CoChat

The DOI matches its Crossref record; 71 of the 72 references it cites were found in a registry or an index; 11 of the 50 sources its claims cite were checked for support.

Turnover number predictions for kinetically uncharacterized enzymes using machine and deep learning

Open access
Kroll, Alexander, Rousset, Yvan, Hu, Xiao-Pan, Liebrand, Nina A., Lercher, Martin J. · 2023 · Nature Communications

Published version of a preprint

Abstract

The turnover number k cat , a measure of enzyme efficiency, is central to understanding cellular physiology and resource allocation. As experimental k cat estimates are unavailable for the vast majority of enzymatic reactions, the development of accurate computational prediction methods is highly desirable. However, existing machine learning models are limited to a single, well-studied organism, or they provide inaccurate predictions except for enzymes that are highly similar to proteins in the training set. Here, we present TurNuP, a general and organism-independent model that successfully predicts turnover numbers for natural reactions of wild-type enzymes. We constructed model inputs by representing complete chemical reactions through differential reaction fingerprints and by representing enzymes through a modified and re-trained Transformer Network model for protein sequences. TurNuP outperforms previous models and generalizes well even to enzymes that are not similar to proteins in the training set. Parameterizing metabolic models with TurNuP-predicted k cat values leads to improved proteome allocation predictions. To provide a powerful and convenient tool for the study of molecular biochemistry and physiology, we implemented a TurNuP web server.

Overview

CoChat Summary

## TL;DR

TurNuP is a machine/deep learning model that predicts enzyme turnover numbers (kcat) from enzyme sequence and full reaction information, and it outperforms prior methods on the paper’s held-out test sets. On the authors’ final test set of 4,271 data points, the combined model reached \(R^2 = 0.44\) with mean absolute deviation 0.69 on the \(\log_{10}\) scale, corresponding to a 4.8-fold average deviation from measured values.

## Key findings

  • The final curated dataset contained 4,271 data points, with 2,977 unique reactions and 2,827 unique enzymes; the train/test split was 80%/20% with no identical enzyme sequences in both splits.
  • Using reaction information alone, the best reaction fingerprint model achieved \(R^2 = 0.38\), MSE = 0.89, and Pearson \(r = 0.62\) on the test set.
  • Using enzyme information alone, the best enzyme representation achieved \(R^2 = 0.37\), MSE = 0.91, and \(r = 0.61\) on the test set.
  • The final TurNuP ensemble of enzyme and reaction models achieved \(R^2 = 0.44\), MSE = 0.81, and \(r = 0.67\) on the test set; adding extra features like \(K_M\) and reaction flux did not improve performance (\(R^2 = 0.39\) with those added features).
  • TurNuP performed better than DLKcat across all enzyme-sequence-identity bins, with reported significance levels of \(p = 4.3 \times 10^{-11}\), 0.0017, 0.00037, and 0.048 for the four sequence-identity groups.
  • Using TurNuP-derived kcat values in yeast proteome allocation modeling improved predictions in 19 of 21 environment-species combinations, with mean squared errors improving by about 18% on average.

## Methods

  • The authors compiled kcat measurements from BRENDA, UniProt, and Sabio-RK, mapped them to enzyme sequences and reaction identities, removed non-wild-type/non-natural reactions, duplicates, incomplete records, and outliers, and log-transformed kcat values.
  • They trained and compared gradient boosting models using different inputs: reaction fingerprints (structural, difference, DRFP), enzyme representations (ESM-1b and a fine-tuned ESM-1b-based representation), and combinations of these.
  • Model development used fivefold cross-validation for hyperparameter search and a held-out test set; performance was assessed with \(R^2\), MSE, Pearson \(r\), and nonparametric tests for model comparisons.

## Limitations

  • The paper states that the training data are small and noisy; it reports an average deviation of 5.7-fold between two kcat measurements for the same enzyme-reaction pair and notes database errors and varying assay conditions as contributors.
  • The authors state they could not include experimental conditions such as pH and temperature for most data points because these are generally unavailable in the databases; this likely limits accuracy.
  • Evident from the text: the model was trained only on natural reactions of wild-type enzymes, so the authors say it should not be used for non-natural reactions or enzyme mutants.

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

Integrity signals

Registry

Crossref · 10.1038/s41467-023-39840-4

Cites

71 of 71 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Machine learning model to predict kcat from numerical enzyme representations and reaction fingerprints. Experimentally measured kcat values are downloaded from three different databases. Enzyme information is represented with numerical vectors obtained from natural language processing (NLP) models that use the linear amino acid sequence as their input. Chemical reactions are represented using integer vectors. Concatenated enzyme-reaction representations are used to train a gradient boosting model to predict kcat. After training, the fitted model can be used to parameterize metabolic networks with kcat values.
Fig. 1 Machine learning model to predict kcat from numerical enzyme representations and reaction fingerprints. Experimentally measured kcat values are downloaded from three different databases. Enzyme information is represented with numerical vectors obtained from natural language processing (NLP) models that use the linear amino acid sequence as their input. Chemical reactions are represented using integer vectors. Concatenated enzyme-reaction representations are used to train a gradient boosting model to predict kcat. After training, the fitted model can be used to parameterize metabolic networks with kcat values.CC BY · PubMed Central
Fig. 2 Calculation of reaction fingerprints for an exemplary reaction. a Structural reaction fingerprints. Binary molecular fingerprints are calculated for each substrate and each product. The bitwise OR-function is applied to all substrates and also to all products. The resulting substrate and the resulting product vector are then concatenated. b Difference reaction fingerprints. Binary molecular fingerprints are calculated for each substrate and each product. All substrate fingerprint vectors are summed, and the same is done for all product fingerprint vectors. To create the difference fingerprint, the resulting product vector is subtracted from the substrate vector.
Fig. 2 Calculation of reaction fingerprints for an exemplary reaction. a Structural reaction fingerprints. Binary molecular fingerprints are calculated for each substrate and each product. The bitwise OR-function is applied to all substrates and also to all products. The resulting substrate and the resulting product vector are then concatenated. b Difference reaction fingerprints. Binary molecular fingerprints are calculated for each substrate and each product. All substrate fingerprint vectors are summed, and the same is done for all product fingerprint vectors. To create the difference fingerprint, the resulting product vector is subtracted from the substrate vector.CC BY · PubMed Central
Fig. 3 Using enzyme and reaction information combined leads to improved kcat predictions. a Coefficients of determination R2 for models with different inputs. b Mean squared errors (MSE) on log10-scale. Boxplots summarize the results of the CV with n = 5 folds on the training set with the best sets of hyperparameters; blue dots show the results on the test set using the optimized models trained on the whole training set. We used a 2× interquartile range for the whiskers, the boxes extend from the lower to upper quartile values, and the red horizontal lines are displaying the median of the data points. Model performances are plotted for the models with structural reaction fingerprints (str. FP), difference reaction fingerprints (diff. FP), differential reaction fingerprints (DRFP), ESM-1b vectors (ESM-1b), task-specific ESM-1b vectors (ESM-1bESP), and with enzyme and reaction information (ESM-1bESP + DRFP). Source data are provided as a Source Data file.
Fig. 3 Using enzyme and reaction information combined leads to improved kcat predictions. a Coefficients of determination R2 for models with different inputs. b Mean squared errors (MSE) on log10-scale. Boxplots summarize the results of the CV with n = 5 folds on the training set with the best sets of hyperparameters; blue dots show the results on the test set using the optimized models trained on the whole training set. We used a 2× interquartile range for the whiskers, the boxes extend from the lower to upper quartile values, and the red horizontal lines are displaying the median of the data points. Model performances are plotted for the models with structural reaction fingerprints (str. FP), difference reaction fingerprints (diff. FP), differential reaction fingerprints (DRFP), ESM-1b vectors (ESM-1b), task-specific ESM-1b vectors (ESM-1bESP), and with enzyme and reaction information (ESM-1bESP + DRFP). Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 4 Comparison of predicted and experimentally measured kcat values. kcat values predicted with the complete TurNuP model, plotted against the corresponding experimental measurements. Each dot is one data point from the test set. The gray dashed line is a diagonal line that indicates perfect correlation. The red dashed line displays the true correlation between predicted and measured values. Source data are provided as a Source Data file.
Fig. 4 Comparison of predicted and experimentally measured kcat values. kcat values predicted with the complete TurNuP model, plotted against the corresponding experimental measurements. Each dot is one data point from the test set. The gray dashed line is a diagonal line that indicates perfect correlation. The red dashed line displays the true correlation between predicted and measured values. Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 5 TurNuP predictions are more accurate for enzymes similar to proteins in the training set and outperform an existing deep learning model. a Coefficients of determination R2 for the test sets for our TurNuP model (black) and the previously published DLKcat model16 (magenta) for different levels of maximal enzyme sequence identity compared to enzymes in the training set. Numbers next to points show how many data points of this category are in the test set. The horizontal dashed line corresponds to a model that predicts the same mean kcat value for all test data points. b Mean squared errors (MSE) for the prediction of absolute proteome data compared to experimental data. Proteome predictions were achieved with enzyme-constrained genome-scale models, parameterized with kcat values predicted with TurNuP (black) or with the DLKcat model (magenta). Proteome data was predicted for four different yeast species (Sce, Saccharomyces cerevisiae; Kla, Kluyveromyces lactis; Kmx, Kluyveromyces marxianus; Yli, Yarrowia lipolytica) in 21 different culture conditions (see Methods for details). Source data are provided as a Source Data file.
Fig. 5 TurNuP predictions are more accurate for enzymes similar to proteins in the training set and outperform an existing deep learning model. a Coefficients of determination R2 for the test sets for our TurNuP model (black) and the previously published DLKcat model16 (magenta) for different levels of maximal enzyme sequence identity compared to enzymes in the training set. Numbers next to points show how many data points of this category are in the test set. The horizontal dashed line corresponds to a model that predicts the same mean kcat value for all test data points. b Mean squared errors (MSE) for the prediction of absolute proteome data compared to experimental data. Proteome predictions were achieved with enzyme-constrained genome-scale models, parameterized with kcat values predicted with TurNuP (black) or with the DLKcat model (magenta). Proteome data was predicted for four different yeast species (Sce, Saccharomyces cerevisiae; Kla, Kluyveromyces lactis; Kmx, Kluyveromyces marxianus; Yli, Yarrowia lipolytica) in 21 different culture conditions (see Methods for details). Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 6 TurNuP makes good predictions even for unseen reactions. Coefficients of determination R2, calculated separately on different splits of the test set. The test set was split according to the maximal reaction similarities compared to reactions in the training set. Source data are provided as a Source Data file.
Fig. 6 TurNuP makes good predictions even for unseen reactions. Coefficients of determination R2, calculated separately on different splits of the test set. The test set was split according to the maximal reaction similarities compared to reactions in the training set. Source data are provided as a Source Data file.CC BY · PubMed Central

Verification

Claims the paper makes

7 Valid · 0 Not valid · 27 Not checked · 101 Own results and interpretations

75 more claims · 1 with their cited sources checked

Create a free account to see all 135 claims and to verify the 39 references still open.

See all claims — free account
  • It may state two separate claims as one. 30
  • Probably not a substantial claim. 24
  • The quote only partly states this claim. 2

Create a free account to see these claims.

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.

2 of 2

Citations

Cites

Show all 72 references

Research works printed but not linked: 5

  1. Friedman, J. H.The Elements Of Statistical Learning: Data Mining, Inference, And Prediction (Springer Open, New York, 2017).
  2. Vaswani, A. et al. Attention is all you need. Adv. Neural Inf. Process. Syst. 30, 5998–6008 (2017).
  3. Bergstra, J., Yamins, D. & Cox, D. Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures. In Proc. 30th International Conference on Machine Learning, vol. 28 of Proceedings of Machine Learning Research, 115–123 (PMLR, Atlanta, 2013).
  4. Pedregosa, F. et al. Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830 (2011).
  5. Abadi, M. et al. Tensorflow: a system for large-scale machine learning. In 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16), 265–283 (2016).

Cited by

See all 186

Metadata

Authors
Kroll, A., Rousset, Y., Hu, X.-P., Liebrand, N. A., & Lercher, M. J.
Affiliations
Heinrich Heine University Düsseldorf (DE)
Corresponding author
Lercher, Martin J.
Venue
Nature Communications
Published
2023-07-12
Volume
14
Issue
1
Pages
4139-4139
ISSN
2041-1723 (ISSN-L)
OpenAlex
W4384023735
Cited by
186
References
72
Citations by year
2026: 53 · 2025: 80 · 2024: 45 · 2023: 8
FWCI
16.43
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Deutsche Forschungsgemeinschaft (CRC 1310 + EXC 2048/1 (Project ID: 390686111))
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Microbial Metabolic Engineering and Bioproduction
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
Keywords
turnover number, proteome allocation, enzyme kinetics, web server
MeSH terms
Machine Learning, Deep Learning, Amino Acid Sequence, Electric Power Supplies, Proteome

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 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

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: Turnover number predictions for kinetically uncharacterized enzymes using machine and deep learning. https://app.cochat.ai/papers/10.1038/s41467-023-39840-4
The paper
Kroll, A., Rousset, Y., Hu, X.-P., Liebrand, N. A., & Lercher, M. J. (2023). Turnover number predictions for kinetically uncharacterized enzymes using machine and deep learning. Nature Communications. https://doi.org/10.1038/s41467-023-39840-4

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.