sequence alignment, to infer residue pair couplings. Surprisingly, we find that the strength of these inferred couplings is an excellent predictor of residue-residue proximity in folded structures. Indeed, the top-scoring residue couplings are sufficiently accurate and well-distributed to define the 3D protein fold with remarkable accuracy.We quantify this observation by computing, from sequence alone, all-atom 3D structures of fifteen test proteins from different fold classes, ranging in size from 50 to 260 residues, including a G-protein coupled receptor. These blinded inferences are de novo, i.e., they do not use homology modeling or sequence-similar fragments from known structures. The co-evolution signals provide sufficient information to determine accurate 3D protein structure to 2.7-4.8 Å C(α)-RMSD error relative to the observed structure, over at least two-thirds of the protein (method called EVfold, details at http://EVfold.org). This discovery provides insight into essential interactions constraining protein evolution and will facilitate a comprehensive survey of the universe of protein structures, new strategies in protein and drug design, and the identification of functional genetic variants in normal and disease genomes.
PSICOV: precise structural contact prediction using sparse inverse covariance estimation on large multiple sequence alignments. Jones et al. (2011). Bioinformatics. doi:10.1093/bioinformatics/btr638
Abstract
Motivation The accurate prediction of residue-residue contacts, critical for maintaining the native fold of a protein, remains an open problem in the field of structural bioinformatics. Interest in this long-standing problem has increased recently with algorithmic improvements and the rapid growth in the sizes of sequence families. Progress could have major impacts in both structure and function prediction to name but two benefits. Sequence-based contact predictions are usually made by identifying correlated mutations within multiple sequence alignments (MSAs), most commonly through the information-theoretic approach of calculating mutual information between pairs of sites in proteins. These predictions are often inaccurate because the true covariation signal in the MSA is often masked by biases from many ancillary indirect-coupling or phylogenetic effects. Here we present a novel method, PSICOV, which introduces the use of sparse inverse covariance estimation to the problem of protein contact prediction. Our method builds on work which had previously demonstrated corrections for phylogenetic and entropic correlation noise and allows accurate discrimination of direct from indirectly coupled mutation correlations in the MSA. Results PSICOV displays a mean precision substantially better than the best performing normalized mutual information approach and Bayesian networks. For 118 out of 150 targets, the L/5 (i.e. top-L/5 predictions for a protein of length L) precision for long-range contacts (sequence separation >23) was ≥ 0.5, which represents an improvement sufficient to be of significant benefit in protein structure prediction or model quality assessment. Availability The PSICOV source code can be downloaded from http://bioinf.cs.ucl.ac.uk/downloads/PSICOV.
Skwark et al. (2014), ★ Wang et al. (2017) — contact prediction with machine learning.
Improved Contact Predictions Using the Recognition of Protein Like Contact Patterns. Skwark et al. (2014). PLoS Computational Biology. doi:10.1371/journal.pcbi.1003889
Abstract
Given sufficient large protein families, and using a global statistical inference approach, it is possible to obtain sufficient accuracy in protein residue contact predictions to predict the structure of many proteins. However, these approaches do not consider the fact that the contacts in a protein are neither randomly, nor independently distributed, but actually follow precise rules governed by the structure of the protein and thus are interdependent. Here, we present PconsC2, a novel method that uses a deep learning approach to identify protein-like contact patterns to improve contact predictions. A substantial enhancement can be seen for all contacts independently on the number of aligned sequences, residue separation or secondary structure type, but is largest for β-sheet containing proteins. In addition to being superior to earlier methods based on statistical inferences, in comparison to state of the art methods using machine learning, PconsC2 is superior for families with more than 100 effective sequence homologs. The improved contact prediction enables improved structure prediction.
Accurate De Novo Prediction of Protein Contact Map by Ultra-Deep Learning Model. Wang et al. (2017). PLOS Computational Biology. doi:10.1371/journal.pcbi.1005324
Abstract
Motivation Protein contacts contain key information for the understanding of protein structure and function and thus, contact prediction from sequence is an important problem. Recently exciting progress has been made on this problem, but the predicted contacts for proteins without many sequence homologs is still of low quality and not very useful for de novo structure prediction. Method This paper presents a new deep learning method that predicts contacts by integrating both evolutionary coupling (EC) and sequence conservation information through an ultra-deep neural network formed by two deep residual neural networks. The first residual network conducts a series of 1-dimensional convolutional transformation of sequential features; the second residual network conducts a series of 2-dimensional convolutional transformation of pairwise information including output of the first residual network, EC information and pairwise potential. By using very deep residual networks, we can accurately model contact occurrence patterns and complex sequence-structure relationship and thus, obtain higher-quality contact prediction regardless of how many sequence homologs are available for proteins in question. Results Our method greatly outperforms existing methods and leads to much more accurate contact-assisted folding. Tested on 105 CASP11 targets, 76 past CAMEO hard targets, and 398 membrane proteins, the average top L long-range prediction accuracy obtained by our method, one representative EC method CCMpred and the CASP11 winner MetaPSICOV is 0.47, 0.21 and 0.30, respectively; the average top L/10 long-range accuracy of our method, CCMpred and MetaPSICOV is 0.77, 0.47 and 0.59, respectively. Ab initio folding using our predicted contacts as restraints but without any force fields can yield correct folds (i.e., TMscore>0.6) for 203 of the 579 test proteins, while that using MetaPSICOV- and CCMpred-predicted contacts can do so for only 79 and 62 of them, respectively. Our contact-assisted models also have much better quality than template-based models especially for membrane proteins. The 3D models built from our contact prediction have TMscore>0.5 for 208 of the 398 membrane proteins, while those from homology modeling have TMscore>0.5 for only 10 of them. Further, even if trained mostly by soluble proteins, our deep learning method works very well on membrane proteins. In the recent blind CAMEO benchmark, our fully-automated web server implementing this method successfully folded 6 targets with a new fold and only 0.3L-2.3L effective sequence homologs, including one β protein of 182 residues, one α+β protein of 125 residues, one α protein of 140 residues, one α protein of 217 residues, one α/β of 260 residues and one α protein of 462 residues. Our method also achieved the highest F1 score on free-modeling targets in the latest CASP (Critical Assessment of Structure Prediction), although it was not fully implemented back then. Availability http://raptorx.uchicago.edu/ContactMap/.
Ovchinnikov et al. (2017) — structures from metagenome sequences.
Protein structure determination using metagenome sequence data. Ovchinnikov et al. (2017). Science. doi:10.1126/science.aah4043
Abstract
Despite decades of work by structural biologists, there are still ~5200 protein families with unknown structure outside the range of comparative modeling. We show that Rosetta structure prediction guided by residue-residue contacts inferred from evolutionary information can accurately model proteins that belong to large families and that metagenome sequence data more than triple the number of protein families with sufficient sequences for accurate modeling. We then integrate metagenome data, contact-based structure matching, and Rosetta structure calculations to generate models for 614 protein families with currently unknown structures; 206 are membrane proteins and 137 have folds not represented in the Protein Data Bank. This approach provides the representative models for large protein families originally envisioned as the goal of the Protein Structure Initiative at a fraction of the cost.
Senior et al. (2020) — AlphaFold at CASP13.
Improved protein structure prediction using potentials from deep learning. Senior et al. (2020). Nature. doi:10.1038/s41586-019-1923-7
Abstract
Protein structure prediction can be used to determine the three-dimensional shape of a protein from its amino acid sequence 1 . This problem is of fundamental importance as the structure of a protein largely determines its function 2 ; however, protein structures can be difficult to determine experimentally. Considerable progress has recently been made by leveraging genetic information. It is possible to infer which amino acid residues are in contact by analysing covariation in homologous sequences, which aids in the prediction of protein structures 3 . Here we show that we can train a neural network to make accurate predictions of the distances between pairs of residues, which convey more information about the structure than contact predictions. Using this information, we construct a potential of mean force 4 that can accurately describe the shape of a protein. We find that the resulting potential can be optimized by a simple gradient descent algorithm to generate structures without complex sampling procedures. The resulting system, named AlphaFold, achieves high accuracy, even for sequences with fewer homologous sequences. In the recent Critical Assessment of Protein Structure Prediction 5 (CASP13)-a blind assessment of the state of the field-AlphaFold created high-accuracy structures (with template modelling (TM) scores 6 of 0.7 or higher) for 24 out of 43 free modelling domains, whereas the next best method, which used sampling and contact information, achieved such accuracy for only 14 out of 43 domains. AlphaFold represents a considerable advance in protein-structure prediction. We expect this increased accuracy to enable insights into the function and malfunction of proteins, especially in cases for which no structures for homologous proteins have been experimentally determined 7 .
Baek et al. (2021) — RoseTTAFold.
Accurate prediction of protein structures and interactions using a three-track neural network. Baek et al. (2021). Science. doi:10.1126/science.abj8754
Abstract
DeepMind presented notably accurate predictions at the recent 14th Critical Assessment of Structure Prediction (CASP14) conference. We explored network architectures that incorporate related ideas and obtained the best performance with a three-track network in which information at the one-dimensional (1D) sequence level, the 2D distance map level, and the 3D coordinate level is successively transformed and integrated. The three-track network produces structure predictions with accuracies approaching those of DeepMind in CASP14, enables the rapid solution of challenging x-ray crystallography and cryo-electron microscopy structure modeling problems, and provides insights into the functions of proteins of currently unknown structure. The network also enables rapid generation of accurate protein-protein complex models from sequence information alone, short-circuiting traditional approaches that require modeling of individual subunits followed by docking. We make the method available to the scientific community to speed biological research.
Mirdita et al. (2022), Ahdritz et al. (2024) — ColabFold and OpenFold.
ColabFold: making protein folding accessible to all. Mirdita et al. (2022). Nature Methods. doi:10.1038/s41592-022-01488-1
Abstract
ColabFold offers accelerated prediction of protein structures and complexes by combining the fast homology search of MMseqs2 with AlphaFold2 or RoseTTAFold. ColabFold's 40-60-fold faster search and optimized model utilization enables prediction of close to 1,000 structures per day on a server with one graphics processing unit. Coupled with Google Colaboratory, ColabFold becomes a free and accessible platform for protein folding. ColabFold is open-source software available at https://github.com/sokrypton/ColabFold and its novel environmental databases are available at https://colabfold.mmseqs.com .
OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization. Ahdritz et al. (2024). Nature Methods. doi:10.1038/s41592-024-02272-z
Abstract
AlphaFold2 revolutionized structural biology with the ability to predict protein structures with exceptionally high accuracy. Its implementation, however, lacks the code and data required to train new models. These are necessary to (1) tackle new tasks, like protein-ligand complex structure prediction, (2) investigate the process by which the model learns and (3) assess the model's capacity to generalize to unseen regions of fold space. Here we report OpenFold, a fast, memory efficient and trainable implementation of AlphaFold2. We train OpenFold from scratch, matching the accuracy of AlphaFold2. Having established parity, we find that OpenFold is remarkably robust at generalizing even when the size and diversity of its training set is deliberately limited, including near-complete elisions of classes of secondary structure elements. By analyzing intermediate structures produced during training, we also gain insights into the hierarchical manner in which OpenFold learns to fold. In sum, our studies demonstrate the power and utility of OpenFold, which we believe will prove to be a crucial resource for the protein modeling community.
Evans et al. (2021), Bryant, Pozzati, and Elofsson (2022) — AlphaFold-Multimer and complex prediction.
Protein complex prediction with AlphaFold-Multimer. Evans et al. (2021). bioRxiv (preprint). doi:10.1101/2021.10.04.463034
Abstract
While the vast majority of well-structured single protein chains can now be predicted to high accuracy due to the recent AlphaFold [1] model, the prediction of multi-chain protein complexes remains a challenge in many cases. In this work, we demonstrate that an AlphaFold model trained specifically for multimeric inputs of known stoichiometry, which we call AlphaFold-Multimer, significantly increases accuracy of predicted multimeric interfaces over input-adapted single-chain AlphaFold while maintaining high intra-chain accuracy. On a benchmark dataset of 17 heterodimer proteins without templates (introduced in [2]) we achieve at least medium accuracy (DockQ [3] ≥ 0.49) on 13 targets and high accuracy (DockQ ≥ 0.8) on 7 targets, compared to 9 targets of at least medium accuracy and 4 of high accuracy for the previous state of the art system (an AlphaFold-based system from [2]). We also predict structures for a large dataset of 4,446 recent protein complexes, from which we score all non-redundant interfaces with low template identity. For heteromeric interfaces we successfully predict the interface (DockQ ≥ 0.23) in 70% of cases, and produce high accuracy predictions (DockQ ≥ 0.8) in 26% of cases, an improvement of +27 and +14 percentage points over the flexible linker modification of AlphaFold [4] respectively. For homomeric inter-faces we successfully predict the interface in 72% of cases, and produce high accuracy predictions in 36% of cases, an improvement of +8 and +7 percentage points respectively.
Improved prediction of protein-protein interactions using AlphaFold2. Bryant et al. (2022). Nature Communications. doi:10.1038/s41467-022-28865-w
Abstract
Predicting the structure of interacting protein chains is a fundamental step towards understanding protein function. Unfortunately, no computational method can produce accurate structures of protein complexes. AlphaFold2, has shown unprecedented levels of accuracy in modelling single chain protein structures. Here, we apply AlphaFold2 for the prediction of heterodimeric protein complexes. We find that the AlphaFold2 protocol together with optimised multiple sequence alignments, generate models with acceptable quality (DockQ ≥ 0.23) for 63% of the dimers. From the predicted interfaces we create a simple function to predict the DockQ score which distinguishes acceptable from incorrect models as well as interacting from non-interacting proteins with state-of-art accuracy. We find that, using the predicted DockQ scores, we can identify 51% of all interacting pairs at 1% FPR.
Boitreaud et al. (2024), Wohlwend et al. (2024), Passaro et al. (2025), Chen et al. (2025) — open AlphaFold3-like models (first released as preprints).
Chai-1: Decoding the molecular interactions of life. Boitreaud et al. (2024). bioRxiv (preprint). doi:10.1101/2024.10.10.615955
Abstract
We introduce Chai-1, a multi-modal foundation model for molecular structure prediction that performs at the state-of-the-art across a variety of tasks relevant to drug discovery. Chai-1 can optionally be prompted with experimental restraints (e.g. derived from wet-lab data) which boosts performance by double-digit percentage points. Chai-1 can also be run in single-sequence mode with-out MSAs while preserving most of its performance. We release Chai-1 model weights and inference code as a Python package for non-commercial use and via a web interface where it can be used for free including for commercial drug discovery purposes.
Boltz-1 Democratizing Biomolecular Interaction Modeling. Wohlwend et al. (2024). bioRxiv (preprint). doi:10.1101/2024.11.19.624167
Abstract
Understanding biomolecular interactions is fundamental to advancing fields like drug discovery and protein design. In this paper, we introduce B oltz -1, an open-source deep learning model incorporating innovations in model architecture, speed optimization, and data processing achieving A lpha F old 3-level accuracy in predicting the 3D structures of biomolecular complexes. B oltz -1 demonstrates a performance on-par with state-of-the-art commercial models on a range of diverse benchmarks, setting a new benchmark for commercially accessible tools in structural biology. Further, we push the boundary of capabilities of these models with B oltz-steering , a new inference time steering technique that is able to fix hallucinations and non-physical predictions from the models. By releasing the training and inference code, model weights, datasets, and benchmarks under the MIT open license, we aim to foster global collaboration, accelerate discoveries, and provide a robust platform for advancing biomolecular modeling.
Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. Passaro et al. (2025). bioRxiv (preprint). doi:10.1101/2025.06.14.659707
Abstract
Accurately modeling biomolecular interactions is a central challenge in modern biology. While recent advances, such as AlphaFold3 and Boltz-1, have substantially improved our ability to predict biomolecular complex structures, these models still fall short in predicting binding affinity, a critical property underlying molecular function and therapeutic efficacy. Here, we present Boltz-2, a new structural biology foundation model that exhibits strong performance for both structure and affinity prediction. Boltz-2 introduces controllability features including experimental method conditioning, distance constraints, and multi-chain template integration for structure prediction, and is, to our knowledge, the first AI model to approach the performance of free-energy perturbation (FEP) methods in estimating small molecule–protein binding affinity. Crucially, it achieves strong correlation with experimental readouts on many benchmarks, while being at least 1000 × more computationally efficient than FEP. By coupling Boltz-2 with a generative model for small molecules, we demonstrate an effective workflow to find diverse, synthesizable, high-affinity binders, as estimated by absolute FEP simulations on the TYK2 target. To foster broad adoption and further innovation at the intersection of machine learning and biology, we are releasing Boltz-2 weights, inference, and training code 1 under a permissive open license, providing a robust and extensible foundation for both academic and industrial research.
Protenix - Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction. Chen et al. (2025). bioRxiv (preprint). doi:10.1101/2025.01.08.631967
Abstract
In this technical report, we present Protenix, a comprehensive reproduction of AlphaFold3 (AF3), aimed at advancing the field of biomolecular structure prediction. Protenix tackles the challenges of predicting complex interactions involving proteins, ligands, and nucleic acids, while enhancing accessibility and reproducibility. Across diverse benchmarks, including PoseBusters V2, low-homology PDB sets, and CASP15 RNA, Protenix achieves state-of-the-art performance in protein-ligand, protein-protein, and protein-nucleic acid predictions. We also address limitations, such as potential memorization effects, and outline future directions for improvement. By open-sourcing Protenix, we aim to democratize advanced structure prediction tools and accelerate interdisciplinary research in computational biology and drug discovery.
Isomorphic Labs Team (2026), OpenDDE contributors (2026) — the IsoDDE technical report (February 2026, not peer reviewed) and the OpenDDE software pipeline (GitHub, 2026).
Accurate Predictions of Novel Biomolecular Interactions with IsoDDE. Isomorphic Labs Team (2026). Isomorphic Labs. doi:10.5281/zenodo.19699685
OpenDDE: Open Drug Discovery Engine. OpenDDE contributors (2026). GitHub repository. https://github.com/contact-ajmal/opendde
Wayment-Steele et al. (2023) — predicting multiple conformations.
Predicting multiple conformations via sequence clustering and AlphaFold2. Wayment-Steele et al. (2023). Nature. doi:10.1038/s41586-023-06832-9
Abstract
AlphaFold2 (ref. 1 ) has revolutionized structural biology by accurately predicting single structures of proteins. However, a protein's biological function often depends on multiple conformational substates 2 , and disease-causing point mutations often cause population changes within these substates 3,4 . We demonstrate that clustering a multiple-sequence alignment by sequence similarity enables AlphaFold2 to sample alternative states of known metamorphic proteins with high confidence. Using this method, named AF-Cluster, we investigated the evolutionary distribution of predicted structures for the metamorphic protein KaiB 5 and found that predictions of both conformations were distributed in clusters across the KaiB family. We used nuclear magnetic resonance spectroscopy to confirm an AF-Cluster prediction: a cyanobacteria KaiB variant is stabilized in the opposite state compared with the more widely studied variant. To test AF-Cluster's sensitivity to point mutations, we designed and experimentally verified a set of three mutations predicted to flip KaiB from Rhodobacter sphaeroides from the ground to the fold-switched state. Finally, screening for alternative states in protein families without known fold switching identified a putative alternative state for the oxidoreductase Mpt53 in Mycobacterium tuberculosis. Further development of such bioinformatic methods in tandem with experiments will probably have a considerable impact on predicting protein energy landscapes, essential for illuminating biological function.
B.9.7 Language models
EvolutionaryScale (2024) — ESM C.
ESM Cambrian: Revealing the mysteries of proteins with unsupervised learning. EvolutionaryScale (2024). https://www.evolutionaryscale.ai/blog/esm-cambrian
Candido et al. (2026) — ESMFold2 (bioRxiv preprint, June 2026).
Language Modeling Materializes a World Model of Protein Biology. Candido et al. (2026). bioRxiv (preprint). doi:10.64898/2026.06.03.729735
Abstract
Proteins are fundamental to life. The full extent of their biology is beyond our ability to characterize with experimental approaches in the physical laboratory. Accurate digital representations could accelerate the discovery of protein biology through virtual experiments. We propose language modeling to learn unified and general representations that can be scaled to all of protein biology. Building on these representations, we develop a structure prediction model that exceeds the performance of established methods for biomolecular complex prediction across benchmarks, including for the interactions of antibodies with their targets. A simple search procedure yields high experimental success rates for the discovery of proteins with nanomolar binding affinities for both miniproteins and single-chain antibodies, a modality critical for therapeutic design. Study of the concepts in the language model’s representation space reveals a systematic organization aligned with the reductionist understanding of proteins developed through empirical science. Leveraging this organization, we generate a comprehensive map of protein biology encompassing over 6.8 billion sequences and 1.1 billion predicted structures, identifying connections across known and unknown biology. As a whole, this shows language modeling as a powerful substrate for representing the biology of proteins, operating across scales from the prediction and design of protein interactions at the atomic level, to identifying properties of proteins at different levels of granularity and abstraction, to the scale of mapping connections between proteins across billions of years of evolution.
Hayes et al. (2025) — ESM3.
Simulating 500 million years of evolution with a language model. Hayes et al. (2025). Science. doi:10.1126/science.ads0018
Abstract
More than 3 billion years of evolution have produced an image of biology encoded into the space of natural proteins. Here, we show that language models trained at scale on evolutionary data can generate functional proteins that are far away from known proteins. We present ESM3, a frontier multimodal generative language model that reasons over the sequence, structure, and function of proteins. ESM3 can follow complex prompts combining its modalities and is highly responsive to alignment to improve its fidelity. We have prompted ESM3 to generate fluorescent proteins. Among the generations that we synthesized, we found a bright fluorescent protein at a far distance (58% sequence identity) from known fluorescent proteins, which we estimate is equivalent to simulating 500 million years of evolution.
Heinzinger et al. (2024) — ProstT5.
Bilingual language model for protein sequence and structure. Heinzinger et al. (2024). NAR Genomics and Bioinformatics. doi:10.1093/nargab/lqae150
Abstract
Adapting language models to protein sequences spawned the development of powerful protein language models (pLMs). Concurrently, AlphaFold2 broke through in protein structure prediction. Now we can systematically and comprehensively explore the dual nature of proteins that act and exist as three-dimensional (3D) machines and evolve as linear strings of one-dimensional (1D) sequences. Here, we leverage pLMs to simultaneously model both modalities in a single model. We encode protein structures as token sequences using the 3Di-alphabet introduced by the 3D-alignment method Foldseek . For training, we built a non-redundant dataset from AlphaFoldDB and fine-tuned an existing pLM (ProtT5) to translate between 3Di and amino acid sequences. As a proof-of-concept for our novel approach, dubbed Protein 'structure-sequence' T5 ( ProstT5 ), we showed improved performance for subsequent, structure-related prediction tasks, leading to three orders of magnitude speedup for deriving 3Di. This will be crucial for future applications trying to search metagenomic sequence databases at the sensitivity of structure comparisons. Our work showcased the potential of pLMs to tap into the information-rich protein structure revolution fueled by AlphaFold2. ProstT5 paves the way to develop new tools integrating the vast resource of 3D predictions and opens new research avenues in the post-AlphaFold2 era.
B.9.8 Membrane proteins
★ Heijne (1992) — hydrophobicity analysis and the positive-inside rule.
Membrane protein structure prediction. von Heijne (1992). Journal of Molecular Biology. doi:10.1016/0022-2836(92)90934-c
Abstract
A new strategy for predicting the topology of bacterial inner membrane proteins is proposed on the basis of hydrophobicity analysis, automatic generation of a set of possible topologies and ranking of these according to the positive-inside rule. A straightforward implementation with no attempts at optimization predicts the correct topology for 23 out of 24 inner membrane proteins with experimentally determined topologies, and correctly identifies 135 transmembrane segments with only one overprediction.
Krogh et al. (2001) — TMHMM.
Predicting transmembrane protein topology with a hidden markov model: application to complete genomes. Krogh et al. (2001). Journal of Molecular Biology. doi:10.1006/jmbi.2000.4315
Abstract
We describe and validate a new membrane protein topology prediction method, TMHMM, based on a hidden Markov model. We present a detailed analysis of TMHMM's performance, and show that it correctly predicts 97-98 % of the transmembrane helices. Additionally, TMHMM can discriminate between soluble and membrane proteins with both specificity and sensitivity better than 99 %, although the accuracy drops when signal peptides are present. This high degree of accuracy allowed us to predict reliably integral membrane proteins in a large collection of genomes. Based on these predictions, we estimate that 20-30 % of all genes in most genomes encode membrane proteins, which is in agreement with previous estimates. We further discovered that proteins with N(in)-C(in) topologies are strongly preferred in all examined organisms, except Caenorhabditis elegans, where the large number of 7TM receptors increases the counts for N(out)-C(in) topologies. We discuss the possible relevance of this finding for our understanding of membrane protein assembly mechanisms. A TMHMM prediction service is available at http://www.cbs.dtu.dk/services/TMHMM/.
Elofsson and Heijne (2007) — membrane protein structure: prediction versus reality.
Membrane Protein Structure: Prediction versus Reality. Elofsson & Heijne (2007). Annual Review of Biochemistry. doi:10.1146/annurev.biochem.76.052705.163539
Abstract
Since high-resolution structural data are still scarce, different kinds of theoretical structure prediction algorithms are of major importance in membrane protein biochemistry. But how well do the current prediction methods perform? Which structural features can be predicted and which cannot? And what can we expect in the next few years?
Hallgren et al. (2022) — DeepTMHMM.
DeepTMHMM predicts alpha and beta transmembrane proteins using deep neural networks. Hallgren et al. (2022). bioRxiv (preprint). doi:10.1101/2022.04.08.487609
Abstract
Transmembrane proteins span the lipid bilayer and are divided into two major structural classes, namely alpha helical and beta barrels. We introduce DeepTMHMM, a deep learning protein language model-based algorithm that can detect and predict the topology of both alpha helical and beta barrels proteins with unprecedented accuracy. DeepTMHMM ( https://dtu.biolib.com/DeepTMHMM ) scales to proteomes and covers all domains of life, which makes it ideal for metagenomics analyses.
Wimley (2003), Bagos et al. (2004), Hayat et al. (2016) — β-barrel membrane proteins and their prediction.
The versatile β-barrel membrane protein. Wimley (2003). Current Opinion in Structural Biology. doi:10.1016/s0959-440x(03)00099-x
Abstract
The beta-barrel membrane protein is found in the outer membranes of bacteria, mitochondria and chloroplasts. Approximately 2-3% of the genes in Gram-negative bacterial genomes encode beta-barrels. Whereas there are fewer than 20 known three-dimensional beta-barrel structures, genomic databases currently contain thousands of beta-barrels belonging to dozens of families. New research is revealing the variety of beta-barrel structures and the variety of functions performed by these versatile proteins.
PRED-TMBB: a web server for predicting the topology of -barrel outer membrane proteins. Bagos et al. (2004). Nucleic Acids Research. doi:10.1093/nar/gkh417
Abstract
The beta-barrel outer membrane proteins constitute one of the two known structural classes of membrane proteins. Whereas there are several different web-based predictors for alpha-helical membrane proteins, currently there is no freely available prediction method for beta-barrel membrane proteins, at least with an acceptable level of accuracy. We present here a web server (PRED-TMBB, http://bioinformatics.biol.uoa.gr/PRED-TMBB) which is capable of predicting the transmembrane strands and the topology of beta-barrel outer membrane proteins of Gram-negative bacteria. The method is based on a Hidden Markov Model, trained according to the Conditional Maximum Likelihood criterion. The model was retrained and the training set now includes 16 non-homologous outer membrane proteins with structures known at atomic resolution. The user may submit one sequence at a time and has the option of choosing between three different decoding methods. The server reports the predicted topology of a given protein, a score indicating the probability of the protein being an outer membrane beta-barrel protein, posterior probabilities for the transmembrane strand prediction and a graphical representation of the assumed position of the transmembrane strands with respect to the lipid bilayer.
Inclusion of dyad-repeat pattern improves topology prediction of transmembrane β-barrel proteins. Hayat et al. (2016). Bioinformatics. doi:10.1093/bioinformatics/btw025
Abstract
Unlabelled : Accurate topology prediction of transmembrane β-barrels is still an open question. Here, we present BOCTOPUS2, an improved topology prediction method for transmembrane β-barrels that can also identify the barrel domain, predict the topology and identify the orientation of residues in transmembrane β-strands. The major novelty of BOCTOPUS2 is the use of the dyad-repeat pattern of lipid and pore facing residues observed in transmembrane β-barrels. In a cross-validation test on a benchmark set of 42 proteins, BOCTOPUS2 predicts the correct topology in 69% of the proteins, an improvement of more than 10% over the best earlier method (BOCTOPUS) and in addition, it produces significantly fewer erroneous predictions on non-transmembrane β-barrel proteins. Availability and implementation BOCTOPUS2 webserver along with full dataset and source code is available at http://boctopus.bioinfo.se/ Contact : arne@bioinfo.se Supplementary information Supplementary data are available at Bioinformatics online.
B.9.9 Dynamics and docking
Levitt and Warshel (1975), McCammon et al. (1977), Warshel and Levitt (1976) — the first simulations and QM/MM.
Computer simulation of protein folding. Levitt & Warshel (1975). Nature. doi:10.1038/253694a0
Abstract
A new and very simple representation of protein conformations has been used together with energy minimisation and thermalisation to simulate protein folding. Under certain conditions, the method succeeds in "renaturing" bovine pancreatic trypsin inhibitor from an open-chain conformation into a folded conformation close to that of the native molecule.
Dynamics of folded proteins. McCammon et al. (1977). Nature. doi:10.1038/267585a0
Abstract
The dynamics of a folded globular protein (bovine pancreatic trypsin inhibitor) have been studied by solving the equations of motion for the atoms with an empirical potential energy function. The results provide the magnitude, correlations and decay of fluctuations about the average structure. These suggest that the protein interior is fluid-like in that the local atom motions have a diffusional character.
Theoretical studies of enzymic reactions: Dielectric, electrostatic and steric stabilization of the carbonium ion in the reaction of lysozyme. Warshel & Levitt (1976). Journal of Molecular Biology. doi:10.1016/0022-2836(76)90311-9
Karplus and McCammon (2002) — molecular dynamics of biomolecules.
Molecular dynamics simulations of biomolecules. Karplus & McCammon (2002). Nature Structural Biology. doi:10.1038/nsb0902-646
Abstract
Molecular dynamics simulations are important tools for understanding the physical basis of the structure and function of biological macromolecules. The early view of proteins as relatively rigid structures has been replaced by a dynamic model in which the internal motions and resulting conformational changes play an essential role in their function. This review presents a brief description of the origin and early uses of biomolecular simulations. It then outlines some recent studies that illustrate the utility of such simulations and closes with a discussion of their ever-increasing potential for contributing to biology.
Lindorff-Larsen et al. (2012) — validating force fields.
Systematic Validation of Protein Force Fields against Experimental Data. Lindorff-Larsen et al. (2012). PLoS ONE. doi:10.1371/journal.pone.0032131
Abstract
Molecular dynamics simulations provide a vehicle for capturing the structures, motions, and interactions of biological macromolecules in full atomic detail. The accuracy of such simulations, however, is critically dependent on the force field--the mathematical model used to approximate the atomic-level forces acting on the simulated molecular system. Here we present a systematic and extensive evaluation of eight different protein force fields based on comparisons of experimental data with molecular dynamics simulations that reach a previously inaccessible timescale. First, through extensive comparisons with experimental NMR data, we examined the force fields' abilities to describe the structure and fluctuations of folded proteins. Second, we quantified potential biases towards different secondary structure types by comparing experimental and simulation data for small peptides that preferentially populate either helical or sheet-like structures. Third, we tested the force fields' abilities to fold two small proteins--one α-helical, the other with β-sheet structure. The results suggest that force fields have improved over time, and that the most recent versions, while not perfect, provide an accurate description of many structural and dynamical properties of proteins.
Tirion (1996), Bahar et al. (1997), Atilgan et al. (2001) — normal modes and elastic network models.
Large Amplitude Elastic Motions in Proteins from a Single-Parameter, Atomic Analysis. Tirion (1996). Physical Review Letters. doi:10.1103/physrevlett.77.1905
Abstract
Normal mode analysis (NMA) is a leading method for studying long-time dynamics and elasticity of biomolecules. The method proceeds from complex semiempirical potentials characterizing the covalent and noncovalent interactions between atoms. It is widely accepted that such detailed potentials are essential to the success of NMA's. We show that a single-parameter potential is sufficient to reproduce the slow dynamics in good detail. Costly and inaccurate energy minimizations are eliminated, permitting direct analysis of crystal coordinates. The technique can be used for new applications, such as mapping of one crystal form to another by means of slow modes, and studying anomalous dynamics of large proteins and complexes.
Direct evaluation of thermal fluctuations in proteins using a single-parameter harmonic potential. Bahar et al. (1997). Folding and Design. doi:10.1016/s1359-0278(97)00024-2
Abstract
Background An elastic network model is proposed for the interactions between closely ( Results The temperature factors of the C alpha atoms of 12 X-ray structures, ranging from a 41 residue subunit to a 633 residue dimer, are accurately predicted. Cross-correlations are also efficiently characterized, in close agreement with results obtained with a normal mode analysis coupled with energy minimization. Conclusions The simple model and method proposed here provide a satisfactory description of the correlations between atomic fluctuations. Furthermore, this is achieved within computation times at least one order of magnitude shorter than commonly used molecular approaches.
Anisotropy of Fluctuation Dynamics of Proteins with an Elastic Network Model. Atilgan et al. (2001). Biophysical Journal. doi:10.1016/s0006-3495(01)76033-x
Abstract
Fluctuations about the native conformation of proteins have proven to be suitably reproduced with a simple elastic network model, which has shown excellent agreement with a number of different properties for a wide variety of proteins. This scalar model simply investigates the magnitudes of motion of individual residues in the structure. To use the elastic model approach further for developing the details of protein mechanisms, it becomes essential to expand this model to include the added details of the directions of individual residue fluctuations. In this paper a new tool is presented for this purpose and applied to the retinol-binding protein, which indicates enhanced flexibility in the region of entry to the ligand binding site and for the portion of the protein binding to its carrier protein.
Kuntz et al. (1982), Meng et al. (1992) — DOCK.
A geometric approach to macromolecule-ligand interactions. Kuntz et al. (1982). Journal of Molecular Biology. doi:10.1016/0022-2836(82)90153-x
Automated docking with grid-based energy evaluation. Meng et al. (1992). Journal of Computational Chemistry. doi:10.1002/jcc.540130412
Abstract
The ability to generate feasible binding orientations of a small molecule within a site of known structure is important for ligand design. We present a method that combines a rapid, geometric docking algorithm with the evaluation of molecular mechanics interaction energies. The computational costs of evaluation are minimal because we precalculate the receptor‐dependent terms in the potential function at points on a three‐dimensional grid. In four test cases where the components of crystallographically determined complexes are redocked, the “force field” score correctly identifies the family of orientations closest to the experimental binding geometry. Scoring functions that consider only steric factors or only electrostatic factors are less successful. The force field function will play an important role in our efforts to search databases for potential lead compounds.
Böhm (1994), ★ Muegge and Martin (1999) — empirical and knowledge-based scoring functions.
The development of a simple empirical scoring function to estimate the binding constant for a protein-ligand complex of known three-dimensional structure. Böhm (1994). Journal of Computer-Aided Molecular Design. doi:10.1007/bf00126743
Abstract
A new simple empirical function has been developed that estimates the free energy of binding for a given protein-ligand complex of known 3D structure. The function takes into account hydrogen bonds, ionic interactions, the lipophilic protein-ligand contact surface and the number of rotatable bonds in the ligand. The dataset for the calibration of the function consists of 45 protein-ligand complexes. The new energy function reproduces the binding constants (ranging from 2.5.10(-2) to 4.10(-14) M, corresponding to binding energies between -9 and -76 kJ/mol) of the dataset with a standard deviation of 7.9 kJ/mol, corresponding to 1.4 orders of magnitude in binding affinity. The individual contributions to protein-ligand binding obtained from the scoring function are: ideal neutral hydrogen bond: -4.7 kJ/mol; ideal ionic interaction: -8.3 kJ/mol; lipophilic contact: -0.17 kJ/mol A2; one rotatable bond in the ligand: +1.4 kJ/mol. The function also contains a constant contribution (+5.4 kJ/mol) which may be rationalized as loss of translational and rotational entropy. The function can be evaluated very fast and is therefore also suitable for application in a 3D database search or de novo ligand design program such as LUDI.
A General and Fast Scoring Function for Protein−Ligand Interactions: A Simplified Potential Approach. Muegge & Martin (1999). Journal of Medicinal Chemistry. doi:10.1021/jm980536j
Abstract
A fast, simplified potential-based approach is presented that estimates the protein-ligand binding affinity based on the given 3D structure of a protein-ligand complex. This general, knowledge-based approach exploits structural information of known protein-ligand complexes extracted from the Brookhaven Protein Data Bank and converts it into distance-dependent Helmholtz free interaction energies of protein-ligand atom pairs (potentials of mean force, PMF). The definition of an appropriate reference state and the introduction of a correction term accounting for the volume taken by the ligand were found to be crucial for deriving the relevant interaction potentials that treat solvation and entropic contributions implicitly. A significant correlation between experimental binding affinities and computed score was found for sets of diverse protein-ligand complexes and for sets of different ligands bound to the same target. For 77 protein-ligand complexes taken from the Brookhaven Protein Data Bank, the calculated score showed a standard deviation from observed binding affinities of 1.8 log Ki units and an R2 value of 0.61. The best results were obtained for the subset of 16 serine protease complexes with a standard deviation of 1.0 log Ki unit and an R2 value of 0.86. A set of 33 inhibitors modeled into a crystal structure of HIV-1 protease yielded a standard deviation of 0.8 log Ki units from measured inhibition constants and an R2 value of 0.74. In contrast to empirical scoring functions that show similar or sometimes better correlation with observed binding affinities, our method does not involve deriving specific parameters that fit the observed binding affinities of protein-ligand complexes of a given training set. We compared the performance of the PMF score, Böhm's score (LUDI), and the SMOG score for eight different test sets of protein-ligand complexes. It was found that for the majority of test sets the PMF score performs best. The strength of the new approach presented here lies in its generality as no knowledge about measured binding affinities is needed to derive atomic interaction potentials. The use of the new scoring function in docking studies is outlined.
Warren et al. (2005) — a critical assessment of docking.
A Critical Assessment of Docking Programs and Scoring Functions. Warren et al. (2005). Journal of Medicinal Chemistry. doi:10.1021/jm050362n
Abstract
Docking is a computational technique that samples conformations of small molecules in protein binding sites; scoring functions are used to assess which of these conformations best complements the protein binding site. An evaluation of 10 docking programs and 37 scoring functions was conducted against eight proteins of seven protein types for three tasks: binding mode prediction, virtual screening for lead identification, and rank-ordering by affinity for lead optimization. All of the docking programs were able to generate ligand conformations similar to crystallographically determined protein/ligand complex structures for at least one of the targets. However, scoring functions were less successful at distinguishing the crystallographic conformation from the set of docked poses. Docking programs identified active compounds from a pharmaceutically relevant pool of decoy compounds; however, no single program performed well for all of the targets. For prediction of compound affinity, none of the docking programs or scoring functions made a useful prediction of ligand binding affinity.
Trott and Olson (2009) — AutoDock Vina.
AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Trott & Olson (2009). Journal of Computational Chemistry. doi:10.1002/jcc.21334
Abstract
AutoDock Vina, a new program for molecular docking and virtual screening, is presented. AutoDock Vina achieves an approximately two orders of magnitude speed-up compared with the molecular docking software previously developed in our lab (AutoDock 4), while also significantly improving the accuracy of the binding mode predictions, judging by our tests on the training set used in AutoDock 4 development. Further speed-up is achieved from parallelism, by using multithreading on multicore machines. AutoDock Vina automatically calculates the grid maps and clusters the results in a way transparent to the user.
Katchalski-Katzir et al. (1992) — FFT-based protein–protein docking.
Molecular surface recognition: determination of geometric fit between proteins and their ligands by correlation techniques.. Katchalski-Katzir et al. (1992). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.89.6.2195
Abstract
A geometric recognition algorithm was developed to identify molecular surface complementarity. It is based on a purely geometric approach and takes advantage of techniques applied in the field of pattern recognition. The algorithm involves an automated procedure including (i) a digital representation of the molecules (derived from atomic coordinates) by three-dimensional discrete functions that distinguishes between the surface and the interior; (ii) the calculation, using Fourier transformation, of a correlation function that assesses the degree of molecular surface overlap and penetration upon relative shifts of the molecules in three dimensions; and (iii) a scan of the relative orientations of the molecules in three dimensions. The algorithm provides a list of correlation values indicating the extent of geometric match between the surfaces of the molecules; each of these values is associated with six numbers describing the relative position (translation and rotation) of the molecules. The procedure is thus equivalent to a six-dimensional search but much faster by design, and the computation time is only moderately dependent on molecular size. The procedure was tested and validated by using five known complexes for which the correct relative position of the molecules in the respective adducts was successfully predicted. The molecular pairs were deoxyhemoglobin and methemoglobin, tRNA synthetase-tyrosinyl adenylate, aspartic proteinase-peptide inhibitor, and trypsin-trypsin inhibitor. A more realistic test was performed with the last two pairs by using the structures of uncomplexed aspartic proteinase and trypsin inhibitor, respectively. The results are indicative of the extent of conformational changes in the molecules tolerated by the algorithm.
Buttenschoen et al. (2024) — PoseBusters.
PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. Buttenschoen et al. (2024). Chemical Science. doi:10.1039/d3sc04185a
Abstract
The last few years have seen the development of numerous deep learning-based protein-ligand docking methods. They offer huge promise in terms of speed and accuracy. However, despite claims of state-of-the-art performance in terms of crystallographic root-mean-square deviation (RMSD), upon closer inspection, it has become apparent that they often produce physically implausible molecular structures. It is therefore not sufficient to evaluate these methods solely by RMSD to a native binding mode. It is vital, particularly for deep learning-based methods, that they are also evaluated on steric and energetic criteria. We present PoseBusters, a Python package that performs a series of standard quality checks using the well-established cheminformatics toolkit RDKit. The PoseBusters test suite validates chemical and geometric consistency of a ligand including its stereochemistry, and the physical plausibility of intra- and intermolecular measurements such as the planarity of aromatic rings, standard bond lengths, and protein-ligand clashes. Only methods that both pass these checks and predict native-like binding modes should be classed as having "state-of-the-art" performance. We use PoseBusters to compare five deep learning-based docking methods (DeepDock, DiffDock, EquiBind, TankBind, and Uni-Mol) and two well-established standard docking methods (AutoDock Vina and CCDC Gold) with and without an additional post-prediction energy minimisation step using a molecular mechanics force field. We show that both in terms of physical plausibility and the ability to generalise to examples that are distinct from the training data, no deep learning-based method yet outperforms classical docking tools. In addition, we find that molecular mechanics force fields contain docking-relevant physics missing from deep-learning methods. PoseBusters allows practitioners to assess docking and molecular generation methods and may inspire new inductive biases still required to improve deep learning-based methods, which will help drive the development of more accurate and more realistic predictions.
B.9.10 Protein design
Kortemme (2024) — de novo protein design, from structures to functions.
De novo protein design—From new structures to programmable functions. Kortemme (2024). Cell. doi:10.1016/j.cell.2023.12.028
Abstract
Methods from artificial intelligence (AI) trained on large datasets of sequences and structures can now "write" proteins with new shapes and molecular functions de novo, without starting from proteins found in nature. In this Perspective, I will discuss the state of the field of de novo protein design at the juncture of physics-based modeling approaches and AI. New protein folds and higher-order assemblies can be designed with considerable experimental success rates, and difficult problems requiring tunable control over protein conformations and precise shape complementarity for molecular recognition are coming into reach. Emerging approaches incorporate engineering principles-tunability, controllability, and modularity-into the design process from the beginning. Exciting frontiers lie in deconstructing cellular functions with de novo proteins and, conversely, constructing synthetic cellular signaling from the ground up. As methods improve, many more challenges are unsolved.
Dahiyat and Mayo (1997) — fully automated design.
De Novo Protein Design: Fully Automated Sequence Selection. Dahiyat & Mayo (1997). Science. doi:10.1126/science.278.5335.82
Abstract
The first fully automated design and experimental validation of a novel sequence for an entire protein is described. A computational design algorithm based on physical chemical potential functions and stereochemical constraints was used to screen a combinatorial library of 1.9 x 10(27) possible amino acid sequences for compatibility with the design target, a betabetaalpha protein motif based on the polypeptide backbone structure of a zinc finger domain. A BLAST search shows that the designed sequence, full sequence design 1 (FSD-1), has very low identity to any known protein sequence. The solution structure of FSD-1 was solved by nuclear magnetic resonance spectroscopy and indicates that FSD-1 forms a compact well-ordered structure, which is in excellent agreement with the design target structure. This result demonstrates that computational methods can perform the immense combinatorial search required for protein design, and it suggests that an unbiased and quantitative algorithm can be used in various structural contexts.
Kuhlman et al. (2003) — Top7, a new fold.
Design of a Novel Globular Protein Fold with Atomic-Level Accuracy. Kuhlman et al. (2003). Science. doi:10.1126/science.1089427
Abstract
A major challenge of computational protein design is the creation of novel proteins with arbitrarily chosen three-dimensional structures. Here, we used a general computational strategy that iterates between sequence design and structure prediction to design a 93-residue alpha/beta protein called Top7 with a novel sequence and topology. Top7 was found experimentally to be folded and extremely stable, and the x-ray crystal structure of Top7 is similar (root mean square deviation equals 1.2 angstroms) to the design model. The ability to design a new protein fold makes possible the exploration of the large regions of the protein universe not yet observed in nature.
Koga et al. (2012), Jiang et al. (2008), ★ Huang et al. (2016) — ideal proteins, enzymes and the coming of age of design.
Principles for designing ideal protein structures. Koga et al. (2012). Nature. doi:10.1038/nature11600
Abstract
Unlike random heteropolymers, natural proteins fold into unique ordered structures. Understanding how these are encoded in amino-acid sequences is complicated by energetically unfavourable non-ideal features--for example kinked α-helices, bulged β-strands, strained loops and buried polar groups--that arise in proteins from evolutionary selection for biological function or from neutral drift. Here we describe an approach to designing ideal protein structures stabilized by completely consistent local and non-local interactions. The approach is based on a set of rules relating secondary structure patterns to protein tertiary motifs, which make possible the design of funnel-shaped protein folding energy landscapes leading into the target folded state. Guided by these rules, we designed sequences predicted to fold into ideal protein structures consisting of α-helices, β-strands and minimal loops. Designs for five different topologies were found to be monomeric and very stable and to adopt structures in solution nearly identical to the computational models. These results illuminate how the folding funnels of natural proteins arise and provide the foundation for engineering a new generation of functional proteins free from natural evolution.
De Novo Computational Design of Retro-Aldol Enzymes. Jiang et al. (2008). Science. doi:10.1126/science.1152692
Abstract
The creation of enzymes capable of catalyzing any desired chemical reaction is a grand challenge for computational protein design. Using new algorithms that rely on hashing techniques to construct active sites for multistep reactions, we designed retro-aldolases that use four different catalytic motifs to catalyze the breaking of a carbon-carbon bond in a nonnatural substrate. Of the 72 designs that were experimentally characterized, 32, spanning a range of protein folds, had detectable retro-aldolase activity. Designs that used an explicit water molecule to mediate proton shuffling were significantly more successful, with rate accelerations of up to four orders of magnitude and multiple turnovers, than those involving charged side-chain networks. The atomic accuracy of the design process was confirmed by the x-ray crystal structure of active designs embedded in two protein scaffolds, both of which were nearly superimposable on the design model.
The coming of age of de novo protein design. Huang et al. (2016). Nature. doi:10.1038/nature19946
Abstract
There are 20(200) possible amino-acid sequences for a 200-residue protein, of which the natural evolutionary process has sampled only an infinitesimal subset. De novo protein design explores the full sequence space, guided by the physical principles that underlie protein folding. Computational methodology has advanced to the point that a wide range of structures can be designed from scratch with atomic-level accuracy. Almost all protein engineering so far has involved the modification of naturally occurring proteins; it should now be possible to design new functional proteins from the ground up to tackle current challenges in biomedicine and nanotechnology.
Dauparas et al. (2022) — ProteinMPNN.
Robust deep learning–based protein sequence design using ProteinMPNN. Dauparas et al. (2022). Science. doi:10.1126/science.add2187
Abstract
Although deep learning has revolutionized protein structure prediction, almost all experimentally characterized de novo protein designs have been generated using physically based approaches such as Rosetta. Here, we describe a deep learning-based protein sequence design method, ProteinMPNN, that has outstanding performance in both in silico and experimental tests. On native protein backbones, ProteinMPNN has a sequence recovery of 52.4% compared with 32.9% for Rosetta. The amino acid sequence at different positions can be coupled between single or multiple chains, enabling application to a wide range of current protein design challenges. We demonstrate the broad utility and high accuracy of ProteinMPNN using x-ray crystallography, cryo-electron microscopy, and functional studies by rescuing previously failed designs, which were made using Rosetta or AlphaFold, of protein monomers, cyclic homo-oligomers, tetrahedral nanoparticles, and target-binding proteins.
Anishchenko et al. (2021), Wang et al. (2022) — hallucination.
De novo protein design by deep network hallucination. Anishchenko et al. (2021). Nature. doi:10.1038/s41586-021-04184-w
Abstract
There has been considerable recent progress in protein structure prediction using deep neural networks to predict inter-residue distances from amino acid sequences 1-3 . Here we investigate whether the information captured by such networks is sufficiently rich to generate new folded proteins with sequences unrelated to those of the naturally occurring proteins used in training the models. We generate random amino acid sequences, and input them into the trRosetta structure prediction network to predict starting residue-residue distance maps, which, as expected, are quite featureless. We then carry out Monte Carlo sampling in amino acid sequence space, optimizing the contrast (Kullback-Leibler divergence) between the inter-residue distance distributions predicted by the network and background distributions averaged over all proteins. Optimization from different random starting points resulted in novel proteins spanning a wide range of sequences and predicted structures. We obtained synthetic genes encoding 129 of the network-'hallucinated' sequences, and expressed and purified the proteins in Escherichia coli; 27 of the proteins yielded monodisperse species with circular dichroism spectra consistent with the hallucinated structures. We determined the three-dimensional structures of three of the hallucinated proteins, two by X-ray crystallography and one by NMR, and these closely matched the hallucinated models. Thus, deep networks trained to predict native protein structures from their sequences can be inverted to design new proteins, and such networks and methods should contribute alongside traditional physics-based models to the de novo design of proteins with new functions.
Scaffolding protein functional sites using deep learning. Wang et al. (2022). Science. doi:10.1126/science.abn2100
Abstract
The binding and catalytic functions of proteins are generally mediated by a small number of functional residues held in place by the overall protein structure. Here, we describe deep learning approaches for scaffolding such functional sites without needing to prespecify the fold or secondary structure of the scaffold. The first approach, "constrained hallucination," optimizes sequences such that their predicted structures contain the desired functional site. The second approach, "inpainting," starts from the functional site and fills in additional sequence and structure to create a viable protein scaffold in a single forward pass through a specifically trained RoseTTAFold network. We use these two methods to design candidate immunogens, receptor traps, metalloproteins, enzymes, and protein-binding proteins and validate the designs using a combination of in silico and experimental tests.
Pacesa et al. (2025) — BindCraft.
One-shot design of functional protein binders with BindCraft. Pacesa et al. (2025). Nature. doi:10.1038/s41586-025-09429-6
Abstract
Protein-protein interactions are at the core of all key biological processes. However, the complexity of the structural features that determine protein-protein interactions makes their design challenging. Here we present BindCraft, an open-source and automated pipeline for de novo protein binder design with experimental success rates of 10-100%. BindCraft leverages the weights of AlphaFold2 (ref. 1 ) to generate binders with nanomolar affinity without the need for high-throughput screening or experimental optimization, even in the absence of known binding sites. We successfully designed binders against a diverse set of challenging targets, including cell-surface receptors, common allergens, de novo designed proteins and multi-domain nucleases, such as CRISPR-Cas9. We showcase the functional and therapeutic potential of designed binders by reducing IgE binding to birch allergen in patient-derived samples, modulating Cas9 gene editing activity and reducing the cytotoxicity of a foodborne bacterial enterotoxin. Last, we use cell-surface-receptor-specific binders to redirect adeno-associated virus capsids for targeted gene delivery. This work represents a significant advancement towards a 'one design-one binder' approach in computational design, with immense potential in therapeutics, diagnostics and biotechnology.
Ingraham et al. (2023), Krishna et al. (2024) — Chroma and RFdiffusion All-Atom.
Illuminating protein space with a programmable generative model. Ingraham et al. (2023). Nature. doi:10.1038/s41586-023-06728-8
Abstract
Three billion years of evolution has produced a tremendous diversity of protein molecules 1 , but the full potential of proteins is likely to be much greater. Accessing this potential has been challenging for both computation and experiments because the space of possible protein molecules is much larger than the space of those likely to have functions. Here we introduce Chroma, a generative model for proteins and protein complexes that can directly sample novel protein structures and sequences, and that can be conditioned to steer the generative process towards desired properties and functions. To enable this, we introduce a diffusion process that respects the conformational statistics of polymer ensembles, an efficient neural architecture for molecular systems that enables long-range reasoning with sub-quadratic scaling, layers for efficiently synthesizing three-dimensional structures of proteins from predicted inter-residue geometries and a general low-temperature sampling algorithm for diffusion models. Chroma achieves protein design as Bayesian inference under external constraints, which can involve symmetries, substructure, shape, semantics and even natural-language prompts. The experimental characterization of 310 proteins shows that sampling from Chroma results in proteins that are highly expressed, fold and have favourable biophysical properties. The crystal structures of two designed proteins exhibit atomistic agreement with Chroma samples (a backbone root-mean-square deviation of around 1.0 Å). With this unified approach to protein design, we hope to accelerate the programming of protein matter to benefit human health, materials science and synthetic biology.
Generalized biomolecular modeling and design with RoseTTAFold All-Atom. Krishna et al. (2024). Science. doi:10.1126/science.adl2528
Abstract
Deep-learning methods have revolutionized protein structure prediction and design but are presently limited to protein-only systems. We describe RoseTTAFold All-Atom (RFAA), which combines a residue-based representation of amino acids and DNA bases with an atomic representation of all other groups to model assemblies that contain proteins, nucleic acids, small molecules, metals, and covalent modifications, given their sequences and chemical structures. By fine-tuning on denoising tasks, we developed RFdiffusion All-Atom (RFdiffusionAA), which builds protein structures around small molecules. Starting from random distributions of amino acid residues surrounding target small molecules, we designed and experimentally validated, through crystallography and binding measurements, proteins that bind the cardiac disease therapeutic digoxigenin, the enzymatic cofactor heme, and the light-harvesting molecule bilin.
Cho et al. (2025) — Protein Hunter (preprint).
Protein Hunter: exploiting structure hallucination within diffusion for protein design. Cho et al. (2025). bioRxiv (preprint). doi:10.1101/2025.10.10.681530
Abstract
1 Interactions between proteins and other biomolecules underlie nearly all biological processes, yet designing such interactions de novo remains challenging. Capturing their specific interactions and co-optimizing sequence and structure are difficult and often require extensive computation. We present Protein Hunter, a fast, fine-tuning-free framework for de novo protein design. Starting from an all-X sequence, we find diffusion-based structure prediction models hallucinate reasonable looking structures that can be further improved through iterative sequence re-design and structure re-prediction. This lightweight strategy achieves high AlphaFold3 in silico success rates across both unconditional and conditional generation tasks, including binders to proteins, cyclic peptides, small molecules, DNA, and RNA. Protein Hunter also supports multi-motif scaffolding and partial redesign, providing a general and efficient platform for de novo protein design across diverse molecular targets.
B.10 Protein function
Further papers for Chapter 10.
Papers marked ★ were mandatory in earlier reading lists. In the HTML version, click Abstract to show the abstract.
B.10.1 Definitions and resources
The Gene Ontology Consortium et al. (2023) — the Gene Ontology knowledgebase in 2023.
The Gene Ontology knowledgebase in 2023. The Gene Ontology Consortium et al. (2023). GENETICS. doi:10.1093/genetics/iyad031
Abstract
The Gene Ontology (GO) knowledgebase (http://geneontology.org) is a comprehensive resource concerning the functions of genes and gene products (proteins and noncoding RNAs). GO annotations cover genes from organisms across the tree of life as well as viruses, though most gene function knowledge currently derives from experiments carried out in a relatively small number of model organisms. Here, we provide an updated overview of the GO knowledgebase, as well as the efforts of the broad, international consortium of scientists that develops, maintains, and updates the GO knowledgebase. The GO knowledgebase consists of three components: (1) the GO-a computational knowledge structure describing the functional characteristics of genes; (2) GO annotations-evidence-supported statements asserting that a specific gene product has a particular functional characteristic; and (3) GO Causal Activity Models (GO-CAMs)-mechanistic models of molecular "pathways" (GO biological processes) created by linking multiple GO annotations using defined relations. Each of these components is continually expanded, revised, and updated in response to newly published discoveries and receives extensive QA checks, reviews, and user feedback. For each of these components, we provide a description of the current contents, recent developments to keep the knowledgebase up to date with new discoveries, and guidance on how users can best make use of the data that we provide. We conclude with future directions for the project.
Chang et al. (2020), Bansal et al. (2021) — BRENDA and Rhea.
BRENDA, the ELIXIR core data resource in 2021: new developments and updates. Chang et al. (2020). Nucleic Acids Research. doi:10.1093/nar/gkaa1025
Abstract
The BRENDA enzyme database (https://www.brenda-enzymes.org), established in 1987, has evolved into the main collection of functional enzyme and metabolism data. In 2018, BRENDA was selected as an ELIXIR Core Data Resource. BRENDA provides reliable data, continuous curation and updates of classified enzymes, and the integration of newly discovered enzymes. The main part contains >5 million data for ∼90 000 enzymes from ∼13 000 organisms, manually extracted from ∼157 000 primary literature references, combined with information of text and data mining, data integration, and prediction algorithms. Supplements comprise disease-related data, protein sequences, 3D structures, genome annotations, ligand information, taxonomic, bibliographic, and kinetic data. BRENDA offers an easy access to enzyme information from quick to advanced searches, text- and structured-based queries for enzyme-ligand interactions, word maps, and visualization of enzyme data. The BRENDA Pathway Maps are completely revised and updated for an enhanced interactive and intuitive usability. The new design of the Enzyme Summary Page provides an improved access to each individual enzyme. A new protein structure 3D viewer was integrated. The prediction of the intracellular localization of eukaryotic enzymes has been implemented. The new EnzymeDetector combines BRENDA enzyme annotations with protein and genome databases for the detection of eukaryotic and prokaryotic enzymes.
Rhea, the reaction knowledgebase in 2022. Bansal et al. (2021). Nucleic Acids Research. doi:10.1093/nar/gkab1016
Abstract
Rhea (https://www.rhea-db.org) is an expert-curated knowledgebase of biochemical reactions based on the chemical ontology ChEBI (Chemical Entities of Biological Interest) (https://www.ebi.ac.uk/chebi). In this paper, we describe a number of key developments in Rhea since our last report in the database issue of Nucleic Acids Research in 2019. These include improved reaction coverage in Rhea, the adoption of Rhea as the reference vocabulary for enzyme annotation in the UniProt knowledgebase UniProtKB (https://www.uniprot.org), the development of a new Rhea website, and the designation of Rhea as an ELIXIR Core Data Resource. We hope that these and other developments will enhance the utility of Rhea as a reference resource to study and engineer enzymes and the metabolic systems in which they function.
Schnoes et al. (2013) — biases in experimental annotations.
Biases in the Experimental Annotations of Protein Function and Their Effect on Our Understanding of Protein Function Space. Schnoes et al. (2013). PLoS Computational Biology. doi:10.1371/journal.pcbi.1003063
Abstract
The ongoing functional annotation of proteins relies upon the work of curators to capture experimental findings from scientific literature and apply them to protein sequence and structure data. However, with the increasing use of high-throughput experimental assays, a small number of experimental studies dominate the functional protein annotations collected in databases. Here, we investigate just how prevalent is the "few articles - many proteins" phenomenon. We examine the experimentally validated annotation of proteins provided by several groups in the GO Consortium, and show that the distribution of proteins per published study is exponential, with 0.14% of articles providing the source of annotations for 25% of the proteins in the UniProt-GOA compilation. Since each of the dominant articles describes the use of an assay that can find only one function or a small group of functions, this leads to substantial biases in what we know about the function of many proteins. Mass-spectrometry, microscopy and RNAi experiments dominate high throughput experiments. Consequently, the functional information derived from these experiments is mostly of the subcellular location of proteins, and of the participation of proteins in embryonic developmental pathways. For some organisms, the information provided by different studies overlap by a large amount. We also show that the information provided by high throughput experiments is less specific than those provided by low throughput experiments. Given the experimental techniques available, certain biases in protein function annotation due to high-throughput experiments are unavoidable. Knowing that these biases exist and understanding their characteristics and extent is important for database curators, developers of function annotation programs, and anyone who uses protein function annotation data to plan experiments.
B.10.2 Homology and families
Devos and Valencia (2000), Rost (2002), Tian and Skolnick (2003) — how well function is conserved between homologues.
Practical limits of function prediction. Devos & Valencia (2000). Proteins: Structure, Function, and Genetics. doi:10.1002/1097-0134(20001001)41:1<98::aid-prot120>3.0.co;2-s
Abstract
The widening gap between known protein sequences and their functions has led to the practice of assigning a potential function to a protein on the basis of sequence similarity to proteins whose function has been experimentally investigated. We present here a critical view of the theoretical and practical bases for this approach. The results obtained by analyzing a significant number of true sequence similarities, derived directly from structural alignments, point to the complexity of function prediction. Different aspects of protein function, including (i) enzymatic function classification, (ii) functional annotations in the form of key words, (iii) classes of cellular function, and (iv) conservation of binding sites can only be reliably transferred between similar sequences to a modest degree. The reason for this difficulty is a combination of the unavoidable database inaccuracies and the plasticity of protein function. In addition, analysis of the relationship between sequence and functional descriptions defines an empirical limit for pairwise-based functional annotations, namely, the three first digits of the six numbers used as descriptors of protein folds in the FSSP database can be predicted at an average level as low as 7.5% sequence identity, two of the four EC digits at 15% identity, half of the SWISS-PROT key words related to protein function would require 20% identity, and the prediction of half of the residues in the binding site can be made at the 30% sequence identity level.
Enzyme Function Less Conserved than Anticipated. Rost (2002). Journal of Molecular Biology. doi:10.1016/s0022-2836(02)00016-5
Abstract
The level of sequence similarity that implies similarity in protein structure is well established. Recently, many groups proposed thresholds for similarity in sequence implying similarity in enzymatic function. All previous results suggest the strong conservation of enzymatic function above levels of 50% pairwise sequence identity. Here, I argue that all groups substantially overestimated the conservation of enzyme function because their data sets were either too biased, or too small. An unbiased analysis suggested that less than 30% of the pair fragments above 50% sequence identity have entirely identical EC numbers. Another surprising finding was that even BLAST E-values below 10(-50) did not suffice to automatically transfer enzyme function without errors. As expected, most misclassifications originated from similarities in relatively short regions and/or from transferring annotations for different domains. Both problems cannot be corrected easily by adjusting the thresholds for automatic transfer of genome annotations. A score relating sequence identity to alignment length (distance from HSSP-threshold) outperformed statistical BLAST scores for high sequence similarity. In particular, the distance score allowed error-free transfer of enzyme function for the 10% most similar enzyme pairs. The results illustrated how difficult it is to assess the conservation of protein function and to guarantee error-free genome annotations, in general: sets with millions of pair comparisons might not suffice to arrive at statistically significant conclusions. In practice, the revised detailed estimates for the sequence conservation of enzyme function may provide important benchmarks for everyday sequence analysis and for more cautious automatic genome annotations.
How Well is Enzyme Function Conserved as a Function of Pairwise Sequence Identity?. Tian & Skolnick (2003). Journal of Molecular Biology. doi:10.1016/j.jmb.2003.08.057
Abstract
Enzyme function conservation has been used to derive the threshold of sequence identity necessary to transfer function from a protein of known function to an unknown protein. Using pairwise sequence comparison, several studies suggested that when the sequence identity is above 40%, enzyme function is well conserved. In contrast, Rost argued that because of database bias, the results from such simple pairwise comparisons might be misleading. Thus, by grouping enzyme sequences into families based on sequence similarity and selecting representative sequences for comparison, he showed that enzyme function starts to diverge quickly when the sequence identity is below 70%. Here, we employ a strategy similar to Rost's to reduce the database bias; however, we classify enzyme families based not only on sequence similarity, but also on functional similarity, i.e. sequences in each family must have the same four digits or the same first three digits of the enzyme commission (EC) number. Furthermore, instead of selecting representative sequences for comparison, we calculate the function conservation of each enzyme family and then average the degree of enzyme function conservation across all enzyme families. Our analysis suggests that for functional transferability, 40% sequence identity can still be used as a confident threshold to transfer the first three digits of an EC number; however, to transfer all four digits of an EC number, above 60% sequence identity is needed to have at least 90% accuracy. Moreover, when PSI-BLAST is used, the magnitude of the E-value is found to be weakly correlated with the extent of enzyme function conservation in the third iteration of PSI-BLAST. As a result, functional annotation based on the E-values from PSI-BLAST should be used with caution. We also show that by employing an enzyme family-specific sequence identity threshold above which 100% functional conservation is required, functional inference of unknown sequences can be accurately accomplished. However, this comes at a cost: those true positive sequences below this threshold cannot be uniquely identified.
Gilks et al. (2002), Schnoes et al. (2009) — error propagation and misannotation.
Modeling the percolation of annotation errors in a database of protein sequences. Gilks et al. (2002). Bioinformatics. doi:10.1093/bioinformatics/18.12.1641
Abstract
Public sequence databases contain information on the sequence, structure and function of proteins. Genome sequencing projects have led to a rapid increase in protein sequence information, but reliable, experimentally verified, information on protein function lags a long way behind. To address this deficit, functional annotation in protein databases is often inferred by sequence similarity to homologous, annotated proteins, with the attendant possibility of error. Now, the functional annotation in these homologous proteins may itself have been acquired through sequence similarity to yet other proteins, and it is generally not possible to determine how the functional annotation of any given protein has been acquired. Thus the possibility of chains of misannotation arises, a process we term 'error percolation'. With some simple assumptions, we develop a dynamical probabilistic model for these misannotation chains. By exploring the consequences of the model for annotation quality it is evident that this iterative approach leads to a systematic deterioration of database quality.
Annotation Error in Public Databases: Misannotation of Molecular Function in Enzyme Superfamilies. Schnoes et al. (2009). PLoS Computational Biology. doi:10.1371/journal.pcbi.1000605
Abstract
Due to the rapid release of new data from genome sequencing projects, the majority of protein sequences in public databases have not been experimentally characterized; rather, sequences are annotated using computational analysis. The level of misannotation and the types of misannotation in large public databases are currently unknown and have not been analyzed in depth. We have investigated the misannotation levels for molecular function in four public protein sequence databases (UniProtKB/Swiss-Prot, GenBank NR, UniProtKB/TrEMBL, and KEGG) for a model set of 37 enzyme families for which extensive experimental information is available. The manually curated database Swiss-Prot shows the lowest annotation error levels (close to 0% for most families); the two other protein sequence databases (GenBank NR and TrEMBL) and the protein sequences in the KEGG pathways database exhibit similar and surprisingly high levels of misannotation that average 5%-63% across the six superfamilies studied. For 10 of the 37 families examined, the level of misannotation in one or more of these databases is >80%. Examination of the NR database over time shows that misannotation has increased from 1993 to 2005. The types of misannotation that were found fall into several categories, most associated with "overprediction" of molecular function. These results suggest that misannotation in enzyme superfamilies containing multiple families that catalyze different reactions is a larger problem than has been recognized. Strategies are suggested for addressing some of the systematic problems contributing to these high levels of misannotation.
Nehrt et al. (2011), Altenhoff et al. (2012) — the ortholog conjecture.
Testing the Ortholog Conjecture with Comparative Functional Genomic Data from Mammals. Nehrt et al. (2011). PLoS Computational Biology. doi:10.1371/journal.pcbi.1002073
Abstract
A common assumption in comparative genomics is that orthologous genes share greater functional similarity than do paralogous genes (the "ortholog conjecture"). Many methods used to computationally predict protein function are based on this assumption, even though it is largely untested. Here we present the first large-scale test of the ortholog conjecture using comparative functional genomic data from human and mouse. We use the experimentally derived functions of more than 8,900 genes, as well as an independent microarray dataset, to directly assess our ability to predict function using both orthologs and paralogs. Both datasets show that paralogs are often a much better predictor of function than are orthologs, even at lower sequence identities. Among paralogs, those found within the same species are consistently more functionally similar than those found in a different species. We also find that paralogous pairs residing on the same chromosome are more functionally similar than those on different chromosomes, perhaps due to higher levels of interlocus gene conversion between these pairs. In addition to offering implications for the computational prediction of protein function, our results shed light on the relationship between sequence divergence and functional divergence. We conclude that the most important factor in the evolution of function is not amino acid sequence, but rather the cellular context in which proteins act.
Resolving the Ortholog Conjecture: Orthologs Tend to Be Weakly, but Significantly, More Similar in Function than Paralogs. Altenhoff et al. (2012). PLoS Computational Biology. doi:10.1371/journal.pcbi.1002514
Abstract
The function of most proteins is not determined experimentally, but is extrapolated from homologs. According to the "ortholog conjecture", or standard model of phylogenomics, protein function changes rapidly after duplication, leading to paralogs with different functions, while orthologs retain the ancestral function. We report here that a comparison of experimentally supported functional annotations among homologs from 13 genomes mostly supports this model. We show that to analyze GO annotation effectively, several confounding factors need to be controlled: authorship bias, variation of GO term frequency among species, variation of background similarity among species pairs, and propagated annotation bias. After controlling for these biases, we observe that orthologs have generally more similar functional annotations than paralogs. This is especially strong for sub-cellular localization. We observe only a weak decrease in functional similarity with increasing sequence divergence. These findings hold over a large diversity of species; notably orthologs from model organisms such as E. coli, yeast or mouse have conserved function with human proteins.
Cantalapiedra et al. (2021) — eggNOG-mapper.
eggNOG-mapper v2: Functional Annotation, Orthology Assignments, and Domain Prediction at the Metagenomic Scale. Cantalapiedra et al. (2021). Molecular Biology and Evolution. doi:10.1093/molbev/msab293
Abstract
Even though automated functional annotation of genes represents a fundamental step in most genomic and metagenomic workflows, it remains challenging at large scales. Here, we describe a major upgrade to eggNOG-mapper, a tool for functional annotation based on precomputed orthology assignments, now optimized for vast (meta)genomic data sets. Improvements in version 2 include a full update of both the genomes and functional databases to those from eggNOG v5, as well as several efficiency enhancements and new features. Most notably, eggNOG-mapper v2 now allows for: 1) de novo gene prediction from raw contigs, 2) built-in pairwise orthology prediction, 3) fast protein domain discovery, and 4) automated GFF decoration. eggNOG-mapper v2 is available as a standalone tool or as an online service at http://eggnog-mapper.embl.de.
Blum et al. (2024), Bileschi et al. (2022) — InterPro and deep-learning family classification.
InterPro: the protein sequence classification resource in 2025. Blum et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae1082
Abstract
InterPro (https://www.ebi.ac.uk/interpro) is a freely accessible resource for the classification of protein sequences into families. It integrates predictive models, known as signatures, from multiple member databases to classify sequences into families and predict the presence of domains and significant sites. The InterPro database provides annotations for over 200 million sequences, ensuring extensive coverage of UniProtKB, the standard repository of protein sequences, and includes mappings to several other major resources, such as Gene Ontology (GO), Protein Data Bank in Europe (PDBe) and the AlphaFold Protein Structure Database. In this publication, we report on the status of InterPro (version 101.0), detailing new developments in the database, associated web interface and software. Notable updates include the increased integration of structures predicted by AlphaFold and the enhanced description of protein families using artificial intelligence. Over the past two years, more than 5000 new InterPro entries have been created. The InterPro website now offers access to 85 000 protein families and domains from its member databases and serves as a long-term archive for retired databases. InterPro data, software and tools are freely available.
Using deep learning to annotate the protein universe. Bileschi et al. (2022). Nature Biotechnology. doi:10.1038/s41587-021-01179-w
Abstract
Understanding the relationship between amino acid sequence and protein function is a long-standing challenge with far-reaching scientific and translational implications. State-of-the-art alignment-based techniques cannot predict function for one-third of microbial protein sequences, hampering our ability to exploit data from diverse organisms. Here, we train deep learning models to accurately predict functional annotations for unaligned amino acid sequences across rigorous benchmark assessments built from the 17,929 families of the protein families database Pfam. The models infer known patterns of evolutionary substitutions and learn representations that accurately cluster sequences from unseen families. Combining deep models with existing methods significantly improves remote homology detection, suggesting that the deep models learn complementary information. This approach extends the coverage of Pfam by >9.5%, exceeding additions made over the last decade, and predicts function for 360 human reference proteome proteins with no previous Pfam annotation. These results suggest that deep learning models will be a core component of future protein annotation tools.
B.10.3 Genomic context and networks
Marcotte et al. (1999) — a combined algorithm for genome-wide function prediction.
A combined algorithm for genome-wide prediction of protein function. Marcotte et al. (1999). Nature. doi:10.1038/47048
Abstract
The availability of over 20 fully sequenced genomes has driven the development of new methods to find protein function and interactions. Here we group proteins by correlated evolution, correlated messenger RNA expression patterns and patterns of domain fusion to determine functional relationships among the 6,217 proteins of the yeast Saccharomyces cerevisiae. Using these methods, we discover over 93,000 pairwise links between functionally related yeast proteins. Links between characterized and uncharacterized proteins allow a general function to be assigned to more than half of the 2,557 previously uncharacterized yeast proteins. Examples of functional links are given for a protein family of previously unknown function, a protein whose human homologues are implicated in colon cancer and the yeast prion Sup35.
Enright and Ouzounis (2001), Tsoka and Ouzounis (2000), Snel et al. (2000), R. F. Doolittle (1999) — gene fusion and fission.
Functional associations of proteins in entire genomes by means of exhaustive detection of gene fusions. Enright & Ouzounis (2001). Genome Biology. doi:10.1186/gb-2001-2-9-research0034
Abstract
Background It has recently been shown that the detection of gene fusion events across genomes can be used for predicting functional associations of proteins, including physical interaction or complex formation. To obtain such predictions we have made an exhaustive search for gene fusion events within 24 available completely sequenced genomes. Results Each genome was used as a query against the remaining 23 complete genomes to detect gene fusion events. Using an improved, fully automatic protocol, a total of 7,224 single-domain proteins that are components of gene fusions in other genomes were detected, many of which were identified for the first time. The total number of predicted pairwise functional associations is 39,730 for all genomes. Component pairs were identified by virtue of their similarity to 2,365 multidomain composite proteins. We also show for the first time that gene fusion is a complex evolutionary process with a number of contributory factors, including paralogy, genome size and phylogenetic distance. On average, 9% of genes in a given genome appear to code for single-domain, component proteins predicted to be functionally associated. These proteins are detected by an additional 4% of genes that code for fused, composite proteins. Conclusions These results provide an exhaustive set of functionally associated genes and also delineate the power of fusion analysis for the prediction of protein interactions.
Prediction of protein interactions: metabolic enzymes are frequently involved in gene fusion. Tsoka & Ouzounis (2000). Nature Genetics. doi:10.1038/79847
Genome evolution. Snel et al. (2000). Trends in Genetics. doi:10.1016/s0168-9525(99)01924-1
Do you dig my groove?. Doolittle (1999). Nature Genetics. doi:10.1038/12597
Dandekar (1998), Overbeek et al. (1999) — conserved gene order.
Conservation of gene order: a fingerprint of proteins that physically interact. Dandekar (1998). Trends in Biochemical Sciences. doi:10.1016/s0968-0004(98)01274-2
Abstract
A systematic comparison of nine bacterial and archaeal genomes reveals a low level of gene-order (and operon architecture) conservation. Nevertheless, a number of gene pairs are conserved. The proteins encoded by conserved gene pairs appear to interact physically. This observation can therefore be used to predict functions of, and interactions between, prokaryotic gene products.
The use of gene clusters to infer functional coupling. Overbeek et al. (1999). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.96.6.2896
Abstract
Previously, we presented evidence that it is possible to predict functional coupling between genes based on conservation of gene clusters between genomes. With the rapid increase in the availability of prokaryotic sequence data, it has become possible to verify and apply the technique. In this paper, we extend our characterization of the parameters that determine the utility of the approach, and we generalize the approach in a way that supports detection of common classes of functionally coupled genes (e.g., transport and signal transduction clusters). Now that the analysis includes over 30 complete or nearly complete genomes, it has become clear that this approach will play a significant role in supporting efforts to assign functionality to the remaining uncharacterized genes in sequenced genomes.
Pellegrini et al. (1999) — phylogenetic profiles.
Assigning protein functions by comparative genome analysis: Protein phylogenetic profiles. Pellegrini et al. (1999). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.96.8.4285
Abstract
Determining protein functions from genomic sequences is a central goal of bioinformatics. We present a method based on the assumption that proteins that function together in a pathway or structural complex are likely to evolve in a correlated fashion. During evolution, all such functionally linked proteins tend to be either preserved or eliminated in a new species. We describe this property of correlated evolution by characterizing each protein by its phylogenetic profile, a string that encodes the presence or absence of a protein in every known genome. We show that proteins having matching or similar profiles strongly tend to be functionally linked. This method of phylogenetic profiling allows us to predict the function of uncharacterized proteins.
Sharan et al. (2007) — network-based prediction of protein function.
Network-based prediction of protein function. Sharan et al. (2007). Molecular Systems Biology. doi:10.1038/msb4100129
Abstract
Functional annotation of proteins is a fundamental problem in the post-genomic era. The recent availability of protein interaction networks for many model species has spurred on the development of computational methods for interpreting such data in order to elucidate protein function. In this review, we describe the current computational approaches for the task, including direct methods, which propagate functional information through the network, and module-assisted methods, which infer functional modules within the network and use those for the annotation task. Although a broad variety of interesting approaches has been developed, further progress in the field will depend on systematic evaluation of the methods and their dissemination in the biological community.
★ Garzón et al. (2016) — a computational interactome and functional annotation of the human proteome.
A computational interactome and functional annotation for the human proteome. Garzón et al. (2016). eLife. doi:10.7554/elife.18715
Abstract
We present a database, PrePPI (Predicting Protein-Protein Interactions), of more than 1.35 million predicted protein-protein interactions (PPIs). Of these at least 127,000 are expected to constitute direct physical interactions although the actual number may be much larger (~500,000). The current PrePPI, which contains predicted interactions for about 85% of the human proteome, is related to an earlier version but is based on additional sources of interaction evidence and is far larger in scope. The use of structural relationships allows PrePPI to infer numerous previously unreported interactions. PrePPI has been subjected to a series of validation tests including reproducing known interactions, recapitulating multi-protein complexes, analysis of disease associated SNPs, and identifying functional relationships between interacting proteins. We show, using Gene Set Enrichment Analysis (GSEA), that predicted interaction partners can be used to annotate a protein's function. We provide annotations for most human proteins, including many annotated as having unknown function.
B.10.4 Structure and machine learning
Laskowski et al. (2005) — ProFunc: function from 3D structure.
ProFunc: a server for predicting protein function from 3D structure. Laskowski et al. (2005). Nucleic Acids Research. doi:10.1093/nar/gki414
Abstract
ProFunc (http://www.ebi.ac.uk/thornton-srv/databases/ProFunc) is a web server for predicting the likely function of proteins whose 3D structure is known but whose function is not. Users submit the coordinates of their structure to the server in PDB format. ProFunc makes use of both existing and novel methods to analyse the protein's sequence and structure identifying functional motifs or close relationships to functionally characterized proteins. A summary of the analyses provides an at-a-glance view of what each of the different methods has found. More detailed results are available on separate pages. Often where one method has failed to find anything useful another may be more forthcoming. The server is likely to be of most use in structural genomics where a large proportion of the proteins whose structures are solved are of hypothetical proteins of unknown function. However, it may also find use in a comparative analysis of members of large protein families. It provides a convenient compendium of sequence and structural information that often hold vital functional clues to be followed up experimentally.
Gligorijević et al. (2021) — DeepFRI: graph convolutional networks on structures.
Structure-based protein function prediction using graph convolutional networks. Gligorijević et al. (2021). Nature Communications. doi:10.1038/s41467-021-23303-9
Abstract
The rapid increase in the number of proteins in sequence databases and the diversity of their functions challenge computational approaches for automated function prediction. Here, we introduce DeepFRI, a Graph Convolutional Network for predicting protein functions by leveraging sequence features extracted from a protein language model and protein structures. It outperforms current leading methods and sequence-based Convolutional Neural Networks and scales to the size of current sequence repositories. Augmenting the training set of experimental structures with homology models allows us to significantly expand the number of predictable functions. DeepFRI has significant de-noising capability, with only a minor drop in performance when experimental structures are replaced by protein models. Class activation mapping allows function predictions at an unprecedented resolution, allowing site-specific annotations at the residue-level in an automated manner. We show the utility and high performance of our method by annotating structures from the PDB and SWISS-MODEL, making several new confident function predictions. DeepFRI is available as a webserver at https://beta.deepfri.flatironinstitute.org/ .
Kulmanov and Hoehndorf (2019) — DeepGOPlus.
DeepGOPlus: improved protein function prediction from sequence. Kulmanov & Hoehndorf (2019). Bioinformatics. doi:10.1093/bioinformatics/btz595
Abstract
Motivation Protein function prediction is one of the major tasks of bioinformatics that can help in wide range of biological problems such as understanding disease mechanisms or finding drug targets. Many methods are available for predicting protein functions from sequence based features, protein-protein interaction networks, protein structure or literature. However, other than sequence, most of the features are difficult to obtain or not available for many proteins thereby limiting their scope. Furthermore, the performance of sequence-based function prediction methods is often lower than methods that incorporate multiple features and predicting protein functions may require a lot of time. Results We developed a novel method for predicting protein functions from sequence alone which combines deep convolutional neural network (CNN) model with sequence similarity based predictions. Our CNN model scans the sequence for motifs which are predictive for protein functions and combines this with functions of similar proteins (if available). We evaluate the performance of DeepGOPlus using the CAFA3 evaluation measures and achieve an Fmax of 0.390, 0.557 and 0.614 for BPO, MFO and CCO evaluations, respectively. These results would have made DeepGOPlus one of the three best predictors in CCO and the second best performing method in the BPO and MFO evaluations. We also compare DeepGOPlus with state-of-the-art methods such as DeepText2GO and GOLabeler on another dataset. DeepGOPlus can annotate around 40 protein sequences per second on common hardware, thereby making fast and accurate function predictions available for a wide range of proteins. Availability and implementation http://deepgoplus.bio2vec.net/ . Supplementary information Supplementary data are available at Bioinformatics online.
Littmann et al. (2021) — transferring GO annotations with embeddings.
Embeddings from deep learning transfer GO annotations beyond homology. Littmann et al. (2021). Scientific Reports. doi:10.1038/s41598-020-80786-0
Abstract
Knowing protein function is crucial to advance molecular and medical biology, yet experimental function annotations through the Gene Ontology (GO) exist for fewer than 0.5% of all known proteins. Computational methods bridge this sequence-annotation gap typically through homology-based annotation transfer by identifying sequence-similar proteins with known function or through prediction methods using evolutionary information. Here, we propose predicting GO terms through annotation transfer based on proximity of proteins in the SeqVec embedding rather than in sequence space. These embeddings originate from deep learned language models (LMs) for protein sequences (SeqVec) transferring the knowledge gained from predicting the next amino acid in 33 million protein sequences. Replicating the conditions of CAFA3, our method reaches an F max of 37 ± 2%, 50 ± 3%, and 57 ± 2% for BPO, MFO, and CCO, respectively. Numerically, this appears close to the top ten CAFA3 methods. When restricting the annotation transfer to proteins with max BPO 33 ± 2%, MFO 43 ± 3%, CCO 53 ± 2%); this still outperforms naïve sequence-based transfer. Preliminary results from CAFA4 appear to confirm these findings. Overall, this new concept is likely to change the annotation of proteins, in particular for proteins from smaller families or proteins with intrinsically disordered regions.
Yu et al. (2023) — CLEAN: enzyme function prediction with contrastive learning.
Enzyme function prediction using contrastive learning. Yu et al. (2023). Science. doi:10.1126/science.adf2465
Abstract
Enzyme function annotation is a fundamental challenge, and numerous computational tools have been developed. However, most of these tools cannot accurately predict functional annotations, such as enzyme commission (EC) number, for less-studied proteins or those with previously uncharacterized functions or multiple activities. We present a machine learning algorithm named CLEAN (contrastive learning-enabled enzyme annotation) to assign EC numbers to enzymes with better accuracy, reliability, and sensitivity compared with the state-of-the-art tool BLASTp. The contrastive learning framework empowers CLEAN to confidently (i) annotate understudied enzymes, (ii) correct mislabeled enzymes, and (iii) identify promiscuous enzymes with two or more EC numbers-functions that we demonstrate by systematic in silico and in vitro experiments. We anticipate that this tool will be widely used for predicting the functions of uncharacterized enzymes, thereby advancing many fields, such as genomics, synthetic biology, and biocatalysis.
B.10.5 Localisation
Blobel and Dobberstein (1975) — the signal hypothesis.
Transfer of proteins across membranes. I. Presence of proteolytically processed and unprocessed nascent immunoglobulin light chains on membrane-bound ribosomes of murine myeloma.. Blobel & Dobberstein (1975). The Journal of cell biology. doi:10.1083/jcb.67.3.835
Abstract
Fractionation of MOPC 41 DL-1 tumors revealed that the mRNA for the light chain of immunoglobulin is localized exclusively in membrane-bound ribosomes. It was shown that the translation product of isolated light chain mRNA in a heterologous protein-synthesizing system in vitro is larger than the authentic secreted light chain; this confirms similar results from several laboratories. The synthesis in vitro of a precursor protein of the light chain is not an artifact of translation in a heterologous system, because it was shown that detached polysomes, isolated from detergent-treated rough microsomes, not only contain nascent light chains which have already been proteolytically processed in vivo but also contain unprocessed nascent light chains. In vitro completion of these nascent light chains thus resulted in the synthesis of some chains having the same mol wt as the authentic secreted light chains, because of completion of in vivo proteolytically processed chains and of other chains which, due to the completion of unprocessed chains, have the same mol wt as the precursor of the light chain. In contrast, completion of the nascent light chains contained in rough microsomes resulted in the synthesis of only processed light chains. Taken together, these results indicate that the processing activity is present in isolated rough microsomes, that it is localized in the membrane moiety of rough microsomes, and, therefore, that it was most likely solubilized during detergent treatment used for the isolation of detached polysomes. Furthermore, these results established that processing in vivo takes place before completion of the nascent chain. The data also indicate that in vitro processing of nascent chains by rough microsomes is dependent on ribosome binding to the membrane. If the latter process is interfered with by aurintricarboxylic acid, rough microsomes also synthesize some unprocessed chains. The data presented in this paper have been interpreted in the light of a recently proposed hypothesis. This hypothesis, referred to as the signal hypothesis, is described in greater detail in the Discussion section.
Heijne (1986) — predicting signal-sequence cleavage sites.
A new method for predicting signal sequence cleavage sites. von Heijne (1986). Nucleic Acids Research. doi:10.1093/nar/14.11.4683
Abstract
A new method for identifying secretory signal sequences and for predicting the site of cleavage between a signal sequence and the mature exported protein is described. The predictive accuracy is estimated to be around 75-80% for both prokaryotic and eukaryotic proteins.
Nielsen et al. (1997), Petersen et al. (2011) — SignalP and the transmembrane problem.
Identification of prokaryotic and eukaryotic signal peptides and prediction of their cleavage sites. Nielsen et al. (1997). Protein Engineering Design and Selection. doi:10.1093/protein/10.1.1
Abstract
We have developed a new method for the identification of signal peptides and their cleavage sites based on neural networks trained on separate sets of prokaryotic and eukaryotic sequence. The method performs significantly better than previous prediction schemes and can easily be applied on genome-wide data sets. Discrimination between cleaved signal peptides and uncleaved N-terminal signal-anchor sequences is also possible, though with lower precision. Predictions can be made on a publicly available WWW server.
SignalP 4.0: discriminating signal peptides from transmembrane regions. Petersen et al. (2011). Nature Methods. doi:10.1038/nmeth.1701
Nakai and Horton (1999) — PSORT.
PSORT: a program for detecting sorting signals in proteins and predicting their subcellular localization. Nakai & Horton (1999). Trends in Biochemical Sciences. doi:10.1016/s0968-0004(98)01336-x
Emanuelsson et al. (2000), Almagro Armenteros et al. (2019) — TargetP.
Predicting Subcellular Localization of Proteins Based on their N-terminal Amino Acid Sequence. Emanuelsson et al. (2000). Journal of Molecular Biology. doi:10.1006/jmbi.2000.3903
Abstract
A neural network-based tool, TargetP, for large-scale subcellular location prediction of newly identified proteins has been developed. Using N-terminal sequence information only, it discriminates between proteins destined for the mitochondrion, the chloroplast, the secretory pathway, and "other" localizations with a success rate of 85% (plant) or 90% (non-plant) on redundancy-reduced test sets. From a TargetP analysis of the recently sequenced Arabidopsis thaliana chromosomes 2 and 4 and the Ensembl Homo sapiens protein set, we estimate that 10% of all plant proteins are mitochondrial and 14% chloroplastic, and that the abundance of secretory proteins, in both Arabidopsis and Homo, is around 10%. TargetP also predicts cleavage sites with levels of correctly predicted sites ranging from approximately 40% to 50% (chloroplastic and mitochondrial presequences) to above 70% (secretory signal peptides). TargetP is available as a web-server at http://www.cbs.dtu.dk/services/TargetP/.
Detecting sequence signals in targeting peptides using deep learning. Almagro Armenteros et al. (2019). Life Science Alliance. doi:10.26508/lsa.201900429
Abstract
In bioinformatics, machine learning methods have been used to predict features embedded in the sequences. In contrast to what is generally assumed, machine learning approaches can also provide new insights into the underlying biology. Here, we demonstrate this by presenting TargetP 2.0, a novel state-of-the-art method to identify N-terminal sorting signals, which direct proteins to the secretory pathway, mitochondria, and chloroplasts or other plastids. By examining the strongest signals from the attention layer in the network, we find that the second residue in the protein, that is, the one following the initial methionine, has a strong influence on the classification. We observe that two-thirds of chloroplast and thylakoid transit peptides have an alanine in position 2, compared with 20% in other plant proteins. We also note that in fungi and single-celled eukaryotes, less than 30% of the targeting peptides have an amino acid that allows the removal of the N-terminal methionine compared with 60% for the proteins without targeting peptide. The importance of this feature for predictions has not been highlighted before.
★ Almagro Armenteros et al. (2017), Thumuluri et al. (2022) — DeepLoc and DeepLoc 2.0.
DeepLoc: prediction of protein subcellular localization using deep learning. Almagro Armenteros et al. (2017). Bioinformatics. doi:10.1093/bioinformatics/btx431
Abstract
Motivation The prediction of eukaryotic protein subcellular localization is a well-studied topic in bioinformatics due to its relevance in proteomics research. Many machine learning methods have been successfully applied in this task, but in most of them, predictions rely on annotation of homologues from knowledge databases. For novel proteins where no annotated homologues exist, and for predicting the effects of sequence variants, it is desirable to have methods for predicting protein properties from sequence information only. Results Here, we present a prediction algorithm using deep neural networks to predict protein subcellular localization relying only on sequence information. At its core, the prediction model uses a recurrent neural network that processes the entire protein sequence and an attention mechanism identifying protein regions important for the subcellular localization. The model was trained and tested on a protein dataset extracted from one of the latest UniProt releases, in which experimentally annotated proteins follow more stringent criteria than previously. We demonstrate that our model achieves a good accuracy (78% for 10 categories; 92% for membrane-bound or soluble), outperforming current state-of-the-art algorithms, including those relying on homology information. Availability and implementation The method is available as a web server at http://www.cbs.dtu.dk/services/DeepLoc. Example code is available at https://github.com/JJAlmagro/subcellular_localization. The dataset is available at http://www.cbs.dtu.dk/services/DeepLoc/data.php. Contact jjalma@dtu.dk.
DeepLoc 2.0: multi-label subcellular localization prediction using protein language models. Thumuluri et al. (2022). Nucleic Acids Research. doi:10.1093/nar/gkac278
Abstract
The prediction of protein subcellular localization is of great relevance for proteomics research. Here, we propose an update to the popular tool DeepLoc with multi-localization prediction and improvements in both performance and interpretability. For training and validation, we curate eukaryotic and human multi-location protein datasets with stringent homology partitioning and enriched with sorting signal information compiled from the literature. We achieve state-of-the-art performance in DeepLoc 2.0 by using a pre-trained protein language model. It has the further advantage that it uses sequence input rather than relying on slower protein profiles. We provide two means of better interpretability: an attention output along the sequence and highly accurate prediction of nine different types of protein sorting signals. We find that the attention output correlates well with the position of sorting signals. The webserver is available at services.healthtech.dtu.dk/service.php?DeepLoc-2.0.
B.10.6 Variant effects
Ng (2003), Adzhubei et al. (2010), Kircher et al. (2014) — SIFT, PolyPhen-2 and CADD.
SIFT: predicting amino acid changes that affect protein function. Ng (2003). Nucleic Acids Research. doi:10.1093/nar/gkg509
Abstract
Single nucleotide polymorphism (SNP) studies and random mutagenesis projects identify amino acid substitutions in protein-coding regions. Each substitution has the potential to affect protein function. SIFT (Sorting Intolerant From Tolerant) is a program that predicts whether an amino acid substitution affects protein function so that users can prioritize substitutions for further study. We have shown that SIFT can distinguish between functionally neutral and deleterious amino acid changes in mutagenesis studies and on human polymorphisms. SIFT is available at http://blocks.fhcrc.org/sift/SIFT.html.
A method and server for predicting damaging missense mutations. Adzhubei et al. (2010). Nature Methods. doi:10.1038/nmeth0410-248
A general framework for estimating the relative pathogenicity of human genetic variants. Kircher et al. (2014). Nature Genetics. doi:10.1038/ng.2892
Abstract
Current methods for annotating and interpreting human genetic variation tend to exploit a single information type (for example, conservation) and/or are restricted in scope (for example, to missense changes). Here we describe Combined Annotation-Dependent Depletion (CADD), a method for objectively integrating many diverse annotations into a single measure (C score) for each variant. We implement CADD as a support vector machine trained to differentiate 14.7 million high-frequency human-derived alleles from 14.7 million simulated variants. We precompute C scores for all 8.6 billion possible human single-nucleotide variants and enable scoring of short insertions-deletions. C scores correlate with allelic diversity, annotations of functionality, pathogenicity, disease severity, experimentally measured regulatory effects and complex trait associations, and they highly rank known pathogenic variants within individual genomes. The ability of CADD to prioritize functional, deleterious and pathogenic variants across many functional categories, effect sizes and genetic architectures is unmatched by any current single-annotation method.
Grimm et al. (2015) — circularity in the evaluation of variant predictors.
The Evaluation of Tools Used to Predict the Impact of Missense Variants Is Hindered by Two Types of Circularity. Grimm et al. (2015). Human Mutation. doi:10.1002/humu.22768
Abstract
Prioritizing missense variants for further experimental investigation is a key challenge in current sequencing studies for exploring complex and Mendelian diseases. A large number of in silico tools have been employed for the task of pathogenicity prediction, including PolyPhen-2, SIFT, FatHMM, MutationTaster-2, MutationAssessor, Combined Annotation Dependent Depletion, LRT, phyloP, and GERP++, as well as optimized methods of combining tool scores, such as Condel and Logit. Due to the wealth of these methods, an important practical question to answer is which of these tools generalize best, that is, correctly predict the pathogenic character of new variants. We here demonstrate in a study of 10 tools on five datasets that such a comparative evaluation of these tools is hindered by two types of circularity: they arise due to (1) the same variants or (2) different variants from the same protein occurring both in the datasets used for training and for evaluation of these tools, which may lead to overly optimistic results. We show that comparative evaluations of predictors that do not address these types of circularity may erroneously conclude that circularity confounded tools are most accurate among all tools, and may even outperform optimized combinations of tools.
Hopf et al. (2017), Riesselman et al. (2018), Frazer et al. (2021) — EVmutation, DeepSequence and EVE.
Mutation effects predicted from sequence co-variation. Hopf et al. (2017). Nature Biotechnology. doi:10.1038/nbt.3769
Abstract
Many high-throughput experimental technologies have been developed to assess the effects of large numbers of mutations (variation) on phenotypes. However, designing functional assays for these methods is challenging, and systematic testing of all combinations is impossible, so robust methods to predict the effects of genetic variation are needed. Most prediction methods exploit evolutionary sequence conservation but do not consider the interdependencies of residues or bases. We present EVmutation, an unsupervised statistical method for predicting the effects of mutations that explicitly captures residue dependencies between positions. We validate EVmutation by comparing its predictions with outcomes of high-throughput mutagenesis experiments and measurements of human disease mutations and show that it outperforms methods that do not account for epistasis. EVmutation can be used to assess the quantitative effects of mutations in genes of any organism. We provide pre-computed predictions for ∼7,000 human proteins at http://evmutation.org/.
Deep generative models of genetic variation capture the effects of mutations. Riesselman et al. (2018). Nature Methods. doi:10.1038/s41592-018-0138-4
Abstract
The functions of proteins and RNAs are defined by the collective interactions of many residues, and yet most statistical models of biological sequences consider sites nearly independently. Recent approaches have demonstrated benefits of including interactions to capture pairwise covariation, but leave higher-order dependencies out of reach. Here we show how it is possible to capture higher-order, context-dependent constraints in biological sequences via latent variable models with nonlinear dependencies. We found that DeepSequence ( https://github.com/debbiemarkslab/DeepSequence ), a probabilistic model for sequence families, predicted the effects of mutations across a variety of deep mutational scanning experiments substantially better than existing methods based on the same evolutionary data. The model, learned in an unsupervised manner solely on the basis of sequence information, is grounded with biologically motivated priors, reveals the latent organization of sequence families, and can be used to explore new parts of sequence space.
Disease variant prediction with deep generative models of evolutionary data. Frazer et al. (2021). Nature. doi:10.1038/s41586-021-04043-8
Abstract
Quantifying the pathogenicity of protein variants in human disease-related genes would have a marked effect on clinical decisions, yet the overwhelming majority (over 98%) of these variants still have unknown consequences 1-3 . In principle, computational methods could support the large-scale interpretation of genetic variants. However, state-of-the-art methods 4-10 have relied on training machine learning models on known disease labels. As these labels are sparse, biased and of variable quality, the resulting models have been considered insufficiently reliable 11 . Here we propose an approach that leverages deep generative models to predict variant pathogenicity without relying on labels. By modelling the distribution of sequence variation across organisms, we implicitly capture constraints on the protein sequences that maintain fitness. Our model EVE (evolutionary model of variant effect) not only outperforms computational approaches that rely on labelled data but also performs on par with, if not better than, predictions from high-throughput experiments, which are increasingly used as evidence for variant classification 12-16 . We predict the pathogenicity of more than 36 million variants across 3,219 disease genes and provide evidence for the classification of more than 256,000 variants of unknown significance. Our work suggests that models of evolutionary information can provide valuable independent evidence for variant interpretation that will be widely useful in research and clinical settings.
Meier et al. (2021), Brandes et al. (2023) — variant effects from protein language models.
Language models enable zero-shot prediction of the effects of mutations on protein function. Meier et al. (2021). bioRxiv (preprint). doi:10.1101/2021.07.09.450648
Abstract
Modeling the effect of sequence variation on function is a fundamental problem for understanding and designing proteins. Since evolution encodes information about function into patterns in protein sequences, unsupervised models of variant effects can be learned from sequence data. The approach to date has been to fit a model to a family of related sequences. The conventional setting is limited, since a new model must be trained for each prediction task. We show that using only zero-shot inference, without any supervision from experimental data or additional training, protein language models capture the functional effects of sequence variation, performing at state-of-the-art.
Genome-wide prediction of disease variant effects with a deep protein language model. Brandes et al. (2023). Nature Genetics. doi:10.1038/s41588-023-01465-0
Abstract
Predicting the effects of coding variants is a major challenge. While recent deep-learning models have improved variant effect prediction accuracy, they cannot analyze all coding variants due to dependency on close homologs or software limitations. Here we developed a workflow using ESM1b, a 650-million-parameter protein language model, to predict all ~450 million possible missense variant effects in the human genome, and made all predictions available on a web portal. ESM1b outperformed existing methods in classifying ~150,000 ClinVar/HGMD missense variants as pathogenic or benign and predicting measurements across 28 deep mutational scan datasets. We further annotated ~2 million variants as damaging only in specific protein isoforms, demonstrating the importance of considering all isoforms when predicting variant effects. Our approach also generalizes to more complex coding variants such as in-frame indels and stop-gains. Together, these results establish protein language models as an effective, accurate and general approach to predicting variant effects.
Fowler and Fields (2014) — deep mutational scanning.
Deep mutational scanning: a new style of protein science. Fowler & Fields (2014). Nature Methods. doi:10.1038/nmeth.3027
Abstract
Mutagenesis provides insight into proteins, but only recently have assays that couple genotype to phenotype been used to assess the activities of as many as 1 million mutant versions of a protein in a single experiment. This approach-'deep mutational scanning'-yields large-scale data sets that can reveal intrinsic protein properties, protein behavior within cells and the consequences of human genetic variation. Deep mutational scanning is transforming the study of proteins, but many challenges must be tackled for it to fulfill its promise.
Landrum et al. (2017), Chen et al. (2023) — ClinVar and gnomAD.
ClinVar: improving access to variant interpretations and supporting evidence. Landrum et al. (2017). Nucleic Acids Research. doi:10.1093/nar/gkx1153
Abstract
ClinVar (https://www.ncbi.nlm.nih.gov/clinvar/) is a freely available, public archive of human genetic variants and interpretations of their significance to disease, maintained at the National Institutes of Health. Interpretations of the clinical significance of variants are submitted by clinical testing laboratories, research laboratories, expert panels and other groups. ClinVar aggregates data by variant-disease pairs, and by variant (or set of variants). Data aggregated by variant are accessible on the website, in an improved set of variant call format files and as a new comprehensive XML report. ClinVar recently started accepting submissions that are focused primarily on providing phenotypic information for individuals who have had genetic testing. Submissions may come from clinical providers providing their own interpretation of the variant ('provider interpretation') or from groups such as patient registries that primarily provide phenotypic information from patients ('phenotyping only'). ClinVar continues to make improvements to its search and retrieval functions. Several new fields are now indexed for more precise searching, and filters allow the user to narrow down a large set of search results.
A genomic mutational constraint map using variation in 76,156 human genomes. Chen et al. (2023). Nature. doi:10.1038/s41586-023-06045-0
Abstract
The depletion of disruptive variation caused by purifying natural selection (constraint) has been widely used to investigate protein-coding genes underlying human disorders 1-4 , but attempts to assess constraint for non-protein-coding regions have proved more difficult. Here we aggregate, process and release a dataset of 76,156 human genomes from the Genome Aggregation Database (gnomAD)-the largest public open-access human genome allele frequency reference dataset-and use it to build a genomic constraint map for the whole genome (genomic non-coding constraint of haploinsufficient variation (Gnocchi)). We present a refined mutational model that incorporates local sequence context and regional genomic features to detect depletions of variation. As expected, the average constraint for protein-coding sequences is stronger than that for non-coding regions. Within the non-coding genome, constrained regions are enriched for known regulatory elements and variants that are implicated in complex human diseases and traits, facilitating the triangulation of biological annotation, disease association and natural selection to non-coding DNA analysis. More constrained regulatory elements tend to regulate more constrained protein-coding genes, which in turn suggests that non-coding constraint can aid the identification of constrained genes that are as yet unrecognized by current gene constraint metrics. We demonstrate that this genome-wide constraint map improves the identification and interpretation of functional human genetic variation.
Richards et al. (2015) — clinical guidelines for variant interpretation.
Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Richards et al. (2015). Genetics in Medicine. doi:10.1038/gim.2015.30
Abstract
The American College of Medical Genetics and Genomics (ACMG) previously developed guidance for the interpretation of sequence variants.(1) In the past decade, sequencing technology has evolved rapidly with the advent of high-throughput next-generation sequencing. By adopting and leveraging next-generation sequencing, clinical laboratories are now performing an ever-increasing catalogue of genetic testing spanning genotyping, single genes, gene panels, exomes, genomes, transcriptomes, and epigenetic assays for genetic disorders. By virtue of increased complexity, this shift in genetic testing has been accompanied by new challenges in sequence interpretation. In this context the ACMG convened a workgroup in 2013 comprising representatives from the ACMG, the Association for Molecular Pathology (AMP), and the College of American Pathologists to revisit and revise the standards and guidelines for the interpretation of sequence variants. The group consisted of clinical laboratory directors and clinicians. This report represents expert opinion of the workgroup with input from ACMG, AMP, and College of American Pathologists stakeholders. These recommendations primarily apply to the breadth of genetic tests used in clinical laboratories, including genotyping, single genes, panels, exomes, and genomes. This report recommends the use of specific standard terminology-"pathogenic," "likely pathogenic," "uncertain significance," "likely benign," and "benign"-to describe variants identified in genes that cause Mendelian disorders. Moreover, this recommendation describes a process for classifying variants into these five categories based on criteria using typical types of variant evidence (e.g., population data, computational data, functional data, segregation data). Because of the increased complexity of analysis and interpretation of clinical genetic testing described in this report, the ACMG strongly recommends that clinical molecular genetic testing should be performed in a Clinical Laboratory Improvement Amendments-approved laboratory, with results interpreted by a board-certified clinical molecular geneticist or molecular genetic pathologist or the equivalent.
The Critical Assessment of Genome Interpretation Consortium (2024) — CAGI.
CAGI, the Critical Assessment of Genome Interpretation, establishes progress and prospects for computational genetic variant interpretation methods. The Critical Assessment of Genome Interpretation Consortium (2024). Genome Biology. doi:10.1186/s13059-023-03113-6
Abstract
Background The Critical Assessment of Genome Interpretation (CAGI) aims to advance the state-of-the-art for computational prediction of genetic variant impact, particularly where relevant to disease. The five complete editions of the CAGI community experiment comprised 50 challenges, in which participants made blind predictions of phenotypes from genetic data, and these were evaluated by independent assessors. Results Performance was particularly strong for clinical pathogenic variants, including some difficult-to-diagnose cases, and extends to interpretation of cancer-related variants. Missense variant interpretation methods were able to estimate biochemical effects with increasing accuracy. Assessment of methods for regulatory variants and complex trait disease risk was less definitive and indicates performance potentially suitable for auxiliary use in the clinic. Conclusions Results show that while current methods are imperfect, they have major utility for research and clinical applications. Emerging methods and increasingly large, robust datasets for training and assessment promise further progress ahead.
B.10.7 Benchmarking
Radivojac et al. (2013), Jiang et al. (2016) — CAFA1 and CAFA2.
A large-scale evaluation of computational protein function prediction. Radivojac et al. (2013). Nature Methods. doi:10.1038/nmeth.2340
Abstract
Automated annotation of protein function is challenging. As the number of sequenced genomes rapidly grows, the overwhelming majority of protein products can only be annotated computationally. If computational predictions are to be relied upon, it is crucial that the accuracy of these methods be high. Here we report the results from the first large-scale community-based critical assessment of protein function annotation (CAFA) experiment. Fifty-four methods representing the state of the art for protein function prediction were evaluated on a target set of 866 proteins from 11 organisms. Two findings stand out: (i) today's best protein function prediction algorithms substantially outperform widely used first-generation methods, with large gains on all types of targets; and (ii) although the top methods perform well enough to guide experiments, there is considerable need for improvement of currently available tools.
An expanded evaluation of protein function prediction methods shows an improvement in accuracy. Jiang et al. (2016). Genome Biology. doi:10.1186/s13059-016-1037-6
Abstract
Background A major bottleneck in our understanding of the molecular underpinnings of life is the assignment of function to proteins. While molecular experiments provide the most reliable annotation of proteins, their relatively low throughput and restricted purview have led to an increasing role for computational function prediction. However, assessing methods for protein function prediction and tracking progress in the field remain challenging. Results We conducted the second critical assessment of functional annotation (CAFA), a timed challenge to assess computational methods that automatically assign protein function. We evaluated 126 methods from 56 research groups for their ability to predict biological functions using Gene Ontology and gene-disease associations using Human Phenotype Ontology on a set of 3681 proteins from 18 species. CAFA2 featured expanded analysis compared with CAFA1, with regards to data set size, variety, and assessment metrics. To review progress in the field, the analysis compared the best methods from CAFA1 to those of CAFA2. Conclusions The top-performing methods in CAFA2 outperformed those from CAFA1. This increased accuracy can be attributed to a combination of the growing number of experimental annotations and improved methods for function prediction. The assessment also revealed that the definition of top-performing algorithms is ontology specific, that different performance metrics can be used to probe the nature of accurate predictions, and the relative diversity of predictions in the biological process and human phenotype ontologies. While there was methodological improvement between CAFA1 and CAFA2, the interpretation of results and usefulness of individual methods remain context-dependent.
De Paolis Kaluza et al. (2026) — CAFA5.
Advances in Protein Function Prediction from the Fifth CAFA Challenge. De Paolis Kaluza et al. (2026). bioRxiv (preprint). doi:10.64898/2026.04.27.716980
Abstract
The Critical Assessment of Functional Annotation (CAFA) is a long-standing community effort to independently assess computational methods for protein function prediction, to highlight well-performing methodologies, to identify bottlenecks in the field, and to provide a forum for the dissemination of results and exchange of ideas. In its fifth round (CAFA5) of triennial challenges, a partnership with Kaggle Inc. facilitated participation from a large community of data scientists and computational biologists through a competitive prospective challenge on the crowdsourcing platform. In this work, we present an in-depth analysis of the submitted predictions and report improvements in accuracy over all methods from the previous CAFA challenges. We further introduce a new evaluation setting for proteins with pre-existing (incomplete) annotations and identify the need for methods that better leverage existing annotations to predict those that will be discovered later. Finally, we characterize the prospective evaluation framework by examining performance on a strict set of unpublished annotations and across intermediate database releases. Our results indicate that recent developments in the field, such as the availability of protein language models and accurately predicted 3D structures, as well as the growth of experimental annotations through biocuration, have all contributed to performance improvements. 1
Clark and Radivojac (2013) — information-theoretic evaluation of ontology predictions.
Information-theoretic evaluation of predicted ontological annotations. Clark & Radivojac (2013). Bioinformatics. doi:10.1093/bioinformatics/btt228
Abstract
Motivation The development of effective methods for the prediction of ontological annotations is an important goal in computational biology, with protein function prediction and disease gene prioritization gaining wide recognition. Although various algorithms have been proposed for these tasks, evaluating their performance is difficult owing to problems caused both by the structure of biomedical ontologies and biased or incomplete experimental annotations of genes and gene products. Results We propose an information-theoretic framework to evaluate the performance of computational protein function prediction. We use a Bayesian network, structured according to the underlying ontology, to model the prior probability of a protein's function. We then define two concepts, misinformation and remaining uncertainty, that can be seen as information-theoretic analogs of precision and recall. Finally, we propose a single statistic, referred to as semantic distance, that can be used to rank classification models. We evaluate our approach by analyzing the performance of three protein function predictors of Gene Ontology terms and provide evidence that it addresses several weaknesses of currently used metrics. We believe this framework provides useful insights into the performance of protein function prediction tools. Supplementary information Supplementary data are available at Bioinformatics online.
B.11 Protein–protein interactions
Further papers for Chapter 11.
Papers marked ★ were mandatory in earlier reading lists. In the HTML version, click Abstract to show the abstract.
B.11.1 Experimental interactomes
Fields and Song (1989) — the yeast two-hybrid system.
A novel genetic system to detect protein–protein interactions. Fields & Song (1989). Nature. doi:10.1038/340245a0
Abstract
Protein-protein interactions between two proteins have generally been studied using biochemical techniques such as crosslinking, co-immunoprecipitation and co-fractionation by chromatography. We have generated a novel genetic system to study these interactions by taking advantage of the properties of the GAL4 protein of the yeast Saccharomyces cerevisiae. This protein is a transcriptional activator required for the expression of genes encoding enzymes of galactose utilization. It consists of two separable and functionally essential domains: an N-terminal domain which binds to specific DNA sequences (UASG); and a C-terminal domain containing acidic regions, which is necessary to activate transcription. We have generated a system of two hybrid proteins containing parts of GAL4: the GAL4 DNA-binding domain fused to a protein 'X' and a GAL4 activating region fused to a protein 'Y'. If X and Y can form a protein-protein complex and reconstitute proximity of the GAL4 domains, transcription of a gene regulated by UASG occurs. We have tested this system using two yeast proteins that are known to interact--SNF1 and SNF4. High transcriptional activity is obtained only when both hybrids are present in a cell. This system may be applicable as a general method to identify proteins that interact with a known protein by the use of a simple galactose selection.
Uetz et al. (2000), Ito et al. (2001) — two-hybrid screens of the yeast proteome.
A comprehensive analysis of protein–protein interactions in Saccharomyces cerevisiae. Uetz et al. (2000). Nature. doi:10.1038/35001009
Abstract
Two large-scale yeast two-hybrid screens were undertaken to identify protein-protein interactions between full-length open reading frames predicted from the Saccharomyces cerevisiae genome sequence. In one approach, we constructed a protein array of about 6,000 yeast transformants, with each transformant expressing one of the open reading frames as a fusion to an activation domain. This array was screened by a simple and automated procedure for 192 yeast proteins, with positive responses identified by their positions in the array. In a second approach, we pooled cells expressing one of about 6,000 activation domain fusions to generate a library. We used a high-throughput screening procedure to screen nearly all of the 6,000 predicted yeast proteins, expressed as Gal4 DNA-binding domain fusion proteins, against the library, and characterized positives by sequence analysis. These approaches resulted in the detection of 957 putative interactions involving 1,004 S. cerevisiae proteins. These data reveal interactions that place functionally unclassified proteins in a biological context, interactions between proteins involved in the same biological function, and interactions that link biological functions together into larger cellular processes. The results of these screens are shown here.
A comprehensive two-hybrid analysis to explore the yeast protein interactome. Ito et al. (2001). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.061034498
Abstract
Protein-protein interactions play crucial roles in the execution of various biological functions. Accordingly, their comprehensive description would contribute considerably to the functional interpretation of fully sequenced genomes, which are flooded with novel genes of unpredictable functions. We previously developed a system to examine two-hybrid interactions in all possible combinations between the approximately 6,000 proteins of the budding yeast Saccharomyces cerevisiae. Here we have completed the comprehensive analysis using this system to identify 4,549 two-hybrid interactions among 3,278 proteins. Unexpectedly, these data do not largely overlap with those obtained by the other project [Uetz, P., et al. (2000) Nature (London) 403, 623-627] and hence have substantially expanded our knowledge on the protein interaction space or interactome of the yeast. Cumulative connection of these binary interactions generates a single huge network linking the vast majority of the proteins. Bioinformatics-aided selection of biologically relevant interactions highlights various intriguing subnetworks. They include, for instance, the one that had successfully foreseen the involvement of a novel protein in spindle pole body function as well as the one that may uncover a hitherto unidentified multiprotein complex potentially participating in the process of vesicular transport. Our data would thus significantly expand and improve the protein interaction map for the exploration of genome functions that eventually leads to thorough understanding of the cell as a molecular system.
Rual et al. (2005), ★ Rolland et al. (2014) — towards the human binary interactome.
Towards a proteome-scale map of the human protein–protein interaction network. Rual et al. (2005). Nature. doi:10.1038/nature04209
Abstract
Systematic mapping of protein-protein interactions, or 'interactome' mapping, was initiated in model organisms, starting with defined biological processes and then expanding to the scale of the proteome. Although far from complete, such maps have revealed global topological and dynamic features of interactome networks that relate to known biological properties, suggesting that a human interactome map will provide insight into development and disease mechanisms at a systems level. Here we describe an initial version of a proteome-scale map of human binary protein-protein interactions. Using a stringent, high-throughput yeast two-hybrid system, we tested pairwise interactions among the products of approximately 8,100 currently available Gateway-cloned open reading frames and detected approximately 2,800 interactions. This data set, called CCSB-HI1, has a verification rate of approximately 78% as revealed by an independent co-affinity purification assay, and correlates significantly with other biological attributes. The CCSB-HI1 data set increases by approximately 70% the set of available binary interactions within the tested space and reveals more than 300 new connections to over 100 disease-associated proteins. This work represents an important step towards a systematic and comprehensive human interactome project.
A Proteome-Scale Map of the Human Interactome Network. Rolland et al. (2014). Cell. doi:10.1016/j.cell.2014.10.050
Abstract
Just as reference genome sequences revolutionized human genetics, reference maps of interactome networks will be critical to fully understand genotype-phenotype relationships. Here, we describe a systematic map of ?14,000 high-quality human binary protein-protein interactions. At equal quality, this map is ?30% larger than what is available from small-scale studies published in the literature in the last few decades. While currently available information is highly biased and only covers a relatively small portion of the proteome, our systematic map appears strikingly more homogeneous, revealing a "broader" human interactome network than currently appreciated. The map also uncovers significant interconnectivity between known and candidate cancer gene products, providing unbiased evidence for an expanded functional cancer landscape, while demonstrating how high-quality interactome models will help "connect the dots" of the genomic revolution.
Yu et al. (2008), Braun et al. (2008), Venkatesan et al. (2008) — high-quality binary maps and how to assess them.
High-Quality Binary Protein Interaction Map of the Yeast Interactome Network. Yu et al. (2008). Science. doi:10.1126/science.1158684
Abstract
Current yeast interactome network maps contain several hundred molecular complexes with limited and somewhat controversial representation of direct binary interactions. We carried out a comparative quality assessment of current yeast interactome data sets, demonstrating that high-throughput yeast two-hybrid (Y2H) screening provides high-quality binary interaction information. Because a large fraction of the yeast binary interactome remains to be mapped, we developed an empirically controlled mapping framework to produce a "second-generation" high-quality, high-throughput Y2H data set covering approximately 20% of all yeast binary interactions. Both Y2H and affinity purification followed by mass spectrometry (AP/MS) data are of equally high quality but of a fundamentally different and complementary nature, resulting in networks with different topological and biological properties. Compared to co-complex interactome models, this binary map is enriched for transient signaling interactions and intercomplex connections with a highly significant clustering between essential proteins. Rather than correlating with essentiality, protein connectivity correlates with genetic pleiotropy.
An experimentally derived confidence score for binary protein-protein interactions. Braun et al. (2008). Nature Methods. doi:10.1038/nmeth.1281
Abstract
Information on protein-protein interactions is of central importance for many areas of biomedical research. At present no method exists to systematically and experimentally assess the quality of individual interactions reported in interaction mapping experiments. To provide a standardized confidence-scoring method that can be applied to tens of thousands of protein interactions, we have developed an interaction tool kit consisting of four complementary, high-throughput protein interaction assays. We benchmarked these assays against positive and random reference sets consisting of well documented pairs of interacting human proteins and randomly chosen protein pairs, respectively. A logistic regression model was trained using the data from these reference sets to combine the assay outputs and calculate the probability that any newly identified interaction pair is a true biophysical interaction once it has been tested in the tool kit. This general approach will allow a systematic and empirical assignment of confidence scores to all individual protein-protein interactions in interactome networks.
An empirical framework for binary interactome mapping. Venkatesan et al. (2008). Nature Methods. doi:10.1038/nmeth.1280
Abstract
Several attempts have been made to systematically map protein-protein interaction, or 'interactome', networks. However, it remains difficult to assess the quality and coverage of existing data sets. Here we describe a framework that uses an empirically-based approach to rigorously dissect quality parameters of currently available human interactome maps. Our results indicate that high-throughput yeast two-hybrid (HT-Y2H) interactions for human proteins are more precise than literature-curated interactions supported by a single publication, suggesting that HT-Y2H is suitable to map a significant portion of the human interactome. We estimate that the human interactome contains approximately 130,000 binary interactions, most of which remain to be mapped. Similar to estimates of DNA sequence data quality and genome size early in the Human Genome Project, estimates of protein interaction data quality and interactome size are crucial to establish the magnitude of the task of comprehensive human interactome mapping and to elucidate a path toward this goal.
Tarassov et al. (2008) — protein-fragment complementation in vivo.
An in Vivo Map of the Yeast Protein Interactome. Tarassov et al. (2008). Science. doi:10.1126/science.1153878
Abstract
Protein interactions regulate the systems-level behavior of cells; thus, deciphering the structure and dynamics of protein interaction networks in their cellular context is a central goal in biology. We have performed a genome-wide in vivo screen for protein-protein interactions in Saccharomyces cerevisiae by means of a protein-fragment complementation assay (PCA). We identified 2770 interactions among 1124 endogenously expressed proteins. Comparison with previous studies confirmed known interactions, but most were not known, revealing a previously unexplored subspace of the yeast protein interactome. The PCA detected structural and topological relationships between proteins, providing an 8-nanometer-resolution map of dynamically interacting complexes in vivo and extended networks that provide insights into fundamental cellular processes, including cell polarization and autophagy, pathways that are evolutionarily conserved and central to both development and human health.
Rigaut et al. (1999) — tandem affinity purification.
A generic protein purification method for protein complex characterization and proteome exploration. Rigaut et al. (1999). Nature Biotechnology. doi:10.1038/13732
Abstract
We have developed a generic procedure to purify proteins expressed at their natural level under native conditions using a novel tandem affinity purification (TAP) tag. The TAP tag allows the rapid purification of complexes from a relatively small number of cells without prior knowledge of the complex composition, activity, or function. Combined with mass spectrometry, the TAP strategy allows for the identification of proteins interacting with a given target protein. The TAP method has been tested in yeast but should be applicable to other cells or organisms.
Gavin et al. (2002), Ho et al. (2002), Krogan et al. (2006) — complexes of the yeast proteome.
Functional organization of the yeast proteome by systematic analysis of protein complexes. Gavin et al. (2002). Nature. doi:10.1038/415141a
Abstract
Most cellular processes are carried out by multiprotein complexes. The identification and analysis of their components provides insight into how the ensemble of expressed proteins (proteome) is organized into functional units. We used tandem-affinity purification (TAP) and mass spectrometry in a large-scale approach to characterize multiprotein complexes in Saccharomyces cerevisiae. We processed 1,739 genes, including 1,143 human orthologues of relevance to human biology, and purified 589 protein assemblies. Bioinformatic analysis of these assemblies defined 232 distinct multiprotein complexes and proposed new cellular roles for 344 proteins, including 231 proteins with no previous functional annotation. Comparison of yeast and human complexes showed that conservation across species extends from single proteins to their molecular environment. Our analysis provides an outline of the eukaryotic proteome as a network of protein complexes at a level of organization beyond binary interactions. This higher-order map contains fundamental biological information and offers the context for a more reasoned and informed approach to drug discovery.
Systematic identification of protein complexes in Saccharomyces cerevisiae by mass spectrometry. Ho et al. (2002). Nature. doi:10.1038/415180a
Abstract
The recent abundance of genome sequence data has brought an urgent need for systematic proteomics to decipher the encoded protein networks that dictate cellular function. To date, generation of large-scale protein-protein interaction maps has relied on the yeast two-hybrid system, which detects binary interactions through activation of reporter gene expression. With the advent of ultrasensitive mass spectrometric protein identification methods, it is feasible to identify directly protein complexes on a proteome-wide scale. Here we report, using the budding yeast Saccharomyces cerevisiae as a test case, an example of this approach, which we term high-throughput mass spectrometric protein complex identification (HMS-PCI). Beginning with 10% of predicted yeast proteins as baits, we detected 3,617 associated proteins covering 25% of the yeast proteome. Numerous protein complexes were identified, including many new interactions in various signalling pathways and in the DNA damage response. Comparison of the HMS-PCI data set with interactions reported in the literature revealed an average threefold higher success rate in detection of known complexes compared with large-scale two-hybrid studies. Given the high degree of connectivity observed in this study, even partial HMS-PCI coverage of complex proteomes, including that of humans, should allow comprehensive identification of cellular networks.
Global landscape of protein complexes in the yeast Saccharomyces cerevisiae. Krogan et al. (2006). Nature. doi:10.1038/nature04670
Abstract
Identification of protein-protein interactions often provides insight into protein function, and many cellular processes are performed by stable protein complexes. We used tandem affinity purification to process 4,562 different tagged proteins of the yeast Saccharomyces cerevisiae. Each preparation was analysed by both matrix-assisted laser desorption/ionization-time of flight mass spectrometry and liquid chromatography tandem mass spectrometry to increase coverage and accuracy. Machine learning was used to integrate the mass spectrometry scores and assign probabilities to the protein-protein interactions. Among 4,087 different proteins identified with high confidence by mass spectrometry from 2,357 successful purifications, our core data set (median precision of 0.69) comprises 7,123 protein-protein interactions involving 2,708 proteins. A Markov clustering algorithm organized these interactions into 547 protein complexes averaging 4.9 subunits per complex, about half of them absent from the MIPS database, as well as 429 additional interactions between pairs of complexes. The data (all of which are available online) will help future studies on individual proteins as well as functional genomics and systems biology.
Huttlin et al. (2015), Huttlin et al. (2021) — BioPlex.
The BioPlex Network: A Systematic Exploration of the Human Interactome. Huttlin et al. (2015). Cell. doi:10.1016/j.cell.2015.06.043
Abstract
Protein interactions form a network whose structure drives cellular function and whose organization informs biological inquiry. Using high-throughput affinity-purification mass spectrometry, we identify interacting partners for 2,594 human proteins in HEK293T cells. The resulting network (BioPlex) contains 23,744 interactions among 7,668 proteins with 86% previously undocumented. BioPlex accurately depicts known complexes, attaining 80%-100% coverage for most CORUM complexes. The network readily subdivides into communities that correspond to complexes or clusters of functionally related proteins. More generally, network architecture reflects cellular localization, biological process, and molecular function, enabling functional characterization of thousands of proteins. Network structure also reveals associations among thousands of protein domains, suggesting a basis for examining structurally related proteins. Finally, BioPlex, in combination with other approaches, can be used to reveal interactions of biological or clinical significance. For example, mutations in the membrane protein VAPB implicated in familial amyotrophic lateral sclerosis perturb a defined community of interactors.
Dual proteome-scale networks reveal cell-specific remodeling of the human interactome. Huttlin et al. (2021). Cell. doi:10.1016/j.cell.2021.04.011
Abstract
Thousands of interactions assemble proteins into modules that impart spatial and functional organization to the cellular proteome. Through affinity-purification mass spectrometry, we have created two proteome-scale, cell-line-specific interaction networks. The first, BioPlex 3.0, results from affinity purification of 10,128 human proteins-half the proteome-in 293T cells and includes 118,162 interactions among 14,586 proteins. The second results from 5,522 immunoprecipitations in HCT116 cells. These networks model the interactome whose structure encodes protein function, localization, and complex membership. Comparison across cell lines validates thousands of interactions and reveals extensive customization. Whereas shared interactions reside in core complexes and involve essential proteins, cell-specific interactions link these complexes, "rewiring" subnetworks within each cell's interactome. Interactions covary among proteins of shared function as the proteome remodels to produce each cell's phenotype. Viewable interactively online through BioPlexExplorer, these networks define principles of proteome organization and enable unknown protein characterization.
Drew et al. (2017), Drew et al. (2021), Fischer et al. (2025) — hu.MAP.
Integration of over 9,000 mass spectrometry experiments builds a global map of human protein complexes. Drew et al. (2017). Molecular Systems Biology. doi:10.15252/msb.20167490
Abstract
Macromolecular protein complexes carry out many of the essential functions of cells, and many genetic diseases arise from disrupting the functions of such complexes. Currently, there is great interest in defining the complete set of human protein complexes, but recent published maps lack comprehensive coverage. Here, through the synthesis of over 9,000 published mass spectrometry experiments, we present hu.MAP, the most comprehensive and accurate human protein complex map to date, containing > 4,600 total complexes, > 7,700 proteins, and > 56,000 unique interactions, including thousands of confident protein interactions not identified by the original publications. hu.MAP accurately recapitulates known complexes withheld from the learning procedure, which was optimized with the aid of a new quantitative metric ( k -cliques) for comparing sets of sets. The vast majority of complexes in our map are significantly enriched with literature annotations, and the map overall shows improved coverage of many disease-associated proteins, as we describe in detail for ciliopathies. Using hu.MAP, we predicted and experimentally validated candidate ciliopathy disease genes in vivo in a model vertebrate, discovering CCDC138, WDR90, and KIAA1328 to be new cilia basal body/centriolar satellite proteins, and identifying ANKRD55 as a novel member of the intraflagellar transport machinery. By offering significant improvements to the accuracy and coverage of human protein complexes, hu.MAP (http://proteincomplexes.org) serves as a valuable resource for better understanding the core cellular functions of human proteins and helping to determine mechanistic foundations of human disease.
hu.MAP 2.0: integration of over 15,000 proteomic experiments builds a global compendium of human multiprotein assemblies. Drew et al. (2021). Molecular Systems Biology. doi:10.15252/msb.202010016
Abstract
A general principle of biology is the self-assembly of proteins into functional complexes. Characterizing their composition is, therefore, required for our understanding of cellular functions. Unfortunately, we lack knowledge of the comprehensive set of identities of protein complexes in human cells. To address this gap, we developed a machine learning framework to identify protein complexes in over 15,000 mass spectrometry experiments which resulted in the identification of nearly 7,000 physical assemblies. We show our resource, hu.MAP 2.0, is more accurate and comprehensive than previous state of the art high-throughput protein complex resources and gives rise to many new hypotheses, including for 274 completely uncharacterized proteins. Further, we identify 253 promiscuous proteins that participate in multiple complexes pointing to possible moonlighting roles. We have made hu.MAP 2.0 easily searchable in a web interface (http://humap2.proteincomplexes.org/), which will be a valuable resource for researchers across a broad range of interests including systems biology, structural biology, and molecular explanations of disease.
hu.MAP3.0: atlas of human protein complexes by integration of >25,000 proteomic experiments. Fischer et al. (2025). Molecular Systems Biology. doi:10.1038/s44320-025-00121-5
Abstract
Macromolecular protein complexes carry out most cellular functions. Unfortunately, we lack the subunit composition for many human protein complexes. To address this gap we integrated >25,000 mass spectrometry experiments using a machine learning approach to identify >15,000 human protein complexes. We show our map of protein complexes is highly accurate and more comprehensive than previous maps, placing nearly 70% of human proteins into their physical contexts. We globally characterize our complexes using mass spectrometry based protein covariation data (ProteomeHD.2) and identify covarying complexes suggesting common functional associations. hu.MAP3.0 generates testable functional hypotheses for 472 uncharacterized proteins which we support using AlphaFold modeling. Additionally, we use AlphaFold modeling to identify 5871 mutually exclusive proteins in hu.MAP3.0 complexes suggesting complexes serve different functional roles depending on their subunit composition. We identify expression as the primary way cells and organisms relieve the conflict of mutually exclusive subunits. Finally, we import our complexes to EMBL-EBI's Complex Portal ( https://www.ebi.ac.uk/complexportal/home ) and provide complexes through our hu.MAP3.0 web interface ( https://humap3.proteincomplexes.org/ ). We expect our resource to be highly impactful to the broader research community.
Costanzo et al. (2016) — the yeast genetic interaction network.
A global genetic interaction network maps a wiring diagram of cellular function. Costanzo et al. (2016). Science. doi:10.1126/science.aaf1420
Abstract
We generated a global genetic interaction network for Saccharomyces cerevisiae, constructing more than 23 million double mutants, identifying about 550,000 negative and about 350,000 positive genetic interactions. This comprehensive network maps genetic interactions for essential gene pairs, highlighting essential genes as densely connected hubs. Genetic interaction profiles enabled assembly of a hierarchical model of cell function, including modules corresponding to protein complexes and pathways, biological processes, and cellular compartments. Negative interactions connected functionally related genes, mapped core bioprocesses, and identified pleiotropic genes, whereas positive interactions often mapped general regulatory connections among gene pairs, rather than shared functionality. The global network illustrates how coherent sets of genetic interactions connect protein complex and pathway modules to map a functional wiring diagram of the cell.
B.11.2 Regulatory networks
Ideker et al. (2000), Ideker et al. (2001) — perturbations and integrated analysis.
Discovery of regulatory interactions through perturbation: inference and experimental design. Ideker et al. (2000). Pacific Symposium on Biocomputing.
Integrated Genomic and Proteomic Analyses of a Systematically Perturbed Metabolic Network. Ideker et al. (2001). Science. doi:10.1126/science.292.5518.929
Abstract
We demonstrate an integrated approach to build, test, and refine a model of a cellular pathway, in which perturbations to critical pathway components are analyzed using DNA microarrays, quantitative proteomics, and databases of known physical interactions. Using this approach, we identify 997 messenger RNAs responding to 20 systematic perturbations of the yeast galactose-utilization pathway, provide evidence that approximately 15 of 289 detected proteins are regulated posttranscriptionally, and identify explicit physical interactions governing the cellular response to each perturbation. We refine the model through further iterations of perturbation and global measurements, suggesting hypotheses about the regulation of galactose utilization and physical interactions between this and a variety of other metabolic pathways.
Gardner et al. (2003), Tegnér et al. (2003) — inferring networks from perturbations.
Inferring Genetic Networks and Identifying Compound Mode of Action via Expression Profiling. Gardner et al. (2003). Science. doi:10.1126/science.1081900
Abstract
The complexity of cellular gene, protein, and metabolite networks can hinder attempts to elucidate their structure and function. To address this problem, we used systematic transcriptional perturbations to construct a first-order model of regulatory interactions in a nine-gene subnetwork of the SOS pathway in Escherichia coli. The model correctly identified the major regulatory genes and the transcriptional targets of mitomycin C activity in the subnetwork. This approach, which is experimentally and computationally scalable, provides a framework for elucidating the functional properties of genetic networks and identifying molecular targets of pharmacological compounds.
Reverse engineering gene networks: Integrating genetic perturbations with dynamical modeling. Tegnér et al. (2003). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.0933416100
Abstract
While the fundamental building blocks of biology are being tabulated by the various genome projects, microarray technology is setting the stage for the task of deducing the connectivity of large-scale gene networks. We show how the perturbation of carefully chosen genes in a microarray experiment can be used in conjunction with a reverse engineering algorithm to reveal the architecture of an underlying gene regulatory network. Our iterative scheme identifies the network topology by analyzing the steady-state changes in gene expression resulting from the systematic perturbation of a particular node in the network. We highlight the validity of our reverse engineering approach through the successful deduction of the topology of a linear in numero gene network and a recently reported model for the segmentation polarity network in Drosophila melanogaster. Our method may prove useful in identifying and validating specific drug targets and in deconvolving the effects of chemical compounds.
Margolin et al. (2006), Huynh-Thu et al. (2010) — ARACNE and GENIE3.
ARACNE: An Algorithm for the Reconstruction of Gene Regulatory Networks in a Mammalian Cellular Context. Margolin et al. (2006). BMC Bioinformatics. doi:10.1186/1471-2105-7-s1-s7
Abstract
Background Elucidating gene regulatory networks is crucial for understanding normal cell physiology and complex pathologic phenotypes. Existing computational methods for the genome-wide "reverse engineering" of such networks have been successful only for lower eukaryotes with simple genomes. Here we present ARACNE, a novel algorithm, using microarray expression profiles, specifically designed to scale up to the complexity of regulatory networks in mammalian cells, yet general enough to address a wider range of network deconvolution problems. This method uses an information theoretic approach to eliminate the majority of indirect interactions inferred by co-expression methods. Results We prove that ARACNE reconstructs the network exactly (asymptotically) if the effect of loops in the network topology is negligible, and we show that the algorithm works well in practice, even in the presence of numerous loops and complex topologies. We assess ARACNE's ability to reconstruct transcriptional regulatory networks using both a realistic synthetic dataset and a microarray dataset from human B cells. On synthetic datasets ARACNE achieves very low error rates and outperforms established methods, such as Relevance Networks and Bayesian Networks. Application to the deconvolution of genetic networks in human B cells demonstrates ARACNE's ability to infer validated transcriptional targets of the cMYC proto-oncogene. We also study the effects of misestimation of mutual information on network reconstruction, and show that algorithms based on mutual information ranking are more resilient to estimation errors. Conclusion ARACNE shows promise in identifying direct transcriptional interactions in mammalian cellular networks, a problem that has challenged existing reverse engineering algorithms. This approach should enhance our ability to use microarray data to elucidate functional mechanisms that underlie cellular processes and to identify molecular targets of pharmacological compounds in mammalian cellular networks.
Inferring Regulatory Networks from Expression Data Using Tree-Based Methods. Huynh-Thu et al. (2010). PLoS ONE. doi:10.1371/journal.pone.0012776
Abstract
One of the pressing open problems of computational systems biology is the elucidation of the topology of genetic regulatory networks (GRNs) using high throughput genomic data, in particular microarray gene expression data. The Dialogue for Reverse Engineering Assessments and Methods (DREAM) challenge aims to evaluate the success of GRN inference algorithms on benchmarks of simulated data. In this article, we present GENIE3, a new algorithm for the inference of GRNs that was best performer in the DREAM4 In Silico Multifactorial challenge. GENIE3 decomposes the prediction of a regulatory network between p genes into p different regression problems. In each of the regression problems, the expression pattern of one of the genes (target gene) is predicted from the expression patterns of all the other genes (input genes), using tree-based ensemble methods Random Forests or Extra-Trees. The importance of an input gene in the prediction of the target gene expression pattern is taken as an indication of a putative regulatory link. Putative regulatory links are then aggregated over all genes to provide a ranking of interactions from which the whole network is reconstructed. In addition to performing well on the DREAM4 In Silico Multifactorial challenge simulated data, we show that GENIE3 compares favorably with existing algorithms to decipher the genetic regulatory network of Escherichia coli. It doesn't make any assumption about the nature of gene regulation, can deal with combinatorial and non-linear interactions, produces directed GRNs, and is fast and scalable. In conclusion, we propose a new algorithm for GRN inference that performs well on both synthetic and real gene expression data. The algorithm, based on feature selection with tree-based ensemble methods, is simple and generic, making it adaptable to other types of genomic data and interactions.
Marbach et al. (2012) — DREAM5: the wisdom of crowds.
Wisdom of crowds for robust gene network inference. Marbach et al. (2012). Nature Methods. doi:10.1038/nmeth.2016
Abstract
Reconstructing gene regulatory networks from high-throughput data is a long-standing challenge. Through the Dialogue on Reverse Engineering Assessment and Methods (DREAM) project, we performed a comprehensive blind assessment of over 30 network inference methods on Escherichia coli, Staphylococcus aureus, Saccharomyces cerevisiae and in silico microarray data. We characterize the performance, data requirements and inherent biases of different inference approaches, and we provide guidelines for algorithm application and development. We observed that no single inference method performs optimally across all data sets. In contrast, integration of predictions from multiple inference methods shows robust and high performance across diverse data sets. We thereby constructed high-confidence networks for E. coli and S. aureus, each comprising ~1,700 transcriptional interactions at a precision of ~50%. We experimentally tested 53 previously unobserved regulatory interactions in E. coli, of which 23 (43%) were supported. Our results establish community-based methods as a powerful and robust tool for the inference of transcriptional gene regulatory networks.
Pratapa et al. (2020) — benchmarking network inference on single-cell data.
Benchmarking algorithms for gene regulatory network inference from single-cell transcriptomic data. Pratapa et al. (2020). Nature Methods. doi:10.1038/s41592-019-0690-6
Abstract
We present a systematic evaluation of state-of-the-art algorithms for inferring gene regulatory networks from single-cell transcriptional data. As the ground truth for assessing accuracy, we use synthetic networks with predictable trajectories, literature-curated Boolean models and diverse transcriptional regulatory networks. We develop a strategy to simulate single-cell transcriptional data from synthetic and Boolean networks that avoids pitfalls of previously used methods. Furthermore, we collect networks from multiple experimental single-cell RNA-seq datasets. We develop an evaluation framework called BEELINE. We find that the area under the precision-recall curve and early precision of the algorithms are moderate. The methods are better in recovering interactions in synthetic networks than Boolean models. The algorithms with the best early precision values for Boolean models also perform well on experimental datasets. Techniques that do not require pseudotime-ordered cells are generally more accurate. Based on these results, we present recommendations to end users. BEELINE will aid the development of gene regulatory network inference algorithms.
Dixit et al. (2016), Replogle et al. (2022) — Perturb-seq.
Perturb-Seq: Dissecting Molecular Circuits with Scalable Single-Cell RNA Profiling of Pooled Genetic Screens. Dixit et al. (2016). Cell. doi:10.1016/j.cell.2016.11.038
Abstract
Genetic screens help infer gene function in mammalian cells, but it has remained difficult to assay complex phenotypes-such as transcriptional profiles-at scale. Here, we develop Perturb-seq, combining single-cell RNA sequencing (RNA-seq) and clustered regularly interspaced short palindromic repeats (CRISPR)-based perturbations to perform many such assays in a pool. We demonstrate Perturb-seq by analyzing 200,000 cells in immune cells and cell lines, focusing on transcription factors regulating the response of dendritic cells to lipopolysaccharide (LPS). Perturb-seq accurately identifies individual gene targets, gene signatures, and cell states affected by individual perturbations and their genetic interactions. We posit new functions for regulators of differentiation, the anti-viral response, and mitochondrial function during immune activation. By decomposing many high content measurements into the effects of perturbations, their interactions, and diverse cell metadata, Perturb-seq dramatically increases the scope of pooled genomic assays.
Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq. Replogle et al. (2022). Cell. doi:10.1016/j.cell.2022.05.013
Abstract
A central goal of genetics is to define the relationships between genotypes and phenotypes. High-content phenotypic screens such as Perturb-seq (CRISPR-based screens with single-cell RNA-sequencing readouts) enable massively parallel functional genomic mapping but, to date, have been used at limited scales. Here, we perform genome-scale Perturb-seq targeting all expressed genes with CRISPR interference (CRISPRi) across >2.5 million human cells. We use transcriptional phenotypes to predict the function of poorly characterized genes, uncovering new regulators of ribosome biogenesis (including CCDC86, ZNF236, and SPATA5L1), transcription (C7orf26), and mitochondrial respiration (TMEM242). In addition to assigning gene function, single-cell transcriptional phenotypes allow for in-depth dissection of complex cellular phenomena-from RNA processing to differentiation. We leverage this ability to systematically identify genetic drivers and consequences of aneuploidy and to discover an unanticipated layer of stress-specific regulation of the mitochondrial genome. Our information-rich genotype-phenotype map reveals a multidimensional portrait of gene and cellular function.
B.11.3 Network concepts
Barabási and Albert (1999), Jeong et al. (2000) — scale-free networks.
Emergence of Scaling in Random Networks. Barabási & Albert (1999). Science. doi:10.1126/science.286.5439.509
Abstract
Systems as diverse as genetic networks or the World Wide Web are best described as networks with complex topology. A common property of many large networks is that the vertex connectivities follow a scale-free power-law distribution. This feature was found to be a consequence of two generic mechanisms: (i) networks expand continuously by the addition of new vertices, and (ii) new vertices attach preferentially to sites that are already well connected. A model based on these two ingredients reproduces the observed stationary scale-free distributions, which indicates that the development of large networks is governed by robust self-organizing phenomena that go beyond the particulars of the individual systems.
The large-scale organization of metabolic networks. Jeong et al. (2000). Nature. doi:10.1038/35036627
Abstract
In a cell or microorganism, the processes that generate mass, energy, information transfer and cell-fate specification are seamlessly integrated through a complex network of cellular constituents and reactions. However, despite the key role of these networks in sustaining cellular functions, their large-scale structure is essentially unknown. Here we present a systematic comparative mathematical analysis of the metabolic networks of 43 organisms representing all three domains of life. We show that, despite significant variation in their individual constituents and pathways, these metabolic networks have the same topological scaling properties and show striking similarities to the inherent organization of complex non-biological systems. This may indicate that metabolic organization is not only identical for all living organisms, but also complies with the design principles of robust and error-tolerant scale-free networks, and may represent a common blueprint for the large-scale organization of interactions among all cellular constituents.
Watts and Strogatz (1998) — small-world networks.
Collective dynamics of ‘small-world’ networks. Watts & Strogatz (1998). Nature. doi:10.1038/30918
Abstract
Networks of coupled dynamical systems have been used to model biological oscillators, Josephson junction arrays, excitable media, neural networks, spatial games, genetic control networks and many other self-organizing systems. Ordinarily, the connection topology is assumed to be either completely regular or completely random. But many biological, technological and social networks lie somewhere between these two extremes. Here we explore simple models of networks that can be tuned through this middle ground: regular networks 'rewired' to introduce increasing amounts of disorder. We find that these systems can be highly clustered, like regular lattices, yet have small characteristic path lengths, like random graphs. We call them 'small-world' networks, by analogy with the small-world phenomenon (popularly known as six degrees of separation. The neural network of the worm Caenorhabditis elegans, the power grid of the western United States, and the collaboration graph of film actors are shown to be small-world networks. Models of dynamical systems with small-world coupling display enhanced signal-propagation speed, computational power, and synchronizability. In particular, infectious diseases spread more easily in small-world networks than in regular lattices.
Jeong et al. (2001), Han et al. (2004), Ekman et al. (2006) — hubs.
Lethality and centrality in protein networks. Jeong et al. (2001). Nature. doi:10.1038/35075138
Evidence for dynamically organized modularity in the yeast protein–protein interaction network. Han et al. (2004). Nature. doi:10.1038/nature02555
Abstract
In apparently scale-free protein-protein interaction networks, or 'interactome' networks, most proteins interact with few partners, whereas a small but significant proportion of proteins, the 'hubs', interact with many partners. Both biological and non-biological scale-free networks are particularly resistant to random node removal but are extremely sensitive to the targeted removal of hubs. A link between the potential scale-free topology of interactome networks and genetic robustness seems to exist, because knockouts of yeast genes encoding hubs are approximately threefold more likely to confer lethality than those of non-hubs. Here we investigate how hubs might contribute to robustness and other cellular properties for protein-protein interactions dynamically regulated both in time and in space. We uncovered two types of hub: 'party' hubs, which interact with most of their partners simultaneously, and 'date' hubs, which bind their different partners at different times or locations. Both in silico studies of network connectivity and genetic interactions described in vivo support a model of organized modularity in which date hubs organize the proteome, connecting biological processes--or modules--to each other, whereas party hubs function inside modules.
What properties characterize the hub proteins of the protein-protein interaction network of Saccharomyces cerevisiae?. Ekman et al. (2006). Genome Biology. doi:10.1186/gb-2006-7-6-r45
Abstract
Background Most proteins interact with only a few other proteins while a small number of proteins (hubs) have many interaction partners. Hub proteins and non-hub proteins differ in several respects; however, understanding is not complete about what properties characterize the hubs and set them apart from proteins of low connectivity. Therefore, we have investigated what differentiates hubs from non-hubs and static hubs (party hubs) from dynamic hubs (date hubs) in the protein-protein interaction network of Saccharomyces cerevisiae. Results The many interactions of hub proteins can only partly be explained by bindings to similar proteins or domains. It is evident that domain repeats, which are associated with binding, are enriched in hubs. Moreover, there is an over representation of multi-domain proteins and long proteins among the hubs. In addition, there are clear differences between party hubs and date hubs. Fewer of the party hubs contain long disordered regions compared to date hubs, indicating that these regions are important for flexible binding but less so for static interactions. Furthermore, party hubs interact to a large extent with each other, supporting the idea of party hubs as the cores of highly clustered functional modules. In addition, hub proteins, and in particular party hubs, are more often ancient. Finally, the more recent paralogs of party hubs are underrepresented. Conclusion Our results indicate that multiple and repeated domains are enriched in hub proteins and, further, that long disordered regions, which are common in date hubs, are particularly important for flexible binding.
Maslov and Sneppen (2002) — degree correlations in protein networks.
Specificity and Stability in Topology of Protein Networks. Maslov & Sneppen (2002). Science. doi:10.1126/science.1065103
Abstract
Molecular networks guide the biochemistry of a living cell on multiple levels: Its metabolic and signaling pathways are shaped by the network of interacting proteins, whose production, in turn, is controlled by the genetic regulatory network. To address topological properties of these two networks, we quantified correlations between connectivities of interacting nodes and compared them to a null model of a network, in which all links were randomly rewired. We found that for both interaction and regulatory networks, links between highly connected proteins are systematically suppressed, whereas those between a highly connected and low-connected pairs of proteins are favored. This effect decreases the likelihood of cross talk between different functional modules of the cell and increases the overall robustness of a network by localizing effects of deleterious perturbations.
Milo et al. (2002), Ravasz et al. (2002) — motifs and hierarchical modularity.
Network Motifs: Simple Building Blocks of Complex Networks. Milo et al. (2002). Science. doi:10.1126/science.298.5594.824
Abstract
Complex networks are studied across many fields of science. To uncover their structural design principles, we defined "network motifs," patterns of interconnections occurring in complex networks at numbers that are significantly higher than those in randomized networks. We found such motifs in networks from biochemistry, neurobiology, ecology, and engineering. The motifs shared by ecological food webs were distinct from the motifs shared by the genetic networks of Escherichia coli and Saccharomyces cerevisiae or from those found in the World Wide Web. Similar motifs were found in networks that perform information processing, even though they describe elements as different as biomolecules within a cell and synaptic connections between neurons in Caenorhabditis elegans. Motifs may thus define universal classes of networks. This approach may uncover the basic building blocks of most networks.
Hierarchical Organization of Modularity in Metabolic Networks. Ravasz et al. (2002). Science. doi:10.1126/science.1073374
Abstract
Spatially or chemically isolated functional modules composed of several cellular components and carrying discrete functions are considered fundamental building blocks of cellular organization, but their presence in highly integrated biochemical networks lacks quantitative support. Here, we show that the metabolic networks of 43 distinct organisms are organized into many small, highly connected topologic modules that combine in a hierarchical manner into larger, less cohesive units, with their number and degree of clustering following a power law. Within Escherichia coli, the uncovered hierarchical modularity closely overlaps with known metabolic functions. The identified network architecture may be generic to system-level cellular organization.
Girvan and Newman (2002), Enright (2002) — community detection and MCL.
Community structure in social and biological networks. Girvan & Newman (2002). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.122653799
Abstract
A number of recent studies have focused on the statistical properties of networked systems such as social networks and the Worldwide Web. Researchers have concentrated particularly on a few properties that seem to be common to many networks: the small-world property, power-law degree distributions, and network transitivity. In this article, we highlight another property that is found in many networks, the property of community structure, in which network nodes are joined together in tightly knit groups, between which there are only looser connections. We propose a method for detecting such communities, built around the idea of using centrality indices to find community boundaries. We test our method on computer-generated and real-world graphs whose community structure is already known and find that the method detects this known structure with high sensitivity and reliability. We also apply the method to two networks whose community structure is not well known--a collaboration network and a food web--and find that it detects significant and informative community divisions in both cases.
An efficient algorithm for large-scale detection of protein families. Enright (2002). Nucleic Acids Research. doi:10.1093/nar/30.7.1575
Abstract
Detection of protein families in large databases is one of the principal research objectives in structural and functional genomics. Protein family classification can significantly contribute to the delineation of functional diversity of homologous proteins, the prediction of function based on domain architecture or the presence of sequence motifs as well as comparative genomics, providing valuable evolutionary insights. We present a novel approach called TRIBE-MCL for rapid and accurate clustering of protein sequences into families. The method relies on the Markov cluster (MCL) algorithm for the assignment of proteins into families based on precomputed sequence similarity information. This novel approach does not suffer from the problems that normally hinder other protein sequence clustering algorithms, such as the presence of multi-domain proteins, promiscuous domains and fragmented proteins. The method has been rigorously tested and validated on a number of very large databases, including SwissProt, InterPro, SCOP and the draft human genome. Our results indicate that the method is ideally suited to the rapid and accurate detection of protein families on a large scale. The method has been used to detect and categorise protein families within the draft human genome and the resulting families have been used to annotate a large proportion of human proteins.
Hakes et al. (2008), Lima-Mendez and Helden (2009), Broido and Clauset (2019) — critical views of network biology.
Protein-protein interaction networks and biology—what’s the connection?. Hakes et al. (2008). Nature Biotechnology. doi:10.1038/nbt0108-69
The powerful law of the power law and other myths in network biology. Lima-Mendez & van Helden (2009). Molecular BioSystems. doi:10.1039/b908681a
Abstract
For almost 10 years, topological analysis of different large-scale biological networks (metabolic reactions, protein interactions, transcriptional regulation) has been highlighting some recurrent properties: power law distribution of degree, scale-freeness, small world, which have been proposed to confer functional advantages such as robustness to environmental changes and tolerance to random mutations. Stochastic generative models inspired different scenarios to explain the growth of interaction networks during evolution. The power law and the associated properties appeared so ubiquitous in complex networks that they were qualified as "universal laws". However, these properties are no longer observed when the data are subjected to statistical tests: in most cases, the data do not fit the expected theoretical models, and the cases of good fitting merely result from sampling artefacts or improper data representation. The field of network biology seems to be founded on a series of myths, i.e. widely believed but false ideas. The weaknesses of these foundations should however not be considered as a failure for the entire domain. Network analysis provides a powerful frame for understanding the function and evolution of biological processes, provided it is brought to an appropriate level of description, by focussing on smaller functional modules and establishing the link between their topological properties and their dynamical behaviour.
Scale-free networks are rare. Broido & Clauset (2019). Nature Communications. doi:10.1038/s41467-019-08746-5
Abstract
Real-world networks are often claimed to be scale free, meaning that the fraction of nodes with degree k follows a power law k -α , a pattern with broad implications for the structure and dynamics of complex systems. However, the universality of scale-free networks remains controversial. Here, we organize different definitions of scale-free networks and construct a severe test of their empirical prevalence using state-of-the-art statistical tools applied to nearly 1000 social, biological, technological, transportation, and information networks. Across these networks, we find robust evidence that strongly scale-free structure is empirically rare, while for most networks, log-normal distributions fit the data as well or better than power laws. Furthermore, social networks are at best weakly scale free, while a handful of technological and biological networks appear strongly scale free. These findings highlight the structural diversity of real-world networks and the need for new theoretical explanations of these non-scale-free patterns.
Yanai and DeLisi (2002), Rhodes et al. (2005) — networks from comparative genomics and probabilistic integration.
The society of genes: networks of functional links between genes from comparative genomics. Yanai & DeLisi (2002). Genome Biology. doi:10.1186/gb-2002-3-11-research0064
Abstract
Background Comparative genomics provides at least three methods beyond traditional sequence similarity for identifying functional links between genes: the examination of common phylogenetic distributions, the analysis of conserved proximity along the chromosomes of multiple genomes, and observations of fusions of genes into a multidomain gene in another organism. We have previously generated the links according to each of these methods individually for 43 known microbial genomes. Here we combine these results to construct networks of functional associations. Results We show that the functional networks obtained by applying these methods have different topologies and that the information they provide is largely additive. In particular, the combined networks of functional links contain an average of 57% of an organism's complete genetic complement, uncover substantial portions of known pathways, and suggest the function of previously unannotated genes. In addition, the combined networks are qualitatively different from the networks obtained using individual methods. They have a dominant cluster that contains approximately 80%-90% of the genes, independent of genome size, and the dominant clusters show the small world behavior expected of a biological system, with global connectivity that is nearly random, and local properties that are highly ordered. Conclusions When the information on functional linkage provided by three emerging computational methods is combined, the integrated network uncovers large numbers of conserved pathways and identifies clusters of functionally related genes. It therefore shows considerable utility and promise as a tool for understanding genomic structure, and for guiding high throughput experimental investigations.
Probabilistic model of the human protein-protein interaction network. Rhodes et al. (2005). Nature Biotechnology. doi:10.1038/nbt1103
Abstract
A catalog of all human protein-protein interactions would provide scientists with a framework to study protein deregulation in complex diseases such as cancer. Here we demonstrate that a probabilistic analysis integrating model organism interactome data, protein domain data, genome-wide gene expression data and functional annotation data predicts nearly 40,000 protein-protein interactions in humans-a result comparable to those obtained with experimental and computational approaches in model organisms. We validated the accuracy of the predictive model on an independent test set of known interactions and also experimentally confirmed two predicted interactions relevant to human cancer, implicating uncharacterized proteins into definitive pathways. We also applied the human interactome network to cancer genomics data and identified several interaction subnetworks activated in cancer. This integrative analysis provides a comprehensive framework for exploring the human protein interaction network.
B.11.4 Functional association networks
★ Guala et al. (2019) — genome-wide functional association networks.
Genome-wide functional association networks: background, data & state-of-the-art resources. Guala et al. (2019). Briefings in Bioinformatics. doi:10.1093/bib/bbz064
Abstract
The vast amount of experimental data from recent advances in the field of high-throughput biology begs for integration into more complex data structures such as genome-wide functional association networks. Such networks have been used for elucidation of the interplay of intra-cellular molecules to make advances ranging from the basic science understanding of evolutionary processes to the more translational field of precision medicine. The allure of the field has resulted in rapid growth of the number of available network resources, each with unique attributes exploitable to answer different biological questions. Unfortunately, the high volume of network resources makes it impossible for the intended user to select an appropriate tool for their particular research question. The aim of this paper is to provide an overview of the underlying data and representative network resources as well as to mention methods of integration, allowing a customized approach to resource selection. Additionally, this report will provide a primer for researchers venturing into the field of network integration.
Jansen et al. (2003), Lee et al. (2004) — Bayesian integration of evidence.
A Bayesian Networks Approach for Predicting Protein-Protein Interactions from Genomic Data. Jansen et al. (2003). Science. doi:10.1126/science.1087361
Abstract
We have developed an approach using Bayesian networks to predict protein-protein interactions genome-wide in yeast. Our method naturally weights and combines into reliable predictions genomic features only weakly associated with interaction (e.g., messenger RNAcoexpression, coessentiality, and colocalization). In addition to de novo predictions, it can integrate often noisy, experimental interaction data sets. We observe that at given levels of sensitivity, our predictions are more accurate than the existing high-throughput experimental data sets. We validate our predictions with TAP (tandem affinity purification) tagging experiments. Our analysis, which gives a comprehensive view of yeast interactions, is available at genecensus.org/intint.
A Probabilistic Functional Network of Yeast Genes. Lee et al. (2004). Science. doi:10.1126/science.1099511
Abstract
A conceptual framework for integrating diverse functional genomics data was developed by reinterpreting experiments to provide numerical likelihoods that genes are functionally linked. This allows direct comparison and integration of different classes of data. The resulting probabilistic gene network estimates the functional coupling between genes. Within this framework, we reconstructed an extensive, high-quality functional gene network for Saccharomyces cerevisiae, consisting of 4681 (approximately 81%) of the known yeast genes linked by approximately 34,000 probabilistic linkages comparable in accuracy to small-scale interaction assays. The integrated linkages distinguish true from false-positive interactions in earlier data sets; new interactions emerge from genes' network contexts, as shown for genes in chromatin modification and ribosome biogenesis.
Alexeyenko and Sonnhammer (2009), Buzzao et al. (2024) — FunCoup.
Global networks of functional coupling in eukaryotes from comprehensive data integration. Alexeyenko & Sonnhammer (2009). Genome Research. doi:10.1101/gr.087528.108
Abstract
No single experimental method can discover all connections in the interactome. A computational approach can help by integrating data from multiple, often unrelated, proteomics and genomics pipelines. Reconstructing global networks of functional coupling (FC) faces the challenges of scale and heterogeneity--how to efficiently integrate huge amounts of diverse data from multiple organisms, yet ensuring high accuracy. We developed FunCoup, an optimized Bayesian framework, to resolve these issues. Because interactomes comprise functional coupling of many types, FunCoup annotates network edges with confidence scores in support of different kinds of interactions: physical interaction, protein complex member, metabolic, or signaling link. This capability boosted overall accuracy. On the whole, the constructed framework was comprehensively tested to optimize the overall confidence and ensure seamless, automated incorporation of new data sets of heterogeneous types. Using over 50 data sets in seven organisms and extensively transferring information between orthologs, FunCoup predicted global networks in eight eukaryotes. For the Ciona intestinalis network, only orthologous information was used, and it recovered a significant number of experimental facts. FunCoup predictions were validated on independent cancer mutation data. We show how FunCoup can be used for discovering candidate members of the Parkinson and Alzheimer pathways. Cross-species pathway conservation analysis provided further support to these observations.
FunCoup 6: advancing functional association networks across species with directed links and improved user experience. Buzzao et al. (2024). Nucleic Acids Research. doi:10.1093/nar/gkae1021
Abstract
FunCoup 6 (https://funcoup.org) represents a significant advancement in global functional association networks, aiming to provide researchers with a comprehensive view of the functional coupling interactome. This update introduces novel methodologies and integrated tools for improved network inference and analysis. Major new developments in FunCoup 6 include vastly expanding the coverage of gene regulatory links, a new framework for bin-free Bayesian training and a new website. FunCoup 6 integrates a new tool for disease and drug target module identification using the TOPAS algorithm. To expand the utility of the resource for biomedical research, it incorporates pathway enrichment analysis using the ANUBIX and EASE algorithms. The unique comparative interactomics analysis in FunCoup provides insights of network conservation, now allowing users to align orthologs only or query each species network independently. Bin-free training was applied to 23 primary species, and in addition, networks were generated for all remaining 618 species in InParanoiDB 9. Accompanying these advancements, FunCoup 6 features a new redesigned website, together with updated API functionalities, and represents a pivotal step forward in functional genomics research, offering unique capabilities for exploring the complex landscape of protein interactions.
Szklarczyk et al. (2022) — STRING.
The STRING database in 2023: protein–protein association networks and functional enrichment analyses for any sequenced genome of interest. Szklarczyk et al. (2022). Nucleic Acids Research. doi:10.1093/nar/gkac1000
Abstract
Much of the complexity within cells arises from functional and regulatory interactions among proteins. The core of these interactions is increasingly known, but novel interactions continue to be discovered, and the information remains scattered across different database resources, experimental modalities and levels of mechanistic detail. The STRING database (https://string-db.org/) systematically collects and integrates protein-protein interactions-both physical interactions as well as functional associations. The data originate from a number of sources: automated text mining of the scientific literature, computational interaction predictions from co-expression, conserved genomic context, databases of interaction experiments and known complexes/pathways from curated sources. All of these interactions are critically assessed, scored, and subsequently automatically transferred to less well-studied organisms using hierarchical orthology information. The data can be accessed via the website, but also programmatically and via bulk downloads. The most recent developments in STRING (version 12.0) are: (i) it is now possible to create, browse and analyze a full interaction network for any novel genome of interest, by submitting its complement of encoded proteins, (ii) the co-expression channel now uses variational auto-encoders to predict interactions, and it covers two new sources, single-cell RNA-seq and experimental proteomics data and (iii) the confidence in each experimentally derived interaction is now estimated based on the detection method used, and communicated to the user in the web-interface. Furthermore, STRING continues to enhance its facilities for functional enrichment analysis, which are now fully available also for user-submitted genomes.
B.11.5 Disease networks and systems biology
Goh et al. (2007), Barabási et al. (2010), Menche et al. (2015) — the human disease network and network medicine.
The human disease network. Goh et al. (2007). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.0701361104
Abstract
A network of disorders and disease genes linked by known disorder-gene associations offers a platform to explore in a single graph-theoretic framework all known phenotype and disease gene associations, indicating the common genetic origin of many diseases. Genes associated with similar disorders show both higher likelihood of physical interactions between their products and higher expression profiling similarity for their transcripts, supporting the existence of distinct disease-specific functional modules. We find that essential human genes are likely to encode hub proteins and are expressed widely in most tissues. This suggests that disease genes also would play a central role in the human interactome. In contrast, we find that the vast majority of disease genes are nonessential and show no tendency to encode hub proteins, and their expression pattern indicates that they are localized in the functional periphery of the network. A selection-based model explains the observed difference between essential and disease genes and also suggests that diseases caused by somatic mutations should not be peripheral, a prediction we confirm for cancer genes.
Network medicine: a network-based approach to human disease. Barabási et al. (2010). Nature Reviews Genetics. doi:10.1038/nrg2918
Abstract
Given the functional interdependencies between the molecular components in a human cell, a disease is rarely a consequence of an abnormality in a single gene, but reflects the perturbations of the complex intracellular and intercellular network that links tissue and organ systems. The emerging tools of network medicine offer a platform to explore systematically not only the molecular complexity of a particular disease, leading to the identification of disease modules and pathways, but also the molecular relationships among apparently distinct (patho)phenotypes. Advances in this direction are essential for identifying new disease genes, for uncovering the biological significance of disease-associated mutations identified by genome-wide association studies and full-genome sequencing, and for identifying drug targets and biomarkers for complex diseases.
Uncovering disease-disease relationships through the incomplete interactome. Menche et al. (2015). Science. doi:10.1126/science.1257601
Abstract
According to the disease module hypothesis, the cellular components associated with a disease segregate in the same neighborhood of the human interactome, the map of biologically relevant molecular interactions. Yet, given the incompleteness of the interactome and the limited knowledge of disease-associated genes, it is not obvious if the available data have sufficient coverage to map out modules associated with each disease. Here we derive mathematical conditions for the identifiability of disease modules and show that the network-based location of each disease module determines its pathobiological relationship to other diseases. For example, diseases with overlapping network modules show significant coexpression patterns, symptom similarity, and comorbidity, whereas diseases residing in separated network neighborhoods are phenotypically distinct. These tools represent an interactome-based platform to predict molecular commonalities between phenotypically related diseases, even if they do not share primary disease genes.
Köhler et al. (2008), Vanunu et al. (2010) — random walks and network propagation.
Walking the Interactome for Prioritization of Candidate Disease Genes. Köhler et al. (2008). The American Journal of Human Genetics. doi:10.1016/j.ajhg.2008.02.013
Abstract
The identification of genes associated with hereditary disorders has contributed to improving medical care and to a better understanding of gene functions, interactions, and pathways. However, there are well over 1500 Mendelian disorders whose molecular basis remains unknown. At present, methods such as linkage analysis can identify the chromosomal region in which unknown disease genes are located, but the regions could contain up to hundreds of candidate genes. In this work, we present a method for prioritization of candidate genes by use of a global network distance measure, random walk analysis, for definition of similarities in protein-protein interaction networks. We tested our method on 110 disease-gene families with a total of 783 genes and achieved an area under the ROC curve of up to 98% on simulated linkage intervals of 100 genes surrounding the disease gene, significantly outperforming previous methods based on local distance measures. Our results not only provide an improved tool for positional-cloning projects but also add weight to the assumption that phenotypically similar diseases are associated with disturbances of subnetworks within the larger protein interactome that extend beyond the disease proteins themselves.
Associating Genes and Protein Complexes with Disease via Network Propagation. Vanunu et al. (2010). PLoS Computational Biology. doi:10.1371/journal.pcbi.1000641
Abstract
A fundamental challenge in human health is the identification of disease-causing genes. Recently, several studies have tackled this challenge via a network-based approach, motivated by the observation that genes causing the same or similar diseases tend to lie close to one another in a network of protein-protein or functional interactions. However, most of these approaches use only local network information in the inference process and are restricted to inferring single gene associations. Here, we provide a global, network-based method for prioritizing disease genes and inferring protein complex associations, which we call PRINCE. The method is based on formulating constraints on the prioritization function that relate to its smoothness over the network and usage of prior information. We exploit this function to predict not only genes but also protein complex associations with a disease of interest. We test our method on gene-disease association data, evaluating both the prioritization achieved and the protein complexes inferred. We show that our method outperforms extant approaches in both tasks. Using data on 1,369 diseases from the OMIM knowledgebase, our method is able (in a cross validation setting) to rank the true causal gene first for 34% of the diseases, and infer 139 disease-related complexes that are highly coherent in terms of the function, expression and conservation of their member proteins. Importantly, we apply our method to study three multi-factorial diseases for which some causal genes have been found already: prostate cancer, alzheimer and type 2 diabetes mellitus. PRINCE's predictions for these diseases highly match the known literature, suggesting several novel causal genes and protein complexes for further investigation.
Ideker et al. (2002), Lee et al. (2008) — active subnetworks and phenotype prediction.
Discovering regulatory and signalling circuits in molecular interaction networks. Ideker et al. (2002). Bioinformatics. doi:10.1093/bioinformatics/18.suppl_1.s233
Abstract
Motivation In model organisms such as yeast, large databases of protein-protein and protein-DNA interactions have become an extremely important resource for the study of protein function, evolution, and gene regulatory dynamics. In this paper we demonstrate that by integrating these interactions with widely-available mRNA expression data, it is possible to generate concrete hypotheses for the underlying mechanisms governing the observed changes in gene expression. To perform this integration systematically and at large scale, we introduce an approach for screening a molecular interaction network to identify active subnetworks, i.e., connected regions of the network that show significant changes in expression over particular subsets of conditions. The method we present here combines a rigorous statistical measure for scoring subnetworks with a search algorithm for identifying subnetworks with high score. Results We evaluated our procedure on a small network of 332 genes and 362 interactions and a large network of 4160 genes containing all 7462 protein-protein and protein-DNA interactions in the yeast public databases. In the case of the small network, we identified five significant subnetworks that covered 41 out of 77 (53%) of all significant changes in expression. Both network analyses returned several top-scoring subnetworks with good correspondence to known regulatory mechanisms in the literature. These results demonstrate how large-scale genomic approaches may be used to uncover signalling and regulatory pathways in a systematic, integrative fashion.
A single gene network accurately predicts phenotypic effects of gene perturbation in Caenorhabditis elegans. Lee et al. (2008). Nature Genetics. doi:10.1038/ng.2007.70
Abstract
The fundamental aim of genetics is to understand how an organism's phenotype is determined by its genotype, and implicit in this is predicting how changes in DNA sequence alter phenotypes. A single network covering all the genes of an organism might guide such predictions down to the level of individual cells and tissues. To validate this approach, we computationally generated a network covering most C. elegans genes and tested its predictive capacity. Connectivity within this network predicts essentiality, identifying this relationship as an evolutionarily conserved biological principle. Critically, the network makes tissue-specific predictions-we accurately identify genes for most systematically assayed loss-of-function phenotypes, which span diverse cellular and developmental processes. Using the network, we identify 16 genes whose inactivation suppresses defects in the retinoblastoma tumor suppressor pathway, and we successfully predict that the dystrophin complex modulates EGF signaling. We conclude that an analogous network for human genes might be similarly predictive and thus facilitate identification of disease genes and rational therapeutic targets.
Kitano (2002) — systems biology.
Systems Biology: A Brief Overview. Kitano (2002). Science. doi:10.1126/science.1069492
Abstract
To understand biology at the system level, we must examine the structure and dynamics of cellular and organismal function, rather than the characteristics of isolated parts of a cell or organism. Properties of systems, such as robustness, emerge as central issues, and understanding these properties may have an impact on the future of medicine. However, many breakthroughs in experimental devices, advanced software, and analytical methods are required before the achievements of systems biology can live up to their much-touted potential.
Edwards and Palsson (2000), Ibarra et al. (2002) — flux balance analysis of E. coli.
The Escherichia coli MG1655 in silico metabolic genotype: Its definition, characteristics, and capabilities. Edwards & Palsson (2000). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.97.10.5528
Abstract
The Escherichia coli MG1655 genome has been completely sequenced. The annotated sequence, biochemical information, and other information were used to reconstruct the E. coli metabolic map. The stoichiometric coefficients for each metabolic enzyme in the E. coli metabolic map were assembled to construct a genome-specific stoichiometric matrix. The E. coli stoichiometric matrix was used to define the system's characteristics and the capabilities of E. coli metabolism. The effects of gene deletions in the central metabolic pathways on the ability of the in silico metabolic network to support growth were assessed, and the in silico predictions were compared with experimental observations. It was shown that based on stoichiometric and capacity constraints the in silico analysis was able to qualitatively predict the growth potential of mutant strains in 86% of the cases examined. Herein, it is demonstrated that the synthesis of in silico metabolic genotypes based on genomic, biochemical, and strain-specific information is possible, and that systems analysis methods are available to analyze and interpret the metabolic phenotype.
Escherichia coli K-12 undergoes adaptive evolution to achieve in silico predicted optimal growth. Ibarra et al. (2002). Nature. doi:10.1038/nature01149
Abstract
Annotated genome sequences can be used to reconstruct whole-cell metabolic networks. These metabolic networks can be modelled and analysed (computed) to study complex biological functions. In particular, constraints-based in silico models have been used to calculate optimal growth rates on common carbon substrates, and the results were found to be consistent with experimental data under many but not all conditions. Optimal biological functions are acquired through an evolutionary process. Thus, incorrect predictions of in silico models based on optimal performance criteria may be due to incomplete adaptive evolution under the conditions examined. Escherichia coli K-12 MG1655 grows sub-optimally on glycerol as the sole carbon source. Here we show that when placed under growth selection pressure, the growth rate of E. coli on glycerol reproducibly evolved over 40 days, or about 700 generations, from a sub-optimal value to the optimal growth rate predicted from a whole-cell in silico model. These results open the possibility of using adaptive evolution of entire metabolic networks to realize metabolic states that have been determined a priori based on in silico analysis.
Thiele and Palsson (2010) — building genome-scale metabolic models.
A protocol for generating a high-quality genome-scale metabolic reconstruction. Thiele & Palsson (2010). Nature Protocols. doi:10.1038/nprot.2009.203
Abstract
Network reconstructions are a common denominator in systems biology. Bottom-up metabolic network reconstructions have been developed over the last 10 years. These reconstructions represent structured knowledge bases that abstract pertinent information on the biochemical transformations taking place within specific target organisms. The conversion of a reconstruction into a mathematical format facilitates a myriad of computational biological studies, including evaluation of network content, hypothesis testing and generation, analysis of phenotypic characteristics and metabolic engineering. To date, genome-scale metabolic reconstructions for more than 30 organisms have been published and this number is expected to increase rapidly. However, these reconstructions differ in quality and coverage that may minimize their predictive potential and use as knowledge bases. Here we present a comprehensive protocol describing each step necessary to build a high-quality genome-scale metabolic reconstruction, as well as the common trials and tribulations. Therefore, this protocol provides a helpful manual for all stages of the reconstruction process.
Brunk et al. (2018), Robinson et al. (2020) — human metabolic models.
Recon3D enables a three-dimensional view of gene variation in human metabolism. Brunk et al. (2018). Nature Biotechnology. doi:10.1038/nbt.4072
Abstract
Genome-scale network reconstructions have helped uncover the molecular basis of metabolism. Here we present Recon3D, a computational resource that includes three-dimensional (3D) metabolite and protein structure data and enables integrated analyses of metabolic functions in humans. We use Recon3D to functionally characterize mutations associated with disease, and identify metabolic response signatures that are caused by exposure to certain drugs. Recon3D represents the most comprehensive human metabolic network model to date, accounting for 3,288 open reading frames (representing 17% of functionally annotated human genes), 13,543 metabolic reactions involving 4,140 unique metabolites, and 12,890 protein structures. These data provide a unique resource for investigating molecular mechanisms of human metabolism. Recon3D is available at http://vmh.life.
An atlas of human metabolism. Robinson et al. (2020). Science Signaling. doi:10.1126/scisignal.aaz1482
Abstract
Genome-scale metabolic models (GEMs) are valuable tools to study metabolism and provide a scaffold for the integrative analysis of omics data. Researchers have developed increasingly comprehensive human GEMs, but the disconnect among different model sources and versions impedes further progress. We therefore integrated and extensively curated the most recent human metabolic models to construct a consensus GEM, Human1. We demonstrated the versatility of Human1 through the generation and analysis of cell- and tissue-specific models using transcriptomic, proteomic, and kinetic data. We also present an accompanying web portal, Metabolic Atlas (https://www.metabolicatlas.org/), which facilitates further exploration and visualization of Human1 content. Human1 was created using a version-controlled, open-source model development framework to enable community-driven curation and refinement. This framework allows Human1 to be an evolving shared resource for future studies of human health and disease.
B.11.6 Prediction
Chothia and Janin (1975), Jones and Thornton (1996), Bogan and Thorn (1998) — principles of interfaces and hot spots.
Principles of protein–protein recognition. Chothia & Janin (1975). Nature. doi:10.1038/256705a0
Abstract
The formation of the protein-protein interface by the insulin dimer, the trypsin-PTI complex and the alphabeta oxyhaemoglobin dimer removes 1,130-1,720 A2 of accessible surface from contact with water. The residues forming the interface are close packed: each occupies the same volume as it does in crystals of amino acids. These results indicate that hydrophobicity is the major factor stabilising protein-protein association, while complementarily plays a selective role in deciding which proteins may associate.
Principles of protein-protein interactions.. Jones & Thornton (1996). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.93.1.13
Abstract
This review examines protein complexes in the Brookhaven Protein Databank to gain a better understanding of the principles governing the interactions involved in protein-protein recognition. The factors that influence the formation of protein-protein complexes are explored in four different types of protein-protein complexes--homodimeric proteins, heterodimeric proteins, enzyme-inhibitor complexes, and antibody-protein complexes. The comparison between the complexes highlights differences that reflect their biological roles.
Anatomy of hot spots in protein interfaces. Bogan & Thorn (1998). Journal of Molecular Biology. doi:10.1006/jmbi.1998.1843
Abstract
Binding of one protein to another is involved in nearly all biological functions, yet the principles governing the interaction of proteins are not fully understood. To analyze the contributions of individual amino acid residues in protein-protein binding we have compiled a database of 2325 alanine mutants for which the change in free energy of binding upon mutation to alanine has been measured (available at http://motorhead. ucsf.edu/thorn/hotspot). Our analysis shows that at the level of side-chains there is little correlation between buried surface area and free energy of binding. We find that the free energy of binding is not evenly distributed across interfaces; instead, there are hot spots of binding energy made up of a small subset of residues in the dimer interface. These hot spots are enriched in tryptophan, tyrosine and arginine, and are surrounded by energetically less important residues that most likely serve to occlude bulk solvent from the hot spot. Occlusion of solvent is found to be a necessary condition for highly energetic interactions.
Burger and Nimwegen (2008), Bitbol et al. (2016) — pairing interaction partners with coevolution.
Accurate prediction of protein–protein interactions from sequence alignments using a Bayesian method. Burger & van Nimwegen (2008). Molecular Systems Biology. doi:10.1038/msb4100203
Abstract
Accurate and large-scale prediction of protein-protein interactions directly from amino-acid sequences is one of the great challenges in computational biology. Here we present a new Bayesian network method that predicts interaction partners using only multiple alignments of amino-acid sequences of interacting protein domains, without tunable parameters, and without the need for any training examples. We first apply the method to bacterial two-component systems and comprehensively reconstruct two-component signaling networks across all sequenced bacteria. Comparisons of our predictions with known interactions show that our method infers interaction partners genome-wide with high accuracy. To demonstrate the general applicability of our method we show that it also accurately predicts interaction partners in a recent dataset of polyketide synthases. Analysis of the predicted genome-wide two-component signaling networks shows that cognates (interacting kinase/regulator pairs, which lie adjacent on the genome) and orphans (which lie isolated) form two relatively independent components of the signaling network in each genome. In addition, while most genes are predicted to have only a small number of interaction partners, we find that 10% of orphans form a separate class of 'hub' nodes that distribute and integrate signals to and from up to tens of different interaction partners.
Inferring interaction partners from protein sequences. Bitbol et al. (2016). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1606762113
Abstract
Specific protein-protein interactions are crucial in the cell, both to ensure the formation and stability of multiprotein complexes and to enable signal transduction in various pathways. Functional interactions between proteins result in coevolution between the interaction partners, causing their sequences to be correlated. Here we exploit these correlations to accurately identify, from sequence data alone, which proteins are specific interaction partners. Our general approach, which employs a pairwise maximum entropy model to infer couplings between residues, has been successfully used to predict the 3D structures of proteins from sequences. Thus inspired, we introduce an iterative algorithm to predict specific interaction partners from two protein families whose members are known to interact. We first assess the algorithm's performance on histidine kinases and response regulators from bacterial two-component signaling systems. We obtain a striking 0.93 true positive fraction on our complete dataset without any a priori knowledge of interaction partners, and we uncover the origin of this success. We then apply the algorithm to proteins from ATP-binding cassette (ABC) transporter complexes, and obtain accurate predictions in these systems as well. Finally, we present two metrics that accurately distinguish interacting protein families from noninteracting ones, using only sequence data.
Ovchinnikov et al. (2014), Hopf et al. (2014) — complexes from coevolution.
Robust and accurate prediction of residue–residue interactions across protein interfaces using evolutionary information. Ovchinnikov et al. (2014). eLife. doi:10.7554/elife.02030
Abstract
Do the amino acid sequence identities of residues that make contact across protein interfaces covary during evolution? If so, such covariance could be used to predict contacts across interfaces and assemble models of biological complexes. We find that residue pairs identified using a pseudo-likelihood-based method to covary across protein-protein interfaces in the 50S ribosomal unit and 28 additional bacterial protein complexes with known structure are almost always in contact in the complex, provided that the number of aligned sequences is greater than the average length of the two proteins. We use this method to make subunit contact predictions for an additional 36 protein complexes with unknown structures, and present models based on these predictions for the tripartite ATP-independent periplasmic (TRAP) transporter, the tripartite efflux system, the pyruvate formate lyase-activating enzyme complex, and the methionine ABC transporter.DOI: http://dx.doi.org/10.7554/eLife.02030.001.
Sequence co-evolution gives 3D contacts and structures of protein complexes. Hopf et al. (2014). eLife. doi:10.7554/elife.03430
Abstract
Protein-protein interactions are fundamental to many biological processes. Experimental screens have identified tens of thousands of interactions, and structural biology has provided detailed functional insight for select 3D protein complexes. An alternative rich source of information about protein interactions is the evolutionary sequence record. Building on earlier work, we show that analysis of correlated evolutionary sequence changes across proteins identifies residues that are close in space with sufficient accuracy to determine the three-dimensional structure of the protein complexes. We evaluate prediction performance in blinded tests on 76 complexes of known 3D structure, predict protein-protein contacts in 32 complexes of unknown structure, and demonstrate how evolutionary couplings can be used to distinguish between interacting and non-interacting protein pairs in a large complex. With the current growth of sequences, we expect that the method can be generalized to genome-wide elucidation of protein-protein interaction networks and used for interaction predictions at residue resolution.
Cong et al. (2019), Green et al. (2021) — proteome-wide coevolution screens.
Protein interaction networks revealed by proteome coevolution. Cong et al. (2019). Science. doi:10.1126/science.aaw6718
Abstract
Residue-residue coevolution has been observed across a number of protein-protein interfaces, but the extent of residue coevolution between protein families on the whole-proteome scale has not been systematically studied. We investigate coevolution between 5.4 million pairs of proteins in Escherichia coli and between 3.9 millions pairs in Mycobacterium tuberculosis We find strong coevolution for binary complexes involved in metabolism and weaker coevolution for larger complexes playing roles in genetic information processing. We take advantage of this coevolution, in combination with structure modeling, to predict protein-protein interactions (PPIs) with an accuracy that benchmark studies suggest is considerably higher than that of proteome-wide two-hybrid and mass spectrometry screens. We identify hundreds of previously uncharacterized PPIs in E. coli and M. tuberculosis that both add components to known protein complexes and networks and establish the existence of new ones.
Large-scale discovery of protein interactions at residue resolution using co-evolution calculated from genomic sequences. Green et al. (2021). Nature Communications. doi:10.1038/s41467-021-21636-z
Abstract
Increasing numbers of protein interactions have been identified in high-throughput experiments, but only a small proportion have solved structures. Recently, sequence coevolution-based approaches have led to a breakthrough in predicting monomer protein structures and protein interaction interfaces. Here, we address the challenges of large-scale interaction prediction at residue resolution with a fast alignment concatenation method and a probabilistic score for the interaction of residues. Importantly, this method (EVcomplex2) is able to assess the likelihood of a protein interaction, as we show here applied to large-scale experimental datasets where the pairwise interactions are unknown. We predict 504 interactions de novo in the E. coli membrane proteome, including 243 that are newly discovered. While EVcomplex2 does not require available structures, coevolving residue pairs can be used to produce structural models of protein interactions, as done here for membrane complexes including the Flagellar Hook-Filament Junction and the Tol/Pal complex.
Humphreys et al. (2021), Humphreys et al. (2024) — computed structures of complexes in yeast and pathogens.
Computed structures of core eukaryotic protein complexes. Humphreys et al. (2021). Science. doi:10.1126/science.abm4805
Abstract
Protein-protein interactions play critical roles in biology, but the structures of many eukaryotic protein complexes are unknown, and there are likely many interactions not yet identified. We take advantage of advances in proteome-wide amino acid coevolution analysis and deep-learning–based structure modeling to systematically identify and build accurate models of core eukaryotic protein complexes within the Saccharomyces cerevisiae proteome. We use a combination of RoseTTAFold and AlphaFold to screen through paired multiple sequence alignments for 8.3 million pairs of yeast proteins, identify 1505 likely to interact, and build structure models for 106 previously unidentified assemblies and 806 that have not been structurally characterized. These complexes, which have as many as five subunits, play roles in almost all key processes in eukaryotic cells and provide broad insights into biological function.
Protein interactions in human pathogens revealed through deep learning. Humphreys et al. (2024). Nature Microbiology. doi:10.1038/s41564-024-01791-x
Abstract
Identification of bacterial protein-protein interactions and predicting the structures of these complexes could aid in the understanding of pathogenicity mechanisms and developing treatments for infectious diseases. Here we developed RoseTTAFold2-Lite, a rapid deep learning model that leverages residue-residue coevolution and protein structure prediction to systematically identify and structurally characterize protein-protein interactions at the proteome-wide scale. Using this pipeline, we searched through 78 million pairs of proteins across 19 human bacterial pathogens and identified 1,923 confidently predicted complexes involving essential genes and 256 involving virulence factors. Many of these complexes were not previously known; we experimentally tested 12 such predictions, and half of them were validated. The predicted interactions span core metabolic and virulence pathways ranging from post-transcriptional modification to acid neutralization to outer-membrane machinery and should contribute to our understanding of the biology of these important pathogens and the design of drugs to combat them.
Katchalski-Katzir et al. (1992), Gray et al. (2003), Pierce et al. (2014), Dominguez et al. (2003), Kozakov et al. (2017) — protein–protein docking.
Molecular surface recognition: determination of geometric fit between proteins and their ligands by correlation techniques.. Katchalski-Katzir et al. (1992). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.89.6.2195
Abstract
A geometric recognition algorithm was developed to identify molecular surface complementarity. It is based on a purely geometric approach and takes advantage of techniques applied in the field of pattern recognition. The algorithm involves an automated procedure including (i) a digital representation of the molecules (derived from atomic coordinates) by three-dimensional discrete functions that distinguishes between the surface and the interior; (ii) the calculation, using Fourier transformation, of a correlation function that assesses the degree of molecular surface overlap and penetration upon relative shifts of the molecules in three dimensions; and (iii) a scan of the relative orientations of the molecules in three dimensions. The algorithm provides a list of correlation values indicating the extent of geometric match between the surfaces of the molecules; each of these values is associated with six numbers describing the relative position (translation and rotation) of the molecules. The procedure is thus equivalent to a six-dimensional search but much faster by design, and the computation time is only moderately dependent on molecular size. The procedure was tested and validated by using five known complexes for which the correct relative position of the molecules in the respective adducts was successfully predicted. The molecular pairs were deoxyhemoglobin and methemoglobin, tRNA synthetase-tyrosinyl adenylate, aspartic proteinase-peptide inhibitor, and trypsin-trypsin inhibitor. A more realistic test was performed with the last two pairs by using the structures of uncomplexed aspartic proteinase and trypsin inhibitor, respectively. The results are indicative of the extent of conformational changes in the molecules tolerated by the algorithm.
Protein–Protein Docking with Simultaneous Optimization of Rigid-body Displacement and Side-chain Conformations. Gray et al. (2003). Journal of Molecular Biology. doi:10.1016/s0022-2836(03)00670-3
Abstract
Protein-protein docking algorithms provide a means to elucidate structural details for presently unknown complexes. Here, we present and evaluate a new method to predict protein-protein complexes from the coordinates of the unbound monomer components. The method employs a low-resolution, rigid-body, Monte Carlo search followed by simultaneous optimization of backbone displacement and side-chain conformations using Monte Carlo minimization. Up to 10(5) independent simulations are carried out, and the resulting "decoys" are ranked using an energy function dominated by van der Waals interactions, an implicit solvation model, and an orientation-dependent hydrogen bonding potential. Top-ranking decoys are clustered to select the final predictions. Small-perturbation studies reveal the formation of binding funnels in 42 of 54 cases using coordinates derived from the bound complexes and in 32 of 54 cases using independently determined coordinates of one or both monomers. Experimental binding affinities correlate with the calculated score function and explain the predictive success or failure of many targets. Global searches using one or both unbound components predict at least 25% of the native residue-residue contacts in 28 of the 32 cases where binding funnels exist. The results suggest that the method may soon be useful for generating models of biologically important complexes from the structures of the isolated components, but they also highlight the challenges that must be met to achieve consistent and accurate prediction of protein-protein interactions.
ZDOCK server: interactive docking prediction of protein–protein complexes and symmetric multimers. Pierce et al. (2014). Bioinformatics. doi:10.1093/bioinformatics/btu097
Abstract
Protein-protein interactions are essential to cellular and immune function, and in many cases, because of the absence of an experimentally determined structure of the complex, these interactions must be modeled to obtain an understanding of their molecular basis. We present a user-friendly protein docking server, based on the rigid-body docking programs ZDOCK and M-ZDOCK, to predict structures of protein-protein complexes and symmetric multimers. With a goal of providing an accessible and intuitive interface, we provide options for users to guide the scoring and the selection of output models, in addition to dynamic visualization of input structures and output docking models. This server enables the research community to easily and quickly produce structural models of protein-protein complexes and symmetric multimers for their own analysis. Availability The ZDOCK server is freely available to all academic and non-profit users at: http://zdock.umassmed.edu. No registration is required.
HADDOCK: A Protein−Protein Docking Approach Based on Biochemical or Biophysical Information. Dominguez et al. (2003). Journal of the American Chemical Society. doi:10.1021/ja026939x
Abstract
The structure determination of protein-protein complexes is a rather tedious and lengthy process, by both NMR and X-ray crystallography. Several methods based on docking to study protein complexes have also been well developed over the past few years. Most of these approaches are not driven by experimental data but are based on a combination of energetics and shape complementarity. Here, we present an approach called HADDOCK (High Ambiguity Driven protein-protein Docking) that makes use of biochemical and/or biophysical interaction data such as chemical shift perturbation data resulting from NMR titration experiments or mutagenesis data. This information is introduced as Ambiguous Interaction Restraints (AIRs) to drive the docking process. An AIR is defined as an ambiguous distance between all residues shown to be involved in the interaction. The accuracy of our approach is demonstrated with three molecular complexes. For two of these complexes, for which both the complex and the free protein structures have been solved, NMR titration data were available. Mutagenesis data were used in the last example. In all cases, the best structures generated by HADDOCK, that is, the structures with the lowest intermolecular energies, were the closest to the published structure of the respective complexes (within 2.0 A backbone RMSD).
The ClusPro web server for protein–protein docking. Kozakov et al. (2017). Nature Protocols. doi:10.1038/nprot.2016.169
Abstract
The ClusPro server (https://cluspro.org) is a widely used tool for protein-protein docking. The server provides a simple home page for basic use, requiring only two files in Protein Data Bank (PDB) format. However, ClusPro also offers a number of advanced options to modify the search; these include the removal of unstructured protein regions, application of attraction or repulsion, accounting for pairwise distance restraints, construction of homo-multimers, consideration of small-angle X-ray scattering (SAXS) data, and location of heparin-binding sites. Six different energy functions can be used, depending on the type of protein. Docking with each energy parameter set results in ten models defined by centers of highly populated clusters of low-energy docked structures. This protocol describes the use of the various options, the construction of auxiliary restraints files, the selection of the energy parameters, and the analysis of the results. Although the server is heavily used, runs are generally completed in <4 h.
Janin et al. (2003), Basu and Wallner (2016) — CAPRI and DockQ.
CAPRI: A Critical Assessment of PRedicted Interactions. Janin et al. (2003). Proteins: Structure, Function, and Bioinformatics. doi:10.1002/prot.10381
Abstract
CAPRI is a communitywide experiment to assess the capacity of protein-docking methods to predict protein-protein interactions. Nineteen groups participated in rounds 1 and 2 of CAPRI and submitted blind structure predictions for seven protein-protein complexes based on the known structure of the component proteins. The predictions were compared to the unpublished X-ray structures of the complexes. We describe here the motivations for launching CAPRI, the rules that we applied to select targets and run the experiment, and some conclusions that can already be drawn. The results stress the need for new scoring functions and for methods handling the conformation changes that were observed in some of the target systems. CAPRI has already been a powerful drive for the community of computational biologists who development docking algorithms. We hope that this issue of Proteins will also be of interest to the community of structural biologists, which we call upon to provide new targets for future rounds of CAPRI, and to all molecular biologists who view protein-protein recognition as an essential process.
DockQ: A Quality Measure for Protein-Protein Docking Models. Basu & Wallner (2016). PLOS ONE. doi:10.1371/journal.pone.0161879
Abstract
The state-of-the-art to assess the structural quality of docking models is currently based on three related yet independent quality measures: Fnat, LRMS, and iRMS as proposed and standardized by CAPRI. These quality measures quantify different aspects of the quality of a particular docking model and need to be viewed together to reveal the true quality, e.g. a model with relatively poor LRMS (>10Å) might still qualify as 'acceptable' with a descent Fnat (>0.50) and iRMS (<3.0Å). This is also the reason why the so called CAPRI criteria for assessing the quality of docking models is defined by applying various ad-hoc cutoffs on these measures to classify a docking model into the four classes: Incorrect, Acceptable, Medium, or High quality. This classification has been useful in CAPRI, but since models are grouped in only four bins it is also rather limiting, making it difficult to rank models, correlate with scoring functions or use it as target function in machine learning algorithms. Here, we present DockQ, a continuous protein-protein docking model quality measure derived by combining Fnat, LRMS, and iRMS to a single score in the range [0, 1] that can be used to assess the quality of protein docking models. By using DockQ on CAPRI models it is possible to almost completely reproduce the original CAPRI classification into Incorrect, Acceptable, Medium and High quality. An average PPV of 94% at 90% Recall demonstrating that there is no need to apply predefined ad-hoc cutoffs to classify docking models. Since DockQ recapitulates the CAPRI classification almost perfectly, it can be viewed as a higher resolution version of the CAPRI classification, making it possible to estimate model quality in a more quantitative way using Z-scores or sum of top ranked models, which has been so valuable for the CASP community. The possibility to directly correlate a quality measure to a scoring function has been crucial for the development of scoring functions for protein structure prediction, and DockQ should be useful in a similar development in the protein docking field. DockQ is available at http://github.com/bjornwallner/DockQ/.
Evans et al. (2021) — AlphaFold-Multimer.
Protein complex prediction with AlphaFold-Multimer. Evans et al. (2021). bioRxiv (preprint). doi:10.1101/2021.10.04.463034
Abstract
While the vast majority of well-structured single protein chains can now be predicted to high accuracy due to the recent AlphaFold [1] model, the prediction of multi-chain protein complexes remains a challenge in many cases. In this work, we demonstrate that an AlphaFold model trained specifically for multimeric inputs of known stoichiometry, which we call AlphaFold-Multimer, significantly increases accuracy of predicted multimeric interfaces over input-adapted single-chain AlphaFold while maintaining high intra-chain accuracy. On a benchmark dataset of 17 heterodimer proteins without templates (introduced in [2]) we achieve at least medium accuracy (DockQ [3] ≥ 0.49) on 13 targets and high accuracy (DockQ ≥ 0.8) on 7 targets, compared to 9 targets of at least medium accuracy and 4 of high accuracy for the previous state of the art system (an AlphaFold-based system from [2]). We also predict structures for a large dataset of 4,446 recent protein complexes, from which we score all non-redundant interfaces with low template identity. For heteromeric interfaces we successfully predict the interface (DockQ ≥ 0.23) in 70% of cases, and produce high accuracy predictions (DockQ ≥ 0.8) in 26% of cases, an improvement of +27 and +14 percentage points over the flexible linker modification of AlphaFold [4] respectively. For homomeric inter-faces we successfully predict the interface in 72% of cases, and produce high accuracy predictions in 36% of cases, an improvement of +8 and +7 percentage points respectively.
Pozzati et al. (2021), Bryant, Pozzati, Zhu, et al. (2022) — folding and docking, and large complexes.
Limits and potential of combined folding and docking. Pozzati et al. (2021). Bioinformatics. doi:10.1093/bioinformatics/btab760
Abstract
Motivation In the last decade, de novo protein structure prediction accuracy for individual proteins has improved significantly by utilising deep learning (DL) methods for harvesting the co-evolution information from large multiple sequence alignments (MSAs). The same approach can, in principle, also be used to extract information about evolutionary-based contacts across protein-protein interfaces. However, most earlier studies have not used the latest DL methods for inter-chain contact distance prediction. This article introduces a fold-and-dock method based on predicted residue-residue distances with trRosetta. Results The method can simultaneously predict the tertiary and quaternary structure of a protein pair, even when the structures of the monomers are not known. The straightforward application of this method to a standard dataset for protein-protein docking yielded limited success. However, using alternative methods for generating MSAs allowed us to dock accurately significantly more proteins. We also introduced a novel scoring function, PconsDock, that accurately separates 98% of correctly and incorrectly folded and docked proteins. The average performance of the method is comparable to the use of traditional, template-based or ab initio shape-complementarity-only docking methods. Moreover, the results of conventional and fold-and-dock approaches are complementary, and thus a combined docking pipeline could increase overall docking success significantly. This methodology contributed to the best model for one of the CASP14 oligomeric targets, H1065. Availability and implementation All scripts for predictions and analysis are available from https://github.com/ElofssonLab/bioinfo-toolbox/ and https://gitlab.com/ElofssonLab/benchmark5/. All models joined alignments, and evaluation results are available from the following figshare repository https://doi.org/10.6084/m9.figshare.14654886.v2. Supplementary information Supplementary data are available at Bioinformatics online.
Predicting the structure of large protein complexes using AlphaFold and Monte Carlo tree search. Bryant et al. (2022). Nature Communications. doi:10.1038/s41467-022-33729-4
Abstract
AlphaFold can predict the structure of single- and multiple-chain proteins with very high accuracy. However, the accuracy decreases with the number of chains, and the available GPU memory limits the size of protein complexes which can be predicted. Here we show that one can predict the structure of large complexes starting from predictions of subcomponents. We assemble 91 out of 175 complexes with 10-30 chains from predicted subcomponents using Monte Carlo tree search, with a median TM-score of 0.51. There are 30 highly accurate complexes (TM-score ≥0.8, 33% of complete assemblies). We create a scoring function, mpDockQ, that can distinguish if assemblies are complete and predict their accuracy. We find that complexes containing symmetry are accurately assembled, while asymmetrical complexes remain challenging. The method is freely available and accesible as a Colab notebook https://colab.research.google.com/github/patrickbryant1/MoLPC/blob/master/MoLPC.ipynb .
Zhu et al. (2023) — evaluation of AlphaFold-Multimer.
Evaluation of AlphaFold-Multimer prediction on multi-chain protein complexes. Zhu et al. (2023). Bioinformatics. doi:10.1093/bioinformatics/btad424
Abstract
Motivation Despite near-experimental accuracy on single-chain predictions, there is still scope for improvement among multimeric predictions. Methods like AlphaFold-Multimer and FoldDock can accurately model dimers. However, how well these methods fare on larger complexes is still unclear. Further, evaluation methods of the quality of multimeric complexes are not well established. Results We analysed the performance of AlphaFold-Multimer on a homology-reduced dataset of homo- and heteromeric protein complexes. We highlight the differences between the pairwise and multi-interface evaluation of chains within a multimer. We describe why certain complexes perform well on one metric (e.g. TM-score) but poorly on another (e.g. DockQ). We propose a new score, Predicted DockQ version 2 (pDockQ2), to estimate the quality of each interface in a multimer. Finally, we modelled protein complexes (from CORUM) and identified two highly confident structures that do not have sequence homology to any existing structures. Availability and implementation All scripts, models, and data used to perform the analysis in this study are freely available at https://gitlab.com/ElofssonLab/afm-benchmark.
B.12 Machine learning
Further papers for Chapter 12.
★ the 2021 version of the exam was based on Sapoval et al. (2022) together with Pevsner (2015). In the HTML version, click Abstract to show the abstract.
B.12.1 Reviews
Tarca et al. (2007) — machine learning and its applications to biology.
Machine Learning and Its Applications to Biology. Tarca et al. (2007). PLoS Computational Biology. doi:10.1371/journal.pcbi.0030116
Abstract
The term machine learning refers to a set of topics dealing with the creation and evaluation of algorithms that facilitate pattern recognition, classification, and prediction, based on models derived from existing data. Two facets of mechanization should be acknowledged when considering machine learning in broad terms. Firstly, it is intended that the classification and prediction tasks can be accomplished by a suitably programmed computing machine. That is, the product of machine learning is a classifier that can be feasibly used on available hardware. Secondly, it is intended that the creation of the classifier should itself be highly mechanized, and should not involve too much human input. This second facet is inevitably vague, but the basic objective is that the use of automatic algorithm construction methods can minimize the possibility that human biases could affect the selection and performance of the algorithm. Both the creation of the algorithm and its operation to classify objects or predict events are to be based on concrete, observable data. The history of relations between biology and the field of machine learning is long and complex. An early technique [1] for machine learning called the perceptron constituted an attempt to model actual neuronal behavior, and the field of artificial neural network (ANN) design emerged from this attempt. Early work on the analysis of translation initiation sequences [2] employed the perceptron to define criteria for start sites in Escherichia coli. Further artificial neural network architectures such as the adaptive resonance theory (ART) [3] and neocognitron [4] were inspired from the organization of the visual nervous system. In the intervening years, the flexibility of machine learning techniques has grown along with mathematical frameworks for measuring their reliability, and it is natural to hope that machine learning methods will improve the efficiency of discovery and understanding in the mounting volume and complexity of biological data. This tutorial is structured in four main components. Firstly, a brief section reviews definitions and mathematical prerequisites. Secondly, the field of supervised learning is described. Thirdly, methods of unsupervised learning are reviewed. Finally, a section reviews methods and examples as implemented in the open source data analysis and visualization language R (http://www.r-project.org).
Ching et al. (2018) — opportunities and obstacles for deep learning in biology and medicine.
Opportunities and obstacles for deep learning in biology and medicine. Ching et al. (2018). Journal of The Royal Society Interface. doi:10.1098/rsif.2017.0387
Abstract
Deep learning describes a class of machine learning algorithms that are capable of combining raw inputs into layers of intermediate features. These algorithms have recently shown impressive results across a variety of domains. Biology and medicine are data-rich disciplines, but the data are complex and often ill-understood. Hence, deep learning techniques may be particularly well suited to solve problems of these fields. We examine applications of deep learning to a variety of biomedical problems-patient classification, fundamental biological processes and treatment of patients-and discuss whether deep learning will be able to transform these tasks or if the biomedical sphere poses unique challenges. Following from an extensive literature review, we find that deep learning has yet to revolutionize biomedicine or definitively resolve any of the most pressing challenges in the field, but promising advances have been made on the prior state of the art. Even though improvements over previous baselines have been modest in general, the recent progress indicates that deep learning methods will provide valuable means for speeding up or aiding human investigation. Though progress has been made linking a specific neural network's prediction to input features, understanding how users should interpret these models to make testable hypotheses about the system under study remains an open challenge. Furthermore, the limited amount of labelled data for training presents problems in some domains, as do legal and privacy constraints on work with sensitive health records. Nonetheless, we foresee deep learning enabling changes at both bench and bedside with the potential to transform several areas of biology and medicine.
Eraslan et al. (2019) — deep learning for genomics.
Deep learning: new computational modelling techniques for genomics. Eraslan et al. (2019). Nature Reviews Genetics. doi:10.1038/s41576-019-0122-6
Abstract
As a data-driven science, genomics largely utilizes machine learning to capture dependencies in data and derive novel biological hypotheses. However, the ability to extract new insights from the exponentially increasing volume of genomics data requires more expressive machine learning models. By effectively leveraging large data sets, deep learning has transformed fields such as computer vision and natural language processing. Now, it is becoming the method of choice for many genomics modelling tasks, including predicting the impact of genetic variation on gene regulatory mechanisms such as DNA accessibility and splicing.
B.12.2 Short primers
Eddy (2004a) — what is a hidden Markov model?
What is a hidden Markov model?. Eddy (2004). Nature Biotechnology. doi:10.1038/nbt1004-1315
Noble (2006) — what is a support vector machine?
What is a support vector machine?. Noble (2006). Nature Biotechnology. doi:10.1038/nbt1206-1565
Abstract
Support vector machines (SVMs) are becoming popular in a wide variety of biological applications. But, what exactly are SVMs and how do they work? And what are their most promising applications in the life sciences?
Lever et al. (2016), Lever et al. (2017) — model selection and overfitting; principal component analysis.
Model selection and overfitting. Lever et al. (2016). Nature Methods. doi:10.1038/nmeth.3968
Principal component analysis. Lever et al. (2017). Nature Methods. doi:10.1038/nmeth.4346
B.12.3 Basics and evaluation
Hobohm et al. (1992), Li and Godzik (2006) — non-redundant data sets.
Selection of representative protein data sets. Hobohm et al. (1992). Protein Science. doi:10.1002/pro.5560010313
Abstract
The Protein Data Bank currently contains about 600 data sets of three-dimensional protein coordinates determined by X-ray crystallography or NMR. There is considerable redundancy in the data base, as many protein pairs are identical or very similar in sequence. However, statistical analyses of protein sequence-structure relations require nonredundant data. We have developed two algorithms to extract from the data base representative sets of protein chains with maximum coverage and minimum redundancy. The first algorithm focuses on optimizing a particular property of the selected proteins and works by successive selection of proteins from an ordered list and exclusion of all neighbors of each selected protein. The other algorithm aims at maximizing the size of the selected set and works by successive thinning out of clusters of similar proteins. Both algorithms are generally applicable to other data bases in which criteria of similarity can be defined and relate to problems in graph theory. The largest nonredundant set extracted from the current release of the Protein Data Bank has 155 protein chains. In this set, no two proteins have sequence similarity higher than a certain cutoff (30% identical residues for aligned subsequences longer than 80 residues), yet all structurally unique protein families are represented. Periodically updated lists of representative data sets are available by electronic mail from the file server "netserv@embl-heidelberg.de." The selection may be useful in statistical approaches to protein folding as well as in the analysis and documentation of the known spectrum of three-dimensional protein structures.
Cd-hit: a fast program for clustering and comparing large sets of protein or nucleotide sequences. Li & Godzik (2006). Bioinformatics. doi:10.1093/bioinformatics/btl158
Abstract
Motivation In 2001 and 2002, we published two papers (Bioinformatics, 17, 282-283, Bioinformatics, 18, 77-82) describing an ultrafast protein sequence clustering program called cd-hit. This program can efficiently cluster a huge protein database with millions of sequences. However, the applications of the underlying algorithm are not limited to only protein sequences clustering, here we present several new programs using the same algorithm including cd-hit-2d, cd-hit-est and cd-hit-est-2d. Cd-hit-2d compares two protein datasets and reports similar matches between them; cd-hit-est clusters a DNA/RNA sequence database and cd-hit-est-2d compares two nucleotide datasets. All these programs can handle huge datasets with millions of sequences and can be hundreds of times faster than methods based on the popular sequence comparison and database search tools, such as BLAST.
Saito and Rehmsmeier (2015) — precision–recall vs. ROC curves.
The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets. Saito & Rehmsmeier (2015). PLOS ONE. doi:10.1371/journal.pone.0118432
Abstract
Binary classifiers are routinely evaluated with performance measures such as sensitivity and specificity, and performance is frequently illustrated with Receiver Operating Characteristics (ROC) plots. Alternative measures such as positive predictive value (PPV) and the associated Precision/Recall (PRC) plots are used less frequently. Many bioinformatics studies develop and evaluate classifiers that are to be applied to strongly imbalanced datasets in which the number of negatives outweighs the number of positives significantly. While ROC plots are visually appealing and provide an overview of a classifier's performance across a wide range of specificities, one can ask whether ROC plots could be misleading when applied in imbalanced classification scenarios. We show here that the visual interpretability of ROC plots in the context of imbalanced datasets can be deceptive with respect to conclusions about the reliability of classification performance, owing to an intuitive but wrong interpretation of specificity. PRC plots, on the other hand, can provide the viewer with an accurate prediction of future classification performance due to the fact that they evaluate the fraction of true positives among positive predictions. Our findings have potential implications for the interpretation of a large number of studies that use ROC plots on imbalanced datasets.
Jones (2019) — setting the standards for machine learning in biology.
Setting the standards for machine learning in biology. Jones (2019). Nature Reviews Molecular Cell Biology. doi:10.1038/s41580-019-0176-5
B.12.4 History and classical methods
Rosenblatt (1958), Minsky and Papert (1969) — the perceptron and its limits.
The perceptron: A probabilistic model for information storage and organization in the brain.. Rosenblatt (1958). Psychological Review. doi:10.1037/h0042519
Abstract
If we are eventually to understand the capability of higher organisms for perceptual recognition, generalization, recall, and thinking, we must first have answers to three fundamental questions: 1. How is information about the physical world sensed, or detected, by the biological system? 2. In what form is information stored, or remembered? 3. How does information contained in storage, or in memory, influence recognition and behavior? The first of these questions is in the
Perceptrons: An Introduction to Computational Geometry. Minsky & Papert (1969). MIT Press.
Rumelhart et al. (1986) — backpropagation.
Learning representations by back-propagating errors. Rumelhart et al. (1986). Nature. doi:10.1038/323533a0
Stormo et al. (1982) — the perceptron applied to translation initiation sites.
Use of the ‘Perceptron’ algorithm to distinguish translational initiation sites in E. coli. Stormo et al. (1982). Nucleic Acids Research. doi:10.1093/nar/10.9.2997
Abstract
We have used a "Perceptron" algorithm to find a weighting function which distinguishes E. coli translational initiation sites from all other sites in a library of over 78,000 nucleotides of mRNA sequence. The "Perceptron" examined sequences as linear representations. The "Perceptron" is more successful at finding gene beginnings than our previous searches using "rules" (see previous paper). We note that the weighting function can find translational initiation sites within sequences that were not included in the training set.
Qian and Sejnowski (1988), Holley and Karplus (1989), Rost and Sander (1993) — neural networks for secondary structure.
Predicting the secondary structure of globular proteins using neural network models. Qian & Sejnowski (1988). Journal of Molecular Biology. doi:10.1016/0022-2836(88)90564-5
Abstract
We present a new method for predicting the secondary structure of globular proteins based on non-linear neural network models. Network models learn from existing protein structures how to predict the secondary structure of local sequences of amino acids. The average success rate of our method on a testing set of proteins non-homologous with the corresponding training set was 64.3% on three types of secondary structure (alpha-helix, beta-sheet, and coil), with correlation coefficients of C alpha = 0.41, C beta = 0.31 and Ccoil = 0.41. These quality indices are all higher than those of previous methods. The prediction accuracy for the first 25 residues of the N-terminal sequence was significantly better. We conclude from computational experiments on real and artificial structures that no method based solely on local information in the protein sequence is likely to produce significantly better results for non-homologous proteins. The performance of our method of homologous proteins is much better than for non-homologous proteins, but is not as good as simply assuming that homologous sequences have identical structures.
Protein secondary structure prediction with a neural network.. Holley & Karplus (1989). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.86.1.152
Abstract
A method is presented for protein secondary structure prediction based on a neural network. A training phase was used to teach the network to recognize the relation between secondary structure and amino acid sequences on a sample set of 48 proteins of known structure. On a separate test set of 14 proteins of known structure, the method achieved a maximum overall predictive accuracy of 63% for three states: helix, sheet, and coil. A numerical measure of helix and sheet tendency for each residue was obtained from the calculations. When predictions were filtered to include only the strongest 31% of predictions, the predictive accuracy rose to 79%.
Prediction of Protein Secondary Structure at Better than 70% Accuracy. Rost & Sander (1993). Journal of Molecular Biology. doi:10.1006/jmbi.1993.1413
Abstract
We have trained a two-layered feed-forward neural network on a non-redundant data base of 130 protein chains to predict the secondary structure of water-soluble proteins. A new key aspect is the use of evolutionary information in the form of multiple sequence alignments that are used as input in place of single sequences. The inclusion of protein family information in this form increases the prediction accuracy by six to eight percentage points. A combination of three levels of networks results in an overall three-state accuracy of 70.8% for globular proteins (sustained performance). If four membrane protein chains are included in the evaluation, the overall accuracy drops to 70.2%. The prediction is well balanced between alpha-helix, beta-strand and loop: 65% of the observed strand residues are predicted correctly. The accuracy in predicting the content of three secondary structure types is comparable to that of circular dichroism spectroscopy. The performance accuracy is verified by a sevenfold cross-validation test, and an additional test on 26 recently solved proteins. Of particular practical importance is the definition of a position-specific reliability index. For half of the residues predicted with a high level of reliability the overall accuracy increases to better than 82%. A further strength of the method is the more realistic prediction of segment length. The protein family prediction method is available for testing by academic researchers via an electronic mail server.
Rabiner (1989), Viterbi (1967), Baum et al. (1970) — HMMs and their algorithms.
A tutorial on hidden Markov models and selected applications in speech recognition. Rabiner (1989). Proceedings of the IEEE. doi:10.1109/5.18626
Abstract
This tutorial provides an overview of the basic theory of hidden Markov models (HMMs) as originated by L.E. Baum and T. Petrie (1966) and gives practical details on methods of implementation of the theory along with a description of selected applications of the theory to distinct problems in speech recognition. Results from a number of original sources are combined to provide a single source of acquiring the background required to pursue further this area of research. The author first reviews the theory of discrete Markov chains and shows how the concept of hidden states, where the observation is a probabilistic function of the state, can be used effectively. The theory is illustrated with two simple examples, namely coin-tossing, and the classic balls-in-urns system. Three fundamental problems of HMMs are noted and several practical techniques for solving these problems are given. The various types of HMMs that have been studied, including ergodic as well as left-right models, are described.>
Error bounds for convolutional codes and an asymptotically optimum decoding algorithm. Viterbi (1967). IEEE Transactions on Information Theory. doi:10.1109/tit.1967.1054010
Abstract
The probability of error in decoding an optimal convolutional code transmitted over a memoryless channel is bounded from above and below as a function of the constraint length of the code. For all but pathological channels the bounds are asymptotically (exponentially) tight for rates aboveR_{0}, the computational cutoff rate of sequential decoding. As a function of constraint length the performance of optimal convolutional codes is shown to be superior to that of block codes of the same length, the relative improvement increasing with rate. The upper bound is obtained for a specific probabilistic nonsequential decoding algorithm which is shown to be asymptotically optimum for rates aboveR_{0}and whose performance bears certain similarities to that of sequential decoding algorithms.
A Maximization Technique Occurring in the Statistical Analysis of Probabilistic Functions of Markov Chains. Baum et al. (1970). The Annals of Mathematical Statistics. doi:10.1214/aoms/1177697196
Churchill (1989), Krogh et al. (1994) — HMMs in biology.
Stochastic models for heterogeneous DNA sequences. Churchill (1989). Bulletin of Mathematical Biology. doi:10.1007/bf02458837
Abstract
The composition of naturally occurring DNA sequences is often strikingly heterogeneous. In this paper, the DNA sequence is viewed as a stochastic process with local compositional properties determined by the states of a hidden Markov chain. The model used is a discrete-state, discrete-outcome version of a general model for non-stationary time series proposed by Kitagawa (1987). A smoothing algorithm is described which can be used to reconstruct the hidden process and produce graphic displays of the compositional structure of a sequence. The problem of parameter estimation is approached using likelihood methods and an EM algorithm for approximating the maximum likelihood estimate is derived. The methods are applied to sequences from yeast mitochondrial DNA, human and mouse mitochondrial DNAs, a human X chromosomal fragment and the complete genome of bacteriophage lambda.
Hidden Markov Models in Computational Biology. Krogh et al. (1994). Journal of Molecular Biology. doi:10.1006/jmbi.1994.1104
Abstract
Hidden Markov Models (HMMs) are applied to the problems of statistical modeling, database searching and multiple sequence alignment of protein families and protein domains. These methods are demonstrated on the globin family, the protein kinase catalytic domain, and the EF-hand calcium binding motif. In each case the parameters of an HMM are estimated from a training set of unaligned sequences. After the HMM is built, it is used to obtain a multiple alignment of all the training sequences. It is also used to search the SWISS-PROT 22 database for other sequences that are members of the given protein family, or contain the given domain. The HMM produces multiple alignments of good quality that agree closely with the alignments produced by programs that incorporate three-dimensional structural information. When employed in discrimination tests (by examining how closely the sequences in a database fit the globin, kinase and EF-hand HMMs), the HMM is able to distinguish members of these families from non-members with a high degree of accuracy. Both the HMM and PROFILESEARCH (a technique used to search for relationships between a protein sequence and multiply aligned sequences) perform better in these tests than PROSITE (a dictionary of sites and patterns in proteins). The HMM appears to have a slight advantage over PROFILESEARCH in terms of lower rates of false negatives and false positives, even though the HMM is trained using only unaligned sequences, whereas PROFILESEARCH requires aligned training sequences. Our results suggest the presence of an EF-hand calcium binding motif in a highly conserved and evolutionary preserved putative intracellular region of 155 residues in the alpha-1 subunit of L-type calcium channels which play an important role in excitation-contraction coupling. This region has been suggested to contain the functional domains that are typical or essential for all L-type calcium channels regardless of whether they couple to ryanodine receptors, conduct ions or both.
Durbin et al. (1998) — Biological Sequence Analysis.
Biological Sequence Analysis: Probabilistic Models of Proteins and Nucleic Acids. Durbin et al. (1998). Cambridge University Press. doi:10.1017/cbo9780511790492
Abstract
Probabilistic models are becoming increasingly important in analysing the huge amount of data being produced by large-scale DNA-sequencing efforts such as the Human Genome Project. For example, hidden Markov models are used for analysing biological sequences, linguistic-grammar-based probabilistic models for identifying RNA secondary structure, and probabilistic evolutionary models for inferring phylogenies of sequences from different organisms. This book gives a unified, up-to-date and self-contained account, with a Bayesian slant, of such methods, and more generally to probabilistic methods of sequence analysis. Written by an interdisciplinary team of authors, it aims to be accessible to molecular biologists, computer scientists, and mathematicians with no formal knowledge of the other fields, and at the same time present the state-of-the-art in this new and highly important field.
Cortes and Vapnik (1995) — support vector machines.
Support-vector networks. Cortes & Vapnik (1995). Machine Learning. doi:10.1007/bf00994018
Breiman (2001), Friedman (2001), Chen and Guestrin (2016) — random forests and gradient boosting.
Random Forests. Breiman (2001). Machine Learning. doi:10.1023/a:1010933404324
Greedy function approximation: A gradient boosting machine.. Friedman (2001). The Annals of Statistics. doi:10.1214/aos/1013203451
Abstract
Function estimation/approximation is viewed from the perspective of numerical optimization in function space, rather than parameter space. A connection is made between stagewise additive expansions and steepest-descent minimization. A general gradient descent “boosting” paradigm is developed for additive expansions based on any fitting criterion.Specific algorithms are presented for least-squares, least absolute deviation, and Huber-M loss functions for regression, and multiclass logistic likelihood for classification. Special enhancements are derived for the particular case where the individual additive components are regression trees, and tools for interpreting such “TreeBoost” models are presented. Gradient boosting of regression trees produces competitive, highly robust, interpretable procedures for both regression and classification, especially appropriate for mining less than clean data. Connections between this approach and the boosting methods of Freund and Shapire and Friedman, Hastie and Tibshirani are discussed.
XGBoost: A Scalable Tree Boosting System. Chen & Guestrin (2016). Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. doi:10.1145/2939672.2939785
Abstract
Over 78 million people will suffer from dementia by 2030, emphasizing the need for early identification of patients with mild cognitive impairment (MCI) at risk, and personalized clinical evaluation steps to diagnose potentially reversible causes. Here, we leverage real-world electronic health records in the observational medical outcomes partnership (OMOP) data model to develop machine learning models to predict MCI up to a year in advance of recorded diagnosis. Our experimental results with logistic regression, random forest, and xgboost models trained and evaluated on more than 531K patient visits show random forest model can predict MCI onset with ROC-AUC of 68.2±0.7. We identify the clinical factors mentioned in clinician notes that are most predictive of MCI. Using similar association mining techniques, we develop a data-driven list of clinical procedures commonly ordered in the workup of MCI cases, that could be used as a basis for guidelines and clinical order set templates.
Hastie et al. (2009) — The Elements of Statistical Learning (free PDF).
The Elements of Statistical Learning. Hastie et al. (2009). Springer. https://hastie.su.domains/ElemStatLearn/
B.12.5 Deep learning
Lecun et al. (1998), Krizhevsky et al. (2017), He et al. (2016) — CNNs, AlexNet and residual networks.
Gradient-based learning applied to document recognition. Lecun et al. (1998). Proceedings of the IEEE. doi:10.1109/5.726791
Abstract
Multilayer neural networks trained with the back-propagation algorithm constitute the best example of a successful gradient based learning technique. Given an appropriate network architecture, gradient-based learning algorithms can be used to synthesize a complex decision surface that can classify high-dimensional patterns, such as handwritten characters, with minimal preprocessing. This paper reviews various methods applied to handwritten character recognition and compares them on a standard handwritten digit recognition task. Convolutional neural networks, which are specifically designed to deal with the variability of 2D shapes, are shown to outperform all other techniques. Real-life document recognition systems are composed of multiple modules including field extraction, segmentation recognition, and language modeling. A new learning paradigm, called graph transformer networks (GTN), allows such multimodule systems to be trained globally using gradient-based methods so as to minimize an overall performance measure. Two systems for online handwriting recognition are described. Experiments demonstrate the advantage of global training, and the flexibility of graph transformer networks. A graph transformer network for reading a bank cheque is also described. It uses convolutional neural network character recognizers combined with global training techniques to provide record accuracy on business and personal cheques. It is deployed commercially and reads several million cheques per day.
ImageNet classification with deep convolutional neural networks. Krizhevsky et al. (2017). Communications of the ACM. doi:10.1145/3065386
Abstract
We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 different classes. On the test data, we achieved top-1 and top-5 error rates of 37.5% and 17.0%, respectively, which is considerably better than the previous state-of-the-art. The neural network, which has 60 million parameters and 650,000 neurons, consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully connected layers with a final 1000-way softmax. To make training faster, we used non-saturating neurons and a very efficient GPU implementation of the convolution operation. To reduce overfitting in the fully connected layers we employed a recently developed regularization method called "dropout" that proved to be very effective. We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry.
Deep Residual Learning for Image Recognition. He et al. (2016). 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). doi:10.1109/cvpr.2016.90
Abstract
Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers - 8× deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation.
Hochreiter and Schmidhuber (1997) — LSTM.
Long Short-Term Memory. Hochreiter & Schmidhuber (1997). Neural Computation. doi:10.1162/neco.1997.9.8.1735
Abstract
Learning to store information over extended time intervals by recurrent backpropagation takes a very long time, mostly because of insufficient, decaying error backflow. We briefly review Hochreiter's (1991) analysis of this problem, then address it by introducing a novel, efficient, gradient-based method called long short-term memory (LSTM). Truncating the gradient where this does not do harm, LSTM can learn to bridge minimal time lags in excess of 1000 discrete-time steps by enforcing constant error flow through constant error carousels within special units. Multiplicative gate units learn to open and close access to the constant error flow. LSTM is local in space and time; its computational complexity per time step and weight is O(1). Our experiments with artificial data involve local, distributed, real-valued, and noisy pattern representations. In comparisons with real-time recurrent learning, back propagation through time, recurrent cascade correlation, Elman nets, and neural sequence chunking, LSTM leads to many more successful runs, and learns much faster. LSTM also solves complex, artificial long-time-lag tasks that have never been solved by previous recurrent network algorithms.
Srivastava et al. (2014) — dropout.
Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Srivastava et al. (2014). Journal of Machine Learning Research. https://jmlr.org/papers/v15/srivastava14a.html
Vaswani et al. (2017) — the transformer.
Attention is all you need. Vaswani et al. (2017). Advances in Neural Information Processing Systems 30 (NeurIPS 2017). https://arxiv.org/abs/1706.03762
Alipanahi et al. (2015), Zhou and Troyanskaya (2015), Kelley et al. (2016) — CNNs for DNA and RNA sequences.
Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning. Alipanahi et al. (2015). Nature Biotechnology. doi:10.1038/nbt.3300
Abstract
Knowing the sequence specificities of DNA- and RNA-binding proteins is essential for developing models of the regulatory processes in biological systems and for identifying causal disease variants. Here we show that sequence specificities can be ascertained from experimental data with 'deep learning' techniques, which offer a scalable, flexible and unified computational approach for pattern discovery. Using a diverse array of experimental data and evaluation metrics, we find that deep learning outperforms other state-of-the-art methods, even when training on in vitro data and testing on in vivo data. We call this approach DeepBind and have built a stand-alone software tool that is fully automatic and handles millions of sequences per experiment. Specificities determined by DeepBind are readily visualized as a weighted ensemble of position weight matrices or as a 'mutation map' that indicates how variations affect binding within a specific sequence.
Predicting effects of noncoding variants with deep learning–based sequence model. Zhou & Troyanskaya (2015). Nature Methods. doi:10.1038/nmeth.3547
Abstract
Identifying functional effects of noncoding variants is a major challenge in human genetics. To predict the noncoding-variant effects de novo from sequence, we developed a deep learning-based algorithmic framework, DeepSEA (http://deepsea.princeton.edu/), that directly learns a regulatory sequence code from large-scale chromatin-profiling data, enabling prediction of chromatin effects of sequence alterations with single-nucleotide sensitivity. We further used this capability to improve prioritization of functional variants including expression quantitative trait loci (eQTLs) and disease-associated variants.
Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks. Kelley et al. (2016). Genome Research. doi:10.1101/gr.200535.115
Abstract
The complex language of eukaryotic gene expression remains incompletely understood. Despite the importance suggested by many noncoding variants statistically associated with human disease, nearly all such variants have unknown mechanisms. Here, we address this challenge using an approach based on a recent machine learning advance-deep convolutional neural networks (CNNs). We introduce the open source package Basset to apply CNNs to learn the functional activity of DNA sequences from genomics data. We trained Basset on a compendium of accessible genomic sites mapped in 164 cell types by DNase-seq, and demonstrate greater predictive accuracy than previous methods. Basset predictions for the change in accessibility between variant alleles were far greater for Genome-wide association study (GWAS) SNPs that are likely to be causal relative to nearby SNPs in linkage disequilibrium with them. With Basset, a researcher can perform a single sequencing assay in their cell type of interest and simultaneously learn that cell's chromatin accessibility code and annotate every mutation in the genome with its influence on present accessibility and latent potential for accessibility. Thus, Basset offers a powerful computational approach to annotate and interpret the noncoding genome.
Rives et al. (2021), Elnaggar et al. (2022) — protein language models.
Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Rives et al. (2021). Proceedings of the National Academy of Sciences. doi:10.1073/pnas.2016239118
Abstract
In the field of artificial intelligence, a combination of scale in data and model capacity enabled by unsupervised learning has led to major advances in representation learning and statistical generation. In the life sciences, the anticipated growth of sequencing promises unprecedented data on natural sequence diversity. Protein language modeling at the scale of evolution is a logical step toward predictive and generative artificial intelligence for biology. To this end, we use unsupervised learning to train a deep contextual language model on 86 billion amino acids across 250 million protein sequences spanning evolutionary diversity. The resulting model contains information about biological properties in its representations. The representations are learned from sequence data alone. The learned representation space has a multiscale organization reflecting structure from the level of biochemical properties of amino acids to remote homology of proteins. Information about secondary and tertiary structure is encoded in the representations and can be identified by linear projections. Representation learning produces features that generalize across a range of applications, enabling state-of-the-art supervised prediction of mutational effect and secondary structure and improving state-of-the-art features for long-range contact prediction.
ProtTrans: Toward Understanding the Language of Life Through Self-Supervised Learning. Elnaggar et al. (2022). IEEE Transactions on Pattern Analysis and Machine Intelligence. doi:10.1109/tpami.2021.3095381
Abstract
Computational biology and bioinformatics provide vast data gold-mines from protein sequences, ideal for Language Models (LMs) taken from Natural Language Processing (NLP). These LMs reach for new prediction frontiers at low inference costs. Here, we trained two auto-regressive models (Transformer-XL, XLNet) and four auto-encoder models (BERT, Albert, Electra, T5) on data from UniRef and BFD containing up to 393 billion amino acids. The protein LMs (pLMs) were trained on the Summit supercomputer using 5616 GPUs and TPU Pod up-to 1024 cores. Dimensionality reduction revealed that the raw pLM-embeddings from unlabeled data captured some biophysical features of protein sequences. We validated the advantage of using the embeddings as exclusive input for several subsequent tasks: (1) a per-residue (per-token) prediction of protein secondary structure (3-state accuracy Q3=81%-87%); (2) per-protein (pooling) predictions of protein sub-cellular location (ten-state accuracy: Q10=81%) and membrane versus water-soluble (2-state accuracy Q2=91%). For secondary structure, the most informative embeddings (ProtT5) for the first time outperformed the state-of-the-art without multiple sequence alignments (MSAs) or evolutionary information thereby bypassing expensive database searches. Taken together, the results implied that pLMs learned some of the grammar of the language of life. All our models are available through https://github.com/agemagician/ProtTrans.
Ji et al. (2021), Dalla-Torre et al. (2024) — DNA language models.
DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome. Ji et al. (2021). Bioinformatics. doi:10.1093/bioinformatics/btab083
Abstract
Motivation Deciphering the language of non-coding DNA is one of the fundamental problems in genome research. Gene regulatory code is highly complex due to the existence of polysemy and distant semantic relationship, which previous informatics methods often fail to capture especially in data-scarce scenarios. Results To address this challenge, we developed a novel pre-trained bidirectional encoder representation, named DNABERT, to capture global and transferrable understanding of genomic DNA sequences based on up and downstream nucleotide contexts. We compared DNABERT to the most widely used programs for genome-wide regulatory elements prediction and demonstrate its ease of use, accuracy and efficiency. We show that the single pre-trained transformers model can simultaneously achieve state-of-the-art performance on prediction of promoters, splice sites and transcription factor binding sites, after easy fine-tuning using small task-specific labeled data. Further, DNABERT enables direct visualization of nucleotide-level importance and semantic relationship within input sequences for better interpretability and accurate identification of conserved sequence motifs and functional genetic variant candidates. Finally, we demonstrate that pre-trained DNABERT with human genome can even be readily applied to other organisms with exceptional performance. We anticipate that the pre-trained DNABERT model can be fined tuned to many other sequence analyses tasks. Availability and implementation The source code, pretrained and finetuned model for DNABERT are available at GitHub (https://github.com/jerryji1993/DNABERT). Supplementary information Supplementary data are available at Bioinformatics online.
Nucleotide Transformer: building and evaluating robust foundation models for human genomics. Dalla-Torre et al. (2024). Nature Methods. doi:10.1038/s41592-024-02523-z
Abstract
The prediction of molecular phenotypes from DNA sequences remains a longstanding challenge in genomics, often driven by limited annotated data and the inability to transfer learnings between tasks. Here, we present an extensive study of foundation models pre-trained on DNA sequences, named Nucleotide Transformer, ranging from 50 million up to 2.5 billion parameters and integrating information from 3,202 human genomes and 850 genomes from diverse species. These transformer models yield context-specific representations of nucleotide sequences, which allow for accurate predictions even in low-data settings. We show that the developed models can be fine-tuned at low cost to solve a variety of genomics applications. Despite no supervision, the models learned to focus attention on key genomic elements and can be used to improve the prioritization of genetic variants. The training and application of foundational models in genomics provides a widely applicable approach for accurate molecular phenotype prediction from DNA sequence.
Nguyen et al. (2024), Brixi et al. (2025) — Evo and Evo 2.
Sequence modeling and design from molecular to genome scale with Evo. Nguyen et al. (2024). Science. doi:10.1126/science.ado9336
Abstract
The genome is a sequence that encodes the DNA, RNA, and proteins that orchestrate an organism's function. We present Evo, a long-context genomic foundation model with a frontier architecture trained on millions of prokaryotic and phage genomes, and report scaling laws on DNA to complement observations in language and vision. Evo generalizes across DNA, RNA, and proteins, enabling zero-shot function prediction competitive with domain-specific language models and the generation of functional CRISPR-Cas and transposon systems, representing the first examples of protein-RNA and protein-DNA codesign with a language model. Evo also learns how small mutations affect whole-organism fitness and generates megabase-scale sequences with plausible genomic architecture. These prediction and generation capabilities span molecular to genomic scales of complexity, advancing our understanding and control of biology.
Genome modeling and design across all domains of life with Evo 2. Brixi et al. (2025). bioRxiv (preprint). doi:10.1101/2025.02.18.638918
Abstract
All of life encodes information with DNA. While tools for sequencing, synthesis, and editing of genomic code have transformed biological research, intelligently composing new biological systems would also require a deep understanding of the immense complexity encoded by genomes. We introduce Evo 2, a biological foundation model trained on 9.3 trillion DNA base pairs from a highly curated genomic atlas spanning all domains of life. We train Evo 2 with 7B and 40B parameters to have an unprecedented 1 million token context window with single-nucleotide resolution. Evo 2 learns from DNA sequence alone to accurately predict the functional impacts of genetic variation—from noncoding pathogenic mutations to clinically significant BRCA1 variants—without task-specific finetuning. Applying mechanistic interpretability analyses, we reveal that Evo 2 autonomously learns a breadth of biological features, including exon–intron boundaries, transcription factor binding sites, protein structural elements, and prophage genomic regions. Beyond its predictive capabilities, Evo 2 generates mitochondrial, prokaryotic, and eukaryotic sequences at genome scale with greater naturalness and coherence than previous methods. Guiding Evo 2 via inference-time search enables controllable generation of epigenomic structure, for which we demonstrate the first inference-time scaling results in biology. We make Evo 2 fully open, including model parameters, training code, inference code, and the OpenGenome2 dataset, to accelerate the exploration and design of biological complexity.
Avsec et al. (2021), Avsec et al. (2026) — Enformer and AlphaGenome.
Effective gene expression prediction from sequence by integrating long-range interactions. Avsec et al. (2021). Nature Methods. doi:10.1038/s41592-021-01252-x
Abstract
How noncoding DNA determines gene expression in different cell types is a major unsolved problem, and critical downstream applications in human genetics depend on improved solutions. Here, we report substantially improved gene expression prediction accuracy from DNA sequences through the use of a deep learning architecture, called Enformer, that is able to integrate information from long-range interactions (up to 100 kb away) in the genome. This improvement yielded more accurate variant effect predictions on gene expression for both natural genetic variants and saturation mutagenesis measured by massively parallel reporter assays. Furthermore, Enformer learned to predict enhancer-promoter interactions directly from the DNA sequence competitively with methods that take direct experimental data as input. We expect that these advances will enable more effective fine-mapping of human disease associations and provide a framework to interpret cis-regulatory evolution.
Advancing regulatory variant effect prediction with AlphaGenome. Avsec et al. (2026). Nature. doi:10.1038/s41586-025-10014-0
Abstract
Deep learning models that predict functional genomic measurements from DNA sequences are powerful tools for deciphering the genetic regulatory code. Existing methods involve a trade-off between input sequence length and prediction resolution, thereby limiting their modality scope and performance 1-5 . We present AlphaGenome, a unified DNA sequence model, which takes as input 1 Mb of DNA sequence and predicts thousands of functional genomic tracks up to single-base-pair resolution across diverse modalities. The modalities include gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription factor binding, chromatin contact maps, splice site usage and splice junction coordinates and strength. Trained on human and mouse genomes, AlphaGenome matches or exceeds the strongest available external models in 25 of 26 evaluations of variant effect prediction. The ability of AlphaGenome to simultaneously score variant effects across all modalities accurately recapitulates the mechanisms of clinically relevant variants near the TAL1 oncogene 6 . To facilitate broader use, we provide tools for making genome track and variant effect predictions from sequence.
Novakovsky et al. (2022) — explainable AI in genetics.
Obtaining genetics insights from deep learning via explainable artificial intelligence. Novakovsky et al. (2022). Nature Reviews Genetics. doi:10.1038/s41576-022-00532-2
Abstract
Artificial intelligence (AI) models based on deep learning now represent the state of the art for making functional predictions in genomics research. However, the underlying basis on which predictive models make such predictions is often unknown. For genomics researchers, this missing explanatory information would frequently be of greater value than the predictions themselves, as it can enable new insights into genetic processes. We review progress in the emerging area of explainable AI (xAI), a field with the potential to empower life science researchers to gain mechanistic insights into complex deep learning models. We discuss and categorize approaches for model interpretation, including an intuitive understanding of how each approach works and their underlying assumptions and limitations in the context of typical high-throughput biological datasets.
B.12.6 Generative and foundation models
Kingma and Welling (2014), Goodfellow et al. (2020), Ho et al. (2020) — VAEs, GANs and diffusion models.
Auto-Encoding Variational Bayes. Kingma & Welling (2014). https://arxiv.org/abs/1312.6114
Generative adversarial networks. Goodfellow et al. (2020). Communications of the ACM. doi:10.1145/3422622
Abstract
Generative adversarial networks are a kind of artificial intelligence algorithm designed to solve the generative modeling problem. The goal of a generative model is to study a collection of training examples and learn the probability distribution that generated them. Generative Adversarial Networks (GANs) are then able to generate more examples from the estimated probability distribution. Generative models based on deep learning are common, but GANs are among the most successful generative models (especially in terms of their ability to generate realistic high-resolution images). GANs have been successfully applied to a wide variety of tasks (mostly in research settings) but continue to present unique challenges and research opportunities because they are based on game theory while most other approaches to generative modeling are based on optimization.
Denoising Diffusion Probabilistic Models. Ho et al. (2020). Advances in Neural Information Processing Systems. https://arxiv.org/abs/2006.11239
Madani et al. (2023), Ferruz et al. (2022) — generating protein sequences with language models.
Large language models generate functional protein sequences across diverse families. Madani et al. (2023). Nature Biotechnology. doi:10.1038/s41587-022-01618-2
Abstract
Deep-learning language models have shown promise in various biotechnological applications, including protein design and engineering. Here we describe ProGen, a language model that can generate protein sequences with a predictable function across large protein families, akin to generating grammatically and semantically correct natural language sentences on diverse topics. The model was trained on 280 million protein sequences from >19,000 families and is augmented with control tags specifying protein properties. ProGen can be further fine-tuned to curated sequences and tags to improve controllable generation performance of proteins from families with sufficient homologous samples. Artificial proteins fine-tuned to five distinct lysozyme families showed similar catalytic efficiencies as natural lysozymes, with sequence identity to natural proteins as low as 31.4%. ProGen is readily adapted to diverse protein families, as we demonstrate with chorismate mutase and malate dehydrogenase.
ProtGPT2 is a deep unsupervised language model for protein design. Ferruz et al. (2022). Nature Communications. doi:10.1038/s41467-022-32007-7
Abstract
Protein design aims to build novel proteins customized for specific purposes, thereby holding the potential to tackle many environmental and biomedical problems. Recent progress in Transformer-based architectures has enabled the implementation of language models capable of generating text with human-like capabilities. Here, motivated by this success, we describe ProtGPT2, a language model trained on the protein space that generates de novo protein sequences following the principles of natural ones. The generated proteins display natural amino acid propensities, while disorder predictions indicate that 88% of ProtGPT2-generated proteins are globular, in line with natural sequences. Sensitive sequence searches in protein databases show that ProtGPT2 sequences are distantly related to natural ones, and similarity networks further demonstrate that ProtGPT2 is sampling unexplored regions of protein space. AlphaFold prediction of ProtGPT2-sequences yields well-folded non-idealized structures with embodiments and large loops and reveals topologies not captured in current structure databases. ProtGPT2 generates sequences in a matter of seconds and is freely available.
Hayes et al. (2025) — ESM3.
Simulating 500 million years of evolution with a language model. Hayes et al. (2025). Science. doi:10.1126/science.ads0018
Abstract
More than 3 billion years of evolution have produced an image of biology encoded into the space of natural proteins. Here, we show that language models trained at scale on evolutionary data can generate functional proteins that are far away from known proteins. We present ESM3, a frontier multimodal generative language model that reasons over the sequence, structure, and function of proteins. ESM3 can follow complex prompts combining its modalities and is highly responsive to alignment to improve its fidelity. We have prompted ESM3 to generate fluorescent proteins. Among the generations that we synthesized, we found a bright fluorescent protein at a far distance (58% sequence identity) from known fluorescent proteins, which we estimate is equivalent to simulating 500 million years of evolution.
Bommasani et al. (2021) — foundation models.
On the Opportunities and Risks of Foundation Models. Bommasani & others (2021). https://arxiv.org/abs/2108.07258
Kaplan et al. (2020) — scaling laws.
Scaling Laws for Neural Language Models. Kaplan et al. (2020). https://arxiv.org/abs/2001.08361
Thorp (2023) — ChatGPT is fun, but not an author.
ChatGPT is fun, but not an author. Thorp (2023). Science. doi:10.1126/science.adg7879
Abstract
In less than 2 months, the artificial intelligence (AI) program ChatGPT has become a cultural sensation. It is freely accessible through a web portal created by the tool's developer, OpenAI. The program-which automatically creates text based on written prompts-is so popular that it's likely to be "at capacity right now" if you attempt to use it. When you do get through, ChatGPT provides endless entertainment. I asked it to rewrite the first scene of the classic American play Death of a Salesman , but to feature Princess Elsa from the animated movie Frozen as the main character instead of Willy Loman. The output was an amusing conversation in which Elsa-who has come home from a tough day of selling-is told by her son Happy, "Come on, Mom. You're Elsa from Frozen . You have ice powers and you're a queen. You're unstoppable." Mash-ups like this are certainly fun, but there are serious implications for generative AI programs like ChatGPT in science and academia.
James et al. (2021) — An Introduction to Statistical Learning (free online).
An Introduction to Statistical Learning. James et al. (2021). Springer. https://www.statlearning.com/
Notin et al. (2023) — ProteinGym.
ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design. Notin et al. (2023). Advances in Neural Information Processing Systems. https://proteingym.org
Kedzierska et al. (2025), Ahlmann-Eltze et al. (2025) — evaluating foundation models against simple baselines.
Zero-shot evaluation reveals limitations of single-cell foundation models. Kedzierska et al. (2025). Genome Biology. doi:10.1186/s13059-025-03574-x
Abstract
Foundation models such as scGPT and Geneformer have not been rigorously evaluated in a setting where they are used without any further training (i.e., zero-shot). Understanding the performance of models in zero-shot settings is critical to applications that exclude the ability to fine-tune, such as discovery settings where labels are unknown. Our evaluation of the zero-shot performance of Geneformer and scGPT suggests that, in some cases, these models may face reliability challenges and could be outperformed by simpler methods. Our findings underscore the importance of zero-shot evaluations in development and deployment of foundation models in single-cell research.
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Ahlmann-Eltze et al. (2025). Nature Methods. doi:10.1038/s41592-025-02772-6
Abstract
Recent research in deep-learning-based foundation models promises to learn representations of single-cell data that enable prediction of the effects of genetic perturbations. Here we compared five foundation models and two other deep learning models against deliberately simple baselines for predicting transcriptome changes after single or double perturbations. None outperformed the baselines, which highlights the importance of critical benchmarking in directing and evaluating method development.
Jumper, Evans, Pritzel, Green, Figurnov, Ronneberger, Tunyasuvunakool, Bates, Žídek, Potapenko, Bridgland, Meyer, Kohl, Ballard, Cowie, Romera-Paredes, Nikolov, Jain, Adler, Back, Petersen, Reiman, Clancy, Zielinski, Steinegger, Pacholska, Berghammer, Bodenstein, et al. (2021) — AlphaFold2 (see Chapter 9).
Highly accurate protein structure prediction with AlphaFold. Jumper et al. (2021). Nature. doi:10.1038/s41586-021-03819-2
Abstract
Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous experimental effort 1-4 , the structures of around 100,000 unique proteins have been determined 5 , but this represents a small fraction of the billions of known protein sequences 6,7 . Structural coverage is bottlenecked by the months to years of painstaking effort required to determine a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the three-dimensional structure that a protein will adopt based solely on its amino acid sequence-the structure prediction component of the 'protein folding problem' 8 -has been an important open research problem for more than 50 years 9 . Despite recent progress 10-14 , existing methods fall far short of atomic accuracy, especially when no homologous structure is available. Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14) 15 , demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates physical and biological knowledge about protein structure, leveraging multi-sequence alignments, into the design of the deep learning algorithm.
B.13 Image analysis and models of cells
Further papers for Chapter 13.
In the HTML version, click Abstract to show the abstract.
B.13.1 Reviews and introductions
Moen et al. (2019) — deep learning for cellular image analysis.
Deep learning for cellular image analysis. Moen et al. (2019). Nature Methods. doi:10.1038/s41592-019-0403-1
Abstract
Recent advances in computer vision and machine learning underpin a collection of algorithms with an impressive ability to decipher the content of images. These deep learning algorithms are being applied to biological images and are transforming the analysis and interpretation of imaging data. These advances are positioned to render difficult analyses routine and to enable researchers to carry out new, previously impossible experiments. Here we review the intersection between deep learning and cellular image analysis and provide an overview of both the mathematical mechanics and the programming frameworks of deep learning that are pertinent to life scientists. We survey the field's progress in four key applications: image classification, image segmentation, object tracking, and augmented microscopy. Last, we relay our labs' experience with three key aspects of implementing deep learning in the laboratory: annotating training data, selecting and training a range of neural network architectures, and deploying solutions. We also highlight existing datasets and implementations for each surveyed application.
Hallou et al. (2021) — deep learning for bioimage analysis, written for biologists.
Deep learning for bioimage analysis in developmental biology. Hallou et al. (2021). Development. doi:10.1242/dev.199616
Abstract
Deep learning has transformed the way large and complex image datasets can be processed, reshaping what is possible in bioimage analysis. As the complexity and size of bioimage data continues to grow, this new analysis paradigm is becoming increasingly ubiquitous. In this Review, we begin by introducing the concepts needed for beginners to understand deep learning. We then review how deep learning has impacted bioimage analysis and explore the open-source resources available to integrate it into a research project. Finally, we discuss the future of deep learning applied to cell and developmental biology. We analyze how state-of-the-art methodologies have the potential to transform our understanding of biological systems through new image-based analysis and modelling that integrate multimodal inputs in space and time.
Laine et al. (2021) — avoiding a replication crisis in deep-learning image analysis.
Avoiding a replication crisis in deep-learning-based bioimage analysis. Laine et al. (2021). Nature Methods. doi:10.1038/s41592-021-01284-3
Lundberg and Borner (2019) — spatial proteomics as a tool for cell biology.
Spatial proteomics: a powerful discovery tool for cell biology. Lundberg & Borner (2019). Nature Reviews Molecular Cell Biology. doi:10.1038/s41580-018-0094-y
Abstract
Protein subcellular localization is tightly controlled and intimately linked to protein function in health and disease. Capturing the spatial proteome - that is, the localizations of proteins and their dynamics at the subcellular level - is therefore essential for a complete understanding of cell biology. Owing to substantial advances in microscopy, mass spectrometry and machine learning applications for data analysis, the field is now mature for proteome-wide investigations of spatial cellular regulation. Studies of the human proteome have begun to reveal a complex architecture, including single-cell variations, dynamic protein translocations, changing interaction networks and proteins localizing to multiple compartments. Furthermore, several studies have successfully harnessed the power of comparative spatial proteomics as a discovery tool to unravel disease mechanisms. We are at the beginning of an era in which spatial proteomics finally integrates with cell biology and medical research, thereby paving the way for unbiased systems-level insights into cellular processes. Here, we discuss current methods for spatial proteomics using imaging or mass spectrometry and specifically highlight global comparative applications. The aim of this Review is to survey the state of the field and also to encourage more cell biologists to apply spatial proteomics approaches.
B.13.2 Deep learning for images
Krizhevsky et al. (2017), He et al. (2016) — AlexNet and residual networks.
ImageNet classification with deep convolutional neural networks. Krizhevsky et al. (2017). Communications of the ACM. doi:10.1145/3065386
Abstract
We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 different classes. On the test data, we achieved top-1 and top-5 error rates of 37.5% and 17.0%, respectively, which is considerably better than the previous state-of-the-art. The neural network, which has 60 million parameters and 650,000 neurons, consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully connected layers with a final 1000-way softmax. To make training faster, we used non-saturating neurons and a very efficient GPU implementation of the convolution operation. To reduce overfitting in the fully connected layers we employed a recently developed regularization method called "dropout" that proved to be very effective. We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry.
Deep Residual Learning for Image Recognition. He et al. (2016). 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). doi:10.1109/cvpr.2016.90
Abstract
Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers - 8× deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation.
Ronneberger et al. (2015), Falk et al. (2018) — U-Net, and U-Net for cell biologists.
U-Net: Convolutional Networks for Biomedical Image Segmentation. Ronneberger et al. (2015). Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. doi:10.1007/978-3-319-24574-4_28
U-Net: deep learning for cell counting, detection, and morphometry. Falk et al. (2018). Nature Methods. doi:10.1038/s41592-018-0261-2
Abstract
U-Net is a generic deep-learning solution for frequently occurring quantification tasks such as cell detection and shape measurements in biomedical image data. We present an ImageJ plugin that enables non-machine-learning experts to analyze their data with U-Net on either a local computer or a remote server/cloud service. The plugin comes with pretrained models for single-cell segmentation and allows for U-Net to be adapted to new tasks on the basis of a few annotated samples.
Dosovitskiy et al. (2021) — the vision transformer.
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. Dosovitskiy et al. (2021). doi:10.48550/arXiv.2010.11929
Abstract
While the Transformer architecture has become the de-facto standard for natural language processing tasks, its applications to computer vision remain limited. In vision, attention is either applied in conjunction with convolutional networks, or used to replace certain components of convolutional networks while keeping their overall structure in place. We show that this reliance on CNNs is not necessary and a pure transformer applied directly to sequences of image patches can perform very well on image classification tasks. When pre-trained on large amounts of data and transferred to multiple mid-sized or small image recognition benchmarks (ImageNet, CIFAR-100, VTAB, etc.), Vision Transformer (ViT) attains excellent results compared to state-of-the-art convolutional networks while requiring substantially fewer computational resources to train.
Caron et al. (2021), Oquab et al. (2023), He et al. (2022) — self-supervised learning of images (DINO, DINOv2, masked autoencoders).
Emerging Properties in Self-Supervised Vision Transformers. Caron et al. (2021). 2021 IEEE/CVF International Conference on Computer Vision (ICCV). doi:10.1109/iccv48922.2021.00951
Abstract
In this paper, we question if self-supervised learning provides new properties to Vision Transformer (ViT) [16] that stand out compared to convolutional networks (convnets). Beyond the fact that adapting self-supervised methods to this architecture works particularly well, we make the following observations: first, self-supervised ViT features contain explicit information about the semantic segmentation of an image, which does not emerge as clearly with supervised ViTs, nor with convnets. Second, these features are also excellent k-NN classifiers, reaching 78.3% top-1 on ImageNet with a small ViT. Our study also underlines the importance of momentum encoder [26], multi-crop training [9], and the use of small patches with ViTs. We implement our findings into a simple self-supervised method, called DINO, which we interpret as a form of self-distillation with no labels. We show the synergy between DINO and ViTs by achieving 80.1% top-1 on ImageNet in linear evaluation with ViT-Base.
DINOv2: Learning Robust Visual Features without Supervision. Oquab et al. (2023). doi:10.48550/arXiv.2304.07193
Abstract
The recent breakthroughs in natural language processing for model pretraining on large quantities of data have opened the way for similar foundation models in computer vision. These models could greatly simplify the use of images in any system by producing all-purpose visual features, i.e., features that work across image distributions and tasks without finetuning. This work shows that existing pretraining methods, especially self-supervised methods, can produce such features if trained on enough curated data from diverse sources. We revisit existing approaches and combine different techniques to scale our pretraining in terms of data and model size. Most of the technical contributions aim at accelerating and stabilizing the training at scale. In terms of data, we propose an automatic pipeline to build a dedicated, diverse, and curated image dataset instead of uncurated data, as typically done in the self-supervised literature. In terms of models, we train a ViT model (Dosovitskiy et al., 2020) with 1B parameters and distill it into a series of smaller models that surpass the best available all-purpose features, OpenCLIP (Ilharco et al., 2021) on most of the benchmarks at image and pixel levels.
Masked Autoencoders Are Scalable Vision Learners. He et al. (2022). 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). doi:10.1109/cvpr52688.2022.01553
Abstract
This paper shows that masked autoencoders (MAE) are scalable self-supervised learners for computer vision. Our MAE approach is simple: we mask random patches of the input image and reconstruct the missing pixels. It is based on two core designs. First, we develop an asymmetric encoder-decoder architecture, with an encoder that operates only on the visible subset of patches (without mask tokens), along with a lightweight decoder that reconstructs the original image from the latent representation and mask tokens. Second, we find that masking a high proportion of the input image, e.g., 75%, yields a nontrivial and meaningful self-supervisory task. Coupling these two designs enables us to train large models efficiently and effectively: we accelerate training (by 3× or more) and improve accuracy. Our scalable approach allows for learning high-capacity models that generalize well: e.g., a vanilla ViT-Huge model achieves the best accuracy (87.8%) among methods that use only ImageNet-1K data. Transfer performance in downstream tasks outperforms supervised pretraining and shows promising scaling behavior.
Kirillov et al. (2023), Archit et al. (2025), Marks et al. (2025) — Segment Anything and its versions for microscopy.
Segment Anything. Kirillov et al. (2023). 2023 IEEE/CVF International Conference on Computer Vision (ICCV). doi:10.1109/iccv51070.2023.00371
Abstract
We introduce the Segment Anything (SA) project: a new task, model, and dataset for image segmentation. Using our efficient model in a data collection loop, we built the largest segmentation dataset to date (by far), with over 1 billion masks on 11M licensed and privacy respecting images. The model is designed and trained to be promptable, so it can transfer zero-shot to new image distributions and tasks. We evaluate its capabilities on numerous tasks and find that its zero-shot performance is impressive – often competitive with or even superior to prior fully supervised results. We are releasing the Segment Anything Model (SAM) and corresponding dataset (SA-1B) of 1B masks and 11M images at segment-anything.com to foster research into foundation models for computer vision. We recommend reading the full paper at: arxiv.org/abs/2304.02643.
Segment Anything for Microscopy. Archit et al. (2025). Nature Methods. doi:10.1038/s41592-024-02580-4
Abstract
Accurate segmentation of objects in microscopy images remains a bottleneck for many researchers despite the number of tools developed for this purpose. Here, we present Segment Anything for Microscopy (μSAM), a tool for segmentation and tracking in multidimensional microscopy data. It is based on Segment Anything, a vision foundation model for image segmentation. We extend it by fine-tuning generalist models for light and electron microscopy that clearly improve segmentation quality for a wide range of imaging conditions. We also implement interactive and automatic segmentation in a napari plugin that can speed up diverse segmentation tasks and provides a unified solution for microscopy annotation across different microscopy modalities. Our work constitutes the application of vision foundation models in microscopy, laying the groundwork for solving image analysis tasks in this domain with a small set of powerful deep learning models.
CellSAM: a foundation model for cell segmentation. Marks et al. (2025). Nature Methods. doi:10.1038/s41592-025-02879-w
Abstract
Cells are a fundamental unit of biological organization, and identifying them in imaging data-cell segmentation-is a critical task for various cellular imaging experiments. Although deep learning methods have led to substantial progress on this problem, most models are specialist models that work well for specific domains but cannot be applied across domains or scale well with large amounts of data. Here we present CellSAM, a universal model for cell segmentation that generalizes across diverse cellular imaging data. CellSAM builds on top of the Segment Anything Model (SAM) by developing a prompt engineering approach for mask generation. We train an object detector, CellFinder, to automatically detect cells and prompt SAM to generate segmentations. We show that this approach allows a single model to achieve human-level performance for segmenting images of mammalian cells, yeast and bacteria collected across various imaging modalities. We show that CellSAM has strong zero-shot performance and can be improved with a few examples via few-shot learning. Additionally, we demonstrate how CellSAM can be applied across diverse bioimage analysis workflows. A deployed version of CellSAM is available at https://cellsam.deepcell.org/ .
Moutakanni et al. (2025) — self-supervised features for fluorescence microscopy (Cell-DINO).
Cell-DINO: Self-supervised image-based embeddings for cell fluorescent microscopy. Moutakanni et al. (2025). PLOS Computational Biology. doi:10.1371/journal.pcbi.1013828
Abstract
Accurately quantifying cellular morphology at scale could substantially empower existing single-cell approaches. However, measuring cell morphology remains an active field of research, which has inspired multiple computer vision algorithms over the years. Here, we show that DINOv2, a vision-transformer based, self-supervised algorithm, has a remarkable ability for learning rich representations of cellular morphology without manual annotations or any other type of supervision. We apply DINOv2 to cell phenotyping problems, and compare the performance of resulting models, called Cell-DINO models, on a wide variety of tasks across two publicly available imaging datasets of diverse specifications and biological focus. Compared to supervised and other self-supervised baselines, Cell-DINO models demonstrate improved performance, especially in low annotation regimes. For instance, to classify protein localization using only 1% of annotations on a challenging single-cell dataset, Cell-DINO performs 70% better than a supervised strategy, and 24% better than another self-supervised alternative. The results show that Cell-DINO can support the study of unknown biological variation, including single-cell heterogeneity and relationships between experimental conditions, making it an excellent tool for image-based biological discovery.
B.13.3 Segmentation and label-free prediction
Schmidt et al. (2018) — StarDist.
Cell Detection with Star-Convex Polygons. Schmidt et al. (2018). Medical Image Computing and Computer Assisted Intervention – MICCAI 2018. doi:10.1007/978-3-030-00934-2_30
Pachitariu and Stringer (2022) — Cellpose 2.0 and human-in-the-loop training.
Cellpose 2.0: how to train your own model. Pachitariu & Stringer (2022). Nature Methods. doi:10.1038/s41592-022-01663-4
Abstract
Pretrained neural network models for biological segmentation can provide good out-of-the-box results for many image types. However, such models do not allow users to adapt the segmentation style to their specific needs and can perform suboptimally for test images that are very different from the training images. Here we introduce Cellpose 2.0, a new package that includes an ensemble of diverse pretrained models as well as a human-in-the-loop pipeline for rapid prototyping of new custom models. We show that models pretrained on the Cellpose dataset can be fine-tuned with only 500-1,000 user-annotated regions of interest (ROI) to perform nearly as well as models trained on entire datasets with up to 200,000 ROI. A human-in-the-loop approach further reduced the required user annotation to 100-200 ROI, while maintaining high-quality segmentations. We provide software tools such as an annotation graphical user interface, a model zoo and a human-in-the-loop pipeline to facilitate the adoption of Cellpose 2.0.
Greenwald et al. (2021) — whole-cell segmentation of tissue images (Mesmer, TissueNet).
Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning. Greenwald et al. (2021). Nature Biotechnology. doi:10.1038/s41587-021-01094-0
Abstract
A principal challenge in the analysis of tissue imaging data is cell segmentation-the task of identifying the precise boundary of every cell in an image. To address this problem we constructed TissueNet, a dataset for training segmentation models that contains more than 1 million manually labeled cells, an order of magnitude more than all previously published segmentation training datasets. We used TissueNet to train Mesmer, a deep-learning-enabled segmentation algorithm. We demonstrated that Mesmer is more accurate than previous methods, generalizes to the full diversity of tissue types and imaging platforms in TissueNet, and achieves human-level performance. Mesmer enabled the automated extraction of key cellular features, such as subcellular localization of protein signal, which was challenging with previous approaches. We then adapted Mesmer to harness cell lineage information in highly multiplexed datasets and used this enhanced version to quantify cell morphology changes during human gestation. All code, data and models are released as a community resource.
Christiansen et al. (2018), Ounkomol et al. (2018) — predicting fluorescence images from transmitted light.
In Silico Labeling: Predicting Fluorescent Labels in Unlabeled Images. Christiansen et al. (2018). Cell. doi:10.1016/j.cell.2018.03.040
Abstract
Microscopy is a central method in life sciences. Many popular methods, such as antibody labeling, are used to add physical fluorescent labels to specific cellular constituents. However, these approaches have significant drawbacks, including inconsistency; limitations in the number of simultaneous labels because of spectral overlap; and necessary perturbations of the experiment, such as fixing the cells, to generate the measurement. Here, we show that a computational machine-learning approach, which we call "in silico labeling" (ISL), reliably predicts some fluorescent labels from transmitted-light images of unlabeled fixed or live biological samples. ISL predicts a range of labels, such as those for nuclei, cell type (e.g., neural), and cell state (e.g., cell death). Because prediction happens in silico, the method is consistent, is not limited by spectral overlap, and does not disturb the experiment. ISL generates biological measurements that would otherwise be problematic or impossible to acquire.
Label-free prediction of three-dimensional fluorescence images from transmitted-light microscopy. Ounkomol et al. (2018). Nature Methods. doi:10.1038/s41592-018-0111-2
Abstract
Understanding cells as integrated systems is central to modern biology. Although fluorescence microscopy can resolve subcellular structure in living cells, it is expensive, is slow, and can damage cells. We present a label-free method for predicting three-dimensional fluorescence directly from transmitted-light images and demonstrate that it can be used to generate multi-structure, integrated images. The method can also predict immunofluorescence (IF) from electron micrograph (EM) inputs, extending the potential applications.
B.13.4 Image-based profiling
Gustafsdottir et al. (2013), M.-A. Bray et al. (2016), Cimini et al. (2023) — the Cell Painting assay and its optimisation.
Multiplex Cytological Profiling Assay to Measure Diverse Cellular States. Gustafsdottir et al. (2013). PLoS ONE. doi:10.1371/journal.pone.0080999
Abstract
Computational methods for image-based profiling are under active development, but their success hinges on assays that can capture a wide range of phenotypes. We have developed a multiplex cytological profiling assay that "paints the cell" with as many fluorescent markers as possible without compromising our ability to extract rich, quantitative profiles in high throughput. The assay detects seven major cellular components. In a pilot screen of bioactive compounds, the assay detected a range of cellular phenotypes and it clustered compounds with similar annotated protein targets or chemical structure based on cytological profiles. The results demonstrate that the assay captures subtle patterns in the combination of morphological labels, thereby detecting the effects of chemical compounds even though their targets are not stained directly. This image-based assay provides an unbiased approach to characterize compound- and disease-associated cell states to support future probe discovery.
Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes. Bray et al. (2016). Nature Protocols. doi:10.1038/nprot.2016.105
Abstract
In morphological profiling, quantitative data are extracted from microscopy images of cells to identify biologically relevant similarities and differences among samples based on these profiles. This protocol describes the design and execution of experiments using Cell Painting, which is a morphological profiling assay that multiplexes six fluorescent dyes, imaged in five channels, to reveal eight broadly relevant cellular components or organelles. Cells are plated in multiwell plates, perturbed with the treatments to be tested, stained, fixed, and imaged on a high-throughput microscope. Next, an automated image analysis software identifies individual cells and measures ∼1,500 morphological features (various measures of size, shape, texture, intensity, and so on) to produce a rich profile that is suitable for the detection of subtle phenotypes. Profiles of cell populations treated with different experimental perturbations can be compared to suit many goals, such as identifying the phenotypic impact of chemical or genetic perturbations, grouping compounds and/or genes into functional pathways, and identifying signatures of disease. Cell culture and image acquisition takes 2 weeks; feature extraction and data analysis take an additional 1-2 weeks.
Optimizing the Cell Painting assay for image-based profiling. Cimini et al. (2023). Nature Protocols. doi:10.1038/s41596-023-00840-9
Abstract
In image-based profiling, software extracts thousands of morphological features of cells from multi-channel fluorescence microscopy images, yielding single-cell profiles that can be used for basic research and drug discovery. Powerful applications have been proven, including clustering chemical and genetic perturbations on the basis of their similar morphological impact, identifying disease phenotypes by observing differences in profiles between healthy and diseased cells and predicting assay outcomes by using machine learning, among many others. Here, we provide an updated protocol for the most popular assay for image-based profiling, Cell Painting. Introduced in 2013, it uses six stains imaged in five channels and labels eight diverse components of the cell: DNA, cytoplasmic RNA, nucleoli, actin, Golgi apparatus, plasma membrane, endoplasmic reticulum and mitochondria. The original protocol was updated in 2016 on the basis of several years' experience running it at two sites, after optimizing it by visual stain quality. Here, we describe the work of the Joint Undertaking for Morphological Profiling Cell Painting Consortium, to improve upon the assay via quantitative optimization by measuring the assay's ability to detect morphological phenotypes and group similar perturbations together. The assay gives very robust outputs despite various changes to the protocol, and two vendors' dyes work equivalently well. We present Cell Painting version 3, in which some steps are simplified and several stain concentrations can be reduced, saving costs. Cell culture and image acquisition take 1-2 weeks for typically sized batches of ≤20 plates; feature extraction and data analysis take an additional 1-2 weeks.This protocol is an update to Nat. Protoc. 11, 1757-1774 (2016): https://doi.org/10.1038/nprot.2016.105.
Caicedo et al. (2017) — data-analysis strategies for image-based profiling.
Data-analysis strategies for image-based cell profiling. Caicedo et al. (2017). Nature Methods. doi:10.1038/nmeth.4397
Abstract
Image-based cell profiling is a high-throughput strategy for the quantification of phenotypic differences among a variety of cell populations. It paves the way to studying biological systems on a large scale by using chemical and genetic perturbations. The general workflow for this technology involves image acquisition with high-throughput microscopy systems and subsequent image processing and analysis. Here, we introduce the steps required to create high-quality image-based (i.e., morphological) profiles from a collection of microscopy images. We recommend techniques that have proven useful in each stage of the data analysis process, on the basis of the experience of 20 laboratories worldwide that are refining their image-based cell-profiling methodologies in pursuit of biological discovery. The recommended techniques cover alternatives that may suit various biological goals, experimental designs, and laboratories' preferences.
Chandrasekaran et al. (2025) — a morphological map of gene over- and underexpression.
Morphological map of under- and overexpression of genes in human cells. Chandrasekaran et al. (2025). Nature Methods. doi:10.1038/s41592-025-02753-9
Abstract
Cell Painting images offer valuable insights into a cell's state and enable many biological applications, but publicly available arrayed datasets only include hundreds of genes perturbed. The JUMP Cell Painting Consortium perturbed roughly 75% of the protein-coding genome in human U-2 OS cells, generating a rich resource of single-cell images and extracted features. These profiles capture the phenotypic impacts of perturbing 15,243 human genes, including overexpressing 12,609 genes (using open reading frames) and knocking out 7,975 genes (using CRISPR-Cas9). Here we mitigated technical artifacts by rigorously evaluating data processing options and validated the dataset's robustness and biological relevance. Analysis of phenotypic profiles revealed previously undiscovered gene clusters and functional relationships, including those associated with mitochondrial function, cancer and neural processes. The JUMP Cell Painting genetic dataset is a valuable resource for exploring gene relationships and uncovering previously unknown functions.
Fay et al. (2023) — RxRx3, a public phenomics map (preprint).
RxRx3: Phenomics Map of Biology. Fay et al. (2023). bioRxiv (preprint). doi:10.1101/2023.02.07.527350
Abstract
The combination of modern genetic perturbation techniques with high content screening has enabled genome-scale cell microscopy experiments that can be leveraged to construct maps of biology . These are built by processing microscopy images to produce readouts in unified and relatable representation space to capture known biological relationships and discover new ones. To further enable the scientific community to develop methods and insights from map-scale data, here we release RxRx3 , the first ever public high-content screening dataset combining genome-scale CRISPR knockouts with multiple-concentration screening of small molecules (a set of FDA approved and commercially available bioactive compounds). The dataset contains 6-channel fluorescent microscopy images and associated deep learning embeddings from over 2.2 million wells that span 17,063 CRISPR knockouts and 1,674 compounds at 8 doses each. RxRx3 is one of the largest collections of cellular screening data, and as far as we know, the largest generated consistently via a common experimental protocol within a single laboratory. Our goal in releasing RxRx3 is to demonstrate the benefits of generating consistent data, enable the development of the machine learning methods on this scale of data and to foster research, methods development, and collaboration. For more information about RxRx3 please visit RxRx.ai/rxrx3
B.13.5 Protein localisation
Thul et al. (2017), Mahdessian et al. (2021) — the subcellular map of the Human Protein Atlas and single-cell variation.
A subcellular map of the human proteome. Thul et al. (2017). Science. doi:10.1126/science.aal3321
Abstract
Resolving the spatial distribution of the human proteome at a subcellular level can greatly increase our understanding of human biology and disease. Here we present a comprehensive image-based map of subcellular protein distribution, the Cell Atlas, built by integrating transcriptomics and antibody-based immunofluorescence microscopy with validation by mass spectrometry. Mapping the in situ localization of 12,003 human proteins at a single-cell level to 30 subcellular structures enabled the definition of the proteomes of 13 major organelles. Exploration of the proteomes revealed single-cell variations in abundance or spatial distribution and localization of about half of the proteins to multiple compartments. This subcellular map can be used to refine existing protein-protein interaction networks and provides an important resource to deconvolute the highly complex architecture of the human cell.
Spatiotemporal dissection of the cell cycle with single-cell proteogenomics. Mahdessian et al. (2021). Nature. doi:10.1038/s41586-021-03232-9
Abstract
The cell cycle, over which cells grow and divide, is a fundamental process of life. Its dysregulation has devastating consequences, including cancer 1-3 . The cell cycle is driven by precise regulation of proteins in time and space, which creates variability between individual proliferating cells. To our knowledge, no systematic investigations of such cell-to-cell proteomic variability exist. Here we present a comprehensive, spatiotemporal map of human proteomic heterogeneity by integrating proteomics at subcellular resolution with single-cell transcriptomics and precise temporal measurements of individual cells in the cell cycle. We show that around one-fifth of the human proteome displays cell-to-cell variability, identify hundreds of proteins with previously unknown associations with mitosis and the cell cycle, and provide evidence that several of these proteins have oncogenic functions. Our results show that cell cycle progression explains less than half of all cell-to-cell variability, and that most cycling proteins are regulated post-translationally, rather than by transcriptomic cycling. These proteins are disproportionately phosphorylated by kinases that regulate cell fate, whereas non-cycling proteins that vary between cells are more likely to be modified by kinases that regulate metabolism. This spatially resolved proteomic map of the cell cycle is integrated into the Human Protein Atlas and will serve as a resource for accelerating molecular studies of the human cell cycle and cell proliferation.
Le et al. (2022) — the weakly supervised single-cell classification competition.
Analysis of the Human Protein Atlas Weakly Supervised Single-Cell Classification competition. Le et al. (2022). Nature Methods. doi:10.1038/s41592-022-01606-z
Abstract
While spatial proteomics by fluorescence imaging has quickly become an essential discovery tool for researchers, fast and scalable methods to classify and embed single-cell protein distributions in such images are lacking. Here, we present the design and analysis of the results from the competition Human Protein Atlas - Single-Cell Classification hosted on the Kaggle platform. This represents a crowd-sourced competition to develop machine learning models trained on limited annotations to label single-cell protein patterns in fluorescent images. The particular challenges of this competition include class imbalance, weak labels and multi-label classification, prompting competitors to apply a wide range of approaches in their solutions. The winning models serve as the first subcellular omics tools that can annotate single-cell locations, extract single-cell features and capture cellular dynamics.
Cho et al. (2022), Kobayashi et al. (2022) — OpenCell and self-supervised localisation features (cytoself).
OpenCell: Endogenous tagging for the cartography of human cellular organization. Cho et al. (2022). Science. doi:10.1126/science.abi6983
Abstract
Elucidating the wiring diagram of the human cell is a central goal of the postgenomic era. We combined genome engineering, confocal live-cell imaging, mass spectrometry, and data science to systematically map the localization and interactions of human proteins. Our approach provides a data-driven description of the molecular and spatial networks that organize the proteome. Unsupervised clustering of these networks delineates functional communities that facilitate biological discovery. We found that remarkably precise functional information can be derived from protein localization patterns, which often contain enough information to identify molecular interactions, and that RNA binding proteins form a specific subgroup defined by unique interaction and localization properties. Paired with a fully interactive website (opencell.czbiohub.org), our work constitutes a resource for the quantitative cartography of human cellular organization.
Self-supervised deep learning encodes high-resolution features of protein subcellular localization. Kobayashi et al. (2022). Nature Methods. doi:10.1038/s41592-022-01541-z
Abstract
Explaining the diversity and complexity of protein localization is essential to fully understand cellular architecture. Here we present cytoself, a deep-learning approach for fully self-supervised protein localization profiling and clustering. Cytoself leverages a self-supervised training scheme that does not require preexisting knowledge, categories or annotations. Training cytoself on images of 1,311 endogenously labeled proteins from the OpenCell database reveals a highly resolved protein localization atlas that recapitulates major scales of cellular organization, from coarse classes, such as nuclear and cytoplasmic, to the subtle localization signatures of individual protein complexes. We quantitatively validate cytoself's ability to cluster proteins into organelles and protein complexes, showing that cytoself outperforms previous self-supervised approaches. Moreover, to better understand the inner workings of our model, we dissect the emergent features from which our clustering is derived, interpret them in the context of the fluorescence images, and analyze the performance contributions of each component of our approach.
Gupta et al. (2024) — SubCell, vision foundation models of the HPA images (preprint).
SubCell: Proteome-aware vision foundation models for microscopy capture single-cell biology. Gupta et al. (2024). bioRxiv (preprint). doi:10.1101/2024.12.06.627299
Abstract
Cell morphology and subcellular protein organization provide important insights into cellular function and behavior. These cellular features can be studied using large-scale fluorescence microscopy, and machine learning has become a powerful tool to interpret the resulting images for biological insights. Here, we introduce SubCell, a deep learning model for fluorescence microscopy designed to accurately capture cellular morphology, protein localization, cellular organization, and biological function beyond what humans can readily perceive. SubCell was trained on the proteome-wide image collection from the Human Protein Atlas with a novel proteome-aware learning objective. SubCell outperforms state-of-the-art methods across a variety of tasks relevant to single-cell biology and generalizes to other fluorescence microscopy datasets without any fine-tuning. Additionally, we construct the first proteome-wide hierarchical map of proteome organization that is directly learned from image data. This vision-based multiscale cell map defines cellular subsystems down to protein complex resolution, reveals proteins with similar functions, and distinguishes dynamic and stable behaviors within cellular compartments. Finally, combining SubCell with a protein sequence model enables a rich multimodal approach to capture gene function better than either vision-only or sequence-only models alone. In conclusion, SubCell creates deep, image-driven representations of cellular architecture that are applicable across diverse biological contexts and datasets.
Xiao et al. (2024) — review of AI methods for protein localisation.
A Review for Artificial Intelligence Based Protein Subcellular Localization. Xiao et al. (2024). Biomolecules. doi:10.3390/biom14040409
Abstract
Proteins need to be located in appropriate spatiotemporal contexts to carry out their diverse biological functions. Mislocalized proteins may lead to a broad range of diseases, such as cancer and Alzheimer's disease. Knowing where a target protein resides within a cell will give insights into tailored drug design for a disease. As the gold validation standard, the conventional wet lab uses fluorescent microscopy imaging, immunoelectron microscopy, and fluorescent biomarker tags for protein subcellular location identification. However, the booming era of proteomics and high-throughput sequencing generates tons of newly discovered proteins, making protein subcellular localization by wet-lab experiments a mission impossible. To tackle this concern, in the past decades, artificial intelligence (AI) and machine learning (ML), especially deep learning methods, have made significant progress in this research area. In this article, we review the latest advances in AI-based method development in three typical types of approaches, including sequence-based, knowledge-based, and image-based methods. We also elaborately discuss existing challenges and future directions in AI-based method development in this research field.
Viana et al. (2023) — integrated intracellular organisation of human stem cells (Allen Institute).
Integrated intracellular organization and its variations in human iPS cells. Viana et al. (2023). Nature. doi:10.1038/s41586-022-05563-7
Abstract
Understanding how a subset of expressed genes dictates cellular phenotype is a considerable challenge owing to the large numbers of molecules involved, their combinatorics and the plethora of cellular behaviours that they determine 1,2 . Here we reduced this complexity by focusing on cellular organization-a key readout and driver of cell behaviour 3,4 -at the level of major cellular structures that represent distinct organelles and functional machines, and generated the WTC-11 hiPSC Single-Cell Image Dataset v1, which contains more than 200,000 live cells in 3D, spanning 25 key cellular structures. The scale and quality of this dataset permitted the creation of a generalizable analysis framework to convert raw image data of cells and their structures into dimensionally reduced, quantitative measurements that can be interpreted by humans, and to facilitate data exploration. This framework embraces the vast cell-to-cell variability that is observed within a normal population, facilitates the integration of cell-by-cell structural data and allows quantitative analyses of distinct, separable aspects of organization within and across different cell populations. We found that the integrated intracellular organization of interphase cells was robust to the wide range of variation in cell shape in the population; that the average locations of some structures became polarized in cells at the edges of colonies while maintaining the 'wiring' of their interactions with other structures; and that, by contrast, changes in the location of structures during early mitotic reorganization were accompanied by changes in their wiring.
B.13.6 Tissues and pathology
Goltsev et al. (2018), Giesen et al. (2014), Angelo et al. (2014) — CODEX, imaging mass cytometry and MIBI.
Deep Profiling of Mouse Splenic Architecture with CODEX Multiplexed Imaging. Goltsev et al. (2018). Cell. doi:10.1016/j.cell.2018.07.010
Abstract
A highly multiplexed cytometric imaging approach, termed co-detection by indexing (CODEX), is used here to create multiplexed datasets of normal and lupus (MRL/lpr) murine spleens. CODEX iteratively visualizes antibody binding events using DNA barcodes, fluorescent dNTP analogs, and an in situ polymerization-based indexing procedure. An algorithmic pipeline for single-cell antigen quantification in tightly packed tissues was developed and used to overlay well-known morphological features with de novo characterization of lymphoid tissue architecture at a single-cell and cellular neighborhood levels. We observed an unexpected, profound impact of the cellular neighborhood on the expression of protein receptors on immune cells. By comparing normal murine spleen to spleens from animals with systemic autoimmune disease (MRL/lpr), extensive and previously uncharacterized splenic cell-interaction dynamics in the healthy versus diseased state was observed. The fidelity of multiplexed spatial cytometry demonstrated here allows for quantitative systemic characterization of tissue architecture in normal and clinically aberrant samples.
Highly multiplexed imaging of tumor tissues with subcellular resolution by mass cytometry. Giesen et al. (2014). Nature Methods. doi:10.1038/nmeth.2869
Abstract
Mass cytometry enables high-dimensional, single-cell analysis of cell type and state. In mass cytometry, rare earth metals are used as reporters on antibodies. Analysis of metal abundances using the mass cytometer allows determination of marker expression in individual cells. Mass cytometry has previously been applied only to cell suspensions. To gain spatial information, we have coupled immunohistochemical and immunocytochemical methods with high-resolution laser ablation to CyTOF mass cytometry. This approach enables the simultaneous imaging of 32 proteins and protein modifications at subcellular resolution; with the availability of additional isotopes, measurement of over 100 markers will be possible. We applied imaging mass cytometry to human breast cancer samples, allowing delineation of cell subpopulations and cell-cell interactions and highlighting tumor heterogeneity. Imaging mass cytometry complements existing imaging approaches. It will enable basic studies of tissue heterogeneity and function and support the transition of medicine toward individualized molecularly targeted diagnosis and therapies.
Multiplexed ion beam imaging of human breast tumors. Angelo et al. (2014). Nature Medicine. doi:10.1038/nm.3488
Abstract
Immunohistochemistry (IHC) is a tool for visualizing protein expression that is employed as part of the diagnostic workup for the majority of solid tissue malignancies. Existing IHC methods use antibodies tagged with fluorophores or enzyme reporters that generate colored pigments. Because these reporters exhibit spectral and spatial overlap when used simultaneously, multiplexed IHC is not routinely used in clinical settings. We have developed a method that uses secondary ion mass spectrometry to image antibodies tagged with isotopically pure elemental metal reporters. Multiplexed ion beam imaging (MIBI) is capable of analyzing up to 100 targets simultaneously over a five-log dynamic range. Here, we used MIBI to analyze formalin-fixed, paraffin-embedded human breast tumor tissue sections stained with ten labels simultaneously. The resulting data suggest that MIBI can provide new insights into disease pathogenesis that will be valuable for basic research, drug discovery and clinical diagnostics.
Janesick et al. (2023), Park et al. (2022) — imaging-based spatial transcriptomics and an overview of spatial omics.
High resolution mapping of the tumor microenvironment using integrated single-cell, spatial and in situ analysis. Janesick et al. (2023). Nature Communications. doi:10.1038/s41467-023-43458-x
Abstract
Single-cell and spatial technologies that profile gene expression across a whole tissue are revolutionizing the resolution of molecular states in clinical samples. Current commercially available technologies provide whole transcriptome single-cell, whole transcriptome spatial, or targeted in situ gene expression analysis. Here, we combine these technologies to explore tissue heterogeneity in large, FFPE human breast cancer sections. This integrative approach allowed us to explore molecular differences that exist between distinct tumor regions and to identify biomarkers involved in the progression towards invasive carcinoma. Further, we study cell neighborhoods and identify rare boundary cells that sit at the critical myoepithelial border confining the spread of malignant cells. Here, we demonstrate that each technology alone provides information about molecular signatures relevant to understanding cancer heterogeneity; however, it is the integration of these technologies that leads to deeper insights, ushering in discoveries that will progress oncology research and the development of diagnostics and therapeutics.
Spatial omics technologies at multimodal and single cell/subcellular level. Park et al. (2022). Genome Biology. doi:10.1186/s13059-022-02824-6
Abstract
Spatial omics technologies enable a deeper understanding of cellular organizations and interactions within a tissue of interest. These assays can identify specific compartments or regions in a tissue with differential transcript or protein abundance, delineate their interactions, and complement other methods in defining cellular phenotypes. A variety of spatial methodologies are being developed and commercialized; however, these techniques differ in spatial resolution, multiplexing capability, scale/throughput, and coverage. Here, we review the current and prospective landscape of single cell to subcellular resolution spatial omics technologies and analysis tools to provide a comprehensive picture for both research and clinical applications.
Amitay et al. (2023) — classifying cells in multiplexed images.
CellSighter: a neural network to classify cells in highly multiplexed images. Amitay et al. (2023). Nature Communications. doi:10.1038/s41467-023-40066-7
Abstract
Multiplexed imaging enables measurement of multiple proteins in situ, offering an unprecedented opportunity to chart various cell types and states in tissues. However, cell classification, the task of identifying the type of individual cells, remains challenging, labor-intensive, and limiting to throughput. Here, we present CellSighter, a deep-learning based pipeline to accelerate cell classification in multiplexed images. Given a small training set of expert-labeled images, CellSighter outputs the label probabilities for all cells in new images. CellSighter achieves over 80% accuracy for major cell types across imaging platforms, which approaches inter-observer concordance. Ablation studies and simulations show that CellSighter is able to generalize its training data and learn features of protein expression levels, as well as spatial features such as subcellular expression patterns. CellSighter's design reduces overfitting, and it can be trained with only thousands or even hundreds of labeled examples. CellSighter also outputs a prediction confidence, allowing downstream experts control over the results. Altogether, CellSighter drastically reduces hands-on time for cell classification in multiplexed images, while improving accuracy and consistency across datasets.
Chen et al. (2024), Lu et al. (2024), Vorontsov et al. (2024), Xu et al. (2024) — pathology foundation models.
Towards a general-purpose foundation model for computational pathology. Chen et al. (2024). Nature Medicine. doi:10.1038/s41591-024-02857-3
Abstract
Quantitative evaluation of tissue images is crucial for computational pathology (CPath) tasks, requiring the objective characterization of histopathological entities from whole-slide images (WSIs). The high resolution of WSIs and the variability of morphological features present significant challenges, complicating the large-scale annotation of data for high-performance applications. To address this challenge, current efforts have proposed the use of pretrained image encoders through transfer learning from natural image datasets or self-supervised learning on publicly available histopathology datasets, but have not been extensively developed and evaluated across diverse tissue types at scale. We introduce UNI, a general-purpose self-supervised model for pathology, pretrained using more than 100 million images from over 100,000 diagnostic H&E-stained WSIs (>77 TB of data) across 20 major tissue types. The model was evaluated on 34 representative CPath tasks of varying diagnostic difficulty. In addition to outperforming previous state-of-the-art models, we demonstrate new modeling capabilities in CPath such as resolution-agnostic tissue classification, slide classification using few-shot class prototypes, and disease subtyping generalization in classifying up to 108 cancer types in the OncoTree classification system. UNI advances unsupervised representation learning at scale in CPath in terms of both pretraining data and downstream evaluation, enabling data-efficient artificial intelligence models that can generalize and transfer to a wide range of diagnostically challenging tasks and clinical workflows in anatomic pathology.
A visual-language foundation model for computational pathology. Lu et al. (2024). Nature Medicine. doi:10.1038/s41591-024-02856-4
Abstract
The accelerated adoption of digital pathology and advances in deep learning have enabled the development of robust models for various pathology tasks across a diverse array of diseases and patient cohorts. However, model training is often difficult due to label scarcity in the medical domain, and a model's usage is limited by the specific task and disease for which it is trained. Additionally, most models in histopathology leverage only image data, a stark contrast to how humans teach each other and reason about histopathologic entities. We introduce CONtrastive learning from Captions for Histopathology (CONCH), a visual-language foundation model developed using diverse sources of histopathology images, biomedical text and, notably, over 1.17 million image-caption pairs through task-agnostic pretraining. Evaluated on a suite of 14 diverse benchmarks, CONCH can be transferred to a wide range of downstream tasks involving histopathology images and/or text, achieving state-of-the-art performance on histology image classification, segmentation, captioning, and text-to-image and image-to-text retrieval. CONCH represents a substantial leap over concurrent visual-language pretrained systems for histopathology, with the potential to directly facilitate a wide array of machine learning-based workflows requiring minimal or no further supervised fine-tuning.
A foundation model for clinical-grade computational pathology and rare cancers detection. Vorontsov et al. (2024). Nature Medicine. doi:10.1038/s41591-024-03141-0
Abstract
The analysis of histopathology images with artificial intelligence aims to enable clinical decision support systems and precision medicine. The success of such applications depends on the ability to model the diverse patterns observed in pathology images. To this end, we present Virchow, the largest foundation model for computational pathology to date. In addition to the evaluation of biomarker prediction and cell identification, we demonstrate that a large foundation model enables pan-cancer detection, achieving 0.95 specimen-level area under the (receiver operating characteristic) curve across nine common and seven rare cancers. Furthermore, we show that with less training data, the pan-cancer detector built on Virchow can achieve similar performance to tissue-specific clinical-grade models in production and outperform them on some rare variants of cancer. Virchow's performance gains highlight the value of a foundation model and open possibilities for many high-impact applications with limited amounts of labeled training data.
A whole-slide foundation model for digital pathology from real-world data. Xu et al. (2024). Nature. doi:10.1038/s41586-024-07441-w
Abstract
Digital pathology poses unique computational challenges, as a standard gigapixel slide may comprise tens of thousands of image tiles 1-3 . Prior models have often resorted to subsampling a small portion of tiles for each slide, thus missing the important slide-level context 4 . Here we present Prov-GigaPath, a whole-slide pathology foundation model pretrained on 1.3 billion 256 × 256 pathology image tiles in 171,189 whole slides from Providence, a large US health network comprising 28 cancer centres. The slides originated from more than 30,000 patients covering 31 major tissue types. To pretrain Prov-GigaPath, we propose GigaPath, a novel vision transformer architecture for pretraining gigapixel pathology slides. To scale GigaPath for slide-level learning with tens of thousands of image tiles, GigaPath adapts the newly developed LongNet 5 method to digital pathology. To evaluate Prov-GigaPath, we construct a digital pathology benchmark comprising 9 cancer subtyping tasks and 17 pathomics tasks, using both Providence and TCGA data 6 . With large-scale pretraining and ultra-large-context modelling, Prov-GigaPath attains state-of-the-art performance on 25 out of 26 tasks, with significant improvement over the second-best method on 18 tasks. We further demonstrate the potential of Prov-GigaPath on vision-language pretraining for pathology 7,8 by incorporating the pathology reports. In sum, Prov-GigaPath is an open-weight foundation model that achieves state-of-the-art performance on various digital pathology tasks, demonstrating the importance of real-world data and whole-slide modelling.
He et al. (2020) — predicting gene expression from histology.
Integrating spatial gene expression and breast tumour morphology via deep learning. He et al. (2020). Nature Biomedical Engineering. doi:10.1038/s41551-020-0578-x
Abstract
Spatial transcriptomics allows for the measurement of RNA abundance at a high spatial resolution, making it possible to systematically link the morphology of cellular neighbourhoods and spatially localized gene expression. Here, we report the development of a deep learning algorithm for the prediction of local gene expression from haematoxylin-and-eosin-stained histopathology images using a new dataset of 30,612 spatially resolved gene expression data matched to histopathology images from 23 patients with breast cancer. We identified over 100 genes, including known breast cancer biomarkers of intratumoral heterogeneity and the co-localization of tumour growth and immune activation, the expression of which can be predicted from the histopathology images at a resolution of 100 µm. We also show that the algorithm generalizes well to The Cancer Genome Atlas and to other breast cancer gene expression datasets without the need for re-training. Predicting the spatially resolved transcriptome of a tissue directly from tissue images may enable image-based screening for molecular biomarkers with spatial variation.
Howard et al. (2021) — site signatures and bias in pathology models.
The impact of site-specific digital histology signatures on deep learning model accuracy and bias. Howard et al. (2021). Nature Communications. doi:10.1038/s41467-021-24698-1
Abstract
The Cancer Genome Atlas (TCGA) is one of the largest biorepositories of digital histology. Deep learning (DL) models have been trained on TCGA to predict numerous features directly from histology, including survival, gene expression patterns, and driver mutations. However, we demonstrate that these features vary substantially across tissue submitting sites in TCGA for over 3,000 patients with six cancer subtypes. Additionally, we show that histologic image differences between submitting sites can easily be identified with DL. Site detection remains possible despite commonly used color normalization and augmentation methods, and we quantify the image characteristics constituting this site-specific digital histology signature. We demonstrate that these site-specific signatures lead to biased accuracy for prediction of features including survival, genomic mutations, and tumor stage. Furthermore, ethnicity can also be inferred from site-specific signatures, which must be accounted for to ensure equitable application of DL. These site-specific signatures can lead to overoptimistic estimates of model performance, and we propose a quadratic programming method that abrogates this bias by ensuring models are not trained and validated on samples from the same site.
B.13.7 Atlases and virtual cells
Regev et al. (2017), Rood et al. (2024) — the Human Cell Atlas.
The Human Cell Atlas. Regev et al. (2017). eLife. doi:10.7554/elife.27041
Abstract
The recent advent of methods for high-throughput single-cell molecular profiling has catalyzed a growing sense in the scientific community that the time is ripe to complete the 150-year-old effort to identify all cell types in the human body. The Human Cell Atlas Project is an international collaborative effort that aims to define all human cell types in terms of distinctive molecular profiles (such as gene expression profiles) and to connect this information with classical cellular descriptions (such as location and morphology). An open comprehensive reference map of the molecular state of cells in healthy human tissues would propel the systematic study of physiological states, developmental trajectories, regulatory circuitry and interactions of cells, and also provide a framework for understanding cellular dysregulation in human disease. Here we describe the idea, its potential utility, early proofs-of-concept, and some design considerations for the Human Cell Atlas, including a commitment to open data, code, and community.
The Human Cell Atlas from a cell census to a unified foundation model. Rood et al. (2024). Nature. doi:10.1038/s41586-024-08338-4
Abstract
With the convergence of notable advances in molecular and spatial profiling methods and new computational approaches taking advantage of artificial intelligence and machine learning, the construction of cell atlases is progressing from data collection to atlas integration and beyond. Here, we explore five ways in which cell atlases, including the Human Cell Atlas, are already revealing valuable biological insights, and how they are poised to provide even greater benefits in the coming years. In particular, we discuss cell atlases as censuses of cells; as 3D maps of cells in the body, across modalities and scales; as maps connecting genotype causes to phenotype effects; as 4D maps of development; and, ultimately, as foundation models of biology unifying all of these aspects and helping to transform medicine.
HuBMAP Consortium (2019), Jain et al. (2023) — HuBMAP.
The human body at cellular resolution: the NIH Human Biomolecular Atlas Program. HuBMAP Consortium (2019). Nature. doi:10.1038/s41586-019-1629-x
Abstract
Transformative technologies are enabling the construction of three-dimensional maps of tissues with unprecedented spatial and molecular resolution. Over the next seven years, the NIH Common Fund Human Biomolecular Atlas Program (HuBMAP) intends to develop a widely accessible framework for comprehensively mapping the human body at single-cell resolution by supporting technology development, data acquisition, and detailed spatial mapping. HuBMAP will integrate its efforts with other funding agencies, programs, consortia, and the biomedical research community at large towards the shared vision of a comprehensive, accessible three-dimensional molecular and cellular atlas of the human body, in health and under various disease conditions.
Advances and prospects for the Human BioMolecular Atlas Program (HuBMAP). Jain et al. (2023). Nature Cell Biology. doi:10.1038/s41556-023-01194-w
Abstract
The Human BioMolecular Atlas Program (HuBMAP) aims to create a multi-scale spatial atlas of the healthy human body at single-cell resolution by applying advanced technologies and disseminating resources to the community. As the HuBMAP moves past its first phase, creating ontologies, protocols and pipelines, this Perspective introduces the production phase: the generation of reference spatial maps of functional tissue units across many organs from diverse populations and the creation of mapping tools and infrastructure to advance biomedical research.
Pearce et al. (2025), Istrate et al. (2025) — TranscriptFormer and rBio (CZI; preprints).
A Cross-Species Generative Cell Atlas Across 1.5 Billion Years of Evolution: The TranscriptFormer Single-cell Model. Pearce et al. (2025). bioRxiv (preprint). doi:10.1101/2025.04.25.650731
Abstract
Single-cell transcriptomics has revolutionized our understanding of cellular diversity, yet our understanding of the transcriptional programs across the tree of life remains limited. Here we present TranscriptFormer, a family of generative foundation models trained on up to 112 million cells spanning 1.53 billion years of evolution across 12 species. By jointly modeling gene identities and expression levels using a novel generative architecture, TranscriptFormer encodes multi-scale biological structure, functioning as a queryable virtual cell atlas. We demonstrate state-of-the-art performance on both in-distribution and out-of-distribution cell type classification, with robust performance even for species separated by over 685 million years of evolution. TranscriptFormer can also perform zero-shot disease state identification in human cells and accurately transfers cell state annotations across species boundaries. As a generative model, TranscriptFormer can be prompted to predict cell type-specific transcription factors and gene-gene interactions that align with independent experimental observations. Developmental trajectories, phylogenetic relationships and cellular hierarchies emerge naturally in TranscriptFormer’s representations without any explicit training on these annotations. This work establishes a powerful framework for quantitative single-cell analysis, and comparative cellular biology, thus demonstrating that universal principles of cellular organization can be learned and predicted across the tree of life.
rbio1 - training scientific reasoning LLMs with biological world models as soft verifiers. Istrate et al. (2025). bioRxiv (preprint). doi:10.1101/2025.08.18.670981
Abstract
Reasoning models are typically trained against verification mechanisms in formally specified systems such as code or symbolic math. In open domains like biology, however, we lack exact rules to enable large-scale formal verification and instead often rely on lab experiments to test predictions. Such experiments are slow, costly, and cannot scale with computation. In this work, we show that world models of biology or other prior knowledge can serve as approximate oracles for soft verification , allowing reasoning systems to be trained without additional experimental data. We present two paradigms of training models with approximate verifiers: RLEMF : reinforcement learning with experimental model feedback and RLPK : reinforcement learning from prior knowledge. Using these paradigms, we introduce rbio1 , a reasoning model for biology post-trained from a pretrained LLM with reinforcement learning, using learned biological models for verification during training. We demonstrate that soft verification can distill biological world models into rbio1 , enabling it to achieve state-of-the-art performance on perturbation prediction in the PerturbQA benchmark. We further show that composing multiple AI-verifiers improves performance and that models trained with soft biological rewards transfer zero-shot to cross-domain tasks such as disease-state prediction. We present rbio1 as a proof of concept that predictions from biological models can train powerful reasoning systems using simulations rather than experimental data, offering a new paradigm for model training.
Roohani et al. (2023), Adduri et al. (2026), Roohani et al. (2025) — GEARS, State and the Virtual Cell Challenge.
Predicting transcriptional outcomes of novel multigene perturbations with GEARS. Roohani et al. (2023). Nature Biotechnology. doi:10.1038/s41587-023-01905-6
Abstract
Understanding cellular responses to genetic perturbation is central to numerous biomedical applications, from identifying genetic interactions involved in cancer to developing methods for regenerative medicine. However, the combinatorial explosion in the number of possible multigene perturbations severely limits experimental interrogation. Here, we present graph-enhanced gene activation and repression simulator (GEARS), a method that integrates deep learning with a knowledge graph of gene-gene relationships to predict transcriptional responses to both single and multigene perturbations using single-cell RNA-sequencing data from perturbational screens. GEARS is able to predict outcomes of perturbing combinations consisting of genes that were never experimentally perturbed. GEARS exhibited 40% higher precision than existing approaches in predicting four distinct genetic interaction subtypes in a combinatorial perturbation screen and identified the strongest interactions twice as well as prior approaches. Overall, GEARS can predict phenotypically distinct effects of multigene perturbations and thus guide the design of perturbational experiments.
Predicting cellular responses to perturbation across diverse contexts with State. Adduri et al. (2026). Cell. doi:10.1016/j.cell.2026.07.052
Abstract
While machine learning models offer potential for predicting transcriptomic effects of perturbation, they currently struggle to generalize across cellular contexts. Here, we introduce State, a machine learning model that predicts perturbation effects while accounting for cellular heterogeneity within and across experiments. State is trained using single-cell gene expression data to predict perturbation effects across sets of cells. State improved discrimination of effects on large datasets by more than 30% and identified differentially expressed genes across genetic, signaling, and chemical perturbations with significantly improved accuracy compared with baselines. Its cell embeddings trained on observational data from 167 million cells enable the identification of strong perturbations in cellular contexts where no perturbations were observed during training. We further introduce Cell-Eval, a comprehensive evaluation framework that can be used to evaluate future models. Overall, the performance and flexibility of State set the stage for scaling the development of AI models of cell state.
Virtual Cell Challenge: Toward a Turing test for the virtual cell. Roohani et al. (2025). Cell. doi:10.1016/j.cell.2025.06.008
Abstract
Virtual cells are an emerging frontier at the intersection of artificial intelligence and biology. A key goal of these cell state models is predicting cellular responses to perturbations. The Virtual Cell Challenge is being established to catalyze progress toward this goal. This recurring and open benchmark competition from the Arc Institute will provide an evaluation framework, purpose-built datasets, and a venue for accelerating model development.
Kedzierska et al. (2025) — zero-shot evaluation of single-cell foundation models.
Zero-shot evaluation reveals limitations of single-cell foundation models. Kedzierska et al. (2025). Genome Biology. doi:10.1186/s13059-025-03574-x
Abstract
Foundation models such as scGPT and Geneformer have not been rigorously evaluated in a setting where they are used without any further training (i.e., zero-shot). Understanding the performance of models in zero-shot settings is critical to applications that exclude the ability to fine-tune, such as discovery settings where labels are unknown. Our evaluation of the zero-shot performance of Geneformer and scGPT suggests that, in some cases, these models may face reliability challenges and could be outperformed by simpler methods. Our findings underscore the importance of zero-shot evaluations in development and deployment of foundation models in single-cell research.
B.13.8 Whole-cell models and cryo-electron tomography
Karr et al. (2012) — the first whole-cell model.
A Whole-Cell Computational Model Predicts Phenotype from Genotype. Karr et al. (2012). Cell. doi:10.1016/j.cell.2012.05.044
Abstract
Understanding how complex phenotypes arise from individual molecules and their interactions is a primary challenge in biology that computational approaches are poised to tackle. We report a whole-cell computational model of the life cycle of the human pathogen Mycoplasma genitalium that includes all of its molecular components and their interactions. An integrative approach to modeling that combines diverse mathematics enabled the simultaneous inclusion of fundamentally different cellular processes and experimental measurements. Our whole-cell model accounts for all annotated gene functions and was validated against a broad range of data. The model provides insights into many previously unobserved cellular behaviors, including in vivo rates of protein-DNA association and an inverse relationship between the durations of DNA replication initiation and replication. In addition, experimental analysis directed by model predictions identified previously undetected kinetic parameters and biological functions. We conclude that comprehensive whole-cell models can be used to facilitate biological discovery.
Thornburg et al. (2022) — a kinetic model of a living minimal cell.
Fundamental behaviors emerge from simulations of a living minimal cell. Thornburg et al. (2022). Cell. doi:10.1016/j.cell.2021.12.025
Abstract
We present a whole-cell fully dynamical kinetic model (WCM) of JCVI-syn3A, a minimal cell with a reduced genome of 493 genes that has retained few regulatory proteins or small RNAs. Cryo-electron tomograms provide the cell geometry and ribosome distributions. Time-dependent behaviors of concentrations and reaction fluxes from stochastic-deterministic simulations over a cell cycle reveal how the cell balances demands of its metabolism, genetic information processes, and growth, and offer insight into the principles of life for this minimal cell. The energy economy of each process including active transport of amino acids, nucleosides, and ions is analyzed. WCM reveals how emergent imbalances lead to slowdowns in the rates of transcription and translation. Integration of experimental data is critical in building a kinetic model from which emerges a genome-wide distribution of mRNA half-lives, multiple DNA replication events that can be compared to qPCR results, and the experimentally observed doubling behavior.
Maritan et al. (2022), Stevens et al. (2023), Yu et al. (2016) — structural models and molecular dynamics of cells.
Building Structural Models of a Whole Mycoplasma Cell. Maritan et al. (2022). Journal of Molecular Biology. doi:10.1016/j.jmb.2021.167351
Abstract
Building structural models of entire cells has been a long-standing cross-discipline challenge for the research community, as it requires an unprecedented level of integration between multiple sources of biological data and enhanced methods for computational modeling and visualization. Here, we present the first 3D structural models of an entire Mycoplasma genitalium (MG) cell, built using the CellPACK suite of computational modeling tools. Our model recapitulates the data described in recent whole-cell system biology simulations and provides a structural representation for all MG proteins, DNA and RNA molecules, obtained by combining experimental and homology-modeled structures and lattice-based models of the genome. We establish a framework for gathering, curating and evaluating these structures, exposing current weaknesses of modeling methods and the boundaries of MG structural knowledge, and visualization methods to explore functional characteristics of the genome and proteome. We compare two approaches for data gathering, a manually-curated workflow and an automated workflow that uses homologous structures, both of which are appropriate for the analysis of mesoscale properties such as crowding and volume occupancy. Analysis of model quality provides estimates of the regularization that will be required when these models are used as starting points for atomic molecular dynamics simulations.
Molecular dynamics simulation of an entire cell. Stevens et al. (2023). Frontiers in Chemistry. doi:10.3389/fchem.2023.1106495
Abstract
The ultimate microscope, directed at a cell, would reveal the dynamics of all the cell's components with atomic resolution. In contrast to their real-world counterparts, computational microscopes are currently on the brink of meeting this challenge. In this perspective, we show how an integrative approach can be employed to model an entire cell, the minimal cell, JCVI-syn3A, at full complexity. This step opens the way to interrogate the cell's spatio-temporal evolution with molecular dynamics simulations, an approach that can be extended to other cell types in the near future.
Biomolecular interactions modulate macromolecular structure and dynamics in atomistic model of a bacterial cytoplasm. Yu et al. (2016). eLife. doi:10.7554/elife.19274
Abstract
Biological macromolecules function in highly crowded cellular environments. The structure and dynamics of proteins and nucleic acids are well characterized in vitro, but in vivo crowding effects remain unclear. Using molecular dynamics simulations of a comprehensive atomistic model cytoplasm we found that protein-protein interactions may destabilize native protein structures, whereas metabolite interactions may induce more compact states due to electrostatic screening. Protein-protein interactions also resulted in significant variations in reduced macromolecular diffusion under crowded conditions, while metabolites exhibited significant two-dimensional surface diffusion and altered protein-ligand binding that may reduce the effective concentration of metabolites and ligands in vivo. Metabolic enzymes showed weak non-specific association in cellular environments attributed to solvation and entropic effects. These effects are expected to have broad implications for the in vivo functioning of biomolecules. This work is a first step towards physically realistic in silico whole-cell models that connect molecular with cellular biology.
Nickell et al. (2006), Beck and Baumeister (2016), Young and Villa (2023) — visual proteomics and cryo-electron tomography.
A visual approach to proteomics. Nickell et al. (2006). Nature Reviews Molecular Cell Biology. doi:10.1038/nrm1861
Abstract
Cryo-electron tomography is an emerging imaging technique that has unique potential for molecular cell biology. At the present resolution of 4-5 nm, large supramolecular structures can be studied in unperturbed cellular environments and, in the future, it will become possible to map molecular landscapes inside cells in a more comprehensive manner. 'Visual proteomics' aims to complement and extend mass-spectrometry-based inventories, and to provide a quantitative description of the macromolecular interactions that underlie cellular functions.
Cryo-Electron Tomography: Can it Reveal the Molecular Sociology of Cells in Atomic Detail?. Beck & Baumeister (2016). Trends in Cell Biology. doi:10.1016/j.tcb.2016.08.006
Abstract
Traditionally, macromolecular structure determination is performed ex situ, that is, with purified materials. But, there are strong incentives to develop approaches to study them in situ in their native functional context. In recent years, cryo-electron tomography (cryo-ET) has emerged as a powerful method for visualizing the molecular organization of unperturbed cellular landscapes with the potential to attain near-atomic resolution. Here, we review recent work on several macromolecular assemblies, demonstrating the power of in situ studies. We also highlight technical challenges and discuss ways to meet them.
Bringing Structure to Cell Biology with Cryo-Electron Tomography. Young & Villa (2023). Annual Review of Biophysics. doi:10.1146/annurev-biophys-111622-091327
Abstract
Recent advances in cryo-electron microscopy have marked only the beginning of the potential of this technique. To bring structure into cell biology, the modality of cryo-electron tomography has fast developed into a bona fide in situ structural biology technique where structures are determined in their native environment, the cell. Nearly every step of the cryo-focused ion beam-assisted electron tomography (cryo-FIB-ET) workflow has been improved upon in the past decade, since the first windows were carved into cells, unveiling macromolecular networks in near-native conditions. By bridging structural and cell biology, cryo-FIB-ET is advancing our understanding of structure-function relationships in their native environment and becoming a tool for discovering new biology.
O’Reilly et al. (2020), Tegunov et al. (2021) — structural biology inside Mycoplasma pneumoniae cells.
In-cell architecture of an actively transcribing-translating expressome. O’Reilly et al. (2020). Science. doi:10.1126/science.abb3758
Abstract
Structural biology studies performed inside cells can capture molecular machines in action within their native context. In this work, we developed an integrative in-cell structural approach using the genome-reduced human pathogen Mycoplasma pneumoniae We combined whole-cell cross-linking mass spectrometry, cellular cryo-electron tomography, and integrative modeling to determine an in-cell architecture of a transcribing and translating expressome at subnanometer resolution. The expressome comprises RNA polymerase (RNAP), the ribosome, and the transcription elongation factors NusG and NusA. We pinpointed NusA at the interface between a NusG-bound elongating RNAP and the ribosome and propose that it can mediate transcription-translation coupling. Translation inhibition dissociated the expressome, whereas transcription inhibition stalled and rearranged it. Thus, the active expressome architecture requires both translation and transcription elongation within the cell.
Multi-particle cryo-EM refinement with M visualizes ribosome-antibiotic complex at 3.5 Å in cells. Tegunov et al. (2021). Nature Methods. doi:10.1038/s41592-020-01054-7
Abstract
Cryo-electron microscopy (cryo-EM) enables macromolecular structure determination in vitro and inside cells. In addition to aligning individual particles, accurate registration of sample motion and three-dimensional deformation during exposures are crucial for achieving high-resolution reconstructions. Here we describe M, a software tool that establishes a reference-based, multi-particle refinement framework for cryo-EM data and couples a comprehensive spatial deformation model to in silico correction of electron-optical aberrations. M provides a unified optimization framework for both frame-series and tomographic tilt-series data. We show that tilt-series data can provide the same resolution as frame-series data on a purified protein specimen, indicating that the alignment step no longer limits the resolution obtainable from tomographic data. In combination with Warp and RELION, M resolves to residue level a 70S ribosome bound to an antibiotic inside intact bacterial cells. Our work provides a computational tool that facilitates structural biology in cells.
Teresa-Trueba et al. (2023), Rice et al. (2023), Lamm et al. (2026) — deep learning for finding particles and membranes in tomograms.
Convolutional networks for supervised mining of molecular patterns within cellular context. de Teresa-Trueba et al. (2023). Nature Methods. doi:10.1038/s41592-022-01746-2
Abstract
Cryo-electron tomograms capture a wealth of structural information on the molecular constituents of cells and tissues. We present DeePiCt (deep picker in context), an open-source deep-learning framework for supervised segmentation and macromolecular complex localization in cryo-electron tomography. To train and benchmark DeePiCt on experimental data, we comprehensively annotated 20 tomograms of Schizosaccharomyces pombe for ribosomes, fatty acid synthases, membranes, nuclear pore complexes, organelles, and cytosol. By comparing DeePiCt to state-of-the-art approaches on this dataset, we show its unique ability to identify low-abundance and low-density complexes. We use DeePiCt to study compositionally distinct subpopulations of cellular ribosomes, with emphasis on their contextual association with mitochondria and the endoplasmic reticulum. Finally, applying pre-trained networks to a HeLa cell tomogram demonstrates that DeePiCt achieves high-quality predictions in unseen datasets from different biological species in a matter of minutes. The comprehensively annotated experimental data and pre-trained networks are provided for immediate use by the community.
TomoTwin: generalized 3D localization of macromolecules in cryo-electron tomograms with structural data mining. Rice et al. (2023). Nature Methods. doi:10.1038/s41592-023-01878-z
Abstract
Cryogenic-electron tomography enables the visualization of cellular environments in extreme detail, however, tools to analyze the full amount of information contained within these densely packed volumes are still needed. Detailed analysis of macromolecules through subtomogram averaging requires particles to first be localized within the tomogram volume, a task complicated by several factors including a low signal to noise ratio and crowding of the cellular space. Available methods for this task suffer either from being error prone or requiring manual annotation of training data. To assist in this crucial particle picking step, we present TomoTwin: an open source general picking model for cryogenic-electron tomograms based on deep metric learning. By embedding tomograms in an information-rich, high-dimensional space that separates macromolecules according to their three-dimensional structure, TomoTwin allows users to identify proteins in tomograms de novo without manually creating training data or retraining the network to locate new proteins.
MemBrain v2: an end-to-end tool for the analysis of membranes in cryo-electron tomography. Lamm et al. (2026). Nature Methods. doi:10.1038/s41592-026-03178-8
Abstract
Cryo-electron tomography provides unique insights into macromolecular complexes in their native environments, yet membrane analysis remains a major bottleneck due to low signal-to-noise ratios, missing wedge artifacts and the complexity of membrane-associated particles. Existing tools often require extensive manual annotation, struggle with generalization across datasets and lack integrated solutions for segmentation, particle localization and quantitative analysis. We introduce MemBrain v2, a deep-learning-enabled framework that unifies these tasks into a streamlined pipeline. MemBrain-seg leverages a diverse, collaboratively generated training dataset and specialized model training strategies to achieve generalizable membrane segmentation across variable tomographic conditions. MemBrain-pick enables data-efficient localization of membrane-bound particles by integrating geometric constraints with deep learning, reducing the need for extensive manual annotation. MemBrain-stats provides quantitative insights into particle distributions, computing spatial metrics to analyze intramembrane particle organization. MemBrain v2 integrates seamlessly into cryo-electron tomography workflows, providing an accessible and structured approach to membrane analysis.
Ermel et al. (2024), Peck et al. (2026) — the CryoET Data Portal and a particle-picking challenge.
A data portal for providing standardized annotations for cryo-electron tomography. Ermel et al. (2024). Nature Methods. doi:10.1038/s41592-024-02477-2
Lessons learned from a Kaggle challenge for particle picking in cryo-electron tomography. Peck et al. (2026). Nature Methods. doi:10.1038/s41592-026-03198-4
Abstract
The difficulty of particle picking in cryo-electron tomography remains a barrier to routine in situ structure determination. Machine learning is well suited to overcome this bottleneck with efficient algorithms that generalize across molecular species. To spur new algorithm development, we held a 3-month Kaggle challenge that tasked contestants with annotating five molecular species across hundreds of experimental tomograms. Here we analyze the results of this competition, which successfully engaged >1,000 participants and delivered particle pickers that outperformed existing state of the art. Systematic comparisons of the contestants' submissions revealed the tolerance of subtomogram averaging to moderate but not severe over-picking and underscored the need for more robust measures of annotation quality. The winning models also highlighted the importance of data augmentation to overcome limited training data. All competition tomograms along with the ground truth and winning teams' annotations have been released on the CryoET Data Portal as a resource to benchmark current and future particle picking algorithms.
B.13.9 Tools, data and standards
Schindelin et al. (2012), Rueden et al. (2017) — Fiji and ImageJ2.
Fiji: an open-source platform for biological-image analysis. Schindelin et al. (2012). Nature Methods. doi:10.1038/nmeth.2019
Abstract
Fiji is a distribution of the popular open-source software ImageJ focused on biological-image analysis. Fiji uses modern software engineering practices to combine powerful software libraries with a broad range of scripting languages to enable rapid prototyping of image-processing algorithms. Fiji facilitates the transformation of new algorithms into ImageJ plugins that can be shared with end users through an integrated update system. We propose Fiji as a platform for productive collaboration between computer science and biology research communities.
ImageJ2: ImageJ for the next generation of scientific image data. Rueden et al. (2017). BMC Bioinformatics. doi:10.1186/s12859-017-1934-z
Abstract
Background ImageJ is an image analysis program extensively used in the biological sciences and beyond. Due to its ease of use, recordable macro language, and extensible plug-in architecture, ImageJ enjoys contributions from non-programmers, amateur programmers, and professional developers alike. Enabling such a diversity of contributors has resulted in a large community that spans the biological and physical sciences. However, a rapidly growing user base, diverging plugin suites, and technical limitations have revealed a clear need for a concerted software engineering effort to support emerging imaging paradigms, to ensure the software's ability to handle the requirements of modern science. Results We rewrote the entire ImageJ codebase, engineering a redesigned plugin mechanism intended to facilitate extensibility at every level, with the goal of creating a more powerful tool that continues to serve the existing community while addressing a wider range of scientific requirements. This next-generation ImageJ, called "ImageJ2" in places where the distinction matters, provides a host of new functionality. It separates concerns, fully decoupling the data model from the user interface. It emphasizes integration with external applications to maximize interoperability. Its robust new plugin framework allows everything from image formats, to scripting languages, to visualization to be extended by the community. The redesigned data model supports arbitrarily large, N-dimensional datasets, which are increasingly common in modern image acquisition. Despite the scope of these changes, backwards compatibility is maintained such that this new functionality can be seamlessly integrated with the classic ImageJ interface, allowing users and developers to migrate to these new methods at their own pace. Conclusions Scientific imaging benefits from open-source programs that advance new method development and deployment to a diverse audience. ImageJ has continuously evolved with this idea in mind; however, new and emerging scientific requirements have posed corresponding challenges for ImageJ's development. The described improvements provide a framework engineered for flexibility, intended to support these requirements as well as accommodate future needs. Future efforts will focus on implementing new algorithms in this framework and expanding collaborations with other popular scientific software suites.
Carpenter et al. (2006), Stirling et al. (2021) — CellProfiler.
CellProfiler: image analysis software for identifying and quantifying cell phenotypes. Carpenter et al. (2006). Genome Biology. doi:10.1186/gb-2006-7-10-r100
Abstract
Biologists can now prepare and image thousands of samples per day using automation, enabling chemical screens and functional genomics (for example, using RNA interference). Here we describe the first free, open-source system designed for flexible, high-throughput cell image analysis, CellProfiler. CellProfiler can address a variety of biological questions quantitatively, including standard assays (for example, cell count, size, per-cell protein levels) and complex morphological assays (for example, cell/organelle shape or subcellular patterns of DNA or protein staining).
CellProfiler 4: improvements in speed, utility and usability. Stirling et al. (2021). BMC Bioinformatics. doi:10.1186/s12859-021-04344-9
Abstract
Background Imaging data contains a substantial amount of information which can be difficult to evaluate by eye. With the expansion of high throughput microscopy methodologies producing increasingly large datasets, automated and objective analysis of the resulting images is essential to effectively extract biological information from this data. CellProfiler is a free, open source image analysis program which enables researchers to generate modular pipelines with which to process microscopy images into interpretable measurements. Results Herein we describe CellProfiler 4, a new version of this software with expanded functionality. Based on user feedback, we have made several user interface refinements to improve the usability of the software. We introduced new modules to expand the capabilities of the software. We also evaluated performance and made targeted optimizations to reduce the time and cost associated with running common large-scale analysis pipelines. Conclusions CellProfiler 4 provides significantly improved performance in complex workflows compared to previous versions. This release will ensure that researchers will have continued access to CellProfiler's powerful computational tools in the coming years.
Berg et al. (2019), Bankhead et al. (2017) — ilastik and QuPath.
ilastik: interactive machine learning for (bio)image analysis. Berg et al. (2019). Nature Methods. doi:10.1038/s41592-019-0582-9
Abstract
We present ilastik, an easy-to-use interactive tool that brings machine-learning-based (bio)image analysis to end users without substantial computational expertise. It contains pre-defined workflows for image segmentation, object classification, counting and tracking. Users adapt the workflows to the problem at hand by interactively providing sparse training annotations for a nonlinear classifier. ilastik can process data in up to five dimensions (3D, time and number of channels). Its computational back end runs operations on-demand wherever possible, allowing for interactive prediction on data larger than RAM. Once the classifiers are trained, ilastik workflows can be applied to new data from the command line without further user interaction. We describe all ilastik workflows in detail, including three case studies and a discussion on the expected performance.
QuPath: Open source software for digital pathology image analysis. Bankhead et al. (2017). Scientific Reports. doi:10.1038/s41598-017-17204-5
Abstract
QuPath is new bioimage analysis software designed to meet the growing need for a user-friendly, extensible, open-source solution for digital pathology and whole slide image analysis. In addition to offering a comprehensive panel of tumor identification and high-throughput biomarker evaluation tools, QuPath provides researchers with powerful batch-processing and scripting functionality, and an extensible platform with which to develop and share new algorithms to analyze complex tissue images. Furthermore, QuPath's flexible design makes it suitable for a wide range of additional image analysis applications across biomedical research.
Chamier et al. (2021), Ouyang et al. (2022) — deep learning without programming (ZeroCostDL4Mic) and the BioImage Model Zoo (preprint).
Democratising deep learning for microscopy with ZeroCostDL4Mic. von Chamier et al. (2021). Nature Communications. doi:10.1038/s41467-021-22518-0
Abstract
Deep Learning (DL) methods are powerful analytical tools for microscopy and can outperform conventional image processing pipelines. Despite the enthusiasm and innovations fuelled by DL technology, the need to access powerful and compatible resources to train DL networks leads to an accessibility barrier that novice users often find difficult to overcome. Here, we present ZeroCostDL4Mic, an entry-level platform simplifying DL access by leveraging the free, cloud-based computational resources of Google Colab. ZeroCostDL4Mic allows researchers with no coding expertise to train and apply key DL networks to perform tasks including segmentation (using U-Net and StarDist), object detection (using YOLOv2), denoising (using CARE and Noise2Void), super-resolution microscopy (using Deep-STORM), and image-to-image translation (using Label-free prediction - fnet, pix2pix and CycleGAN). Importantly, we provide suitable quantitative tools for each network to evaluate model performance, allowing model optimisation. We demonstrate the application of the platform to study multiple biological processes.
BioImage Model Zoo: A Community-Driven Resource for Accessible Deep Learning in BioImage Analysis. Ouyang et al. (2022). bioRxiv (preprint). doi:10.1101/2022.06.07.495102
Abstract
Deep learning-based approaches are revolutionizing imaging-driven scientific research. However, the accessibility and reproducibility of deep learning-based workflows for imaging scientists remain far from sufficient. Several tools have recently risen to the challenge of democratizing deep learning by providing user-friendly interfaces to analyze new data with pre-trained or fine-tuned models. Still, few of the existing pre-trained models are interoperable between these tools, critically restricting a model’s overall utility and the possibility of validating and reproducing scientific analyses. Here, we present the BioImage Model Zoo ( https://bioimage.io ): a community-driven, fully open resource where standardized pre-trained models can be shared, explored, tested, and downloaded for further adaptation or direct deployment in multiple end user-facing tools (e.g., ilastik, deepImageJ, QuPath, StarDist, ImJoy, ZeroCostDL4Mic, CSBDeep). To enable everyone to contribute and consume the Zoo resources, we provide a model standard to enable cross-compatibility, a rich list of example models and practical use-cases, developer tools, documentation, and the accompanying infrastructure for model upload, download and testing. Our contribution aims to lay the groundwork to make deep learning methods for microscopy imaging findable, accessible, interoperable, and reusable (FAIR) across software tools and platforms.
Ljosa et al. (2012), Williams et al. (2017), Hartley et al. (2022) — BBBC, the Image Data Resource and the BioImage Archive.
Annotated high-throughput microscopy image sets for validation. Ljosa et al. (2012). Nature Methods. doi:10.1038/nmeth.2083
Image Data Resource: a bioimage data integration and publication platform. Williams et al. (2017). Nature Methods. doi:10.1038/nmeth.4326
Abstract
Access to primary research data is vital for the advancement of science. To extend the data types supported by community repositories, we built a prototype Image Data Resource (IDR) that collects and integrates imaging data acquired across many different imaging modalities. IDR links data from several imaging modalities, including high-content screening, super-resolution and time-lapse microscopy, digital pathology, public genetic or chemical databases, and cell and tissue phenotypes expressed using controlled ontologies. Using this integration, IDR facilitates the analysis of gene networks and reveals functional interactions that are inaccessible to individual studies. To enable re-analysis, we also established a computational resource based on Jupyter notebooks that allows remote access to the entire IDR. IDR is also an open source platform that others can use to publish their own image data. Thus IDR provides both a novel on-line resource and a software infrastructure that promotes and extends publication and re-analysis of scientific image data.
The BioImage Archive – Building a Home for Life-Sciences Microscopy Data. Hartley et al. (2022). Journal of Molecular Biology. doi:10.1016/j.jmb.2022.167505
Abstract
Despite the huge impact of data resources in genomics and structural biology, until now there has been no central archive for biological data for all imaging modalities. The BioImage Archive is a new data resource at the European Bioinformatics Institute (EMBL-EBI) designed to fill this gap. In its initial development BioImage Archive accepts bioimaging data associated with publications, in any format, from any imaging modality from the molecular to the organism scale, excluding medical imaging. The BioImage Archive will ensure reproducibility of published studies that derive results from image data and reduce duplication of effort. Most importantly, the BioImage Archive will help scientists to generate new insights through reuse of existing data to answer new biological questions, and provision of training, testing and benchmarking data for development of tools for image analysis. The archive is available at https://www.ebi.ac.uk/bioimage-archive/.
Sarkans et al. (2021), Moore et al. (2021), Moore et al. (2023) — REMBI metadata and the OME-Zarr format.
REMBI: Recommended Metadata for Biological Images—enabling reuse of microscopy data in biology. Sarkans et al. (2021). Nature Methods. doi:10.1038/s41592-021-01166-8
OME-NGFF: a next-generation file format for expanding bioimaging data-access strategies. Moore et al. (2021). Nature Methods. doi:10.1038/s41592-021-01326-w
Abstract
The rapid pace of innovation in biological imaging and the diversity of its applications have prevented the establishment of a community-agreed standardized data format. We propose that complementing established open formats such as OME-TIFF and HDF5 with a next-generation file format such as Zarr will satisfy the majority of use cases in bioimaging. Critically, a common metadata format used in all these vessels can deliver truly findable, accessible, interoperable and reusable bioimaging data.
OME-Zarr: a cloud-optimized bioimaging file format with international community support. Moore et al. (2023). Histochemistry and Cell Biology. doi:10.1007/s00418-023-02209-1
Abstract
A growing community is constructing a next-generation file format (NGFF) for bioimaging to overcome problems of scalability and heterogeneity. Organized by the Open Microscopy Environment (OME), individuals and institutes across diverse modalities facing these problems have designed a format specification process (OME-NGFF) to address these needs. This paper brings together a wide range of those community members to describe the cloud-optimized format itself-OME-Zarr-along with tools and data resources available today to increase FAIR access and remove barriers in the scientific process. The current momentum offers an opportunity to unify a key component of the bioimaging domain-the file format that underlies so many personal, institutional, and global data management and analysis tasks.
Maška et al. (2023) — the Cell Tracking Challenge.
The Cell Tracking Challenge: 10 years of objective benchmarking. Maška et al. (2023). Nature Methods. doi:10.1038/s41592-023-01879-y
Abstract
The Cell Tracking Challenge is an ongoing benchmarking initiative that has become a reference in cell segmentation and tracking algorithm development. Here, we present a significant number of improvements introduced in the challenge since our 2017 report. These include the creation of a new segmentation-only benchmark, the enrichment of the dataset repository with new datasets that increase its diversity and complexity, and the creation of a silver standard reference corpus based on the most competitive results, which will be of particular interest for data-hungry deep learning-based strategies. Furthermore, we present the up-to-date cell segmentation and tracking leaderboards, an in-depth analysis of the relationship between the performance of the state-of-the-art methods and the properties of the datasets and annotations, and two novel, insightful studies about the generalizability and the reusability of top-performing methods. These studies provide critical practical conclusions for both developers and users of traditional and machine learning-based cell segmentation and tracking algorithms.