Deep learning based kcat prediction enables improved enzyme constrained model reconstruction
Feiran LiiD, Le YuaniD, Hongzhong LuiD, Gang LiiD, Yu CheniD, Martin K. M. EngqvistiD, Eduard Johannes KerkhoveniD, Jens NielseniD
Abstract
Enzyme turnover numbers (kcat values) are key parameters to understand cell metabolism, proteome allocation and physiological diversity, but experimentally measured kcat data are sparse and noisy. Here we provide a deep learning approach to predict kcat values for metabolic enzymes in a high-throughput manner with the input of substrate structures and protein sequences. Our approach can capture kcat changes for mutated enzymes and identify amino acid residues with great impact on kcat values. Furthermore, we applied the approach to predict genome scale kcat values for over 300 yeast species, demonstrating that the predicted kcat values are consistent with current evolutional understanding. Additionally, we designed an automatic pipeline using the predicted kcat values to parameterize enzyme-constrained genome scale metabolic models (ecGEMs) facilitated by a Bayesian approach, which outperformed the default ecGEMs in predicting phenotypes and proteomes and enabled to explain phenotype differences among yeast species. The deep learning kcat prediction approach and automatic ecGEM construction pipeline would thus be a valuable tool to uncover the global trend of enzyme kinetics and physiological diversity, and to further elucidate cell metabolism on a large scale.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 466
- FWCI
- 27.49
- Field percentile
- top 1%
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (8 Aug 2021) · CC BY 4.0
A substrate graph neural network plus protein-sequence CNN predicts enzyme kcat (test Pearson r 0.71), tracks mutation effects, and parameterizes enzyme-constrained models for 343 yeast/fungi species.
Measured kcat values in BRENDA and SABIO-RK are sparse and assay-dependent, so enzyme-constrained genome-scale metabolic models (ecGEMs) borrow values from other substrates or organisms; an earlier machine-learning predictor needed feature and absolute proteome data and was restricted to E. coli. A predictor needing only protein sequence and substrate SMILES allows automatic ecGEM construction for poorly characterized species.
Key results
ShownThe GNN-CNN model predicts in vitro kcat to within roughly one order of magnitude, holds up on test entries whose substrates and enzymes were absent from training, and ranks preferred above alternative substrates of promiscuous enzymes.
RMSE 0.99 (validation) and 1.06 (test); Pearson r 0.88 on the whole 16,838-entry dataset, 0.71 on the test set, 0.70 on test entries with substrates and enzymes unseen in training. Predicted median kcat 6.45 /s for preferred vs 1.49 /s for alternative substrates (P < 1e-10; test set P = 0.009). · Construction of a deep learning framework for…; Deep learning model performance for kcat…
ShownPredictions capture amino acid substitution effects: mutant kcat is predicted at least as well as wildtype, mutants with decreased kcat score lower than wildtype-like ones, and attention weights mark residues whose mutation lowers kcat in human PNP.
Pearson r 0.87 (whole) / 0.65 (test) for wildtype and 0.90 / 0.78 for mutants; r = 0.94 on enzyme-substrate pairs with at least 25 substitutions each; for human PNP with inosine, decreased-kcat mutation sites had higher attention weights than wildtype-like sites (P = 0.0014). · Prediction and interpretation of kcat of mutated…; Figure legends
ShownBayesian calibration by sequential Monte Carlo approximate Bayesian computation (SMC-ABC) against growth data mostly narrows kcat variance rather than shifting means, and held-out S. cerevisiae phenotype error falls alongside training error.
S. cerevisiae (50% train, 50% test): after 30 generations RMSE 0.5 training vs 1 test. In S. cerevisiae, variance fell for 2,644 enzyme-substrate pairs vs mean changed for 146 (Šidák-adjusted P < 0.01); Y. lipolytica 2,721 vs 159. Posterior vs deep learning kcat Pearson r = 0.83 in both species. · Bayesian approach for 343 ecGEMs reconstruction; Statistical tests for comparison between sampled…
ShownDeep-learning-parameterized ecGEMs raise kcat coverage over database-matched Classical-ecGEMs and better predict exchange rates and proteomes, with Bayesian posterior-mean models performing best.
Across 343 species, Classical-ecGEMs cover ca. 40% of enzymes and ca. 60% of enzymatic reactions vs ca. 80% and ca. 90% for DL-ecGEMs and Posterior-mean-ecGEMs. Against proteomics from three species, DL-ecGEMs lowered RMSE by 30% vs Classical-ecGEMs, and Posterior-mean-ecGEMs gave the lowest RMSE. · Deep learning and Bayesian approach improve…; Figure legends
SuggestedPosterior-mean ecGEMs mostly reproduce a protein-efficiency explanation of the Crabtree effect across yeasts, and predicted kcats point to pyruvate kinase and citrate synthase as phenotype-associated enzymes.
102 yeast species (25 Crabtree positive, 77 negative): 89% follow the expected pattern (higher low-yield pathway efficiency in positive species, higher high-yield efficiency in negative ones). Pyruvate kinase kcat higher in positive species (P = 0.009); citrate synthase higher in negative species (P = 0.008); two-sided Wilcoxon tests. · kcat profile comparison enables to identify…; Figure legends
Methods, data and reproducibility
Substrate SMILES become molecular graphs encoded by a GNN using r-radius subgraphs, and protein sequences become overlapping 3-gram words encoded by a CNN; the concatenated vectors regress kcat. Predicted kcats set log-normal priors that SMC-ABC updates against measured growth and exchange rates to build ecGEMs.
- kcat entries merged from BRENDA and SABIO-RK (queried 10 July 2020) with PubChem SMILES and UniProt sequences: 7,822 sequences, 2,672 substrates, 851 organisms · multiple (851 organisms) · n = 16,838 entries (9,411 wildtype, 7,427 mutant)
- Enzyme-substrate pairs from literature, BRENDA and SABIO-RK, each with kcat for at least 25 amino acid substitutions (e.g., DHFR, PNP, GGPPs) · multiple; PNP is Homo sapiens
- Published GEMs for 332 yeast species plus 11 outgroup fungi, with per-species protein FASTA files, used as prediction input · yeasts and fungi · n = 343 species; around three million protein-substrate pairs
- Literature-curated batch and chemostat growth and exchange-rate measurements on 16 carbon sources (Supplementary Table 5) · yeast/fungi (76 species) · n = 445 entries
- Published quantitative proteomics (absolute, or relative normalized to absolute) across carbon sources, culture modes and media · yeast/fungi (species not named in text) · n = three species
- Yeast species with experimentally reported Crabtree phenotype, compared via Posterior-mean-ecGEMs and predicted kcats · yeast · n = 102 species (25 positive, 77 negative)
- Single-copy ortholog groups from the annotated genomes, used for gene-level dN/dS versus predicted kcat · yeasts and fungi · n = 343 species (332 yeasts in 12 clades + 11 outgroup fungi)
- Supplementary dataset: protein FASTA files, predicted kcats, and Classical-, DL- and Posterior-mean-ecGEMs · yeasts and fungi · n = 343 species · 10.5281/zenodo.5164210
- Methods & tools
- GNN-CNN kcat predictor with neural attention (DLKcat repository), PyTorch v1.4.0 (CUDA/10.1.243), RDKit, BRENDA, SABIO-RK, PubChem, UniProt, KEGG, MetaNetX, PAML v4.7 (yn00), GECKO and sMOMENT ecGEM formats, SMC-ABC (sequential Monte Carlo approximate Bayesian computation)
- Code
- https://github.com/SysBioChalmers/DLKcathttps://github.com/SysBioChalmers/DLKcat/tree/master/DeeplearningApproach/Code/example
- Data & accessions
- https://github.com/SysBioChalmers/DLKcat/tree/master/DeeplearningApproach/Data/databasehttps://github.com/SysBioChalmers/DLKcat/tree/master/BayesianApporach/Data/Proteome_ref.xlsxhttps://doi.org/10.5281/zenodo.5164210
- Software & environment
- PyTorch v1.4.0 through the Python interface under CUDA/10.1.243; PAML v4.7 yn00 for dN/dS; sMOMENT ecGEM format during the SMC-ABC search and GECKO format for evaluation.
- Reproducibility gaps
- Source GEMs for the 343 species are cited only as a GitHub repository with no URL; growth phenotype and Crabtree compilations are in supplementary tables rather than a deposited dataset.
Limitations (6)
- Training kcats were mostly measured in vitro, so the model outputs in vitro-like values that may differ substantially from in vivo turnover; the Bayesian step is introduced to compensate. · stated by the authors
- No kcat can be predicted for generic substrates that lack SMILES, which accounts for the remaining coverage gaps in DL-ecGEMs and Posterior-mean-ecGEMs. · stated by the authors
- Too few mutants with kcat above 2-fold of wildtype existed to define an increased-kcat group, so predictions for activity-enhancing substitutions were not evaluated. · stated by the authors
- Species whose HY/LY protein-efficiency ratio does not follow the Crabtree phenotype may reflect regulation absent from ecGEMs; fumarase and phosphoglucose isomerase differences were not interpreted because these enzymes are reversible. · stated by the authors
- Entries were split randomly 80/10/10, so point mutants and close homologs of training enzymes can sit in the test set; the 'unseen' subset (r 0.70) appears to exclude only exact matches, and the headline r 0.88 includes training data. · noted in this summary
- Growth data cover 76 species, yet Posterior-mean-ecGEMs are reported for all 343 without explaining how posteriors were set for unmeasured species; phenotype comparisons also reuse data that informed the posterior. · noted in this summary
Who should read:Read this if you build enzyme-constrained metabolic models for poorly characterized organisms or want kcat predictions, including mutation effects, from only protein sequence and substrate SMILES.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- KcatNet: A Geometric Deep Learning Framework for Genome-Wide Prediction of Enzyme Catalytic Efficiency
Tong Pan, Xin Cui, … Jiangning Song·bioRxiv·Mar 9, 2025
- EnzymeTuning: a Deep-learning-based Toolbox for Optimizing Enzyme-constrained Metabolic Modeling with Enhanced Proteome Abundance Prediction
Xueting Wang, Yingping Zhuang, … Hongzhong Lu·bioRxiv·Aug 23, 2025
- kinGEMs: A Robust and Scalable Framework for Resource-Constraint Models through Stochastic Tuning of Deep Learning-Predicted Kinetic Parameters
Rana Ahmed Barghout, Lya Chinas, … Radhakrishnan Mahadevan·bioRxiv·Oct 22, 2026
- ENKIE: A package for predicting enzyme kinetic parameter values and their uncertainties
Mattia G. Gollub, Thierry Backes, … Jörg Stelling·BioinformaticsPeer-reviewed·Oct 30, 2024
- Pretrained language models and weight redistributionachieve precise kcat prediction
- DLKcat cannot predict meaningful kcat values for mutants and unfamiliar enzymes
Alexander Kroll, Martin J. Lercher·Biology Methods and ProtocolsPeer-reviewed·Jan 1, 2024