Enzyme Kinetic Parameter Prediction
Summary¶
Enzyme kinetic parameters — the turnover number (kcat), Michaelis constant (Km), inhibition constant (Ki), and catalytic efficiency (kcat/Km) — are experimentally scarce relative to the space of known and predicted enzymes, since determining them requires slow, low-throughput in vitro assays. Classical machine learning on structural and network features (e.g., the Heckmann Enzyme Turnover ML Model, 2018) preceded a wave of deep learning models beginning with DLKcat (2022) that predict these parameters directly from protein sequence and substrate structure, enabling genome-scale parameterization of enzyme-constrained metabolic models (ecGEMs) and accelerating enzyme discovery and engineering. The field is defined by an unresolved methodological tension: benchmark accuracy figures are highly sensitive to how train/test splits handle near-duplicate and highly similar sequences, and a direct critique of DLKcat showed its headline results were substantially inflated by data leakage — a finding that reshaped how every subsequent model in this cluster is evaluated.
Two Paradigms: Bottom-Up Sequence Prediction vs. Top-Down Model Calibration¶
- Bottom-up (sequence-based) prediction: models that predict a kinetic parameter directly from an enzyme's amino acid sequence and a substrate's structure, independent of any particular metabolic model or physiological context. This is the dominant paradigm and includes DLKcat, TurNuP, the Structural Km Predictor, UniKP, CatPred, CataPro, and EnzGFM.
- Top-down (model-based) calibration: rather than predicting kcat de novo, these methods calibrate or refine kcat values within an existing enzyme-constrained genome-scale metabolic model against system-level experimental data (growth phenotypes, proteomics), correcting for the fact that a physicochemically plausible point-estimate kcat may not reflect realized in vivo behavior. DLKcat itself includes a Bayesian SMC-ABC refinement step in this spirit, and EnzymeTuning is a dedicated GAN-based framework for this task.
The DLKcat Generalization Dispute¶
Disputed: DLKcat's original paper reported strong mutation-sensitivity (Pearson r=0.94 on a curated mutant panel) and broad applicability across 343 yeast/fungal species. Kroll & Lercher (2024) subsequently showed that 67.9% of DLKcat's nominal test-set enzymes were also in its training data, that DLKcat's R² is negative for enzymes below 60% sequence identity to the training set, and that its mutation-effect predictions on genuinely held-out mutants achieve R²=-0.18 to -0.21 — directly contradicting the original r=0.94 claim. This dispute is preserved rather than resolved in favor of either side on this page: DLKcat's architecture and its genome-scale ecGEM application remain widely used and cited, but its specific generalization claims should be read alongside the critique. The dispute reshaped the field's evaluation norms — every subsequent model reviewed here (TurNuP, CatPred, CataPro) explicitly stratifies evaluation by sequence identity to the training set, reports performance on strictly out-of-distribution/unbiased splits, and in CataPro's and CatPred's case, re-benchmarks DLKcat and UniKP under these stricter conditions.
Benchmark Landscape (kcat prediction, approximate)¶
| Model | Held-out / in-distribution R² | Out-of-distribution / low-similarity R² | Notes |
|---|---|---|---|
| DLKcat | 0.445 (barely above a 0.420 trivial baseline) | negative below 60% sequence identity | severe train/test leakage identified post hoc |
| TurNuP | R²=0.44 (test set) | R²=0.33 at <40% sequence identity | reaction fingerprints + retrained ESM-1b |
| UniKP | R²=0.65 (Extra Trees) | R²=0.260 at ≤40% identity (per CatPred's re-evaluation) | best on in-distribution/random splits |
| CatPred | R²=0.607 (held-out) | R²=0.390 (OOD); R²=0.33 at <40% identity | explicit uncertainty quantification |
| CataPro | competitive with TurNuP when retrained on the same data (PCC=0.672 vs. 0.67) | R²=0.333 at 0–40% identity | unbiased, sequence-clustered 10-fold benchmark |
These figures are drawn from different papers' own reported evaluations (not a single head-to-head study) and are not perfectly comparable across differing dataset compositions and identity-cutoff definitions; they are included to give an approximate sense of relative robustness to sequence novelty, which is the central axis of disagreement in this field. See each tool's page for the exact evaluation protocol.
Complementary Experimental Angle¶
Evolutionary-Scale Enzyme Kinetics Landscape Mapping approaches the same underlying limitation — sparse, heterogeneous public kinetic databases — from the data-generation side rather than the modeling side, using high-throughput microfluidic enzymology to produce a large, internally consistent kinetic dataset for one enzyme family (adenylate kinase) across evolutionary breadth, and showing that models trained on this denser, consistent dataset outperform models trained only on public databases — a complementary line of evidence to the DLKcat critique's conclusion that data quality, not just model architecture, limits this field.
Key Implementations¶
- Heckmann Enzyme Turnover ML Model — an early (2018) classical-ML precedent using structural, network, and biochemical features rather than end-to-end sequence deep learning; validated via downstream proteome-allocation model accuracy.
- DLKcat — the original GNN+CNN kcat predictor and ecGEM-parameterization pipeline; its generalization claims are disputed (see above).
- TurNuP — reaction-fingerprint + Transformer model designed for robustness to unfamiliar enzymes.
- Structural Km Predictor — GNN-based Km prediction, precursor to TurNuP's architecture.
- UniKP — unified kcat/Km/kcat-Km framework with an environmental-factor extension (EF-UniKP).
- CatPred — kcat/Km/Ki prediction with per-query uncertainty quantification.
- CataPro — introduces unbiased, sequence-clustered benchmark datasets; applied to vanillin-pathway enzyme discovery.
- EnzGFM — enzyme-domain-adapted protein language model (Mamba-Transformer hybrid) underlying an agentic enzyme-engineering pipeline.
- EnzymeTuning — GAN-based top-down kcat calibration against in vivo proteomics/growth data.
See Also¶
- DLKcat Generalization Failure and Train/Test Leakage — full detail on the central methodological dispute in this field.
- Evolutionary-Scale Enzyme Kinetics Landscape Mapping — a large, consistent experimental kinetic dataset complementing the ML models above.
Citations¶
[1] Li, F., Yuan, L., Lu, H., Li, G., Chen, Y., Engqvist, M.K.M., Kerkhoven, E.J., & Nielsen, J. (2022). Deep learning-based kcat prediction enables improved enzyme-constrained model reconstruction. Nature Catalysis, 5, 662–672. [2] Kroll, A., & Lercher, M. J. (2024). DLKcat cannot predict meaningful kcat values for mutants and unfamiliar enzymes. Biology Methods and Protocols, 9(1), bpae061.