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

The DOI matches its Crossref record; 77 of the 77 references it cites were found in a registry or an index; 44 of the 59 sources its claims cite were checked for support.

Deep learning-based kcat prediction enables improved enzyme-constrained model reconstruction

Open access

Published version of a preprint

Abstract

Enzyme turnover numbers ( k cat ) are key to understanding cellular metabolism, proteome allocation and physiological diversity, but experimentally measured k cat data are sparse and noisy. Here we provide a deep learning approach (DLKcat) for high-throughput k cat prediction for metabolic enzymes from any organism merely from substrate structures and protein sequences. DLKcat can capture k cat changes for mutated enzymes and identify amino acid residues with a strong impact on k cat values. We applied this approach to predict genome-scale k cat values for more than 300 yeast species. Additionally, we designed a Bayesian pipeline to parameterize enzyme-constrained genome-scale metabolic models from predicted k cat values. The resulting models outperformed the corresponding original enzyme-constrained genome-scale metabolic models from previous pipelines in predicting phenotypes and proteomes, and enabled us to explain phenotypic differences. DLKcat and the enzyme-constrained genome-scale metabolic model construction pipeline are valuable tools to uncover global trends of enzyme kinetics and physiological diversity, and to further elucidate cellular metabolism on a large scale.

Overview

CoChat Summary

TL;DR

The authors developed DLKcat, a deep learning model that predicts enzyme turnover numbers (kcat) from substrate structures and protein sequences, and used it to reconstruct enzyme-constrained genome-scale metabolic models for 343 yeast/fungi species. The resulting models improved kcat coverage and phenotype/proteome predictions compared with previous ecGEM pipelines.

Key findings

  • DLKcat was trained on 16,838 unique entries covering 7,822 unique protein sequences, 851 organisms and 2,672 unique substrates; it achieved a test-set r.m.s.e. of 1.06 and Pearson’s r = 0.88 on the whole dataset (r = 0.71 on the test dataset).
  • For promiscuous enzymes, predicted kcat values were higher for preferred substrates (median 11.07 s⁻¹) than alternative substrates (median 6.01 s⁻¹; P = 1.3 × 10⁻¹²) and random substrates (median 3.51 s⁻¹; P = 9.3 × 10⁻⁶).
  • DLKcat captured mutation effects: whole-dataset Pearson’s r = 0.90 for mutated enzymes and r = 0.94 for well-studied enzyme–substrate pairs with rich mutagenesis data; attention weights were significantly higher for residues mutated in the decreased-kcat group (P = 0.0014).
  • DL-ecGEMs and posterior-mean-ecGEMs covered kcat values for ~80% of enzymes and defined enzymatic constraints for ~90% of enzymatic reactions, versus ~40% and ~60% for original-ecGEMs; DL-ecGEMs reduced proteome prediction r.m.s.e. by 30% compared with original-ecGEMs.

Methods

  • DLKcat combines a graph neural network for substrates, represented as molecular graphs from SMILES, with a convolutional neural network for protein sequences split into overlapping n-gram amino acids; it was trained on data extracted from BRENDA and SABIO-RK, randomly split 80% training, 10% validation and 10% test.
  • A Bayesian SMC-ABC pipeline used predicted kcat values as prior means with log-normal distributions N(kcat_i, 1) and updated them against experimental growth data collated as 371 entries for 53 species with 16 carbon sources; posterior-mean kcat values were used to parameterize posterior-mean-ecGEMs.
  • The study reconstructed and compared original-ecGEMs, DL-ecGEMs and posterior-mean-ecGEMs for 343 yeast/fungi species, evaluating kcat coverage, growth/exchange-rate predictions, proteome predictions and Crabtree-phenotype-related enzyme differences.

Limitations

  • The paper states that DLKcat still predicts a level of kinetic activity towards random substrates that is likely too high, due to limited negative data; it also did not consider environmental factors such as pH and temperature.
  • The paper states that for reactions with multiple substrates or heteromeric enzyme complexes, multiple predicted kcat values can arise and the current approach selects the maximum, which is not ideal.
  • The paper states that DL-ecGEMs and posterior-mean-ecGEMs inherit limitations from basic GEMs, including the steady-state assumption and lack of regulatory behaviour; kcat is not the only kinetic parameter, and affinity constants and internal metabolite concentrations are not readily considered.

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

Integrity signals

Registry

Crossref · 10.1038/s41929-022-00798-z

Cites

77 of 77 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Deep learning of k cat for ecGEM parameterization. a , DLKcat, the approach developed for k cat prediction by combining a GNN for substrates and a CNN for proteins. b , Information extraction from GEMs as the input for the deep learning model to predict k cat values. c , The developed Bayesian facilitated pipeline to reconstruct ecGEMs using the predicted k cat profiles from the deep learning model.
Fig. 1 Deep learning of k cat for ecGEM parameterization. a , DLKcat, the approach developed for k cat prediction by combining a GNN for substrates and a CNN for proteins. b , Information extraction from GEMs as the input for the deep learning model to predict k cat values. c , The developed Bayesian facilitated pipeline to reconstruct ecGEMs using the predicted k cat profiles from the deep learning model.From the PDF
Fig. 2 Deep learning model performance for k cat prediction. a , The r.m.s.e. of k cat prediction during the training process. b , Performance of the final deep learning model. The correlation between predicted k cat values and those present in the whole dataset (training, validation and test datasets) was evaluated. The brightness of colour represents the density of data points. Student's t -test was used to calculate the P value for Pearson's correlation. c , Enzyme promiscuity analysis on the whole dataset. For enzymes with multiple substrates, we divided the substrates into preferred and alternative by their experimentally measured k cat value, and then used the predicted k cat values for this box plot. Random substrates were randomly chosen from the compound dataset in our training data, except for the documented substrates and products for the tested enzyme. We evaluated 945 promiscuous enzymes in the whole dataset ( n = 945 for preferred substrates, n = 4,238 for alternative substrates, n = 945 for random substrates). d , Comparison of the predicted k cat values for the native substrates and the underground substrates with the human aldo-keto reductase enzyme as a case study. Here, we defined those substrates with the top 10% catalytic ability (experimental k cat value) as the native substrates ( n = 6), while those with the last 10% catalytic ability (experimental k cat value) were considered as the underground substrates as defined in the reference ( n = 6) 24 . In each box plot ( c and d ), the central band represents the median value, the box represents the upper and lower quartiles and the whiskers extend up to 1.5 times the interquartile range beyond the box range. A two-sided Wilcoxon rank sum test was used to calculate the P values in c and d .
Fig. 2 Deep learning model performance for k cat prediction. a , The r.m.s.e. of k cat prediction during the training process. b , Performance of the final deep learning model. The correlation between predicted k cat values and those present in the whole dataset (training, validation and test datasets) was evaluated. The brightness of colour represents the density of data points. Student's t -test was used to calculate the P value for Pearson's correlation. c , Enzyme promiscuity analysis on the whole dataset. For enzymes with multiple substrates, we divided the substrates into preferred and alternative by their experimentally measured k cat value, and then used the predicted k cat values for this box plot. Random substrates were randomly chosen from the compound dataset in our training data, except for the documented substrates and products for the tested enzyme. We evaluated 945 promiscuous enzymes in the whole dataset ( n = 945 for preferred substrates, n = 4,238 for alternative substrates, n = 945 for random substrates). d , Comparison of the predicted k cat values for the native substrates and the underground substrates with the human aldo-keto reductase enzyme as a case study. Here, we defined those substrates with the top 10% catalytic ability (experimental k cat value) as the native substrates ( n = 6), while those with the last 10% catalytic ability (experimental k cat value) were considered as the underground substrates as defined in the reference ( n = 6) 24 . In each box plot ( c and d ), the central band represents the median value, the box represents the upper and lower quartiles and the whiskers extend up to 1.5 times the interquartile range beyond the box range. A two-sided Wilcoxon rank sum test was used to calculate the P values in c and d .From the PDF
Fig. 3 Deep learning model for the prediction and interpretation of k cat of mutated enzymes. a , b , Prediction performance of k cat values for all wild-type ( a ) and mutated ( b ) enzymes. Colour brightness represents data density. c , Comparison between predicted and measured k cat values for several well-studied enzyme-substrate pairs with rich experimental mutagenesis data. Enzyme abbreviations: DHFR, dihydrofolate reductase; PGDH, d -3-phosphoglycerate dehydrogenase; AKIII, aspartokinase III; DAOCS, deacetoxycephalosporin C synthase; PNP, purine nucleoside phosphorylase; GGPPs, geranylgeranyl pyrophosphate synthase. Substrate abbreviations: G3P, glycerate 3-phosphate; L-Asp, l -aspartate; IPP, isopentenyl diphosphate. In a -c , the student's t -test was used to calculate the P value for the Pearson's correlation. d , Comparison of predicted k cat values on several mutated enzyme-substrate pairs between enzymes with wild-type-like k cat and decreased k cat . P < 0.05 (*), P < 0.01 (**) and P < 0.001 (***), two-sided Wilcoxon rank sum test. Detailed information and sample numbers can be found in Supplementary Table 2. e , Attention weight of residue position in the wild-type PNP enzyme, using inosine as substrate. The mutated residues in each of the mutated enzymes (with both wild-type-like k cat and decreased k cat ) were marked on the curve according to their mutated residue. Dot size indicates the number of mutated enzymes with mutations of that residue. f , Overall attention weights for the PNP-inosine pair, comparing enzymes with wild-type-like k cat and decreased k cat by two-sided Wilcoxon rank sum test. n = 15 for wild-type-like k cat ; n = 72 for decreased k cat . In each box plot ( d and f ), the central band represents the median value, the box represents the upper and lower quartiles and the whiskers extend up to 1.5 times the interquartile range beyond the box range.
Fig. 3 Deep learning model for the prediction and interpretation of k cat of mutated enzymes. a , b , Prediction performance of k cat values for all wild-type ( a ) and mutated ( b ) enzymes. Colour brightness represents data density. c , Comparison between predicted and measured k cat values for several well-studied enzyme-substrate pairs with rich experimental mutagenesis data. Enzyme abbreviations: DHFR, dihydrofolate reductase; PGDH, d -3-phosphoglycerate dehydrogenase; AKIII, aspartokinase III; DAOCS, deacetoxycephalosporin C synthase; PNP, purine nucleoside phosphorylase; GGPPs, geranylgeranyl pyrophosphate synthase. Substrate abbreviations: G3P, glycerate 3-phosphate; L-Asp, l -aspartate; IPP, isopentenyl diphosphate. In a -c , the student's t -test was used to calculate the P value for the Pearson's correlation. d , Comparison of predicted k cat values on several mutated enzyme-substrate pairs between enzymes with wild-type-like k cat and decreased k cat . P < 0.05 (*), P < 0.01 (**) and P < 0.001 (***), two-sided Wilcoxon rank sum test. Detailed information and sample numbers can be found in Supplementary Table 2. e , Attention weight of residue position in the wild-type PNP enzyme, using inosine as substrate. The mutated residues in each of the mutated enzymes (with both wild-type-like k cat and decreased k cat ) were marked on the curve according to their mutated residue. Dot size indicates the number of mutated enzymes with mutations of that residue. f , Overall attention weights for the PNP-inosine pair, comparing enzymes with wild-type-like k cat and decreased k cat by two-sided Wilcoxon rank sum test. n = 15 for wild-type-like k cat ; n = 72 for decreased k cat . In each box plot ( d and f ), the central band represents the median value, the box represents the upper and lower quartiles and the whiskers extend up to 1.5 times the interquartile range beyond the box range.From the PDF
Fig. 4 Bayesian modelling training performance for S. cerevisiae ecGEM. a , The r.m.s.e. for phenotype measurement and prediction during the Bayesian training process. b , Simulated exchange rates by posterior-mean-ecGEM (lines) compared with experimental data (dots). The k cat values in the posterior-mean-ecGEMs are mean values from 100 sampled posterior datasets obtained from the Bayesian training process. c , Principal component analysis for k cat datasets sampled during the Bayesian training approach, showing the progression from prior to posterior dataset. Each parameter in the set was standardized by subtracting the mean and then divided by the standard deviation before the principal component analysis. In blue are 100 sampled prior datasets; in red are 100 sampled posterior datasets; in grey are all other intermediate datasets. PC, principal component. d , The number of enzymes with a significantly changed mean value (Šidák-adjusted Welch's t -test, P < 0.01, two-sided) and variance (Šidák-adjusted one-tailed F -test, P < 0.01) between the sampled prior and posterior k cat datasets. Parameters from 126 prior and 100 posterior ecGEMs were used for statistical tests. e , Variance distribution comparison for prior and posterior distribution. f , Correlation between deep learning-predicted k cat and posterior mean k cat . Student's t -test was used to calculate P value for Pearson's correlation.
Fig. 4 Bayesian modelling training performance for S. cerevisiae ecGEM. a , The r.m.s.e. for phenotype measurement and prediction during the Bayesian training process. b , Simulated exchange rates by posterior-mean-ecGEM (lines) compared with experimental data (dots). The k cat values in the posterior-mean-ecGEMs are mean values from 100 sampled posterior datasets obtained from the Bayesian training process. c , Principal component analysis for k cat datasets sampled during the Bayesian training approach, showing the progression from prior to posterior dataset. Each parameter in the set was standardized by subtracting the mean and then divided by the standard deviation before the principal component analysis. In blue are 100 sampled prior datasets; in red are 100 sampled posterior datasets; in grey are all other intermediate datasets. PC, principal component. d , The number of enzymes with a significantly changed mean value (Šidák-adjusted Welch's t -test, P < 0.01, two-sided) and variance (Šidák-adjusted one-tailed F -test, P < 0.01) between the sampled prior and posterior k cat datasets. Parameters from 126 prior and 100 posterior ecGEMs were used for statistical tests. e , Variance distribution comparison for prior and posterior distribution. f , Correlation between deep learning-predicted k cat and posterior mean k cat . Student's t -test was used to calculate P value for Pearson's correlation.From the PDF
Fig. 5 Evaluation of three ecGEM modelling pipelines including original-ecGEM, DL-ecGEM and posterior-mean-ecGEM reconstruction.
Fig. 5 Evaluation of three ecGEM modelling pipelines including original-ecGEM, DL-ecGEM and posterior-mean-ecGEM reconstruction.From the PDF
Fig. 6 Explanation of the Crabtree effect by energy metabolism. a , HY and LY pathway definitions. PEP, phosphoenolpyruvate; PYR, pyruvate; OAA, oxaloacetate; CIT, citrate; ETC, electron transport chain. b , Model-inferred protein efficiency of energy metabolism in several common yeast species. Protein efficiency is the ATP produced per protein mass per unit time, in units of millimoles ATP per grams protein per hour. L. kluyveri , Lachancea kluyveri . c , Enzymes with significantly different k cat values between Crabtree positive and negative species. A two-sided Wilcoxon rank sum test was used to calculate P values. Crabtree positive species ( n = 25) and Crabtree negative species ( n = 77) were examined in this analysis. In the box plot, the central band represents the median value, the box represents the upper and lower quartiles and the whiskers extend up to 1.5 times the interquartile range beyond the box range.
Fig. 6 Explanation of the Crabtree effect by energy metabolism. a , HY and LY pathway definitions. PEP, phosphoenolpyruvate; PYR, pyruvate; OAA, oxaloacetate; CIT, citrate; ETC, electron transport chain. b , Model-inferred protein efficiency of energy metabolism in several common yeast species. Protein efficiency is the ATP produced per protein mass per unit time, in units of millimoles ATP per grams protein per hour. L. kluyveri , Lachancea kluyveri . c , Enzymes with significantly different k cat values between Crabtree positive and negative species. A two-sided Wilcoxon rank sum test was used to calculate P values. Crabtree positive species ( n = 25) and Crabtree negative species ( n = 77) were examined in this analysis. In the box plot, the central band represents the median value, the box represents the upper and lower quartiles and the whiskers extend up to 1.5 times the interquartile range beyond the box range.From the PDF

Verification

Claims the paper makes

21 Valid · 4 Not valid · 9 Not checked · 71 Own results and interpretations

52 more claims · 10 with their cited sources checked

Create a free account to see all 105 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 21
  • Probably not a substantial claim. 30
  • The quote only partly states this claim. 1

Create a free account to see these claims.

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?

3 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

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.

1 of 1

Citations

Cites

Show all 68 references

Cited by

See all 470

Metadata

Authors
Li, F., Yuan, L., Lu, H., Li, G., Chen, Y., Engqvist, M. K. M., Kerkhoven, E. J., & Nielsen, J.
Affiliations
  • Chalmers University of Technology (SE)
  • BioInnovation Institute (DK)
Corresponding author
Kerkhoven, Eduard J.
Venue
Nature Catalysis
Published
2022-06-16
Volume
5
Issue
8
Pages
662-672
ISSN
2520-1158 (ISSN-L)
OpenAlex
W4283032119
Cited by
466
References
68
Citations by year
2026: 128 · 2025: 146 · 2024: 104 · 2023: 65 · 2022: 22
FWCI
27.49
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Catalysis (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by)
  • Chalmers Research (Chalmers University of Technology) (submittedVersion)
Funders
  • European Commission (HORIZON2020, 686070)
  • Chalmers Tekniska Högskola
  • Knut och Alice Wallenbergs Stiftelse (NNF10CC1016517)
  • Novo Nordisk (NNF10CC1016517)
  • Vetenskapsrådet (05973, 2018-, 201805973)
  • Novo Nordisk Fonden (NNF10CC1016517)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Microbial Metabolic Engineering and Bioproduction
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Bioinformatics and Genomic Networks
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
Keywords
turnover number, metabolic enzymes, deep learning, genome-scale metabolic models, enzyme kinetics, yeast, substrate structure, protein sequences, amino acid residues, phenotype prediction, physiological diversity

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.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

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: Deep learning-based kcat prediction enables improved enzyme-constrained model reconstruction. https://app.cochat.ai/papers/10.1038/s41929-022-00798-z
The paper
Li, F., Yuan, L., Lu, H., Li, G., Chen, Y., Engqvist, M. K. M., Kerkhoven, E. J., & Nielsen, J. (2022). Deep learning-based kcat prediction enables improved enzyme-constrained model reconstruction. Nature Catalysis. https://doi.org/10.1038/s41929-022-00798-z

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.