Skip to content

Structural Km Predictor

Summary

Kroll, Engqvist, Heckmann & Lercher (2021) present a deep-learning-derived model for predicting Michaelis constants (Km) organism-independently, genome-scale, from structural features alone. A graph neural network learns a task-specific substrate fingerprint (outperforming predefined fingerprints such as ECFP), which is combined with an enzyme-sequence representation (UniRep, plus domain-content features) and fed into a gradient-boosting regressor, achieving a test-set R²=0.53 — this model and its underlying dataset preprocessing pipeline were the direct precursor to the authors' later kcat model, TurNuP.

Method

  • Substrate features: a graph neural network (GNN) trained end-to-end on the prediction task learns a substrate fingerprint from 2D molecular connectivity graphs, outperforming predefined expert-crafted fingerprints (ECFP) as an input representation; adding molecular weight (MW) and LogP as auxiliary features had only a minor additional effect.
  • Enzyme features: a 1,900-dimensional UniRep representation of the enzyme's amino acid sequence, plus domain-content annotations.
  • Model: gradient boosting regression combining substrate and enzyme representations.
  • Data: entries curated from BRENDA and SABIO-RK.

Performance

  • Combined substrate+enzyme model: MSE=0.65 (log10 scale), R²=0.53 on the test set, rm²=0.53 — substantially better than substrate-only or enzyme-only ablations.
  • Substrate-only vs. enzyme-only: neither single-modality model matched the combined model's accuracy, confirming both structural inputs carry complementary information.
  • Generalization: performance was also assessed specifically on test entries where neither the enzyme nor the substrate appeared in the training set, and on the training-set-size sensitivity of the gradient boosting model, both reported as part of the validation.

Availability

Code and Jupyter notebooks (data download/preprocessing, training, and validation): github.com/AlexanderKroll/KM_prediction. UniRep implementation used: github.com/EngqvistLab/UniRep50.

See Also

  • TurNuP — the same authors' later kcat prediction model, building on this Km model's GNN substrate-fingerprint approach.
  • UniKP — cites this dataset as the source of its own Km benchmark (R²=0.530).
  • Enzyme Kinetic Parameter Prediction — field overview.

Citations

[1] Kroll, A., Engqvist, M.K.M., Heckmann, D., & Lercher, M.J. (2021). Deep learning allows genome-scale prediction of Michaelis constants from structural features. PLoS Biology, 19(10), e3001402.