Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat
Sign Up
PreprintCited by 2Protein Structure and DynamicsMachine Learning in BioinformaticsGenomics and Rare Diseases

The DOI matches its Crossref record; its reference list is not available.

Data-efficient protein mutational effect prediction with weak supervision by molecular simulation and protein language models

Open access
Deguchi, Teppei, Ab Ghani, Nur Syatila, Kurumida, Yoichi, Iida, Shinji, Kobayashi, Kaito, Saito, Yutaka·2025·Briefings in Bioinformatics
DOI 10.1101/2025.04.08.647800PMC PMC12513171

Preprint — published as 10.1093/bib/bbaf536 · Paper page

Open this paper in CoChat — free accountOpen PDF

Abstract

Machine learning-based protein mutational effect prediction is widely used in protein engineering and pathogenicity prediction, but training data scarcity remains a major challenge due to high costs of experimental measurements. A previous study proposed data augmentation using computational estimates by molecular simulation. However, this approach has been limited to predicting mutational effects on thermostability. Here, we present a new data augmentation method that combines molecular simulation with zero-shot prediction computed by protein language models. These computational estimates serve as “weak” training data to supplement experimental training data. Our method dynamically adjusts the weight and inclusion of weak training data based on available experimental training data. This reduces potential negative impacts of weak training data while extending applicability to diverse protein properties such as binding affinity and enzymatic activity. Benchmark tests demonstrate that our method improves prediction accuracy particularly when experimental training data are scarce. These results indicate the capability of our approach to advance protein engineering and pathogenicity prediction in small data regimes.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1101/2025.04.08.647800

Cites

Statistics

Figures

Figure 1 (a) Overview of the data-efficient protein mutational effect prediction method. Mutational effects are estimated by molecular simulation and pLM’s zero-shot prediction. These computational estimates are used as weak training data integrated with experimental training data for ML models. (b) The weight of weak training data relative to experimental training data are determined by the weight adjustment algorithm. (c) The decision algorithm determines whether or not to include weak training data based on cross validation.
Figure 1 (a) Overview of the data-efficient protein mutational effect prediction method. Mutational effects are estimated by molecular simulation and pLM’s zero-shot prediction. These computational estimates are used as weak training data integrated with experimental training data for ML models. (b) The weight of weak training data relative to experimental training data are determined by the weight adjustment algorithm. (c) The decision algorithm determines whether or not to include weak training data based on cross validation.CC BY · PubMed Central
Figure 2 Impact of weak supervision with differing amounts of available experimental training data. Prediction accuracy measured by Spearman’s correlation coefficient (CC) was compared between ML models trained using experimental data only (dashed line), those integrated with computational estimates by Rosetta, ESM-2, and their hybrid score (solid lines). As the baseline, prediction accuracy obtained by computational estimates alone was shown (dotted lines). esm2_t6_8M_UR50D was used for embedding. Single-residue mutants were used for both training and test splits. (a) APH(3’)II enzymatic activity. (b) CYP2C9 enzymatic activity. (c) DAOx enzymatic activity. (d) CcdB cytotoxicity. (e) GRB2-SH3 binding affinity. (f) GRB2-SH3 protein abundance. (g) PSD95-PDZ binding affinity. (h) PSD95-PDZ protein abundance. (i) GB1 binding affinity. (j) KRAS protein abundance. (k) CreiLOV fluorescence. (l) DHFR enzymatic activity.
Figure 2 Impact of weak supervision with differing amounts of available experimental training data. Prediction accuracy measured by Spearman’s correlation coefficient (CC) was compared between ML models trained using experimental data only (dashed line), those integrated with computational estimates by Rosetta, ESM-2, and their hybrid score (solid lines). As the baseline, prediction accuracy obtained by computational estimates alone was shown (dotted lines). esm2_t6_8M_UR50D was used for embedding. Single-residue mutants were used for both training and test splits. (a) APH(3’)II enzymatic activity. (b) CYP2C9 enzymatic activity. (c) DAOx enzymatic activity. (d) CcdB cytotoxicity. (e) GRB2-SH3 binding affinity. (f) GRB2-SH3 protein abundance. (g) PSD95-PDZ binding affinity. (h) PSD95-PDZ protein abundance. (i) GB1 binding affinity. (j) KRAS protein abundance. (k) CreiLOV fluorescence. (l) DHFR enzymatic activity.CC BY · PubMed Central
Figure 3 Ablation experiments for the weight adjustment algorithm and the decision algorithm for data inclusion. The prediction accuracy of ML model was compared among the hybrid score with both methods (hybrid), without the weight adjustment algorithm, without the decision algorithm, and without both methods. esm2_t6_8M_UR50D was used for embedding. Single-residue mutants were used for both training and test splits. (a) APH(3’)II enzymatic activity. (b) PSD95-PDZ binding affinity.
Figure 3 Ablation experiments for the weight adjustment algorithm and the decision algorithm for data inclusion. The prediction accuracy of ML model was compared among the hybrid score with both methods (hybrid), without the weight adjustment algorithm, without the decision algorithm, and without both methods. esm2_t6_8M_UR50D was used for embedding. Single-residue mutants were used for both training and test splits. (a) APH(3’)II enzymatic activity. (b) PSD95-PDZ binding affinity.CC BY · PubMed Central
Figure 4 Prediction accuracy for double-residue mutational effects only using single-residue mutants as training data. ML models trained using experimental data only (dashed line), those integrated with computational estimates by Rosetta, ESM-2, and their hybrid score (solid lines) were compared. As the baseline, prediction accuracy obtained by computational estimates alone was shown (dotted lines). esm2_t6_8M_UR50D was used for amino acid embedding. (a) GRB2-SH3 binding affinity (b) GRB2-SH3 protein abundance. (c) PSD95-PDZ binding affinity. (d) PSD95-PDZ protein abundance. (e) GB1 binding affinity.
Figure 4 Prediction accuracy for double-residue mutational effects only using single-residue mutants as training data. ML models trained using experimental data only (dashed line), those integrated with computational estimates by Rosetta, ESM-2, and their hybrid score (solid lines) were compared. As the baseline, prediction accuracy obtained by computational estimates alone was shown (dotted lines). esm2_t6_8M_UR50D was used for amino acid embedding. (a) GRB2-SH3 binding affinity (b) GRB2-SH3 protein abundance. (c) PSD95-PDZ binding affinity. (d) PSD95-PDZ protein abundance. (e) GB1 binding affinity.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Self-check

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

The self-check needs the paper’s full text.

Citations

No citation lists for this paper

No citation lists for this paper yet — checked again daily.

References 59 · Cited by 2

Research works printed but not linked: 55

  1. 10.1038/s41592-021-01100-y · DOI · Paper page
  2. 10.1021/acssynbio.8b00155 · DOI
  3. 10.1016/j.copbio.2022.102713 · DOI
  4. 10.1186/1471-2105-7-405 · DOI
  5. 10.1007/s00726-012-1407-7 · DOI
  6. 10.1016/j.sbi.2023.102600 · DOI
  7. 2019, Machine-learning-guided directed evolution for protein engineering, Nature methods, 16, 687, 10.1038/s41592-019-0496-6 · DOI · Paper page
  8. 10.1021/acscatal.3c02743 · DOI
  9. 10.1021/acscatal.9b04321 · DOI
  10. 10.1093/bioinformatics/bty238 · DOI
  11. 10.1016/B978-0-12-381270-4.00019-6 · DOI
  12. 10.1126/science.ade2574 · DOI · Paper page
  13. 2021, Language models enable zero-shot prediction of the effects of mutations on protein function, Advances in neural information processing systems, 34, 29287
  14. 10.1038/s41586-022-04586-4 · DOI
  15. 10.1093/nar/gku511 · DOI
  16. 10.1016/j.ajhg.2021.07.001 · DOI
  17. 10.1038/s41467-024-45630-3 · DOI
  18. 10.1093/molbev/msw182 · DOI
  19. 10.1038/s41586-023-06954-0 · DOI
  20. 10.1021/acssynbio.2c00662 · DOI
  21. 10.1038/s41467-024-47671-0 · DOI
  22. 2023, Predicting a protein’s stability under a million mutations, Advances in Neural Information Processing Systems, 36, 76229
  23. 10.1038/s41467-024-51776-x · DOI · Paper page
  24. 2024, Improving the prediction of protein stability changes upon mutations by geometric learning and a pre-training strategy, Nat Comput Sci, 4, 840, 10.1038/s43588-024-00716-2 · DOI · Paper page
  25. Savojardo C , Manfredi M , Martelli PL , Casadio R. DDGemb: predicting protein stability change upon single- and multi-point variations with embeddings and deep learning. Bioinformatics. 2024;41(1). · DOI
  26. 10.1126/science.adg7492 · DOI · Paper page
  27. 2025, Structure-based self-supervised learning enables ultrafast protein stability prediction upon mutation, Innovation (Camb), 6, 100750
  28. 10.1093/nar/gki387 · DOI · Paper page
  29. 10.1517/17460441.2015.1032936 · DOI
  30. 2023, Proteingym: Large-scale benchmarks for protein fitness prediction and design, Advances in Neural Information Processing Systems, 36, 64331
  31. 10.1002/prot.22921 · DOI
  32. 2018, Flex ddG: Rosetta ensemble-based estimation of changes in protein–protein binding affinity upon mutation, The Journal of Physical Chemistry B, 122, 5389, 10.1021/acs.jpcb.7b11367 · DOI · Paper page
  33. 10.1093/nar/28.1.235 · DOI · Paper page
  34. 10.1016/j.str.2009.03.017 · DOI
  35. 10.1016/S0092-8674(00)81307-0 · DOI · Paper page
  36. 10.1016/S0969-2126(01)00157-5 · DOI · Paper page
  37. 10.1016/S0022-2836(03)00121-9 · DOI
  38. 10.1074/jbc.M405427200 · DOI · Paper page
  39. 10.1038/s41467-021-21422-x · DOI
  40. 10.1006/jmbi.1998.2395 · DOI
  41. 10.1073/pnas.97.23.12463 · DOI
  42. 10.1021/bi962337c · DOI
  43. 2016, Comparative protein structure modeling using MODELLER, Current protocols in bioinformatics, 54, 5
  44. 10.1089/cmb.2008.0173 · DOI
  45. 10.1111/J.1747-0285.2008.00641.X · DOI
  46. 10.1021/ci980211b · DOI
  47. 10.1016/j.molstruc.2006.07.004 · DOI
  48. 10.1007/s00726-009-0287-y · DOI
  49. 1998, New chemical descriptors relevant for the design of biologically active peptides, A multivariate characterization of 87 amino acids. Journal of medicinal chemistry, 41, 2481
  50. 2005, A new set of amino acid descriptors and its application in peptide QSARs, Peptide Science: Original Research on Biomolecules, 80, 775, 10.1002/bip.20296 · DOI
  51. 2013, Benchmarking of protein descriptor sets in proteochemometric modeling (part 2): modeling performance of 13 amino acid descriptor sets, Journal of cheminformatics, 5, 1
  52. 2001, Random forests, Machine learning, 45, 5, 10.1023/A:1010933404324 · DOI · Paper page
  53. 10.1023/B:STCO.0000035301.49549.88 · DOI · Paper page
  54. 10.1016/j.jmp.2018.03.001 · DOI · Paper page
  55. 2011, Scikit-learn: Machine learning in Python, the Journal of machine Learning research, 12, 2825

Metadata

Authors
Deguchi, T., Ab Ghani, N. S., Kurumida, Y., Iida, S., Kobayashi, K., & Saito, Y.
Affiliations
  • Artificial Intelligence Research Center (JP)
  • School of Frontier Engineering, Kitasato University, 1-15-1 Kitazato, Minami-ku, Sagamihara, Kanagawa 252-0373, Japan
  • Kitasato University (JP)
  • Graduate School of Frontier Sciences, The University of Tokyo, 5-1-5, Kashiwanoha, Kashiwa, Chiba 277-0882, Japan
ORCID
Venue
Briefings in Bioinformatics
Publisher
openRxiv
Published
2025-04-14
Volume
26
Issue
5
ISSN
1467-5463 (ISSN-L)
DOI
10.1101/2025.04.08.647800
PMC
PMC12513171
OpenAlex
W4415046037
Cited by
2
References
59
Citations by year
2026: 2
FWCI
0.60
Type
journal article
Language
en
Open access
Licence
Locations
  • Briefings in Bioinformatics (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
  • Japan Agency for Medical Research and Development (JP19ak0101122)
  • National Institute of Advanced Industrial Science and Technology
  • Japan Association for Chemical Innovation
  • Japan Society for the Promotion of Science London (22H03691)
  • New Energy and Industrial Technology Development Organization (JPNP14004)
  • Japan Society for the Promotion of Science (22H03691, JPMJSP2108, JPNP14004)
  • Cybermedia Center, Osaka University
  • Institute of Genetics
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Genomics and Rare Diseases
Keywords
weakly supervised learning, data augmentation, molecular simulation, protein language models, zero-shot prediction, data scarcity, protein thermal stability, binding affinity prediction, pathogenicity prediction, protein engineering
MeSH terms
Machine Learning, Mutation, Proteins, Computational Biology, Molecular Dynamics Simulation

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.

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

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Data-efficient protein mutational effect prediction with weak supervision by molecular simulation and protein language models. https://app.cochat.ai/papers/10.1101/2025.04.08.647800
The paper
Deguchi, T., Ab Ghani, N. S., Kurumida, Y., Iida, S., Kobayashi, K., & Saito, Y. (2025). Data-efficient protein mutational effect prediction with weak supervision by molecular simulation and protein language models. Briefings in Bioinformatics. https://doi.org/10.1101/2025.04.08.647800

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.