DLKcat
Summary¶
Li, Yuan, Lu, et al. (2022) present DLKcat, a deep learning model combining a graph neural network (substrate structure) and a convolutional neural network (enzyme sequence) to predict enzyme turnover numbers (kcat) for arbitrary enzyme-substrate pairs. Trained on 16,838 entries curated from BRENDA and SABIO-RK (7,822 proteins, 851 organisms, 2,672 substrates), DLKcat was used to parameterize enzyme-constrained genome-scale metabolic models ("DL-ecGEMs") for 343 yeast and fungal species. Disputed: subsequent analysis by Kroll & Lercher found DLKcat's benchmark performance is driven largely by train/test data leakage, with accuracy collapsing for enzymes and mutants genuinely dissimilar to the training set — see that page for details.
Method and Training¶
- Architecture: a graph neural network (GNN) encodes the 2D connectivity graph of a substrate molecule; a convolutional neural network (CNN) encodes the enzyme's amino acid sequence; the two representations are combined to predict log-scale kcat.
- Training data: 16,838 kcat entries from BRENDA and SABIO-RK spanning 7,822 proteins, 851 organisms, and 2,672 substrates.
- Reported accuracy: Pearson r=0.88 across the whole dataset; r=0.71 on a held-out test set (r=0.70 for unseen substrate/enzyme combinations); RMSE=1.06 on a log10 scale (within one order of magnitude on average).
Reported Findings¶
- Enzyme promiscuity: DLKcat's predictions distinguish preferred, alternative, and random substrates for a given enzyme, and were used to compare native versus "underground" (non-physiological) substrates for a human aldo-keto reductase.
- Mutation sensitivity: the original paper reports Pearson r=0.94 on a curated panel of well-studied enzyme mutants — a claim directly disputed by Kroll & Lercher's re-analysis, which found this panel was 84.4% contaminated with training-set entries and that R² on the genuinely held-out subset is -0.21.
- Genome-scale application: DLKcat kcat predictions were applied to ~3 million enzyme-substrate pairs across 343 yeast/fungal species to build enzyme-constrained genome-scale metabolic models (DL-ecGEMs).
- Bayesian refinement: a sequential Monte Carlo approximate Bayesian computation (SMC-ABC) pipeline updates the DLKcat-predicted (prior) kcat values against experimental growth-phenotype data (371 entries, 53 species, 16 carbon sources), producing "posterior-mean-DL-ecGEMs" that outperform the original DL-ecGEM pipeline on phenotype and proteome-allocation prediction.
Disputed: Generalization and Data Leakage¶
Disputed: Kroll & Lercher (2024) report that 67.9% of DLKcat's nominal test-set enzymes were also present in its training data, with a further 23.3% at ≥99% sequence identity to a training enzyme. After correcting for this, DLKcat's R² is negative (worse than a constant-mean baseline) for enzymes under 60% maximal sequence identity to the training set, and its mutation-effect predictions on truly held-out mutants achieve R²=-0.18 to -0.21 — contradicting the original paper's r=0.94 mutation-sensitivity claim. This dispute is unresolved in the sense that DLKcat's architecture and genome-scale application remain widely used, but its claimed generalization to novel enzymes and mutants should be treated with caution; see TurNuP and CatPred for models designed with sequence-identity-stratified evaluation from the outset.
Availability¶
Code, training data, and Bayesian pipeline: github.com/SysBioChalmers/DLKcat. Supplementary dataset (predicted kcat values, GEMs, DL-ecGEMs, posterior-mean-ecGEMs for 343 species): Zenodo 10.5281/zenodo.6438262.
See Also¶
- DLKcat Generalization Failure and Train/Test Leakage — the critique disputing DLKcat's headline generalization claims.
- Enzyme Kinetic Parameter Prediction — field overview.
- TurNuP — successor model addressing DLKcat's generalization weaknesses.
- EnzymeTuning — uses DLKcat-style predictions as one input to a top-down GAN-based kcat refinement pipeline for ecGEMs.