Skip to content

ESP (Enzyme Substrate Prediction)

Summary

ESP predicts whether a given small molecule is a substrate of a given enzyme, addressing a problem distinct from kinetic-parameter prediction: most training data for enzymes contains only positive (confirmed-substrate) examples, with almost no confirmed non-substrates. ESP creates negative examples by data augmentation. It reports 91.5% accuracy on independent, sequence-diverse test data.[1]

Method

  • Enzyme representation: a task-specific fine-tuned variant of the ESM-1b protein language model. An extra 1,280-dimensional "whole-enzyme" token is added to the standard per-residue representations and trained end-to-end (on ~287,000 data points, including phylogenetically inferred evidence) to store information relevant to substrate prediction; after training, this token is extracted as the "ESM-1bts" enzyme vector.[1]
  • Substrate representation: task-specific molecular fingerprints generated by a graph neural network (GNN), rather than generic ECFP fingerprints.[1]
  • Negative-data augmentation: for each confirmed positive enzyme-substrate pair, three "non-substrate" molecules are sampled with fingerprint similarity 0.75–0.95 to the true substrate, drawn only from the ~1,400 substrates that occur among all confirmed pairs in the dataset (rather than the space of all possible small molecules) — reasoning that most such molecules are unlikely to be true secondary substrates for any given enzyme, keeping the false-negative rate low.[1]
  • Classifier: a gradient-boosted decision tree trained on the concatenated ESM-1bts and GNN-fingerprint vectors, over a final dataset of 69,365 entries (80/20 train/test split, with no test-set enzyme exceeding 80% sequence identity to any training-set enzyme).[1]

Performance

The best-performing configuration (ESM-1bts + GNN fingerprints) achieved 91.5% accuracy, ROC-AUC of 0.956, and a Matthews correlation coefficient of 0.78 on the held-out test set — significantly better (McNemar's test, p<10⁻⁵) than the next-best configuration (ESM-1bts + ECFP), and better than models using un-fine-tuned ESM-1b or logistic-regression/random-forest classifiers on the same features. ESP generalizes across enzyme families broadly, rather than being tuned to one enzyme class.[1]

Availability

A web server for testing candidate substrates in silico is described in the paper; the model and code are associated with the publication (no separate repository link verified in this pass).

See Also

  • EZSpecificity — a later (2025), structurally explicit model for the related but distinct problem of ranking which substrate among candidates an enzyme prefers (specificity), rather than binary substrate/non-substrate classification.
  • Enzyme Kinetic Parameter Prediction — the adjacent field of predicting kcat/Km once a substrate pair is established.
  • Predicting Catalytic Competence of Enzyme-Ligand Complexes — framework page situating ESP as a substrate-scope triage step, upstream of pose generation and mechanistic validation.

Citations

[1] Kroll, A., Ranjan, S., Engqvist, M.K.M., Lercher, M.J. (2023). A general model to predict small molecule substrates of enzymes based on machine and deep learning. Nature Communications, 14, 2787. Supports: method and all performance figures above. Location: Abstract; Results (task-specific representations; negative-sampling strategy; model comparison).