Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 26 of the 26 references it cites were found in a registry or an index.

Enhancing kcat prediction through residue-aware attention mechanism and pre-trained representations

Open access

Published version of a preprint

Abstract

The turnover number (kcat) is a critical parameter in enzyme kinetics, representing the maximum number of substrate molecules an enzyme can convert to product per unit time under optimal conditions. While kcat reflects an enzyme's catalytic efficiency and plays a central role in understanding enzyme activity, existing computational models for predicting this parameter often face significant limitations. Most models rely solely on substrate information, overlooking the integral role of the reaction itself in predicting kcat. The only model that attempts to incorporate reaction information, TurNuP, still falls short in both predictive accuracy and interpretability. To address these challenges, we propose PMAK, a deep learning model that leverages pre-trained representation learning and a residue-aware attention mechanism. By incorporating both enzyme sequences and reaction details, PMAK generates robust representations, effectively capturing the complex interplay between enzyme and reaction, and offering superior accuracy and generalization compared to existing methods. Specifically, PMAK achieved an average R2 improvement of approximately 16.9% and 10.0% over TurNup under new reaction setting and new enzyme setting in five-fold cross-validation, respectively. Besides, our approach provides valuable insights into the key features driving enzyme catalysis, making it a more interpretable solution for enzymatic reaction prediction.

Overview

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

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

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

Integrity signals

Registry

Crossref · 10.1038/s42003-026-09551-9

Cites

26 of 26 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 The overview of PMAK. The model generates embeddings from enzyme amino acid sequences and reaction SMILES using two pre-trained models ProT5 and RXNFP. These embeddings are then processed through a convolutional neural network and a fully connected network to extract features and align the dimensions of both representations. A residue-aware attention block is subsequently employed to capture the feature interactions between the enzyme and reaction representations. Finally, after applying max pooling along the residue dimension to the enzyme representation, both embeddings are concatenated and passed through a multilayer perceptron for kcat value prediction.
Fig. 1 The overview of PMAK. The model generates embeddings from enzyme amino acid sequences and reaction SMILES using two pre-trained models ProT5 and RXNFP. These embeddings are then processed through a convolutional neural network and a fully connected network to extract features and align the dimensions of both representations. A residue-aware attention block is subsequently employed to capture the feature interactions between the enzyme and reaction representations. Finally, after applying max pooling along the residue dimension to the enzyme representation, both embeddings are concatenated and passed through a multilayer perceptron for kcat value prediction.CC BY-NC-ND · PubMed Central
Fig. 2 Performance of PMAK under warm-start conditions. a. Pearson correlation coefficient(PCC) for PMAK on the test set of Warm-start experiments. b.Comparison of the coefficient of determination (R2) between PMAK, TurNup, CatPred, CataPro, EITLEM, and Unikp on the test sets, generated through five random splits under warm-start conditions (n = 5 independent experiments).
Fig. 2 Performance of PMAK under warm-start conditions. a. Pearson correlation coefficient(PCC) for PMAK on the test set of Warm-start experiments. b.Comparison of the coefficient of determination (R2) between PMAK, TurNup, CatPred, CataPro, EITLEM, and Unikp on the test sets, generated through five random splits under warm-start conditions (n = 5 independent experiments).CC BY-NC-ND · PubMed Central
Fig. 3 Model performance under enzyme and reaction cold-start settings. a, b Comparison of coefficient of determination(R2) and mean squared error (MSE) for PMAK and TurNup on the test set under the cold-start experiments for reactions. c, d Comparison of coefficient of determination(R2) and mean squared error (MSE) for PMAK, TurNup, CatPred, CataPro, EITLEM, and Unikp on the test set under the cold-start experiments for enzymes. Boxplots present the results of fivefold cross-validation on the training set, using the optimal hyperparameter configurations (n = 5 independent experiments). Triangular markers represent the performance of test set, obtained from models trained on the complete training set. White horizontal lines indicate the median values of the data points. e Comparison of coefficient of determination(R2) on the test set for different levels of maximal reaction similarities under the cold-start experiments for reactions. f Comparison of coefficient of determination(R2) on the test set for different levels of maximal enzyme sequence identity under the cold-start experiments for enzymes. The numbers in ‘[]’ represent the sample size for each interval.
Fig. 3 Model performance under enzyme and reaction cold-start settings. a, b Comparison of coefficient of determination(R2) and mean squared error (MSE) for PMAK and TurNup on the test set under the cold-start experiments for reactions. c, d Comparison of coefficient of determination(R2) and mean squared error (MSE) for PMAK, TurNup, CatPred, CataPro, EITLEM, and Unikp on the test set under the cold-start experiments for enzymes. Boxplots present the results of fivefold cross-validation on the training set, using the optimal hyperparameter configurations (n = 5 independent experiments). Triangular markers represent the performance of test set, obtained from models trained on the complete training set. White horizontal lines indicate the median values of the data points. e Comparison of coefficient of determination(R2) on the test set for different levels of maximal reaction similarities under the cold-start experiments for reactions. f Comparison of coefficient of determination(R2) on the test set for different levels of maximal enzyme sequence identity under the cold-start experiments for enzymes. The numbers in ‘[]’ represent the sample size for each interval.CC BY-NC-ND · PubMed Central
Fig. 4 Functional residue enrichment revealed by residue-aware attention. a Hit rate comparison across different top-k% residue thresholds (1–20%). The blue line represents the attention-based method, the gray line denotes the random baseline, and the dashed line indicates pooled-embedding results. Asterisks mark statistical significance from permutation tests: *p < 0.05, **p < 0.01, ***p < 0.001, while “ns” indicates non-significant differences (p≥0.05) (n = 1000 independent experiments). b Hit rate of top 10% attention residues under an expanded structural definition of functional sites, as the required number of matches per enzyme increases.
Fig. 4 Functional residue enrichment revealed by residue-aware attention. a Hit rate comparison across different top-k% residue thresholds (1–20%). The blue line represents the attention-based method, the gray line denotes the random baseline, and the dashed line indicates pooled-embedding results. Asterisks mark statistical significance from permutation tests: *p < 0.05, **p < 0.01, ***p < 0.001, while “ns” indicates non-significant differences (p≥0.05) (n = 1000 independent experiments). b Hit rate of top 10% attention residues under an expanded structural definition of functional sites, as the required number of matches per enzyme increases.CC BY-NC-ND · PubMed Central
Fig. 5 Attention score enrichment at binding and active sites. a,b Comparison of attention scores between binding/active sites and general sites in AdoMetDC (UniProt ID: Q9WZC3 ; n = 130 residues) and HTPA synthase (UniProt ID: Q6GH13 ; n = 295 residues). The table summarizes the amino acids of active and binding sites from UniProt alongside those predicted by EasIFA. Boxplots and line charts compare attention scores for ’General sites’ and ’Binding/Active sites’. P values, derived from a one-tailed independent samples t test, indicate significantly higher attention scores for ’Binding/Active sites’. Green dots in the line chart represent attention scores for binding/active sites. The 3D structure comparison highlights the spatial relationship between binding/active sites and high-weight sites (top 5% residues with the highest attention scores). Red, blue, and green markers represent binding/active site amino acids, high-weight site amino acids, and overlapping residues, respectively.
Fig. 5 Attention score enrichment at binding and active sites. a,b Comparison of attention scores between binding/active sites and general sites in AdoMetDC (UniProt ID: Q9WZC3 ; n = 130 residues) and HTPA synthase (UniProt ID: Q6GH13 ; n = 295 residues). The table summarizes the amino acids of active and binding sites from UniProt alongside those predicted by EasIFA. Boxplots and line charts compare attention scores for ’General sites’ and ’Binding/Active sites’. P values, derived from a one-tailed independent samples t test, indicate significantly higher attention scores for ’Binding/Active sites’. Green dots in the line chart represent attention scores for binding/active sites. The 3D structure comparison highlights the spatial relationship between binding/active sites and high-weight sites (top 5% residues with the highest attention scores). Red, blue, and green markers represent binding/active site amino acids, high-weight site amino acids, and overlapping residues, respectively.CC BY-NC-ND · PubMed Central
Fig. 6 Prediction accuracy of absolute proteome allocation. Mean squared errors (MSE) for the prediction of absolute proteome data compared to experimental data.
Fig. 6 Prediction accuracy of absolute proteome allocation. Mean squared errors (MSE) for the prediction of absolute proteome data compared to experimental data.CC BY-NC-ND · PubMed Central
Fig. 7 Generalization performance on held-out and out-of-distribution datasets. a Comparison of R2 values for PMAK and baseline methods on the CatPred-DB. Solid bars represent results on the held-out test set, while hatched bars represent results on the out-of-distribution (OOD) test set. b Pearson Correlation Coefficient (PCC) values under the same evaluation settings, with solid bars indicating held-out results and hatched bars indicating OOD results.
Fig. 7 Generalization performance on held-out and out-of-distribution datasets. a Comparison of R2 values for PMAK and baseline methods on the CatPred-DB. Solid bars represent results on the held-out test set, while hatched bars represent results on the out-of-distribution (OOD) test set. b Pearson Correlation Coefficient (PCC) values under the same evaluation settings, with solid bars indicating held-out results and hatched bars indicating OOD results.CC BY-NC-ND · PubMed Central
Fig. 8 Performance on deep mutational scanning datasets. Comparison of Spearman correlation coefficients (SCC) for PMAK and baseline methods across four mutant enzyme datasets (SsIGPS, TmIGPS, TtIGPS, and HIS3). All models were trained on the CatPred-DB containing only wild-type enzymes and directly tested on mutant datasets without fine-tuning. PMAK consistently demonstrates superior or comparable performance, highlighting its robustness and transferability to mutation landscapes.
Fig. 8 Performance on deep mutational scanning datasets. Comparison of Spearman correlation coefficients (SCC) for PMAK and baseline methods across four mutant enzyme datasets (SsIGPS, TmIGPS, TtIGPS, and HIS3). All models were trained on the CatPred-DB containing only wild-type enzymes and directly tested on mutant datasets without fine-tuning. PMAK consistently demonstrates superior or comparable performance, highlighting its robustness and transferability to mutation landscapes.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

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

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

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

Cites

Show all 26 references

Research works printed but not linked: 11

  1. Stephanopoulos, G., Nielsen, J. & Lee, S. Y. (Eds.). Metabolic Engineering: Concepts and Applications. (Wiley - VCH: Weinheim, Germany, 2021).
  2. Heckmann, D. et al. Machine learning applied to enzyme turnover numbers reveals protein structural correlates and improves metabolic models. Nat. Commun. 9, 5252 (2018). · DOI · Paper page
  3. Yu, H., Deng, H., He, J., Keasling, J. D. & Luo, X. Unikp: A unified framework for the prediction of enzyme kinetic parameters. Nat. Commun. 14, 8211 (2023). · DOI · Paper page
  4. Wang, Z. et al. Robust enzyme discovery and engineering with deep learning using catapro. Nat. Commun. 16, 2736 (2025). · DOI
  5. Kroll, A., Rousset, Y., Hu, X.-P., Liebrand, N. A. & Lercher, M. J. Turnover number predictions for kinetically uncharacterized enzymes using machine and deep learning. Nat. Commun. 14, 4139 (2023). · DOI · Paper page
  6. Kroll, A., Ranjan, S., Engqvist, M. K. & Lercher, M. J. A general model to predict small molecule substrates of enzymes based on machine and deep learning. Nat. Commun. 14, 2787 (2023). · DOI · Paper page
  7. Meier, J. et al. Language models enable zero-shot prediction of the effects of mutations on protein function. Adv. Neural Inf. Process. Syst. 34, 29287–29303 (2021).
  8. Lin, Z. et al. Language models of protein sequences at the scale of evolution enable accurate structure prediction. BioRxiv 2022, 500902 (2022).
  9. Landrum, G. Rdkit documentation. Release 1, 4 (2013).
  10. Kenton, J. D. M.-W. C. & Toutanova, L. K. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, vol. 1, 2 (Minneapolis, Minnesota, 2019).
  11. Chan, Y. H., Venev, S. V., Zeldovich, K. B. & Matthews, C. R. Correlation of fitness landscapes from three orthologous tim barrels originates from sequence and structure constraints. Nat. Commun. 8, 14614 (2017). · DOI

Metadata

Authors
Cai, Y., Ge, F., Zhang, C., Chen, H., Qi, X., Liao, X., & Wang, L.
Affiliations
  • Tianjin University of Science and Technology (CN)
  • University of Science and Technology of China (CN)
  • Chinese Academy of Sciences (CN)
Corresponding author
  • Liao, Xiaoping
  • Wang, Lin
Venue
Communications Biology
Published
2026-01-22
Volume
9
Issue
1
Pages
273-273
ISSN
2399-3642 (ISSN-L)
OpenAlex
W7125399703
Cited by
2
References
26
Citations by year
2026: 2
FWCI
4.13
Type
journal article
Language
en
Open access
Licence
Locations
  • Communications Biology (publishedVersion, cc-by-nc-nd)
  • Research Square (acceptedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Natural Science Foundation of China (12326611, 61603273)
  • Chinese Academy of Sciences (XDC0120201)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Enzyme Catalysis and Immobilization
Keywords
enzyme kinetics, catalytic efficiency, pre-trained representations
MeSH terms
Deep Learning, Enzymes, Kinetics, Computational Biology

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: Enhancing kcat prediction through residue-aware attention mechanism and pre-trained representations. https://app.cochat.ai/papers/10.1038/s42003-026-09551-9
The paper
Cai, Y., Ge, F., Zhang, C., Chen, H., Qi, X., Liao, X., & Wang, L. (2026). Enhancing kcat prediction through residue-aware attention mechanism and pre-trained representations. Communications Biology. https://doi.org/10.1038/s42003-026-09551-9

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.