CoChat
Sign Up
Cited by 542Protein Structure and DynamicsMachine Learning in BioinformaticsBioinformatics and Genomic Networks

Computed structures of core eukaryotic protein complexes

Open access
Humphreys, Ian R., Pei, Jimin, Baek, Minkyung, Krishnakumar, Aditya, Anishchenko, Ivan, Ovchinnikov, Sergey, Zhang, Jing, Ness, Travis J. +22 more·2021·Science
DOI 10.1126/science.abm4805PMC PMC7612107

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Deep learning for protein interactions: The use of deep learning has revolutionized the field of protein modeling. Humphreys et al . combined this approach with proteome-wide, coevolution-guided protein interaction identification to conduct a large-scale screen of protein-protein interactions in yeast (see the Perspective by Pereira and Schwede). The authors generated predicted interactions and accurate structures for complexes spanning key biological processes in Saccharomyces cerevisiae . The complexes include larger protein assemblies such as trimers, tetramers, and pentamers and provide insights into biological function. —VV

AI snapshot

Written from the paper’s full text.

TL;DR

The paper combines proteome-wide coevolution analysis with deep-learning structure prediction (RoseTTAFold then AlphaFold) to identify and model core protein complexes in *Saccharomyces cerevisiae*. It reports 1,505 likely interacting protein pairs and structural models for 106 previously unidentified assemblies and 806 that had not been structurally characterized.

Key findings

  • Screening 8.3 million yeast protein pairs identified 1,505 likely interactions; among these, 106 were previously unidentified and 806 had not been structurally characterized, with complexes containing as many as 5 subunits.
  • RoseTTAFold substantially outperformed Direct Coupling Analysis in distinguishing gold-standard interactions from random pairs (AUC 0.219 vs. 0.016; corrected RF AUC 0.248), and the RF→AF pipeline selected 715 de novo candidate PPIs at an AF contact-probability cutoff of 0.67 with expected precision of 95%.
  • Adding AF analysis of pooled experimental PPI sets yielded 1,251 additional candidates, 461 overlapping with the de novo set, for a total of 1,505 PPIs.
  • In benchmarking against 699 pairs with experimental PDB structures, 92% had at least 50% of confident AF-predicted contacts present in the experimental structures.

Methods

  • Orthologous groups were built from OrthoDB plus 4,325 assembled Eukaryote proteomes (2,026 fungal), using reciprocal best hits; paired multiple sequence alignments were generated for 4,090 yeast proteins across all 8,362,005 pairwise combinations, focusing on 4,286,433 pairs with >200 sequences and <1,300 amino acids.
  • A lightweight RoseTTAFold two-track model (10.7 million parameters) scored all pMSAs, requiring about 11 seconds per 1,000-amino-acid pMSA on an NVIDIA TITAN RTX; AlphaFold was then applied to the top 5,495 RF-supported pairs and to literature/experimental PPI sets, with AF contact probability >0.67 used as the confident-interaction cutoff.
  • For assemblies with three or more mutually interacting proteins, full-complex models were generated from concatenated sequence alignments; models were benchmarked against 699 PDB-characterized pairs.

Limitations

  • The paper states the study is not comprehensive: it surveyed only about 2/3 of the yeast proteome, excluded recent duplicates, likely misses organism-restricted or rapidly evolving interactions, and works less well for transient interactions and intrinsically disordered regions.
  • The paper also states that interactions between single hydrophobic or amphipathic helices may be overpredicted, and that binary complex models can be inaccurate for high-order obligate complexes, as illustrated by the SNARE example.
  • Evident from the text: AlphaFold was too slow to apply to all 4.3 million pMSAs, the decameric GPI-T complex was too large to model, and some reported models are explicitly lower confidence (e.g., Slx8–Cue3); most predictions were not experimentally validated in this study.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1126/science.abm4805

Cites

120 of 120 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 74 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    AlphaFold (AF) was too slow to be applied to the entire set of 4.3 million pMSAs (requiring 0.1-1 million GPU hours).

    “AF was too slow to be applied to the entire set of 4.3 million pMSAs (this would require 0.1-1 million GPU hours); instead we applied AF to the 5,495 protein pairs with the highest RF support (indicated by the black vertical line in Fig. 1B).”
    Quote in the paperSubstantial claim
  • Findingown data

    The RF method considerably outperformed DCA in distinguishing gold standard interactions from random pairs (area under the curve: 0.219), using the highest predicted contact probability over all pairs of residues in the two proteins as a measure of the propensity for two proteins to interact.

    “This method considerably outperformed DCA in distinguishing gold standard interactions from random pairs (Fig. 1B, blue curve, area under the curve: 0.219), using the highest predicted contact probability over all pairs of residues in the two proteins as a measure of the propensity for two proteins to interact (fig. S6).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The 'gold standard' set of known yeast interactions is accurate and suitable for evaluating the accuracy of predicted interactions and structures.
  • Findingown data

    The Direct Coupling Analysis (DCA) method could not effectively distinguish a gold standard set of 768 yeast protein pairs known to interact from a much larger set of 768,000 primarily non-interacting pairs (area under the curve: 0.016).

    “In a first set of calculations, we found that even with the advantages of S. cerevisiae and improved ortholog identification, the statistical method (Direct Coupling Analysis, DCA) we had used in our previous coevolution-guided protein-protein interaction (PPI) screen in Prokaryotes (9) (the more accurate GREMLIN (11) method is too slow for this) could not effectively distinguish a “gold standard””
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [9] Cong Q Anishchenko I Ovchinnikov S Baker D Protein interaction networks revealed by proteome coevolution Science 2019 365 185 189 31296772 PMC6948103 source

      Not checked in the free run.

    • cites [11] Ovchinnikov S Kamisetty H Baker D Robust and accurate prediction of residue-residue interactions across protein interfaces using evolutionary information Elife 2014 3 e02030 24842992 PMC4034769 source

      Not checked in the free run.

    • cites [5] Yu H Braun P Yildirim MA Lemmens I Venkatesan K Sahalie J Hirozane-Kishikawa T Gebreab F Li N Simonis N Hao T High-quality binary protein interaction map of the yeast interactome network Science 2008 322 104 110 18719252 PMC2746753 source

      Not checked in the free run.

    • Rests on an assumption: The 'gold standard' set of known yeast interactions is accurate and suitable for evaluating the accuracy of predicted interactions and structures.
  • Findingcitation

    The predicted Spo11-Ski8 structure is supported by crosslinking and mutagenesis data.

    “The predicted Spo11–Ski8 structure is supported by crosslinking and mutagenesis data (26,27).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [26] Arora C Kee K Maleki S Keeney S Antiviral protein Ski8 is a direct partner of Spo11 in meiotic DNA break formation, independent of its cytoplasmic role in RNA metabolism Mol Cell 2004 13 549 559 14992724 source

      Not checked in the free run.

    • cites [27] Bouuaert CC Tischfield SE Pu S Mimitou EP Arias-Palomo E Berger JM Keeney S Structural and functional characterization of the Spo11 core complex Nat Struct Mol Biol 2021 28 92 102 33398171 PMC7855791 source

      Not checked in the free run.

  • Findingcitation

    The ratio of confidently predicted PPIs is even lower for protein pairs identified by Y2H (18%) or APMS (14%) screens, consistent with the known larger fraction of false positives in large-scale experimental screens.

    “The ratio of confidently predicted PPIs is even lower for protein pairs identified by Y2H (18%) or APMS (14%) screens (table S1), consistent with the known larger fraction of false positives in large-scale experimental screens (8,22).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [8] Mackay JP Sunde M Lowry JA Crossley M Matthews JM Protein interactions: is seeing believing? Trends Biochem Sci 2007 32 530 531 17980603 source

      Not checked in the free run.

    • cites [22] Huang H Jedynak BM Bader JS Where have all the interactions gone? Estimating the coverage of two-hybrid protein interaction maps PLoS Comput Biol 2007 3 e214 18039026 PMC2082503 source

      Not checked in the free run.

  • May combine two claims. 34
  • Classified as probably not a substantial claim. 8

  • May combine two claims.

    Findingreasoning

    They focused on 4,286,433 pairs with alignments containing over 200 sequences and less than 1,300 amino acids.

    “We focused on 4,286,433 pairs with alignments containing over 200 sequences to increase prediction accuracy and less than 1,300 amino acids to accelerate computation (fig. S3).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    A lighter-weight (10.7 million parameters) RoseTTAFold (RF) two-track model requires 11 seconds (about 100 times faster than AlphaFold) to process a pMSA of 1,000 amino acids on a NVIDIA TITAN RTX GPU and can effectively distinguish gold standard PPIs amongst much larger sets of randomly paired proteins.

    “We found that a lighter-weight (10.7 million parameters) RF two-track model (figs. S4, S5) provided a good tradeoff between compute time and accuracy: the model requires 11 seconds (about 100 times faster than AF) to process a pMSA of 1,000 amino acids on a NVIDIA TITAN RTX graphic processing unit, and it can effectively distinguish gold standard PPIs amongst much larger sets of randomly paired”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The 'gold standard' set of known yeast interactions is accurate and suitable for evaluating the accuracy of predicted interactions and structures.

  • May combine two claims.

    Findingown data

    The models do miss many contacts observed in the experimental structures, likely due to lower residue-residue co-evolution.

    “The models do miss many contacts observed in the experimental structures however, likely due to lower residue-residue co-evolution (fig. S15).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Out of the 1,505 PPIs, 699 have been structurally characterized, 700 have some supporting experimental data from literature and databases, and 106 are not to our knowledge previously described.

    “Out of these, 699 have been structurally characterized, 700 have some supporting experimental data from literature and databases, and 106 are not to our knowledge previously described.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The combination of RF followed by AF provided excellent performance, with almost all gold standard pairs ranked higher than negative controls, allowing selection of 715 candidate PPIs with an expected precision of 95% at an AF contact probability cutoff of 0.67.

    “Using the highest AF contact probability over all residue pairs as a measure of interaction strength, we found that the combination of RF followed by AF provided excellent performance (Fig. 1C and figs. S9, S11). Almost all the gold standard pairs were ranked higher than the negative controls, allowing selection of a set of 715 candidate PPIs with an expected precision of 95% at an AF contact”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The 'gold standard' set of known yeast interactions is accurate and suitable for evaluating the accuracy of predicted interactions and structures.

  • May combine two claims.

    Findingown data

    In total, 715 likely interacting pairs were identified from the 'de novo RF → AF' screen, and 1,251 from the 'pooled experimental sets → AF' screen, of which 461 overlap, resulting in a total of 1,505 PPIs.

    “In total, we identified 715 likely interacting pairs from the “de novo RF → AF” screen, and 1,251 from the “pooled experimental sets → AF” screen, of which 461 overlap, resulting in a total of 1,505 PPIs (see figs. S11-S13 for interface size and secondary structure distributions for the predicted complex structures).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    47% of the gold standard PPIs were confidently predicted (AF contact probability > 0.67), with lower ratios (31% and 24%) for candidate PPIs from the literature or supported by low-throughput experiments according to BIOGRID.

    “We found that 47% of the gold standard PPIs were confidently predicted, with lower ratios (31% and 24%) for candidate PPIs from the literature (http://interactome.dfci.harvard.edu/S_cerevisiae/download/LC_multiple.txt) (3) or supported by low-throughput experiments according to BIOGRID (21) (Fig. 1D).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [3] Reguly T Breitkreutz A Boucher L Breitkreutz B-J Hon GC Myers CL Parsons A Friesen H Oughtred R Tong A Stark C Comprehensive curation and analysis of global interaction networks in Saccharomyces cerevisiae J Biol 2006 5 11 16762047 PMC1561585 source

      Not checked in the free run.

    • cites [21] Oughtred R Rust J Chang C Breitkreutz B-J Stark C Willems A Boucher L Leung G Kolas N Zhang F Dolma S The BioGRID database: A comprehensive biomedical resource of curated protein, genetic, and chemical interactions Protein Sci 2021 30 187 200 33070389 PMC7737760 source

      Not checked in the free run.

    • Rests on an assumption: The 'gold standard' set of known yeast interactions is accurate and suitable for evaluating the accuracy of predicted interactions and structures.

  • May combine two claims.

    Findingown data

    The model of Spo11-Ski8 resembles a previous model based on the Ski3-Ski8 complex, with Ski8 contacting a sequence in Ski3 similar to QREIF380 in Spo11, but suggests a more extensive interaction surface involving an insertion in Ski8 present in Saccharomyces species but not in Schizosaccharomyces pombe and Sordaria macrospora.

    “Our model resembles a previous model based on the Ski3–Ski8 complex, with Ski8 contacting a sequence in Ski3 that is similar to the sequence QREIF380 in Spo11 (27,28) (fig. S17A), but suggests a more extensive interaction surface than previously appreciated, involving an insertion in Ski8 that is present in Saccharomyces species but not in Schizosaccharomyces pombe and Sordaria macrospora, where”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [27] Bouuaert CC Tischfield SE Pu S Mimitou EP Arias-Palomo E Berger JM Keeney S Structural and functional characterization of the Spo11 core complex Nat Struct Mol Biol 2021 28 92 102 33398171 PMC7855791 source

      Not checked in the free run.

    • cites [28] Halbach F Reichelt P Rode M Conti E The yeast ski complex: crystal structure and RNA channeling to the exosome complex Cell 2013 154 814 826 23953113 source

      Not checked in the free run.

    • cites [29] Steiner S Kohli J Ludin K Functional interactions among members of the meiotic initiation complex in fission yeast Curr Genet 2010 56 237 249 20364342 source

      Not checked in the free run.

    • cites [30] Tessé S Storlazzi A Kleckner N Gargano S Zickler D Localization and roles of Ski8p protein in Sordaria meiosis and delineation of three mechanistically distinct steps of meiotic homolog juxtaposition Proc Natl Acad Sci U S A 2003 100 12865 12870 14563920 PMC240710 source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    The long central helix observed in the Centrin2 (Rad33) structure is kinked about 90° in the Rad33-Rad14 complex model; both conformations are feasible and compatible for the interaction with Rad14.

    “The long central helix observed in the Centrin2 (Rad33) structure (40) is kinked about 90° in our Rad33-Rad14 complex model (fig. S19B); both conformations are feasible and are compatible for the interaction with Rad14.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [40] Thompson JR Ryan ZC Salisbury JL Kumar R The Structure of the Human Centrin 2-Xeroderma Pigmentosum Group C Protein Complex Journal of Biological Chemistry 2006 281 18746 18752 16627479 source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    The model predicts a different angle for the long helical arm of Rec102 and contains a kink corresponding to a conserved sequence motif EYPMVF192 in Saccharomyces that is missing in both archaeal TopoVI and mammals.

    “Both Rec102 and Top6B have long, helical arms that feed into the Spo11 interface; our model predicts a different angle for this arm and contains a kink that corresponds to a conserved sequence motif EYPMVF192 in Saccharomyces that is missing in both archaeal TopoVI and mammals (fig. S17D, E).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In most cases, pairwise interactions in models of larger assemblies are quite similar to those for independently built binary complexes, but simultaneous modeling of the full complex allows conformational changes that could accompany full assembly.

    “in most cases the pairwise interactions are quite similar to those for the independently built binary complexes, but simultaneous modeling of the full complex has the advantage of allowing conformational changes that could accompany full assembly.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    The model of the entire GARP complex shows the four subunits assemble through a four-helix bundle, with C-terminal 'CATCHR' folds in Vps52, Vps53, and Vps54 emanating from the bundle; this architecture resembles portions of the cryo-EM structure of the Exocyst complex.

    “In this model, the four subunits assemble through a four-helix bundle. In each of the three larger subunits, Vps52, Vps53, and Vps54, C-terminal domains comprising “CATCHR” folds emanate from the bundle. This architecture resembles portions of the cryo-EM structure of the Exocyst complex, a distinct MTC that mediates fusion of vesicles at the plasma membrane (76), which possesses two separate four”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [76] Mei K Li Y Wang S Shao G Wang J Ding Y Luo G Yue P Liu J-J Wang X Dong M-Q Cryo-EM structure of the exocyst complex Nat Struct Mol Biol 2018 25 139 146 29335562 PMC5971111 source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Previously unpublished mutations in Vps24 that prevent ESCRT function in multivesicular body sorting are located on the predicted interface between Vps2 and Vps24, supporting the model and the functional importance of the Vps2-Vps24 interaction.

    “Notably, previously unpublished mutations (fig. S23) in Vps24 that prevent ESCRT function in multivesicular body sorting are located on the predicted interface between Vps2 and Vps24, supporting our model and the functional importance of the Vps2–Vps24 interaction.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Structure models of nine binary complexes (Dad2-Ask1, Dad2-Hsk3, Dad2-Spc1, Dad4-Hsk3, Dam1-Duo1, Duo1-Dad1, Spc19-Dad1, Spc34-Duo1, and Spc34-Spc19) encompassing several members of Dam1c are largely consistent with the Dam1c structure, suggesting that the findings from the thermophile structure can likely be extended to S. cerevisiae.

    “We generated structure models of nine binary complexes (Dad2-Ask1, Dad2-Hsk3, Dad2-Spc1, Dad4-Hsk3, Dam1-Duo1, Duo1-Dad1, Spc19-Dad1, Spc34-Duo1, and Spc34-Spc19) that encompass several members of Dam1c (fig. S22). These complexes are largely consistent with the Dam1c structure, suggesting that the findings from the thermophile structure can likely be extended to S. cerevisiae.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    The model of the Sec23-Grh1 complex contains an interface between the Sec23 gelsolin domain and the PPP motif of Grh1, and an interface between the Grh1 IDR and Sec23 involving a disorder-to-helical transition.

    “Our model of the Sec23–Grh1 complex contains an interface between the Sec23 gelsolin domain and the PPP motif of Grh1 (80), and an interface between the Grh1 IDR and Sec23 involving a disorder-to-helical transition (fig. S25C).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [80] Ma W Goldberg J TANGO1/cTAGE5 receptor as a polyvalent template for assembly of large COPII coats Proc Natl Acad Sci U S A 2016 113 10061 10066 27551091 PMC5018777 source

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    The predicted interface between the Vps2 and Vps24 subunits of the ESCRT-III complex resembles the polymerization interface of a different ESCRT-III subunit Snf7, providing insight into the roles of these previously uncharacterized ESCRT-III subunits and highlighting the generality of this mode of interaction in ESCRT-III complexes.

    “Our predicted interface between the Vps2 and Vps24 subunits of the ESCRT-III complex resembles the polymerization interface of a different ESCRT-III subunit Snf7 (73), providing insight into the roles of these previously uncharacterized ESCRT-III subunits, and highlighting the generality of this mode of interaction in ESCRT-III complexes.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [73] Tang S Henne WM Borbat PP Buchkovich NJ Freed JH Mao Y Fromme JC Emr SD Structural basis for activation, assembly and membrane binding of ESCRT-III Snf7 filaments Elife 2015 4 10.7554/eLife.12548 PMC4720517 26670543 source

      Not checked in the free run.

  • May combine two claims.

    Findingreasoning

    The pentameric GPI-T complex presented leaves open the dimerization face of Gpi8, consistent with probable dimerization.

    “the pentameric complex we present here leaves open the dimerization face of Gpi8, consistent with probable dimerization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    The majority of known interactions identified by the approach are likely obligate assemblies and involve ordered structural elements.

    “The majority of known interactions identified by our approach are likely obligate assemblies and involve ordered structural elements.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In the model, the catalytic dyad in Gpi8 (Cys199 and His157) faces the transmembrane domains of Gaa1 and Gab1 and abuts a highly conserved face of Gaa1.

    “The catalytic dyad in Gpi8 (Cys199 and His157) faces these transmembrane domains, and abuts against a highly conserved face of Gaa1”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Y2H, APMS, and other high-throughput experimental approaches have identified many pairs of interacting proteins in yeast and other organisms, but there are discrepancies between sets generated using different methods and considerable false positive and false negative rates.

    “Yeast two hybrid (Y2H), affinity-purification mass spectrometry (APMS), and other high-throughput experimental approaches have identified many pairs of interacting proteins in yeast and other organisms (1–5), but there are discrepancies between sets generated using the different methods and considerable false positive and false negative rates (6–8).”
    Quote in the paperSubstantial claim0 of 8 sources checked
    • cites [1] Ito T Chiba T Ozawa R Yoshida M Hattori M Sakaki Y A comprehensive two-hybrid analysis to explore the yeast protein interactome Proc Natl Acad Sci U S A 2001 98 4569 4574 11283351 PMC31875 source

      Not checked in the free run.

    • cites [2] Collins SR Kemmeren P Zhao X-C Greenblatt JF Spencer F Holstege FCP Weissman JS Krogan NJ Toward a comprehensive atlas of the physical interactome of Saccharomyces cerevisiae Mol Cell Proteomics 2007 6 439 450 17200106 source

      Not checked in the free run.

    • cites [3] Reguly T Breitkreutz A Boucher L Breitkreutz B-J Hon GC Myers CL Parsons A Friesen H Oughtred R Tong A Stark C Comprehensive curation and analysis of global interaction networks in Saccharomyces cerevisiae J Biol 2006 5 11 16762047 PMC1561585 source

      Not checked in the free run.

    • cites [4] Uetz P Giot L Cagney G Mansfield TA Judson RS Knight JR Lockshon D Narayan V Srinivasan M Pochart P Qureshi-Emili A A comprehensive analysis of protein-protein interactions in Saccharomyces cerevisiae Nature 2000 403 623 627 10688190 source

      Not checked in the free run.

    • cites [5] Yu H Braun P Yildirim MA Lemmens I Venkatesan K Sahalie J Hirozane-Kishikawa T Gebreab F Li N Simonis N Hao T High-quality binary protein interaction map of the yeast interactome network Science 2008 322 104 110 18719252 PMC2746753 source

      Not checked in the free run.

    • cites [6] Kuchaiev O Rasajski M Higham DJ Przulj N Geometric de-noising of protein-protein interaction networks PLoS Comput Biol 2009 5 e1000454 19662157 PMC2711306 source

      Not checked in the free run.

    • cites [7] Edwards AM Kus B Jansen R Greenbaum D Greenblatt J Gerstein M Bridging structural biology and genomics: assessing protein interaction data with known complexes Trends Genet 2002 18 529 536 12350343 source

      Not checked in the free run.

    • cites [8] Mackay JP Sunde M Lowry JA Crossley M Matthews JM Protein interactions: is seeing believing? Trends Biochem Sci 2007 32 530 531 17980603 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The much better performance of RF than DCA likely stems from the extensive information on protein sequence-structure relationships embedded in the RF deep neural network; DCA operates solely on protein sequences with no underlying protein structure model.

    “The much better performance of RF than DCA likely stems from the extensive information on protein sequence-structure relationships embedded in the RF deep neural network; DCA by contrast operates solely on protein sequences with no underlying protein structure model.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Yeast Saccharomyces cerevisiae has a large number of fungal genomes, a relatively small genome (6,000 genes total), and relatively little mRNA splicing.

    “To mitigate the first three challenges, we chose to predict protein complexes for yeast Saccharomyces cerevisiae as the starting point because there are a large number of fungal genomes (16), the genome is relatively small (6,000 genes total), and there is relatively little mRNA splicing (17).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [16] Grigoriev IV Nikitin R Haridas S Kuo A Ohm R Otillar R Riley R Salamov A Zhao X Korzeniewski F Smirnova T MycoCosm portal: gearing up for 1000 fungal genomes Nucleic Acids Res 2014 42 D699-704 24297253 PMC3965089 source

      Not checked in the free run.

    • cites [17] Spingola M Grate L Haussler D Ares M Jr Genome-wide bioinformatic and molecular analysis of introns in Saccharomyces cerevisiae RNA 1999 5 221 234 10024174 PMC1369754 source

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning

    The interactome of yeast has been extensively studied, and there is a 'gold standard' set of known interactions to evaluate the accuracy of predicted interactions and structures.

    “Furthermore, because the interactome of yeast has been extensively studied, there is a “gold standard” set (see supplemental Methods) of known interactions to evaluate the accuracy of predicted interactions and structures.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites supplemental Methods

      Not checked in the free run.

    • Rests on an assumption: The 'gold standard' set of known yeast interactions is accurate and suitable for evaluating the accuracy of predicted interactions and structures.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Amino acid coevolution has been used to guide modeling of complexes for cases in which the structures of the partners are known, and to systematically identify pairs of interacting proteins in Prokaryotes with accuracy higher than experimental screens.

    “Such amino acid coevolution has been used to guide modeling of complexes for cases in which the structures of the partners are known (11,12), and to systematically identify pairs of interacting proteins in Prokaryotes with accuracy higher than experimental screens (9).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [11] Ovchinnikov S Kamisetty H Baker D Robust and accurate prediction of residue-residue interactions across protein interfaces using evolutionary information Elife 2014 3 e02030 24842992 PMC4034769 source

      Not checked in the free run.

    • cites [12] Hopf TA Schärfe CPI Rodrigues JPGLM Green AG Kohlbacher O Sander C Bonvin AMJJ Marks DS Sequence co-evolution gives 3D contacts and structures of protein complexes Elife 2014 3 10.7554/eLife.03430 PMC4360534 25255213 source

      Not checked in the free run.

    • cites [9] Cong Q Anishchenko I Ovchinnikov S Baker D Protein interaction networks revealed by proteome coevolution Science 2019 365 185 189 31296772 PMC6948103 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The model of the complex between full length Rad14 and Rad33 resolves much of the current structural ambiguity in this system, shedding light on how Rad14 may be recruited to the Rad4-Rad23-Rad33 complex.

    “We generated a model of the complex between full length Rad14 and Rad33 that resolves much of the current structural ambiguity in this system (Fig. 2 and fig S19B), shedding light on how Rad14 may be recruited to the Rad4–Rad23–Rad33 complex.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The model of the Siz2 and Mms21 SUMO ligase complex suggests that both E3s could act jointly to modify DNA associated substrates perhaps through the DNA binding SAP domain of Siz2 or involving the Mms21 (Nse2) containing Smc5-6 complex.

    “Our model of the Siz2 and Mms21 SUMO ligase complex (fig. S21A) suggests that both E3s could act jointly to modify DNA associated substrates perhaps through the DNA binding SAP domain of Siz2 (53) or involving the Mms21 (Nse2) containing Smc5–6 complex which modulates DNA recombination, replication and repair (54,55).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [53] Psakhye I Jentsch S Protein group modification and synergy in the SUMO pathway as exemplified in DNA repair Cell 2012 151 807 820 23122649 source

      Not checked in the free run.

    • cites [54] Menolfi D Delamarre A Lengronne A Pasero P Branzei D Essential Roles of the Smc5/6 Complex in Replication through Natural Pausing Sites and Endogenous DNA Damage Tolerance Mol Cell 2015 60 835 846 26698660 PMC4691243 source

      Not checked in the free run.

    • cites [55] Agashe S Joseph CR Reyes TAC Menolfi D Giannattasio M Waizenegger A Szakal B Branzei D Smc5/6 functions with Sgs1-Top3-Rmi1 to complete chromosome replication at natural pause sites Nat Commun 2021 12 2111 33833229 PMC8032827 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation

    Two complexes, Rpl12B-Rmt2 and Rpl7A-Fpr4, involving enzymes that introduce protein modifications such as arginine methylations or proline isomerizations, provide insight into mechanisms that expand the chemical diversity of ribosomal proteins at functional sites and possibly regulate translation.

    “Two complexes, Rpl12B–Rmt2 and Rpl7A–Fpr4, involving enzymes that introduce protein modifications such as arginine methylations or proline isomerizations (47), provide insight into mechanisms that expand the chemical diversity of ribosomal proteins at functional sites (48) and possibly regulate translation (49).”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [47] Mulvaney KM Blomquist C Acharya N Li R Ranaghan MJ O’Keefe M Rodriguez DJ Young MJ Kesar D Pal D Stokes M Molecular basis for substrate recruitment to the PRMT5 methylosome Mol Cell 2021 81 3481 3495 e7 34358446 PMC9016627 source

      Not checked in the free run.

    • cites [48] Watson ZL Ward FR Méheust R Ad O Schepartz A Banfield JF Cate JH Structure of the bacterial ribosome at 2 Å resolution Elife 2020 9 10.7554/eLife.60482 PMC7550191 32924932 source

      Not checked in the free run.

    • cites [49] Małecki JM Odonohue M-F Kim Y Jakobsson ME Gessa L Pinto R Wu J Davydova E Moen A Olsen JV Thiede B Human METTL18 is a histidine-specific methyltransferase that targets RPL3 and affects ribosome biogenesis and function Nucleic Acids Res 2021 49 3185 3203 33693809 PMC8034639 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The predicted structures of complexes between Ksh1 and Yos1 and Yip1 suggest Ksh1 is a fourth member of this enigmatic complex essential to the secretory pathway, and explain how Ksh1 can play a role in secretion despite its small size of 72 amino acids.

    “These structures suggest Ksh1 is a fourth member of this enigmatic complex essential to the secretory pathway, and explains how Ksh1 can play a role in secretion despite its small size of 72 amino acids.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    SNARE proteins drive intracellular membrane fusion between transport vesicles and organelles.

    “SNARE proteins drive intracellular membrane fusion between transport vesicles and organelles (82).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [82] Südhof TC Rothman JE Membrane fusion: grappling with SNARE and SM proteins Science 2009 323 474 477 19164740 PMC3736821 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The model suggests that the combinatorial multivalent interaction between Grh1 and Sec23 may compete with the interaction between Sec31 and Sec23 to promote vesicle uncoating; consistent with this model, Grh1 is recruited to GST-Sec23, dependent on the IDR, and competes for Sec31 binding.

    “Our model suggests that the combinatorial multivalent interaction between Grh1 and Sec23 may compete with the interaction between Sec31 and Sec23 to promote vesicle uncoating; consistent with this model, Grh1 is recruited to GST-Sec23, dependent on the IDR, and competes for Sec31 binding (fig. S25D).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The model suggests that Rad14 and Rad4 can be present at the same time in the repair cascade; crosstalk between these important proteins could modulate downstream events.

    “Our model suggests that Rad14 and Rad4 can be present at the same time in the repair cascade; crosstalk between these important proteins could modulate downstream events.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Membrane fusion requires formation of a 4-helix bundle (the SNARE complex) between the vesicle SNARE and the target membrane SNAREs.

    “Membrane fusion requires the formation of a 4-helix bundle (called the SNARE complex) between the vesicle SNARE and the target membrane SNAREs (84,85).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [84] Sutton RB Fasshauer D Jahn R Brunger AT Crystal structure of a SNARE complex involved in synaptic exocytosis at 2.4 A resolution Nature 1998 395 347 353 9759724 source

      Not checked in the free run.

    • cites [85] Jahn R Scheller RH SNAREs--engines for membrane fusion Nat Rev Mol Cell Biol 2006 7 631 643 16912714 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    GPI-T catalyzes attachment of GPI anchors to the C-terminus of specific substrate proteins based on recognition of a C-terminal signal peptide.

    “which catalyzes the attachment of GPI anchors to the C-terminus of specific substrate proteins, based on recognition of a C-terminal signal peptide (91).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [91] Moran P Caras IW A nonfunctional sequence converted to a signal for glycophosphatidylinositol membrane anchor attachment J Cell Biol 1991 115 329 336 1717483 PMC2289147 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Gaa1 is essential for binding of the GPI anchor to GPI-T.

    “Gaa1 is essential for binding of the GPI anchor to GPI-T (95)”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [95] Vainauskas S Menon AK A conserved proline in the last transmembrane segment of Gaa1 is required for glycosylphosphatidylinositol (GPI) recognition by GPI transamidase J Biol Chem 2004 279 6540 6545 14660601 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The disulfide bond between Gpi8 and Gpi16 was previously characterized for human GPI-T.

    “previously characterized for human GPI-T (94)”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [94] Ohishi K Nagamune K Maeda Y Kinoshita T Two subunits of glycosylphosphatidylinositol transamidase, GPI8 and PIG-T, form a functionally important intermolecular disulfide bridge J Biol Chem 2003 278 13959 13967 12582175 source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    GPI-T is a pentameric enzyme complex of unknown structure.

    “Glycosylphosphatidylinositol transamidase (GPI-T) is a pentameric enzyme complex of unknown structure (88–90)”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [88] Hong Y Ohishi K Kang JY Tanaka S Inoue N Nishimura J-I Maeda Y Kinoshita T Human PIG-U and yeast Cdc91p are the fifth subunit of GPI transamidase that attaches GPI-anchors to proteins Mol Biol Cell 2003 14 1780 1789 12802054 PMC165076 source

      Not checked in the free run.

    • cites [89] Gamage DG Hendrickson TL GPI transamidase and GPI anchored proteins: oncogenes and biomarkers for cancer Crit Rev Biochem Mol Biol 2013 48 446 464 23978072 source

      Not checked in the free run.

    • cites [90] Yi L Bozkurt G Li Q Lo S Menon AK Wu H Disulfide Bond Formation and N-Glycosylation Modulate Protein-Protein Interactions in GPI-Transamidase (GPIT) Sci Rep 2017 8 45912 28374821 PMC5379210 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    In vivo, GPI-T is expected to be a dimer of pentamers, with dimerization occurring on one face of the caspase-like Gpi8 subunit.

    “In vivo, GPI-T is expected to be a dimer of pentamers, with dimerization occurring on one face of the caspase-like Gpi8 subunit (92,97,98).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [92] Fraering P Imhof I Meyer U Strub JM van Dorsselaer A Vionnet C Conzelmann A The GPI transamidase complex of Saccharomyces cerevisiae contains Gaa1p, Gpi8p, and Gpi16p Mol Biol Cell 2001 12 3295 3306 11598210 PMC60174 source

      Not checked in the free run.

    • cites [97] Gamage DG Varma Y Meitzler JL Morissette R Ness TJ Hendrickson TL The soluble domains of Gpi8 and Gaa 1 two subunits of glycosylphosphatidylinositol transamidase (GPI-T), assemble into a complex Arch Biochem Biophys 2017 633 58 67 28893510 source

      Not checked in the free run.

    • cites [98] Meitzler JL Gray JJ Hendrickson TL Truncation of the caspase-related subunit (Gpi8p) of Saccharomyces cerevisiae GPI transamidase: dimerization revealed Arch Biochem Biophys 2007 462 83 93 17475206 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationcitation

    Training RF and AF on protein complexes should further improve performance of both methods, particularly for protein pairs with fewer homologues and/or weaker and more transient interactions, and reduce dependence on ortholog identification.

    “Training RF and AF on protein complexes should further improve performance of both methods (100), particularly for protein pairs with fewer homologues and/or weaker and more transient interactions, and reduce the dependence on ortholog identification.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [100] Evans R O’Neill M Pritzel A Antropova N Senior A Green T Žídek A Bates R Blackwell S Yim J Ronneberger O Protein complex prediction with AlphaFold-Multimer bioRxiv 2021 2021.10.04.463034

      Not checked in the free run.

  • May combine two claims.

    Study design and conductown data

    The study is not comprehensive: it eliminated proteins from recent duplication and surveyed only 2/3 of the yeast proteome.

    “First, our study is not comprehensive, so conclusions should not be drawn about absences; in particular we eliminated proteins that arose from recent duplication due to difficulty in identifying orthologs in other organisms, and thus only surveyed 2/3 of the entire yeast proteome.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Gpi16 is immediately adjacent to these interactions and is likely also involved in anchor recognition.

    “Gpi16 is immediately adjacent to these interactions and is likely to also be involved in anchor recognition.”
    Quote in the paperNot a substantial claim
40 of 40

  • The paper relies on this without support.

    Far fewer genome sequences are available for eukaryotes than prokaryotes, and the average number of orthologous sequences is on the order of 10,000 for bacterial proteins but 1,000 for eukaryotic proteins.

    “First, far fewer genome sequences are available for eukaryotes than prokaryotes, and the average number of orthologous sequences (excluding nearly identical copies with > 95% sequence identity) is on the order of 10,000 for bacterial proteins, but 1,000 for eukaryotic proteins.”

  • The paper relies on this without support.

    Eukaryotes in general have a larger number of genes, making comprehensive pairwise analysis more computationally intensive and increasing background noise.

    “Second, eukaryotes in general have a larger number of genes, making comprehensive pairwise analysis more computationally intensive, and increasing the background noise.”

  • The paper relies on this without support.

    mRNA splicing in eukaryotes further increases the number of protein species, resulting in errors in gene predictions and complicating sequence alignments.

    “Third, mRNA splicing in eukaryotes further increases the number of protein species, resulting in errors in gene predictions and complicating sequence alignments.”

  • The paper relies on this without support.

    The 'gold standard' set of known yeast interactions is accurate and suitable for evaluating the accuracy of predicted interactions and structures.

    “there is a “gold standard” set (see supplemental Methods) of known interactions to evaluate the accuracy of predicted interactions and structures.”
4 of 4

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Combining standard protein-protein docking methods with co-evolutionary signals or deep learning models has been carried out on full proteomes.

    “Combining standard protein-protein docking methods with co-evolutionary signals28 or even with deep learning models29 has also been carried out on full proteomes.”

    From AF2Complex predicts direct physical interactions in multimeric proteins with deep learning (2022) · cites this paper as [29]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Rad55-Rad57 is called a homodimer

    Minor

    A homodimer contains two copies of the same protein, whereas the paper describes the complex as containing the distinct proteins Rad55 and Rad57.

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. M. D. Nichols K. A. DeAngelis J. L. Keck J. M. Berger Structure of the DNA topoisomerase vi a subunit (1999). doi:10.2210/pdb1d3y/pdb · DOI

Cited by

Sorted by citations

Related

Authors
Humphreys, I. R., Pei, J., Baek, M., Krishnakumar, A., Anishchenko, I., Ovchinnikov, S., Zhang, J., Ness, T. J., Banjade, S., Bagde, S. R., Stancheva, V. G., Li, X.-H., Liu, K., Zheng, Z., Barrero, D. J., Roy, U., Kuper, J., Fernández, I. S., Szakal, B., Branzei, D., Rizo, J., Kisker, C., Greene, E. C., Biggins, S., Keeney, S., Miller, E. A., Fromme, J. C., Hendrickson, T. L., Cong, Q., & Baker, D.
Affiliations
  • University of Washington (US)
  • The University of Texas Southwestern Medical Center (US)
  • Harvard University (US)
  • Wayne State University (US)
  • Cornell University (US)
  • MRC Laboratory of Molecular Biology (GB)
  • MRC Laboratory of Molecular Biology, Cambridge CB2 0QH, UK.
  • Memorial Sloan Kettering Cancer Center (US)
  • Howard Hughes Medical Institute (US)
  • Columbia University (US)
  • University of Würzburg (DE)
  • St. Jude Children's Research Hospital (US)
  • IFOM, the FIRC Institute of Molecular Oncology, Via Adamello 16, 20139, Milan, Italy.
  • IFOM (IT)
Corresponding author
  • Humphreys, Ian R.
  • Pei, Jimin
  • Baek, Minkyung
  • Krishnakumar, Aditya
  • Cong, Qian
  • Baker, David
ORCID
Venue
Science
Publisher
American Association for the Advancement of Science (AAAS)
Published
2021-12-10
Volume
374
Issue
6573
Pages
eabm4805-eabm4805
ISSN
  • 0036-8075 (ISSN-L)
  • 1095-9203
DOI
10.1126/science.abm4805
PMC
PMC7612107
OpenAlex
W3212912034
Semantic Scholar
7095b3ee1447af4f429313b912031bd81689c623
Cited by
542
References
120
Citations by year
2026: 59 · 2025: 82 · 2024: 124 · 2023: 117 · 2022: 146 · 2021: 12 · 2013: 1 · 2012: 1
FWCI
31.36
Type
journal article
Language
en
Open access
Locations
  • Science (publishedVersion)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
  • Discovery Research Portal (University of Dundee) (submittedVersion)
  • Universitätsbibliographie, Universität Duisburg-Essen (submittedVersion)
  • eScholarship (California Digital Library) (submittedVersion)
  • Seoul National University Open Repository (Seoul National University) (submittedVersion)
Funders
  • Howard Hughes Medical Institute
  • Damon Runyon Cancer Research Foundation
  • Washington Research Foundation
  • Southwestern Medical Foundation
  • Associazione Italiana per la Ricerca sul Cancro
  • National Institutes of Health
  • Medical Research Council (MRC_UP_1201/10)
  • National Cancer Institute (R01CA221858, P30CA008748)
  • National Institute of General Medical Sciences (R35GM136258, R35GM118026)
  • National Institute of Allergy and Infectious Diseases (R21AI156595)
  • National Institute of Neurological Disorders and Stroke (R35NS097333)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Bioinformatics and Genomic Networks
Keywords
protein-protein interaction, protein complexes, deep learning, coevolution, protein structure prediction, Saccharomyces cerevisiae
MeSH terms
Deep Learning, Acyltransferases, Computer Simulation, DNA Repair, Ligases, Membrane Proteins, Models, Molecular, Protein Conformation, Ribosomes, Saccharomyces cerevisiae, Protein Biosynthesis, Evolution, Molecular, Computational Biology, Chromosome Segregation, Proteome, Ubiquitin, Protein Interaction Mapping, Saccharomyces cerevisiae Proteins, Multiprotein Complexes, Homologous Recombination, Protein Interaction Maps

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review