CoChat
Sign Up
Cited by 193Machine Learning in BioinformaticsComputational Drug Discovery MethodsEnzyme Catalysis and Immobilization

The DOI matches its Crossref record; 71 of the 73 references it cites were found in a registry or an index.

A general model to predict small molecule substrates of enzymes based on machine and deep learning

Open access
Kroll, Alexander, Ranjan, Sahasra, Engqvist, Martin K. M., Lercher, Martin J.·2023·Nature Communications
DOI 10.1038/s41467-023-38347-2PMC PMC10185530
Open this paper in CoChat — free accountOpen PDF

Abstract

For most proteins annotated as enzymes, it is unknown which primary and/or secondary reactions they catalyze. Experimental characterizations of potential substrates are time-consuming and costly. Machine learning predictions could provide an efficient alternative, but are hampered by a lack of information regarding enzyme non-substrates, as available training data comprises mainly positive examples. Here, we present ESP, a general machine-learning model for the prediction of enzyme-substrate pairs with an accuracy of over 91% on independent and diverse test data. ESP can be applied successfully across widely different enzymes and a broad range of metabolites included in the training data, outperforming models designed for individual, well-studied enzyme families. ESP represents enzymes through a modified transformer model, and is trained on data augmented with randomly sampled small molecules assigned as non-substrates. By facilitating easy in silico testing of potential substrates, the ESP web server may support both basic and applied science.

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 · 10.1038/s41467-023-38347-2

Cites

71 of 71 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Model overview. Experimentally validated enzyme-substrate pairs and sampled negative enzyme-small metabolite pairs are numerically represented with task-specific enzyme and small molecule representations. Concatenated enzyme-small molecule representations are used to train a gradient-boosting model. After training, the fitted model can be used to predict promising candidate substrates for enzymes.
Fig. 1 Model overview. Experimentally validated enzyme-substrate pairs and sampled negative enzyme-small metabolite pairs are numerically represented with task-specific enzyme and small molecule representations. Concatenated enzyme-small molecule representations are used to train a gradient-boosting model. After training, the fitted model can be used to predict promising candidate substrates for enzymes.CC BY · PubMed Central
Fig. 2 A task-specific enzyme representation developed from the ESM-1b model. a ESM-1b model. Amino acids of a protein sequence are represented with numerical vectors and passed through a transformer network. Some amino acid representations are masked. All representations are iteratively updated 33 times, using information about neighboring and distant amino acids. The ESM-1b model is trained to predict the masked amino acids. ESM-1b vectors are calculated by taking the element-wise mean of all representations in the last layer. b Modified ESM-1b model. An additional representation for the whole enzyme is added to the amino acid representations. After updating all representations 33 times, the enzyme representation is concatenated with a small molecule representation. The network is trained to predict whether the small molecule is a substrate for the given enzyme. After training, the ESM-1bts vector is extracted as the enzyme representation before adding the small molecule representation.
Fig. 2 A task-specific enzyme representation developed from the ESM-1b model. a ESM-1b model. Amino acids of a protein sequence are represented with numerical vectors and passed through a transformer network. Some amino acid representations are masked. All representations are iteratively updated 33 times, using information about neighboring and distant amino acids. The ESM-1b model is trained to predict the masked amino acids. ESM-1b vectors are calculated by taking the element-wise mean of all representations in the last layer. b Modified ESM-1b model. An additional representation for the whole enzyme is added to the amino acid representations. After updating all representations 33 times, the enzyme representation is concatenated with a small molecule representation. The network is trained to predict whether the small molecule is a substrate for the given enzyme. After training, the ESM-1bts vector is extracted as the enzyme representation before adding the small molecule representation.CC BY · PubMed Central
Fig. 3 Optimized models provide accurate predictions of enzyme-substrate pairs. a Accuracies. Boxplots summarize the results of the CV with n = 5 folds on the training set with the best sets of hyperparameters. 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. Blue dots display the accuracies on the test set, using the optimized models trained on the whole training set. b ROC curves for the test set. The dotted line displays the ROC curve expected for a completely random model. Source data are provided as a Source Data file.
Fig. 3 Optimized models provide accurate predictions of enzyme-substrate pairs. a Accuracies. Boxplots summarize the results of the CV with n = 5 folds on the training set with the best sets of hyperparameters. 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. Blue dots display the accuracies on the test set, using the optimized models trained on the whole training set. b ROC curves for the test set. The dotted line displays the ROC curve expected for a completely random model. Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 4 Accurate predictions even for enzymes with distinct sequence similarity compared to enzymes in the training data. We divided the test set into subsets with different levels of enzyme sequence identity compared to enzymes in the training set. a ESP accuracies, calculated separately for enzyme-small molecule pairs where the small molecule occurred in the training set and where it did not occur in the training set. b ESP ROC curves. The dotted line displays the ROC curve expected for a completely random model. Source data are provided as a Source Data file.
Fig. 4 Accurate predictions even for enzymes with distinct sequence similarity compared to enzymes in the training data. We divided the test set into subsets with different levels of enzyme sequence identity compared to enzymes in the training set. a ESP accuracies, calculated separately for enzyme-small molecule pairs where the small molecule occurred in the training set and where it did not occur in the training set. b ESP ROC curves. The dotted line displays the ROC curve expected for a completely random model. Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 5 Model performance increases with training set size. Points show accuracies and ROC-AUC scores for the test set versus the fraction of the available training data used for training the gradient-boosting model. Source data are provided as a Source Data file.
Fig. 5 Model performance increases with training set size. Points show accuracies and ROC-AUC scores for the test set versus the fraction of the available training data used for training the gradient-boosting model. Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 6 Prediction scores around 0.5 indicate model uncertainty. Stacked histogram bars display the prediction score distributions of true predictions (blue) and false predictions (red). The inset shows a blow-up of the interval [0.2, 0.8]. Source data are provided as a Source Data file.
Fig. 6 Prediction scores around 0.5 indicate model uncertainty. Stacked histogram bars display the prediction score distributions of true predictions (blue) and false predictions (red). The inset shows a blow-up of the interval [0.2, 0.8]. Source data are provided as a Source Data file.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

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?

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 73 references

Research works printed but not linked: 4

  1. Cooper, G. M., Hausman, R. E. & Hausman, R. E.The Cell: A Molecular Approach, vol. 4 (ASM press, Washington DC, 2007).
  2. Zhou, J. et al. Graph neural networks: A review of methods and applications. AI Open 1, 57–81 (2020). · DOI · Paper page
  3. Van Rossum, G. & Drake, F. L. Python 3 Reference Manual. (CreateSpace, Scotts Valley, 2009).
  4. Bergstra, J., Yamins, D. & Cox, D. Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures. In International conference on machine learning, 115-123 (PMLR, 2013).

Cited by

Sorted by citations

Related

See all

Metadata

Authors
Kroll, A., Ranjan, S., Engqvist, M. K. M., & Lercher, M. J.
Affiliations
  • Heinrich Heine University Düsseldorf (DE)
  • Indian Institute of Technology Bombay (IN)
  • EnginZyme (Sweden) (SE)
Corresponding author
Lercher, Martin J.
ORCID
Venue
Nature Communications
Publisher
Springer Science and Business Media LLC
Published
2023-05-15
Volume
14
Issue
1
Pages
2787-2787
ISSN
2041-1723 (ISSN-L)
DOI
10.1038/s41467-023-38347-2
PMC
PMC10185530
OpenAlex
W4376601927
Cited by
193
References
73
Citations by year
2026: 65 · 2025: 64 · 2024: 51 · 2023: 13
FWCI
17.23
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Chalmers Research (Chalmers University of Technology) (submittedVersion, other-oa)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • Univ. Duesseldorf: Duesseldorfer Dokumenten- und Publikationsserver (submittedVersion, cc-by)
Funders
  • Deutsche Forschungsgemeinschaft (CRC 1310 + 390686111)
  • Volkswagen Foundation
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 catalysis, transformer models, deep learning, metabolites
MeSH terms
Machine Learning, Deep Learning, Catalysis, Proteins, Support Vector Machine

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