CoChat
Sign Up
Cited by 252Protein Structure and DynamicsMachine Learning in BioinformaticsEnzyme Structure and Function

AF2Complex predicts direct physical interactions in multimeric proteins with deep learning

Open access
Gao, Mu, Nakajima An, Davi, Parks, Jerry M., Skolnick, Jeffrey·2022·Nature Communications
DOI 10.1038/s41467-022-29394-2PMC PMC8975832
Open this paper in CoChat — free accountOpen PDF

Abstract

Accurate descriptions of protein-protein interactions are essential for understanding biological systems. Remarkably accurate atomic structures have been recently computed for individual proteins by AlphaFold2 (AF2). Here, we demonstrate that the same neural network models from AF2 developed for single protein sequences can be adapted to predict the structures of multimeric protein complexes without retraining. In contrast to common approaches, our method, AF2Complex, does not require paired multiple sequence alignments. It achieves higher accuracy than some complex protein-protein docking strategies and provides a significant improvement over AF-Multimer, a development of AlphaFold for multimeric proteins. Moreover, we introduce metrics for predicting direct protein-protein interactions between arbitrary protein pairs and validate AF2Complex on some challenging benchmark sets and the E. coli proteome. Lastly, using the cytochrome c biogenesis system I as an example, we present high-confidence models of three sought-after assemblies formed by eight members of this system.

AI snapshot

Written from the paper’s full text.

TL;DR

AF2Complex adapts AlphaFold2 deep learning models, without retraining and without paired multiple sequence alignments, to predict multimeric protein complex structures and direct protein-protein interactions. It outperforms docking-centric approaches and improves on AF-Multimer on several benchmarks, and it produces high-confidence models for E. coli cytochrome c biogenesis system I assemblies.

Key findings

  • On the CP17 heterodimer benchmark, AF2Complex top models were acceptable or better in 15 of 17 cases (88%) and medium/high quality in 13 of 17 (76%), with mean DockQ 0.62 versus 0.25 for a docking-centric approach; using AF-Multimer DL models raised the mean to 0.74, and combining monomer and multimer runs gave acceptable or better models for all 17 targets.
  • On 440 heterodimers from the larger Dimer1193 set, AF2Complex performed better than AF-Multimer (median/mean DockQ 0.69/0.56 versus 0.65/0.55, p = 5 × 10−3, Wilcoxon signed-rank test), improving to 0.70/0.57 when top monomer and multimer runs were combined (p = 6 × 10−4); homodimer performance was comparable.
  • For identifying true interacting pairs among all 561 pairwise combinations of the 34 CP17 sequences, interface-score and piTM achieved AUC0.1 values of 0.72 and 0.69, versus 0.49 for pTM and 0.10 for pLDDT, with random guessing at 0.05; precision-recall reached ~45% recall at ~90% precision.
  • In the E. coli proteome test using 701 positive and 6,849 negative pairs, AUC0.1 was 0.60 when each monomer appeared only once (C = 1), dropping to 0.50 at C = 3, 0.40 at C = 5, and 0.22 for the full set; among confident predictions (interface-score > 0.45), about 40% had not been experimentally characterized. High-confidence Ccm models had interface scores of 0.82 for A2B2CDE, 0.56 for EFH, and ~0.72 for FGH.

Methods

  • AF2Complex reuses the AF2 data pipeline, assembles complex MSAs by padding individual monomer MSAs with gaps, and increases the residue_index feature by 200 × (p − 1) to distinguish chains; it uses unpaired MSAs and AF2 monomer DL models (v2.0.1) or AF-Multimer DL models (v2.1.1), with up to 20 recycles.
  • Benchmarks included CASP14 multimeric targets, the CP17 heterodimer set, curated Dimer1193/Oligomer562 sets from the PDB, and E. coli K12 proteome sets of 701 positive and 6,849 randomly selected negative pairs; Ccm system models were built for about two dozen component combinations, with heme placement via Rosetta.
  • Main analyses used interface-score and piTM for complex confidence, DockQ/TM-score/IS-score for structural accuracy, ROC and precision-recall curves with AUC0.1, and paired one-tailed Wilcoxon signed-rank tests.

Limitations

  • The paper states that AF2 does not directly incorporate cofactors or prosthetic groups, so heme molecules were modeled separately with Rosetta; it also notes that the biological context of a hypothetical complex is often unavailable a priori, post-translational modifications such as CcmH signal peptide cleavage are not modeled, and computational limits restricted targets to <3000 residues for CASP14 and <1480 residues for E. coli/benchmark runs.
  • The paper states that reliable statistical analysis of CASP14 assembly predictions was impossible because many experimental structures were unreleased, and that automated mapping of chains for high-order oligomers remains an open issue.
  • Evident from the text: performance depends on MSA depth, as the two CP17 failures each had a monomer with only single-digit MSA depth; the E. coli “gold standard” set likely contains indirect interactions within large assemblies, lowering accuracy as C increases, and the GroL example shows that missing assembly context can prevent confident modeling.

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

Integrity signals

Registry

Crossref · 10.1038/s41467-022-29394-2

Cites

67 of 67 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 60 cited sources checked · 0 supported · 0 flagged

  • Findingnone given

    AF2Complex achieves higher accuracy than some complex protein-protein docking strategies.

    “It achieves higher accuracy than some complex protein-protein docking strategies”
    Quote in the paperSubstantial claim
  • Findingnone given

    AF2Complex provides a significant improvement over AF-Multimer, a development of AlphaFold for multimeric proteins.

    “provides a significant improvement over AF-Multimer, a development of AlphaFold for multimeric proteins.”
    Quote in the paperSubstantial claim
  • Findingnone given

    AF2Complex adapts the same neural network models from AF2 developed for single protein sequences to predict the structures of multimeric protein complexes without retraining.

    “Here, we demonstrate that the same neural network models from AF2 developed for single protein sequences can be adapted to predict the structures of multimeric protein complexes without retraining.”
    Quote in the paperSubstantial claim
  • Findingown data

    AF2Complex produces high-confidence models of three sought-after assemblies formed by eight members of cytochrome c biogenesis system I.

    “using the cytochrome c biogenesis system I as an example, we present high-confidence models of three sought-after assemblies formed by eight members of this system.”
    Quote in the paperSubstantial claim
  • Findingnone given

    AF2Complex does not require paired multiple sequence alignments.

    “In contrast to common approaches, our method, AF2Complex, does not require paired multiple sequence alignments.”
    Quote in the paperSubstantial claim
  • May combine two claims. 41
  • Classified as probably not a substantial claim. 39
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingown data

    Using AF-Multimer DL models, AF2Complex obtained a single model with the expected C12 symmetry using unpaired MSAs but failed to build a physical model with paired MSAs.

    “Using AF-Multimer DL models, we obtained a single model with the expected C12 symmetry using unpaired MSAs but failed to build a physical model with paired MSAs (Supplementary Fig. 1).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    For T1070-D1, AF2Complex's top complex model yields a TM-score of 0.74, which is a significant improvement over 0.62 by AF2 in its official CASP14 assessment.

    “our model for T1070-D1 yields a TM-score of 0.74, which is a significant improvement over 0.62 by AF2 in its official CASP14 assessment.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    AF2Complex's overall top models are acceptable or better in 15 of 17 (88%) CP17 cases, and 13 (76%) models are of medium or high quality according to DockQ, with a mean of 0.62.

    “the overall top models from AF2Complex are acceptable or better in 15 of 17 (88%) cases, and 13 (76%) models are of medium or high quality according to the DockQ score42 (Fig. 3a, Supplementary Data 1), with a mean of 0.62”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [42] Basu, S. & Wallner, B. DockQ: A quality measure for protein-protein docking models. PLoS One 11, e0161879 (2016). source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    AF2Complex mitigates the unphysical model issue with unpaired MSAs using the same multimer DL models and deals much more effectively with unpaired MSAs and the original monomer DL models of AF2.

    “In contrast, AF2Complex mitigates this issue with unpaired MSAs using the same multimer DL models and deals much more effectively with unpaired MSAs and the original monomer DL models of AF2 (Supplementary Fig. 3).”
    Quote in the paperSubstantial claim

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

    Findingown data

    GroL is the most frequent monomer in the E. coli gold standard set, appearing in 79 pairs.

    “the chaperonin protein GroL is the most frequent monomer in this set, appears in 79 pairs.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    pTM is much more discriminating than pLDDT but still much worse than interface-score or piTM.

    “Although the pTM metric is much more discriminating than pLDDT, it is still much worse than interface-score or piTM in this regard.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The normalized AUC0.1 is 0.72 and 0.69 for interface-score and piTM, versus 0.49 and 0.10 for pTM and pLDDT.

    “The normalized area under the curve (AUC) of this plot, AUC0.1, is 0.72 and 0.69 for interface-score and piTM, versus 0.49 and 0.10 for pTM and pLDDT, respectively.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All protein pairs other than the CP17 pairs are non-interacting.
    • Rests on an assumption: Most pairs of proteins are non-interacting.

  • May combine two claims.

    Findingreasoning

    The AUC0.1 drops to 0.50 at C=3, 0.40 at C=5, and 0.22 for the full set, likely due to inclusion of more non-direct interacting pairs as C increases.

    “The AUC0.1 drops to 0.50 at C = 3, 0.40 at C = 5, and 0.22 for the full set, likely due to the inclusion of more non-direct interacting pairs as C increases.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On 440 heterodimers, AF2Complex performs significantly better than AF-Multimer, with median/mean DockQ 0.69/0.56 versus 0.65/0.55, p=5e-3.

    “On 440 heterodimers obtained using the same DL models, AF2Complex performs significantly better than AF-Multimer, albeit at smaller advantage (median/mean 0.69/0.56 versus 0.65/0.55, p value = 5 × 10−3, Wilcoxon signed-rank test, paired, one-tailed, n = 440, see Supplementary Text and Supplementary Fig. 2)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the precision-recall plot, ~45% recall was achieved at ~90% precision, and recall increases to ~70% at ~45% precision.

    “The same trend is also displayed in the precision-recall plot, whereby we achieved ~45% recall (equivalent to the true positive rate) at ~90% precision, and the recall increases to ~70% at ~45% precision.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All protein pairs other than the CP17 pairs are non-interacting.
    • Rests on an assumption: Most pairs of proteins are non-interacting.

  • May combine two claims.

    Findingown data

    The mean score of AF2Complex increases to 0.74 using AF-Multimer DL models, due to unpaired MSAs, increased recycles and the interface-score metric.

    “The mean score of AF2Complex increases to 0.74 using AF-Multimer DL models (Fig. 3d), thanks to unpaired MSAs, increased recycles and the interface-score metric (Supplementary Data 2).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The interface-score and piTM values are 0.55/0.59 and 0.45/0.50, respectively.

    “Correspondingly, the interface-score and piTM values are 0.55/0.59 and 0.45/0.50, respectively.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All protein pairs other than the CP17 pairs are non-interacting.
    • Rests on an assumption: Most pairs of proteins are non-interacting.

  • May combine two claims.

    Findingown data

    In the other A2B2CDE model, the heme-binding loop of CcmE is oriented away from the HBS of CcmCD, interactions between CcmC and E disappear, and the only remaining inter-protein residue-residue contacts are between a pair of transmembrane helices of CcmC and CcmE.

    “The heme-binding loop of CcmE is oriented away from the HBS of CcmCD in the other model (Fig. 6b right panel), in which the interactions between CcmC and E disappear, and the only remaining inter-protein residue-residue contacts are between a pair of transmembrane helices of CcmC and CcmE.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    AF2Complex produced confident models of three assemblies, A2B2CDE, EFH, and FGH, the last one with and without apocyt peptides, all with high interface-score (0.82, 0.56, and ~0.72, respectively).

    “we present the confident models of three assemblies, A2B2CDE, EFH, and FGH, the last one with and without apocyt peptides, all with high interface-score (0.82, 0.56, and ~0.72, respectively).”
    Quote in the paperSubstantial claim

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

    Findingown data

    Two E. coli targets with high-confidence models, CcmE/CcmF and CcmF/CcmH, caught attention.

    “Two E. coli targets with high-confidence models, CcmE/CcmF and CcmF/CcmH, caught our attention.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Using CcmC as reference to superimpose the two complexes, movement is evident in CcmA2B2 but not in CcmD, which appears tightly coupled with CcmC.

    “Using CcmC as the reference to superimpose the two complexes, movement is evident in CcmA2B2 but not in CcmD, which appears tightly coupled with CcmC.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    CcmE interacts with CcmF such that the HBS of CcmF faces the HBS of CcmE, which has a heme-handling motif like that in CcmC.

    “CcmE interacts with CcmF such that the HBS of CcmF faces the HBS of CcmE, which has a heme-handling motif like that in CcmC.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    A pore is visible between transmembrane helices 2 and 5 of CcmC and exposes a heme molecule bound to CcmC.

    “A pore is visible between transmembrane helices 2 and 5 of CcmC and exposes a heme molecule bound to CcmC (Fig. 6d).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The heme-bound model displays the P-heme coordination role of His303CcmF and His173CcmF, and the TM-heme is coordinated by His261CcmF and His461CcmF.

    “the heme-bound model displays the P-heme coordination role of His303CcmF and His173CcmF, and the TM-heme is coordinated by His261CcmF and His461CcmF (Fig. 7b).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The CcmH N-terminal domain moves closer to the heme-binding sites of CcmF, leaving space to accommodate CcmG that now binds CcmF with CcmH.

    “The CcmH N-terminal domain moves closer to the heme-binding sites of CcmF, leaving space to accommodate CcmG that now binds CcmF with CcmH.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the absence of CcmE, the N-terminus of CcmH occupies the site otherwise occupied by CcmE, moving closer to the HBS of CcmF.

    “In the absence of CcmE, the N-terminus of CcmH occupies the site otherwise occupied by CcmE, essentially moving closer to the HBS of CcmF.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Proper modeling of CcmEFH and CcmFHI requires the cleavage of the N-terminal signal peptide of CcmH to obtain biologically accurate models.

    “For instance, proper modeling of CcmEFH and CcmFHI requires the cleavage of the N-terminal signal peptide of CcmH to obtain biologically accurate models.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    The power of a deep learning-based approach for predicting direct protein-protein interactions has been demonstrated.

    “Nevertheless, the power of a deep learning-based approach for predicting direct protein-protein interactions has been demonstrated.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    CcmF in our model does not expose a pore as observed in the TtCcmF structure, in which the pore was speculated to enable P-heme access.

    “CcmF in our model does not expose a pore as observed in the TtCcmF structure, in which the pore was speculated to enable P-heme access49.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [49] Brausemann, A., Zhang, L., Ilcu, L. & Einsle, O. Architecture of the membrane-bound cytochrome c heme lyase CcmF. Nat. Chem. Biol. 17, 800–805 (2021). source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Using AF2Complex without paired MSAs significantly enhances AF-Multimer and returns more physical complex models.

    “which also significantly enhances AF-Multimer and returns more physical complex models.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The two CcmH domains, encoded in two ORFs in some organisms but fused together in E. coli, are linked by a long loop without direct interactions.

    “Interestingly, the two CcmH domains, encoded in two ORFs in some organisms but fused together in E. coli, are linked by a long loop without direct interactions.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The lengths of the apocyt c peptides range from 8 to 52 residues, mostly around 18 AAs.

    “The lengths of these apocyt c peptides range from 8 to 52 residues, mostly around 18 AAs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Findingown data

    Filtering out pairs whose total size is longer than 1,480 residues led to 701 pairs from the positive set.

    “Filtering led to 701 pairs from the positive set.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    For an individual target of fewer than 1000 residues, models may be obtained within 20 min for each deep learning model using "super" mode.

    “For an individual target of fewer than 1000 residues, models may be obtained within 20 min for each deep learning model using “super” mode”
    Quote in the paperNot a substantial claim

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

    Findingown data

    An AF2Complex run of ~7000 pairs of proteins using 923 nodes required about 2 hours in wall clock time.

    “An AF2Complex run of ~7000 pairs of proteins using 923 nodes required about 2 hours in wall clock time.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    The score distributions do not follow normal distributions.

    “because the score distributions do not follow normal distributions.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    For experimental structures, the interface clash indicator χ is typically between 0.6 and 1.2.

    “For experimental structures, χ is typically between 0.6 and 1.2.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    All targets of Dimer1275 have identified four templates in the multimer protocol, but 82 targets have none or fewer than four templates identified in the monomer protocol that AF2Complex relies on for input generation.

    “All targets of Dimer1275 have identified four templates in the multimer protocol, but 82 targets have none or fewer than four templates identified in the monomer protocol that AF2Complex relies on for input generation.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Two key factors contributing to AF2's success are the completeness of the structural space of single-domain proteins and an abundance of sequences in sequence databases.

    “two key factors contributing to the success of AF2 are the completeness of the structural space of single-domain proteins and an abundance of sequences in sequence databases3.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [3] Skolnick, J., Gao, M., Zhou, H. & Singh, S. AlphaFold 2: Why It Works and Its Implications for Understanding the Relationships of Protein Sequence, Structure, and Function. J. Chem. Inf. Model. 61, 4827–4831 (2021). source

      Not checked in the free run.

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

    Premisecitation

    AF2 achieved remarkable performance on individual domains of target protein sequences in CASP14.

    “In CASP14, a blind test for protein structure prediction, AF2 achieved remarkable performance when assessed on the individual domains of target protein sequences1.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [1] Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). source

      Not checked in the free run.

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

    Premisecitation

    Template-based methods for predicting protein complex structures utilize experimentally resolved multimeric structures.

    “A second class of approaches is template-based methods, which utilize experimentally resolved multimeric structures12–16.”
    Quote in the paperNot a substantial claim0 of 5 sources checked
    • cites [12] Aloy, P. et al. Structure-based assembly of protein complexes in yeast. Science 303, 2026–2029 (2004). source

      Not checked in the free run.

    • cites [13] Chen, H. L. & Skolnick, J. M-TASSER: An algorithm for protein quaternary structure prediction. Biophys. J. 94, 918–928 (2008). source

      Not checked in the free run.

    • cites [14] Zhang, Q. C., Petrey, D., Norel, R. & Honig, B. H. Protein interface conservation across structure space. Proc. Natl Acad. Sci. USA 107, 10896–10901 (2010). source

      Not checked in the free run.

    • cites [15] Keskin, O., Nussinov, R. & Gursoy, A. PRISM: protein-protein interaction prediction by structural matching. Methods Mol. Biol. 484, 505–521 (2008). source

      Not checked in the free run.

    • cites [16] Mukherjee, S. & Zhang, Y. Protein-Protein complex structure predictions by multimeric threading and template recombination. Structure 19, 955–966 (2011). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    AF2 delivered excellent results on multidomain proteins and has been applied to such proteins in several model organisms.

    “AF2 also delivered excellent results on multidomain proteins1 and has been applied to such proteins in several model organisms2.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [1] Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). source

      Not checked in the free run.

    • cites [2] Tunyasuvunakool, K. et al. Highly accurate protein structure prediction for the human proteome. Nature 596, 590–596 (2021). source

      Not checked in the free run.

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

    Premisecitation

    Some examples of using a neural network model to predict complex structures were demonstrated with RoseTTAFold, whose design was inspired by AF2.

    “some examples of using such a neural network model to predict complex structures were demonstrated with RoseTTAFold7, whose design was inspired by AF2”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [7] Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373, 871–876 (2021). source

      Not checked in the free run.

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

    Premisecitation

    More recent methods incorporate inter-protein residue-residue contact predictions obtained via co-evolutionary analysis and a deep learning algorithm.

    “More recent methods incorporate inter-protein residue-residue contact predictions obtained via co-evolutionary analysis17 and a deep learning algorithm18.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [17] Szurmant, H. & Weigt, M. Inter-residue, inter-protein and inter-family coevolution: bridging the scales. Curr. Opin. Struct. Biol. 50, 26–32 (2018). source

      Not checked in the free run.

    • cites [18] Zhou, T., Wang, S. & Xu, J. Deep learning reveals many more inter-protein residue-residue contacts than direct coupling analysis. bioRxiv, 240754 (2017). source

      Not checked in the free run.

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

    Premisecitation

    Studies using AF2-generated monomer models with docking methods are based on the idea that AF2 generates high-quality monomeric models that could improve the chance of native-like poses in docking.

    “studies have been carried out whereby models of single proteins generated with AF2 are used with docking methods. They are based on the idea that AF2 generates high-quality monomeric models that could improve the chance of native-like poses in docking22.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] Ghani, U. et al. Improved docking of protein models by a combination of AlphaFold2 and ClusPro. bioRxiv, 2021.2009.2007.459290 (2021). source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    One main consequence of evolution is the diversification of protein functions by producing paralogs.

    “one main consequence of evolution is the diversification of protein functions by producing paralogs30.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [30] Koonin, E. V. Orthologs, paralogs, and evolutionary genomics. Annu. Rev. Genet. 39, 309–338 (2005). source

      Not checked in the free run.

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

    Premisecitation

    An early study adapted AF2 to complex prediction by connecting two protein sequences with a poly-glycine linker.

    “One early such study involved simply connecting two protein sequences with a poly-glycine linker, thus converting it into a single-chain structure prediction problem19.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [19] Ko, J. & Lee, J. Can AlphaFold2 predict protein-peptide complex structures accurately? bioRxiv, 2021.2007.2027.453972 (2021). source

      Not checked in the free run.

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

    Interpretationcitation

    In docking studies using AF2 monomers, the benchmark set includes protein structures used to train AF2 deep learning models, and use of holo monomers compromises rigor because AF2 likely provides an observed holo-structure for docking.

    “One issue with these studies, as some authors pointed out, is that the benchmark set tested includes protein structures used to train the AF2 deep learning models. Although the AF2 models were not trained on protein complex structures, the use of the holo monomers in training compromises rigor because AF2 likely provides an “observed” holo-structure for docking.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] Ghani, U. et al. Improved docking of protein models by a combination of AlphaFold2 and ClusPro. bioRxiv, 2021.2009.2007.459290 (2021). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    High-throughput experimental techniques designed to identify interacting protein partners have results that are far from complete and often mutually inconsistent.

    “Several high-throughput experimental techniques have been designed to identify interacting protein partners23–26, but their results are far from complete and are often mutually inconsistent.”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [23] Uetz, P. et al. A comprehensive analysis of protein-protein interactions in Saccharomyces cerevisiae. Nature 403, 623–627 (2000). source

      Not checked in the free run.

    • cites [24] Butland, G. et al. Interaction network containing conserved and essential protein complexes in Escherichia coli. Nature 433, 531–537 (2005). source

      Not checked in the free run.

    • cites [25] Arifuzzaman, M. et al. Large-scale identification of protein–protein interaction of Escherichia coli K-12. Genome Res. 16, 686–691 (2006). source

      Not checked in the free run.

    • cites [26] Yu, H. et al. High-quality binary protein interaction map of the yeast interactome network. Science 322, 104–110 (2008). source

      Not checked in the free run.

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

    Premisenone given

    Random guessing yields an AUC0.1 of 0.05.

    “For reference, random guessing yields an AUC0.1 of 0.05.”
    Quote in the paperNot a substantial claim

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

    Interpretationown data

    In the two CP17 cases where AF2Complex failed, each has one monomer with only single-digit depth in its monomeric MSAs, which may explain these failures.

    “In these two failed cases, each has one monomer with only single-digit depth in its monomeric MSAs, which may explain these failures.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    The Ccm system I consists of two modules: module 1 includes CcmABCD and is responsible for acquiring and loading a heme molecule onto the heme chaperone CcmE.

    “it is thought that the Ccm system I consists of two modules: module 1 includes CcmABCD and is responsible for acquiring and loading a heme molecule onto the heme chaperone CcmE43.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [43] Feissner, R. E., Richard-Fogal, C. L., Frawley, E. R. & Kranz, R. G. ABC transporter-mediated release of a haem chaperone allows cytochrome c biogenesis. Mol. Microbiol. 61, 219–231 (2006). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    These models are likely biologically relevant, and heme was modeled in the expected sites in all these models.

    “These models are likely biologically relevant, and we were able to model the heme in the expected sites in all these models.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    GroL is part of a large assembly that requires seven GroL copies forming a ring stacked with another heptameric GroE ring.

    “It is part of a large assembly that requires seven GroL copies forming a ring stacked with another heptameric GroE ring.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: GroL is part of a large assembly that requires seven GroL copies forming a ring stacked with another heptameric GroE ring.

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

    Premisecitation

    CcmE shuttles the heme from module 1 to module 2, composed of CcmFGH, where the heme is delivered to CcmF.

    “CcmE then shuttles the heme from module 1 to module 2, composed of CcmFGH, where the heme is delivered to CcmF44.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [44] San Francisco, B. & Kranz, R. G. Interaction of HoloCcmE with CcmF in Heme Trafficking and Cytochrome c Biosynthesis. J. Mol. Biol. 426, 570–585 (2014). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    E. coli CcmH has a fused C-terminal domain that appears as the standalone protein CcmI in other species with a similar Ccm system.

    “E. coli CcmH has a fused C-terminal domain that appears as the standalone protein CcmI in other species with a similar Ccm system.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    These proteins belong to the cytochrome c maturation (Ccm) system I, which is composed of eight constituents (CcmABCDEFGH).

    “These proteins belong to the cytochrome c maturation (Ccm) system I, which is composed of eight constituents (CcmABCDEFGH)33,34.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [33] Sanders, C., Turkarslan, S., Lee, D. W. & Daldal, F. Cytochrome c biogenesis: the Ccm system. Trends Microbiol. 18, 266–274 (2010). source

      Not checked in the free run.

    • cites [34] Kranz, R. G., Richard-Fogal, C., Taylor, J. S. & Frawley, E. R. Cytochrome c biogenesis: mechanisms for covalent modifications and trafficking of heme and for heme-iron redox control. Microbiol. Mol. Biol. Rev. 73, 510–528 (2009). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The main reason for the knowledge gap is that the assemblies involve transient but essential interactions among membrane proteins (except for CcmA) that are difficult to capture experimentally.

    “The main reason for this knowledge gap is that the assemblies involve transient but essential interactions among membrane proteins (except for CcmA) that are difficult to capture experimentally.”
    Quote in the paperNot a substantial claim

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

    Premisereasoning

    Many mechanistic details of the Ccm system are still unclear, in part because there are no structures of the assembled modules.

    “Many mechanistic details of the Ccm system are still unclear, in part because there are no structures of the assembled modules.”
    Quote in the paperSubstantial claim

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

    Interpretationown data

    This conformation presumably corresponds to a resting state before heme loading, after heme unloading, or both.

    “This conformation presumably corresponds to a resting state before heme loading, after heme unloading, or both.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    CcmFGH covalently attaches the heme to nascent cytochrome c-type proteins.

    “Subsequently, CcmFGH covalently attaches the heme to nascent cytochrome c-type proteins45,46.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [45] San Francisco, B., Sutherland, M. C. & Kranz, R. G. The CcmFH complex is the system I holocytochrome c synthetase: engineering cytochrome c maturation independent of CcmABCDE. Mol. Microbiol. 91, 996–1008 (2014). source

      Not checked in the free run.

    • cites [46] Verissimo, A. F. et al. The thioreduction component CcmG confers efficiency and the heme ligation component CcmH ensures stereo-specificity during cytochrome c maturation. J. Biol. Chem. 292, 13154–13167 (2017). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation

    A critical loop containing the equivalent histidine of E. coli His303CcmF is missing in the TtCcmF structure, suggesting that the P-heme binding pocket might be flexible.

    “a critical loop containing the equivalent histidine of E. coli His303CcmF is missing in the TtCcmF structure, suggesting that the P-heme binding pocket might be flexible.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    These large conformational changes could be the result of ATP hydrolysis within CcmA.

    “These large conformational changes could be the result of ATP hydrolysis within CcmA.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

    Premisecitation

    As previously proposed, the energy of ATP hydrolysis could be harnessed to release the cargo-loaded CcmE.

    “As previously proposed43, the energy of the hydrolysis could be harnessed to release the cargo-loaded CcmE.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [43] Feissner, R. E., Richard-Fogal, C. L., Frawley, E. R. & Kranz, R. G. ABC transporter-mediated release of a haem chaperone allows cytochrome c biogenesis. Mol. Microbiol. 61, 219–231 (2006). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    CcmF has two heme-binding sites: one (P-heme) for the cytochrome attachment and the other accessory heme (TM-heme) that assists in the ejection of the P-heme.

    “CcmF has two heme-binding sites: one (P-heme) for the cytochrome attachment and the other accessory heme (TM-heme) that assists in the ejection of the P-heme34.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [34] Kranz, R. G., Richard-Fogal, C., Taylor, J. S. & Frawley, E. R. Cytochrome c biogenesis: mechanisms for covalent modifications and trafficking of heme and for heme-iron redox control. Microbiol. Mol. Biol. Rev. 73, 510–528 (2009). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given

    His303 is highly conserved and is known to coordinate the Fe center of P-heme.

    “His303 is highly conserved and is known to coordinate the Fe center of P-heme.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationcitation

    The pore may allow the access of heme from the outer leaflet, but no channel for potential heme trafficking is present within CcmC, as previously speculated.

    “The pore may allow the access of heme from the outer leaflet, but no channel for potential heme trafficking is present within CcmC, as previously speculated48.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [48] Richard-Fogal, C. & Kranz, R. G. The CcmC:heme:CcmE complex in heme trafficking and cytochrome c biosynthesis. J. Mol. Biol. 401, 350–362 (2010). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    One major hurdle to this bottom-up approach for predicting protein interactions is that the context of such a hypothetical complex is often unavailable a priori.

    “One major hurdle to this bottom-up approach for predicting protein interactions is that the context of such a hypothetical complex is often unavailable a priori.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Docking studies suggest that our models are consistent with known biochemical evidence, though other conformations are also expected.

    “Although the approach is currently limited to structural models without directly incorporating a heme, docking studies suggest that our models are consistent with known biochemical evidence, though other conformations are also expected.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    The structural difference may be attributed to a lipid molecule occupying the pore in the crystal structure but absent in our computational model.

    “The structural difference may be attributed to a lipid molecule occupying the pore in the crystal structure but absent in our computational model.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The CXXC motif of CcmH then reverts to the oxidized state, and CcmG dissociates to be reduced by the thiol-disulfide interchange protein DsbD.

    “The CXXC motif of CcmH then reverts to the oxidized state, and CcmG dissociates to be reduced by the thiol-disulfide interchange protein DsbD.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    The Wilcoxon signed-rank test was used to test whether the model quality of one model is better than another in terms of a scoring metric, with paired and one-tailed tests.

    “To test the hypothesis that the model quality of one model is better than another one in terms of a scoring metric, e.g., the DockQ score, we used the Wilcoxon signed-rank test, a non-parametric statistical test, because the score distributions do not follow normal distributions. The tests were paired, as every method makes a prediction on the same set of targets. All tests are one-tailed.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    In the benchmark test of the 17 targets, three independent runs were carried out, and each generated five top models.

    “In the benchmark test of the 17 targets, three independent runs were carried out, and each generated five top models.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    NrfA contains five CXXC motifs for covalent heme attachment.

    “NrfA contains five CXXC motifs for covalent heme attachment59.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [59] Clarke, T. A. et al. Role of a Conserved Glutamine Residue in Tuning the Catalytic Activity of Escherichia coli Cytochrome c Nitrite Reductase. Biochemistry 47, 3789–3799 (2008). source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The structural templates from the PDB were restricted to those dated before the same cutoff date for the AF2 training structures.

    “We used the same sequence library as above but restricted the structural templates from the PDB to those dated before the same cutoff date for the AF2 training structures.”
    Quote in the paperSubstantial claim
70 of 70

  • The paper relies on this without support.

    Intra-protein domain-domain interactions are not physically different from interprotein interactions.

    “Considering that intra-protein domain-domain interactions are not physically different from interprotein interactions, these AF2 results are very intriguing.”

  • The authors state this as an assumption.

    All protein pairs other than the CP17 pairs are non-interacting.

    “Here, we naively assumed that all protein pairs other than the CP17 pairs are non-interacting, and any hit above a cutoff value of a metric adopted for evaluation is a false positive.”

  • The paper relies on this without support.

    Most pairs of proteins are non-interacting.

    “Because we expect that most pairs of proteins are non-interacting, we focus on the regime of low false-positive rate (i.e., FP < 0.1) in the receiver operating characteristic (ROC) curve.”

  • The paper relies on this without support.

    GroL is part of a large assembly that requires seven GroL copies forming a ring stacked with another heptameric GroE ring.

    “It is part of a large assembly that requires seven GroL copies forming a ring stacked with another heptameric GroE ring.”

  • The authors state this as an assumption.

    All assemblies in the Earlier Set were used for DL training.

    “We assume all assemblies in the Earlier Set were used for DL training and removed all the “easy” homologs in the Later Set by using the 30% sequence identity clusters of all protein sequences provided by the PDB.”

  • The paper relies on this without support.

    Protein-protein interactions are not physicochemically different from what drives protein folding.

    “After all, protein-protein interactions are not physicochemically different from what drives protein folding in the first place.”
6 of 6

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?

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.

  • Figure 8 assigns CcmF histidines to CcmE

    Minor

    The Figure 8 caption labels His173 and His303 as residues of CcmE, but the text explicitly identifies both as CcmF residues coordinating P-heme.

  • Figure 8 names the wrong complex for the apocyt models

    Minor

    The Figure 8 caption calls the apocyt-containing models CcmEFH complexes, while the accompanying text says that apocyt peptides were modeled with CcmFGH.

  • Assembly subsets do not sum to the stated total

    Minor

    The two subsets contain 92,047 + 24,853 = 116,900 assemblies, not 120,703. The stated split leaves 3,803 assemblies unaccounted for.

    Reported 120703 · recomputed 116900

  • Benchmark set sizes differ between Results and Methods

    Minor

    The Results describe a benchmark of 1,192 dimers and 593 oligomers, whereas the Methods name the final benchmark sets as 1,193 dimers and 562 oligomers.

  • CcmFGH passage refers to a binding site on absent CcmE

    Minor

    The CcmFGH description says the substrate reaches the heme-binding site of CcmE, although the same passage specifies that CcmE is absent from this complex and identifies CcmF as having the relevant site.

5 of 5

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Gao, M. et al. in 2021 IEEE/ACM Workshop on Machine Learning in High Performance Computing Environments (MLHPC). 46-57.

Cited by

Sorted by citations

Related

Authors
Gao, M., Nakajima An, D., Parks, J. M., & Skolnick, J.
Affiliations
  • Georgia Institute of Technology (US)
  • Oak Ridge National Laboratory (US)
Corresponding author
  • Gao, Mu
  • Skolnick, Jeffrey
ORCID
Venue
Nature Communications
Publisher
Springer Science and Business Media LLC
Published
2022-04-01
Volume
13
Issue
1
Pages
1744-1744
ISSN
2041-1723 (ISSN-L)
DOI
10.1038/s41467-022-29394-2
PMC
PMC8975832
OpenAlex
W4220990251
Cited by
252
References
67
Citations by year
2026: 23 · 2025: 41 · 2024: 78 · 2023: 81 · 2022: 28 · 2021: 1
FWCI
18.61
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information) (submittedVersion)
Funders
  • U.S. Department of Health & Human Services | NIH | National Institute of General Medical Sciences (NIH R35GM118039, R35GM118039)
  • DOE | SC | Biological and Environmental Research (DOE DE-SC0021303)
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
  • Physical Sciences › Materials Science › Materials Chemistry › Enzyme Structure and Function
Keywords
protein-protein interaction, multimeric proteins, AlphaFold2, protein structure prediction, protein-protein docking, Escherichia coli proteome, deep learning
MeSH terms
Deep Learning, Escherichia coli, Sequence Alignment, Neural Networks, Computer, Proteome

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