CoChat
Sign Up
Cited by 369Machine Learning in BioinformaticsProtein Structure and DynamicsGenomics and Rare Diseases

mCSM-PPI2: predicting the effects of mutations on protein–protein interactions

Open access
Rodrigues, Carlos H M, Myung, Yoochan, Pires, Douglas E V, Ascher, David B·2019·Nucleic Acids Research
DOI 10.1093/nar/gkz383PMC PMC6602427
Open this paper in CoChat — free accountOpen PDF

Abstract

Protein–protein Interactions are involved in most fundamental biological processes, with disease causing mutations enriched at their interfaces. Here we present mCSM-PPI2, a novel machine learning computational tool designed to more accurately predict the effects of missense mutations on protein–protein interaction binding affinity. mCSM-PPI2 uses graph-based structural signatures to model effects of variations on the inter-residue interaction network, evolutionary information, complex network metrics and energetic terms to generate an optimised predictor. We demonstrate that our method outperforms previous methods, ranking first among 26 others on CAPRI blind tests. mCSM-PPI2 is freely available as a user friendly webserver at http://biosig.unimelb.edu.au/mcsm_ppi2/.

AI snapshot

Written from the paper’s full text.

TL;DR

mCSM-PPI2 is a machine learning webserver that predicts the effects of missense mutations on protein–protein binding affinity by combining graph-based structural signatures, evolutionary information, non-covalent interaction network metrics, and energetic terms. It outperformed previous methods on CAPRI blind tests and cross-validation, ranking first among 26 methods.

Key findings

  • On 10-times stratified 10-fold cross-validation using the S8338 dataset (8338 single-point mutations), mCSM-PPI2 achieved an average Pearson correlation of 0.82 (σ = 0.06) and RMSE of 1.18 kcal/mol; training only on S4169 gave ρ = 0.76 and RMSE = 1.19 kcal/mol.
  • In low-redundancy tests, leave-one-complex-out gave ρ = 0.75 and RMSE = 1.30 kcal/mol versus MutaBind’s ρ = 0.68 and RMSE = 1.57 kcal/mol; leave-one-binding-site-out gave ρ = 0.67 and RMSE = 1.39 kcal/mol versus MutaBind’s ρ = 0.57 and RMSE = 1.57 kcal/mol (p < 0.0001).
  • On CAPRI round 26, mCSM-PPI2 achieved Kendall scores of up to 0.42 for T55 and 0.32 for T56, ranking first among 26 other methods; on the MDM2-p53 blind test of 26 variants, it achieved ρ = 0.40 and RMSE = 0.36 kcal/mol.
  • For hotspot identification across 378 alanine-scanning mutations, it correctly predicted 95% of hotspots and 92% of non-hotspots, with ρ = 0.95 and RMSE = 0.25 kcal/mol.

Methods

  • Data were derived from SKEMPI 2.0, filtered to single-point mutations with available wild-type crystal structures, yielding 4169 variants in 319 complexes; hypothetical reverse mutations were added to balance 901 increased and 3268 decreased binding-affinity variants, producing the final S8338 dataset.
  • Features included graph-based structural signatures; wild-type residue environment descriptors (RSA, PHI, residue depth, amino acid content); glycine/proline binary terms; PSSM evolutionary information; Arpeggio-derived non-covalent contacts and network centrality metrics; AAindex contact potentials; FoldX energetic terms; and Bio3D atomic fluctuations.
  • Validation used 10-times stratified 10-fold cross-validation, leave-one-complex-out, leave-one-binding-site-out, CAPRI round 26, an MDM2-p53 blind test, and hotspot analysis on alanine-scanning mutations.

Limitations

  • The paper does not state a dedicated limitations section. Evident from the text: CAPRI T55 and T56 lacked experimental 3D structures, so homology models generated with Modeller were used, and performance on these targets was lower (Kendall 0.42 and 0.32; RMSE 2.55 and 4.06 kcal/mol).
  • Evident from the text: the original SKEMPI 2.0 dataset was unbalanced (901 increased versus 3268 decreased binding-affinity variants), and hypothetical reverse mutations were added to improve robustness; these reverse mutations are not independent experimental measurements.
  • Evident from the text: performance dropped in low-redundancy settings (leave-one-binding-site-out ρ = 0.67; S472 not in SKEMPI1 ρ = 0.63) and on the MDM2-p53 blind test (ρ = 0.40), indicating limited generalization to unseen complexes or binding sites.

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

Integrity signals

Registry

Crossref · 10.1093/nar/gkz383

Cites

59 of 59 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 25 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    mCSM-PPI2 trained only on the original subset of mutations from SKEMPI2 (S4169) achieved a correlation of 0.76 and RMSE of 1.19 kcal/mol.

    “We also evaluate the performance of mCSM-PPI2 when trained only on the original subset of mutations from SKEMPI2 (S4169) using the same procedure and obtained a correlation of 0.76 and RMSE of 1.19 kcal/mol.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The experimental binding affinity changes in the SKEMPI2 dataset are accurate.
  • Findingown data

    mCSM-PPI2 achieved an average Pearson correlation of 0.82 with a standard deviation of 0.06 across 10 runs of 10-times stratified 10-fold cross-validation on the S8338 dataset.

    “Our method was able to achieve an average Pearson correlation (ρ) of 0.82 with a standard deviation (σ) of 0.06 across the 10 runs (Figure 2A)”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The experimental binding affinity changes in the SKEMPI2 dataset are accurate.
    • Rests on an assumption: The effect of a reverse mutation is the negative of the effect of the original mutation (i.e., binding affinity changes are symmetric).
  • Findingown data

    mCSM-PPI2 achieved an RMSE of 1.18 kcal/mol on the S8338 dataset under 10-times 10-fold cross-validation.

    “mCSM-PPI2 was able to achieve a Pearson's correlation of 0.82 and RMSE 1.18 kcal/mol when trained on the S8338 dataset applying 10-fold cross-validation 10 times (A).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The experimental binding affinity changes in the SKEMPI2 dataset are accurate.
    • Rests on an assumption: The effect of a reverse mutation is the negative of the effect of the original mutation (i.e., binding affinity changes are symmetric).
  • Findingown data

    The original dataset from SKEMPI 2.0 is unbalanced, containing 901 variants that increased binding affinity and 3268 variants that decreased binding affinity.

    “Given the unbalanced nature of the original dataset collected from SKEMPI 2.0, 901 variants increased (ΔΔGwt-mt ≥ 0) and 3268 decreased (ΔΔGwt-mt < 0) binding affinity”
    Quote in the paperSubstantial claim
  • Findingown data

    mCSM-PPI2 outperformed MutaBind on leave-one-complex-out cross-validation (MutaBind ρ = 0.68 and RMSE = 1.57 kcal/mol).

    “outperforming MutaBind (25) (ρ = 0.68 and RMSE = 1.57 kcal/mol).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [25] Li, 2016, MutaBind estimates and interprets the effects of sequence variants on protein–protein interactions, Nucleic Acids Res. source

      Not checked in the free run.

    • Rests on an assumption: The experimental binding affinity changes in the SKEMPI2 dataset are accurate.
    • Rests on an assumption: The effect of a reverse mutation is the negative of the effect of the original mutation (i.e., binding affinity changes are symmetric).
  • May combine two claims. 9
  • Classified as probably not a substantial claim. 1

  • May combine two claims.

    Findingown data

    The final dataset for building mCSM-PPI2 includes 8338 single-point mutations, an increase of up to three-fold compared to previous methods that used data from the first version of SKEMPI.

    “Therefore, the final dataset for building mCSM-PPI2 predictive model includes 8338 single-point mutations (S8338), which represents an increase of up to three-fold in datapoints in comparison with previous methods that used data from the first version of SKEMPI with 2007 (S2007) (20,23), 1964 (S1964) (25), 1102 (S1102) (24) and 1327 (S1327) mutations (41).”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [20] Pires, 2014, mCSM: Predicting the effects of mutations in proteins using graph-based signatures, Bioinformatics source

      Not checked in the free run.

    • cites [23] Dehouck, 2013, BeAtMuSiC: Prediction of changes in protein–protein binding affinity on mutations, Nucleic Acids Res. source

      Not checked in the free run.

    • cites [25] Li, 2016, MutaBind estimates and interprets the effects of sequence variants on protein–protein interactions, Nucleic Acids Res. source

      Not checked in the free run.

    • cites [24] Geng, 2019, iSEE: Interface structure, evolution, and energy-based machine learning predictor of binding affinity changes upon mutations, Proteins source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    mCSM-PPI2 achieved ρ = 0.75 and RMSE = 1.30 kcal/mol on leave-one-complex-out cross-validation.

    “we achieved ρ = 0.75 (Figure 2B) and Root Mean Square Error (RMSE) of 1.30 kcal/mol”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The experimental binding affinity changes in the SKEMPI2 dataset are accurate.
    • Rests on an assumption: The effect of a reverse mutation is the negative of the effect of the original mutation (i.e., binding affinity changes are symmetric).

  • May combine two claims.

    Findingown data

    mCSM-PPI2 outperforms previous methods, ranking first among 26 others on CAPRI blind tests.

    “We demonstrate that our method outperforms previous methods, ranking first among 26 others on CAPRI blind tests.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    mCSM-PPI2 correctly predicted 95% of hotspots and 92% of non-hotspots in a set of 378 alanine-scanning mutations.

    “Our predictive model was was able to accurately distinguish hot and not-hot spots (95% of hotspots and 92% of non-hotspots were correctly predicted)”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The 378 alanine-scanning experimental mutations are accurate measures of binding energy changes.

  • May combine two claims.

    Findingown data

    mCSM-PPI2 achieved a Kendall's score of up to 0.42 for mutations in T55 and 0.32 for mutations in T56 on CAPRI round 26.

    “mCSM-PPI2 was able to achieve a Kendall's score of up to 0.42 and 0.32 for mutations in T55 and T56, respectively”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The homology models generated for T55 and T56 from close homologues accurately represent the structures of these complexes for predicting mutation effects.
    • Rests on an assumption: The deep sequencing enrichment values used as experimental measurements for CAPRI accurately reflect changes in binding affinity.

  • May combine two claims.

    Findingown data

    mCSM-PPI2 achieved a correlation of 0.63 and RMSE of 1.11 kcal/mol on the S472 subset not present in SKEMPI1.

    “Our method achieved a correlation of 0.63 (RMSE = 1.11 kcal/mol).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The experimental binding affinity changes in the SKEMPI2 dataset are accurate.

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

    Premisecitation

    mCSM was one of the first scalable computational tools to accurately predict the effects of mutations on binding affinity.

    “mCSM (20) was one of the first scalable computational tools to accurately predict the effects of mutations on binding affinity.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [20] Pires, 2014, mCSM: Predicting the effects of mutations in proteins using graph-based signatures, Bioinformatics source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The use of hypothetical reverse mutations improves the performance and robustness of the mCSM-PPI2 predictive model.

    “These results corroborate the use of reverse mutations in order to improve performance and robustness of our predictive model.”
    • Also stated in: § CONCLUSION · boosted
      “In addition, the use of hypothetical reverse mutations has shown to improve the robustness of our predictive model allowing for a more balanced prediction.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The effect of a reverse mutation is the negative of the effect of the original mutation (i.e., binding affinity changes are symmetric).

  • May combine two claims.

    Premisecitation

    Significant efforts have been devoted to computationally study the effects of mutations on protein complexes, but their poor predictive performance on new variants, particularly mutations that lead to increased binding affinity, has limited their use.

    “Since then, significant efforts have been devoted to computationally study the effects of mutations on protein complexes (24,25) but their poor predictive performance on new variants, particularly mutations that lead to increased binding affinity of the complex, has limited their use.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [24] Geng, 2019, iSEE: Interface structure, evolution, and energy-based machine learning predictor of binding affinity changes upon mutations, Proteins source

      Not checked in the free run.

    • cites [25] Li, 2016, MutaBind estimates and interprets the effects of sequence variants on protein–protein interactions, Nucleic Acids Res. source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    mCSM-PPI2 performs better than its previous version and other methods.

    “our updated method has shown to perform better than its previous version and other methods.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The experimental binding affinity changes in the SKEMPI2 dataset are accurate.
    • Rests on an assumption: The homology models generated for T55 and T56 from close homologues accurately represent the structures of these complexes for predicting mutation effects.
    • Rests on an assumption: The deep sequencing enrichment values used as experimental measurements for CAPRI accurately reflect changes in binding affinity.
    • Rests on an assumption: The effect of a reverse mutation is the negative of the effect of the original mutation (i.e., binding affinity changes are symmetric).
    • Rests on an assumption: The 378 alanine-scanning experimental mutations are accurate measures of binding energy changes.
10 of 10

  • The paper relies on this without support.

    Binding regions are evolutionarily conserved.

    “Binding regions are known to be evolutionarily conserved, which has been exploited in a variety of studies to identify potential protein interaction interfaces.”

  • The paper relies on this without support.

    The conformational flexibility of glycine side chains and the rigidity of proline side chains are important for defining the backbone flexibility.

    “The conformational flexibility of glycine side chains and the rigidity of proline side chains are important for defining the backbone flexibility.”

  • The paper relies on this without support.

    Mutations from and to glycine or proline can lead to large structural effects.

    “Mutations from and to these two amino acids can lead to large structural effects.”

  • The paper relies on this without support.

    The experimental binding affinity changes in the SKEMPI2 dataset are accurate.

    “using 90% of our original dataset (S8338) for training and the remaining as a blind test”
    “trained only on the original subset of mutations from SKEMPI2 (S4169)”

  • The paper relies on this without support.

    The homology models generated for T55 and T56 from close homologues accurately represent the structures of these complexes for predicting mutation effects.

    “Although the 3D structures for these two complexes were not available, structures of close homologues have been described (53,54) and were used for generating homology model structures by introducing point mutations using Modeller (55)”

  • The paper relies on this without support.

    The deep sequencing enrichment values used as experimental measurements for CAPRI accurately reflect changes in binding affinity.

    “The in vitro experimental measurements used the enrichment values generated from deep sequencing and were calculated based on the binary logarithm of the ratio of number of times the variant sequence was observed after and before the selection for binding.”

  • The paper relies on this without support.

    The effect of a reverse mutation is the negative of the effect of the original mutation (i.e., binding affinity changes are symmetric).

    “the hypothetical reverse mutations were kept in either training or test sets during the splits according to its counterpart original mutation.”

  • The paper relies on this without support.

    The 378 alanine-scanning experimental mutations are accurate measures of binding energy changes.

    “a previously proposed set of 378 alanine-scanning experimental mutations within 19 different protein–protein complexes (56,57)”
8 of 8

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.

  • Dataset increase exceeds the stated three-fold maximum

    Minor

    The stated 8338 mutations are about 7.56 times the cited 1102-mutation dataset, not an increase of up to three-fold. Even measured as the additional mutations relative to the earlier dataset, the increase is about 6.56-fold.

    Reported 3 · recomputed 7.566243

1 of 1

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Rodrigues, C. H. M., Myung, Y., Pires, D. E. V., & Ascher, D. B.
Affiliations
Baker Heart and Diabetes Institute (AU)
Corresponding author
  • Pires, Douglas E V
  • Ascher, David B
ORCID
Venue
Nucleic Acids Research
Publisher
Oxford University Press (OUP)
Published
2019-05-22
Volume
47
Issue
W1
Pages
W338-W344
ISSN
  • 0305-1048 (ISSN-L)
  • 1362-4962
DOI
10.1093/nar/gkz383
PMC
PMC6602427
OpenAlex
W2945403072
Semantic Scholar
c26f9b66730796c5ba3f86287c44fdbc482f0051
Cited by
369
References
59
Citations by year
2026: 33 · 2025: 43 · 2024: 51 · 2023: 65 · 2022: 61 · 2021: 69 · 2020: 42 · 2019: 3 · 2012: 1
FWCI
14.40
Type
journal article
Language
en
Open access
Licence
Locations
  • Nucleic Acids Research (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Minerva Access (University of Melbourne) (submittedVersion, cc-by)
  • Apollo (University of Cambridge) (submittedVersion, cc-by)
  • Apollo (cc-by)
Funders
  • Jack Brockhoff Foundation (JBF 4186)
  • Fundação de Amparo à Pesquisa do Estado de Minas Gerais (MR/M026302/1)
  • National Health and Medical Research Council (APP1072476)
  • University of Melbourne (UOM0017)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Genomics and Rare Diseases
Keywords
protein-protein interaction, missense mutations, binding affinity prediction, evolutionary information, web server
MeSH terms
Machine Learning, Protein Conformation, alpha-Helical, Protein Conformation, beta-Strand, Binding Sites, Humans, Protein Binding, Proteins, Software, Thermodynamics, Crystallography, X-Ray, Benchmarking, Mutation, Missense, Internet, Protein Interaction Mapping, Protein Interaction Domains and Motifs, Datasets as Topic

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