CoChat
Sign Up
Cited by 141Protein Structure and DynamicsMachine Learning in BioinformaticsComputational Drug Discovery Methods

Deep geometric representations for modeling effects of mutations on protein-protein binding affinity

Open access
Liu, Xianggen, Luo, Yunan, Li, Pengyong, Song, Sen, Peng, Jian·2021·PLOS Computational Biology
DOI 10.1371/journal.pcbi.1009284PMC PMC8366979
Open this paper in CoChat — free accountOpen PDF

Abstract

Modeling the impact of amino acid mutations on protein-protein interaction plays a crucial role in protein engineering and drug design. In this study, we develop GeoPPI, a novel structure-based deep-learning framework to predict the change of binding affinity upon mutations. Based on the three-dimensional structure of a protein, GeoPPI first learns a geometric representation that encodes topology features of the protein structure via a self-supervised learning scheme. These representations are then used as features for training gradient-boosting trees to predict the changes of protein-protein binding affinity upon mutations. We find that GeoPPI is able to learn meaningful features that characterize interactions between atoms in protein structures. In addition, through extensive experiments, we show that GeoPPI achieves new state-of-the-art performance in predicting the binding affinity changes upon both single- and multi-point mutations on six benchmark datasets. Moreover, we show that GeoPPI can accurately estimate the difference of binding affinities between a few recently identified SARS-CoV-2 antibodies and the receptor-binding domain (RBD) of the S protein. These results demonstrate the potential of GeoPPI as a powerful and useful computational tool in protein design and engineering. Our code and datasets are available at: https://github.com/Liuxg16/GeoPPI.

AI snapshot

Written from the paper’s full text.

TL;DR

GeoPPI is a structure-based deep-learning framework that learns self-supervised geometric representations of protein complexes and uses them with gradient-boosting trees to predict mutation-induced changes in protein-protein binding affinity. It reports state-of-the-art performance across six benchmark datasets and in SARS-CoV-2 antibody case studies.

Key findings

  • GeoPPI achieved Pearson correlations of 0.51, 0.58, 0.52, and 0.68 on single-point mutation datasets S645, S1131, S4169, and S8338, respectively, and 0.53 and 0.74 on multi-point datasets M1101 and M1707. On S1131 it improved Pearson correlation by 45% over the previous best method, TopGBT.
  • In leave-one-structure-out cross-validation, GeoPPI reached 0.57 on S645 versus 0.39 for TopGBT, and 0.76 on M1707 versus 0.72 for MutaBind2. On the independent S641 test set, GeoPPI obtained regression RP = 0.37, classification ACC = 0.83, AUC = 0.68, and MCC = 0.19.
  • For SARS-CoV-2 antibody case studies, GeoPPI achieved a correlation of 0.62 on pairwise affinity changes among 17 structurally similar antibodies, compared with 0.29 for MutaBind2, and reached RP = 0.83 on C110 variants. It also predicted A107W and D103Y as affinity-enhancing mutations in C110.
  • GeoPPI required 17.2 seconds per single-mutant prediction, accelerating the previously fastest method, MutaBind2, by 151%.

Methods

  • GeoPPI combines a message-passing geometric encoder with a gradient-boosting tree. The encoder is trained self-supervised to reconstruct original side-chain coordinates after random side-chain torsion perturbation, then produces atom-level representations that are pooled and fed to the GBT for ΔΔG prediction.
  • The geometric encoder was trained on 13,590 unlabeled complexes from PDB-BIND and 3DComplex, with each complex randomly perturbed 2,000 times, yielding 27,180,000 training data points. Evaluation used six benchmark datasets: S645, S1131, S4169, S8338, M1101, and M1707.
  • Main analyses included split-by-structure cross-validation using ECOD homology to separate folds, leave-one-structure-out cross-validation, an independent S641 test set, binary classification of stabilizing versus destabilizing mutations, inference-speed comparison, and SARS-CoV-2 antibody/NTD mutational scanning.

Limitations

  • The paper states that self-supervised learning requires substantial computational resources: training took around 12 days on an NVIDIA GeForce GTX TITAN X GPU, though it notes this can be performed offline.
  • The paper reports that all methods performed poorly on the independent S641 test set, with GeoPPI’s correlation dropping to 0.37, reflecting the difficulty of generalizing to dissimilar structures.
  • For the SARS-CoV-2 antibody case study, most antibody structures were not solved; the authors used Rosetta3 homology modeling and ZDOCK to approximate structures and binding orientations, which the text presents as making the prediction task more difficult. Evident from the text: the proposed affinity-enhancing mutations (e.g., A107W, D103Y) are computationally predicted and structurally rationalized, but no wet-lab experimental validation is reported in the paper.

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

Integrity signals

Registry

Crossref · 10.1371/journal.pcbi.1009284

Cites

78 of 78 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 32 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    GeoPPI learns meaningful features that characterize interactions between atoms in protein structures.

    “We find that GeoPPI is able to learn meaningful features that characterize interactions between atoms in protein structures.”
    • Also stated in: § GeoPPI captures meaningful patterns in the structure of the protein complex · boosted
      “providing the evidence that GeoPPI captures meaningful patterns of the complex structure.”
    Quote in the paperSubstantial claim
  • Findingown data

    GeoPPI achieves new state-of-the-art performance in predicting binding affinity changes upon both single- and multi-point mutations on six benchmark datasets.

    “In addition, through extensive experiments, we show that GeoPPI achieves new state-of-the-art performance in predicting the binding affinity changes upon both single- and multi-point mutations on six benchmark datasets.”
    • Also stated in: § Introduction · plain
      “GeoPPI obtains the new state-of-the-art performance on all of these datasets with the fastest inference speed, showing its effectiveness and efficiency.”
    • Also stated in: § GeoPPI advances the state of the art in estimating the effects of mutations on binding affinity · plain
      “As for GeoPPI, it outperforms all of the baselines by a large margin on all datasets.”
    Quote in the paperSubstantial claim
  • Findingown data

    GeoPPI can accurately estimate or predict binding affinity changes between recently identified SARS-CoV-2 antibodies and the RBD of the S protein, even when trained with low-order mutants and applied to higher-order mutants.

    “Moreover, we show that GeoPPI can accurately estimate the difference of binding affinities between a few recently identified SARS-CoV-2 antibodies and the receptor-binding domain (RBD) of the S protein.”
    • Also stated in: § Introduction · plain
      “GeoPPI is able to accurately predict the binding affinity changes between these complexes, even when it is trained with low-order mutants and applied to higher-order mutants.”
    Quote in the paperSubstantial claim
  • Findingown data

    The most important component of geometric representations produced by the trained encoder can clearly indicate the perturbation level of atoms.

    “the most important component of the geometric representations produced by the trained geometric encoder, can clearly indicate the perturbation level of atoms.”
    Quote in the paperSubstantial claim
  • Findingown data

    TopGBT yields a correlation of 0.21 on the most conservative-mutation data points from the S645 dataset.

    “We found that GeoPPI obtains a correlation of 0.66 while TopGBT only yields a correlation of 0.21 on these data points.”
    Quote in the paperSubstantial claim
  • May combine two claims. 16
  • The quote only partly states this claim. 1
  • Classified as probably not a substantial claim. 13

  • May combine two claims.

    Findingown data

    In leave-one-structure-out cross-validation, TopGBT obtains correlation 0.39 on S645 while GeoPPI achieves 0.57.

    “TopGBT obtains a correlation of 0.39 on S645 while GeoPPI achieves 0.57 (Fig 3C).”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    GeoPPI achieves the highest correlation on S641 among tested methods.

    “However, our model GeoPPI still achieves the highest correlation, showing the generalizability of our method.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    FoldX yields a correlation of 0.46 on S1131 and 0.16 on S641.

    “FoldX yields a correlation of 0.46 on S1131 but only obtains 0.16 in this test set, which reflects the challenge of the prediction in this test dataset.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In leave-one-structure-out cross-validation, MutaBind2 obtains correlation 0.72 on M1707 while GeoPPI yields 0.76.

    “MutaBind2 obtains a correlation of 0.72 on M1707 while GeoPPI yields 0.76 (Fig 3D).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    FoldX and BeAtMuSic yield the lowest correlations and highest RMSE on single-point mutation datasets.

    “We observed FoldX [8] and BeAtMuSic [38] yield the lowest correlations and the highest root-mean-square error (RMSE) on these datasets, indicating the difficulty of gauging the impacts of mutations via empirical energy-based methods.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [8] J Schymkowitz, 2005, The FoldX web server: an online force field, Nucleic Acids Research source

      Not checked in the free run.

    • cites [38] Y Dehouck, 2013, BeAtMuSiC: prediction of changes in protein–protein binding affinity on mutations, Nucleic Acids Research source

      Not checked in the free run.

  • May combine two claims.

    Findingnone given

    Most structures of the 17 structurally similar SARS-CoV-2 neutralizing Abs are not determined, making prediction more difficult than on solved structures.

    “Most of the structures of these Abs are not determined, making the prediction task more difficult than that on the solved structures.”
    Quote in the paperSubstantial claim

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

    Findingnone given

    GeoPPI is efficient in the prediction stage compared to existing methods that rely on computation-heavy biophysical simulations.

    “GeoPPI is efficient in prediction stage compared to existing methods that rely on computation-heavy biophysical simulations.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    A large region centered on residue A27 on the SARS-CoV-2 NTD is mutationally constrained by its binding with ACE2 and is evolutionarily conserved.

    “Intriguingly, we found a large region centered on residue A27 that is mutationally constrained by its binding with ACE2 and is also evolutionarily conserved (S2 Text, S6 Fig).”
    Quote in the paperSubstantial claim

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

    Findingnone given

    GeoPPI training takes around 12 days on a GPU (NVIDIA GeForce GTX TITIAN X).

    “In the case of GeoPPI, it takes around 12 days on a GPU (NVIDIA GeForce GTX TITIAN X).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Empirical energy-based methods have accelerated prediction but suffer from insufficient conformational sampling, especially for mutations in flexible regions.

    “The empirical energy-based methods, exemplified by STATIUM [12], FoldX [8] and Discovery Studio [13], have accelerated the prediction but suffered from insufficient conformational sampling, especially for mutations in flexible regions.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [12] J DeBartolo, 2014, Genome-wide prediction and validation of peptides that bind human prosurvival Bcl-2 proteins, PLoS Computational Biology source

      Not checked in the free run.

    • cites [8] J Schymkowitz, 2005, The FoldX web server: an online force field, Nucleic Acids Research source

      Not checked in the free run.

    • cites [13] Inc AS. Discovery Studio Modeling Environment, Release 4.0. 2013;.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Traditional molecular modeling methods usually provide reliable simulation results but require heavy computational resources, limiting their applicability.

    “The methods in the first category, such as coarse-grained molecular dynamics simulations [9, 10], usually provide reliable simulation results but require heavy computational resources, which limits their applicability [11].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [9] AY Shih, 2006, Coarse grained protein- lipid model with application to lipoprotein particles, The Journal of Physical Chemistry B source

      Not checked in the free run.

    • cites [10] A Arkhipov, 2006, Stability and dynamics of virus capsids described by coarse-grained modeling, Structure source

      Not checked in the free run.

    • cites [11] RC Bernardi, 2015, Enhanced sampling techniques in molecular dynamics simulations of biological systems, Biochimica et Biophysica Acta-General Subjects source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Some complexes in the benchmark datasets are highly related, so machine learning methods tend to be overtrained on these datasets, as evidenced by drops in prediction performance when meeting unseen structures.

    “As some complexes in the above benchmark datasets are highly related (S2 and S3 Figs), machine learning methods tend to be overtrained on these datasets, as evidenced by the drops of the prediction performance when machine learning methods meet unseen structures [31].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [31] M Wang, 2020, A topology-based network tree for the prediction of protein–protein binding affinity changes following mutation, Nature Machine Intelligence source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    The ECOD homology level is a strict similarity criterion because it clusters domains even if they only have distant relationships.

    “The ECOD homology level (“H level”) is a strict similarity criterion since it clusters domains even if they only have distant relationships.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [37] H Cheng, 2014, ECOD: an evolutionary classification of protein domains, PLoS Computational Biology source

      Not checked in the free run.

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

    Premisereasoning

    The seven features used in MutaBind2 are manually designed and may not comprehensively characterize the impacts of mutations.

    “Considering that the seven features used in MutaBind2 are manually designed, these features may not comprehensively characterize the impacts of mutations.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [15] N Zhang, 2020, MutaBind2: predicting the impacts of single and multiple mutations on protein-protein interactions, Iscience source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The features produced by the geometric encoder are learned to describe differences between unstable and stable structures, leading to better predictive power of GeoPPI than MutaBind2.

    “However, as the features produced by the geometric encoder are learned to describe the differences between the unstable structure and the stable one, leading to the better predictive power of GeoPPI than MutaBind2.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The self-supervised learning scheme in GeoPPI is built to explicitly learn interactions between atoms, thus leading to better prediction results.

    “By contrast, the self-supervised learning scheme in GeoPPI is built to explicitly learn the interactions between atoms, thus leading to better prediction results.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    The six benchmark datasets have been widely used to test the predictive power of PPI prediction methods.

    “These datasets have been widely used to test the predictive power of the PPI prediction methods [12, 13, 27–29, 29–35].”
    Quote in the paperNot a substantial claim0 of 10 sources checked
    • cites [12] J DeBartolo, 2014, Genome-wide prediction and validation of peptides that bind human prosurvival Bcl-2 proteins, PLoS Computational Biology source

      Not checked in the free run.

    • cites [13] Inc AS. Discovery Studio Modeling Environment, Release 4.0. 2013;.

      Not checked in the free run.

    • cites [27] Y Yang, 2008, Specific interactions for ab initio folding of protein terminal regions with secondary structures, Proteins: Structure, Function, and Bioinformatics source

      Not checked in the free run.

    • cites [28] H Zhou, 2002, Distance-scaled, finite ideal-gas reference state improves structure-derived potentials of mean force for structure selection and stability prediction, Protein Science source

      Not checked in the free run.

    • cites [29] S Liu, 2004, A physical reference state unifies the structure-derived potential of mean force for protein folding and binding, Proteins: Structure, Function, and Bioinformatics source

      Not checked in the free run.

    • cites [30] P Xiong, 2017, BindProfX: assessing mutation-induced binding affinity change by protein interface profiles with pseudo-counts, Journal of Molecular Biology source

      Not checked in the free run.

    • cites [31] M Wang, 2020, A topology-based network tree for the prediction of protein–protein binding affinity changes following mutation, Nature Machine Intelligence source

      Not checked in the free run.

    • cites [32] MF Lensink, 2013, Docking, scoring, and affinity prediction in CAPRI, Proteins: Structure, Function, and Bioinformatics source

      Not checked in the free run.

    • cites [33] M Petukh, 2016, SAAMBE: webserver to predict the charge of binding free energy caused by amino acids mutations, International Journal of Molecular Sciences source

      Not checked in the free run.

    • cites [34] S Sirin, 2016, AB-Bind: antibody binding mutational database for computational affinity predictions, Protein Science source

      Not checked in the free run.

  • May combine two claims.

    Study design and conductnone given

    Both the geometric encoder and the t-SNE algorithm were not informed of the group information.

    “Note that both the geometric encoder and the t-SNE algorithm were not informed of the group information.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductreasoning

    The test and training data (S641 and S1131) are from different datasets and share little similarity at the homology level, so test performance reflects prediction generalizability.

    “As the test and training data (i.e., S1131) are from different datasets and share little similarity in the homology level, the test performance can reflect the prediction generalizability of the methods.”
    Quote in the paperSubstantial claim

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

    Study design and conductown data

    For a cross-validation fold, 10% of the training data was held out as development data, and GBT hyperparameters were chosen to maximize performance on that development data.

    “Taking a fold of the cross-validation experiment as an example, we held out 10% of the training data as the development data; We chose the hyper-parameters of the GBT that yield the highest performance on the development data.”
    Quote in the paperNot a substantial claim

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

    Study design and conductown data

    For each dataset, the GBT was trained using the training data of each fold in cross-validation tests.

    “As for the learning of GBT for each dataset, we trained the GBT using the training data of each fold in the cross-validation tests.”
    Quote in the paperNot a substantial claim

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

    Study design and conductown data

    The best hyperparameters of GeoPPI were calibrated through a grid search procedure on the development set.

    “The best hyperparameters of GeoPPI were calibrated through a grid search procedure on the development set (S11 Table).”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    By reconstructing original conformations, a model is expected to capture patterns of biomolecular interactions between atoms and residues in three-dimensional space.

    “By reconstructing the original conformations, a model is expected to capture the patterns of the biomolecular interactions between atoms and those between residues in the three-dimensional space.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: For a particular protein complex, only a few conformations can lead to the lowest free energy.
    • Rests on an assumption: Most side-chain perturbations will increase the free energy and make the complex less stable.

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

    Study design and conductreasoning

    For complexes not covered by ECOD, TMalign with a cutoff of 0.5 was used to further remove similar complexes.

    “For these complexes, we leveraged the TMalign software (with a cutoff of 0.5) [58] to further remove the similar complexes.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [58] Y Zhang, 2005, TM-align: a protein structure alignment algorithm based on the TM-score, Nucleic Acids Research source

      Not checked in the free run.

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

    Premisecitation

    The geometric encoder is a message passing neural network that shares the basic idea of the graph attention network (GAT).

    “The geometric encoder is a message passing neural network and shares the basic idea of the graph attention network (GAT) [19]:”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [19] Veličković P, Cucurull G, Casanova A, Romero A, Lio P, Bengio Y. Graph attention networks. ICLR. 2018;.

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    The joint distribution of side-chain torsion angles of a residue r, conditional on its backbone dihedral angles, is approximated by a protein-dependent side-chain rotamer library.

    “p(⋅|ϕr, ψr, r) stands for the joint distribution of side-chain torsion angles of the residue r, which is approximated by a protein-dependent side-chain rotamer library [64].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [64] MV Shapovalov, 2011, A smoothed backbone-dependent rotamer library for proteins derived from adaptive kernel density estimates and regressions, Structure source

      Not checked in the free run.

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

    Study design and conductown data

    The selection of hyperparameters in all experiments was based on the training data.

    “The selection of the hyper-parameters in all the experiments of this paper was also based on the training data.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The new hydrogen bond accounts for GeoPPI's prediction of A107W's high positive impact and confirms the reliability of the prediction.

    “We found that it gives rise to a new hydrogen bond between C110 and the SARS-CoV-2 RBD (Fig 5F and 5G), which accounts for the prediction of GeoPPI and thus further confirms the reliability of the prediction by GeoPPI.”
    Quote in the paperSubstantial claim
28 of 28

  • The paper relies on this without support.

    Experimental measurement of mutation effects is labor-intensive and time-consuming.

    “On the one hand, the experimental measurement of the mutation effects is a labor-intensive and time-consuming process.”

  • The paper relies on this without support.

    The space of possible antibody mutants is combinatorially large, so existing methods cannot explore it exhaustively in a reasonable timeframe.

    “On the other hand, the space of possible Ab mutants is combinatorially large, for which existing methods cannot explore exhaustively in a reasonable timeframe.”

  • The paper relies on this without support.

    Existing machine-learning input features are mostly manually engineered based on known rules in protein structures, often limiting predictive generalization across various protein structures.

    “In addition, these input features are mostly manually engineered based on the known rules in protein structures, often limiting their predictive generalization across various protein structures.”

  • The paper relies on this without support.

    GeoPPI is the first method that employs a self-supervised learning scheme to learn representations of protein structure and utilizes message passing neural networks to model interactions between atoms in this task.

    “To our best knowledge, GeoPPI is the first method that employs a self-supervised learning scheme to learn representations of the protein structure and utilizes message passing neural networks to model the interactions between atoms in this task, serving as a novel method for estimating the protein-protein binding affinity changes upon mutations.”

  • The paper relies on this without support.

    The binding affinity of two proteins in a complex is largely determined by the interaction strengths between atoms in their interface.

    “As the binding affinity of two proteins in a complex is largely determined by the interaction strengths between atoms in their interface”

  • The paper relies on this without support.

    Mutations lead to different conformations at the atom level.

    “Since the mutations also lead to different conformations in the atom level, the sensitivity to the interactions between atoms is helpful in the prediction of the binding affinity changes upon mutations.”

  • The paper relies on this without support.

    For a particular protein complex, only a few conformations can lead to the lowest free energy.

    “This idea stems from the observation that, for a particular complex, only a few conformations can lead to the lowest free energy.”

  • The paper relies on this without support.

    Most side-chain perturbations will increase the free energy and make the complex less stable.

    “Most of the side-chain perturbations will increase the free energy and make the complex less stable.”

  • The authors state this as an assumption.

    If the distance between two atoms is shorter than 3Å, there exists an edge between them.

    “if the distance of two nodes is shorter than a threshold (i.e., 3Å), we assume there exists an edge between them.”

  • The paper relies on this without support.

    No self-supervised learning scheme has been studied in the field of protein structure dynamics.

    “However, due to the complexity of the dynamics in protein structure, no self-supervised learning scheme has been studied in this field.”

  • The paper relies on this without support.

    GeoPPI is the first study to introduce self-supervised learning for representing protein structure.

    “To our best knowledge, GeoPPI serves as the first study to introduce the self-supervised learning technique for representing the protein structure.”

  • The paper relies on this without support.

    The complexes in S645 are not diverse enough for training.

    “we took S1131 as the training dataset, as the complexes in S645 are not diverse enough for training and S4169 contains most of the known protein domains among the benchmark datasets (difficult to construct independent test data).”

  • The paper relies on this without support.

    S4169 contains most of the known protein domains among the benchmark datasets, making it difficult to construct independent test data.

    “we took S1131 as the training dataset, as the complexes in S645 are not diverse enough for training and S4169 contains most of the known protein domains among the benchmark datasets (difficult to construct independent test data).”

  • The paper relies on this without support.

    Binding affinity is usually characterized by the interface residues of the complex.

    “As our downstream task is to model the binding affinity, which is usually characterized by the interface residues of the complex, we set p(r) to be the uniform distribution over the interface residues and the probabilities of other non-interface residues are zero.”

  • The paper relies on this without support.

    The ranges of coordinates differ a lot in individual complexes, making direct prediction of absolute coordinate values more difficult.

    “However, as the ranges of the coordinates differ a lot in individual complexes, directly predicting the absolute values of the coordinates increases the difficulty of reconstruction.”
15 of 15

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?

  • The model makes predictions with moderate to high correlation with experimental binding data and outperforms GeoPPI and other recently proposed methods for predicting single-mutation effects.

    “We found that our model is able to make predictions with moderate to high correlation with experimental binding data and also outperforms GeoPPI (41), the current state-of-the-art method, as well as a few other recently proposed methods for predicting single-mutation effects.”

    From Deep learning guided optimization of human antibody against SARS-CoV-2 variants with broad neutralization (2022) · cites this paper as [41]

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.

  • FoldX is not among the lowest-correlation methods on S1131

    Major

    The text says FoldX yields one of the lowest correlations on these datasets, but Table 1 gives it a correlation of 0.46 on S1131, higher than TopGBT's 0.32 and TopNetTree's 0.29.

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. A Leaver-Fay, 2011, Methods in enzymology, 545

Cited by

Sorted by citations

Related

Authors
Liu, X., Luo, Y., Li, P., Song, S., & Peng, J.
Affiliations
  • Southwest Jiaotong University (CN)
  • University of Illinois Urbana-Champaign (US)
  • Tsinghua University (CN)
Corresponding author
  • Liu, Xianggen
  • Luo, Yunan
  • Li, Pengyong
  • Song, Sen
  • Peng, Jian
ORCID
Venue
PLOS Computational Biology
Publisher
Public Library of Science (PLoS)
Published
2021-08-04
Volume
17
Issue
8
Pages
e1009284
ISSN
  • 1553-734X (ISSN-L)
  • 1553-7358
DOI
10.1371/journal.pcbi.1009284
PMC
PMC8366979
OpenAlex
W3082498841
Semantic Scholar
baa0d23702f7d094848b1f82f94d536259e2bf08
Cited by
140
References
77
Citations by year
2026: 17 · 2025: 39 · 2024: 42 · 2023: 20 · 2022: 19 · 2021: 3
FWCI
6.67
Type
journal article
Language
en
Update notices
new_version (2021-08-16)
Open access
Licence
Locations
  • PLoS Computational Biology (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University)
Funders
  • china scholarship council
  • institute guoqiang at tsinghua university
  • National Science Foundation (1652815)
  • compgen fellowship
  • baidu fellowship
  • nsf career award
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 › Computer Science › Computational Theory and Mathematics › Computational Drug Discovery Methods
Keywords
protein-protein binding affinity, amino acid mutations, geometric representation, self-supervised learning, gradient boosting decision tree, SARS-CoV-2 antibodies, receptor binding domain, protein engineering, protein design
MeSH terms
Models, Chemical, Protein Binding, Proteins, Point Mutation, Amino Acid Substitution

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