CatPred
Summary¶
Boorla & Maranas (2025) present CatPred, a deep learning framework predicting three enzyme kinetic parameters — turnover number (kcat), Michaelis constant (Km), and inhibition constant (Ki) — with per-prediction uncertainty estimates, evaluated using explicit sequence-identity-stratified out-of-distribution test sets designed to expose the kind of generalization failure identified in DLKcat by Kroll & Lercher. CatPred introduces new benchmark datasets (CatPred-DB: ~23k kcat, ~41k Km, ~12k Ki entries) and shows that protein language model (pLM) features are particularly important for out-of-distribution accuracy.
Method¶
- Inputs: substrate structural features, a self-attention sequence encoder ("Seq-Attn") over the enzyme sequence, pretrained protein language model (pLM) embeddings, and optionally 3D structural features from an equivariant graph neural network (E-GNN, "EGNN").
- Uncertainty quantification: a probabilistic regression framework separates aleatoric (data-noise) and epistemic (model-knowledge) uncertainty, giving each prediction a per-query variance estimate.
- Evaluation design: two test sets per parameter — a random 10% "held-out" split, and a stricter "out-of-distribution" subset (~12–15% of held-out) from which all enzymes with ≥99% sequence identity to any training enzyme are additionally excluded — directly addressing the leakage problem identified in DLKcat.
- Ensembling: ten replicate models per parameter are trained and averaged; production models use Substrate+Seq-Attn+pLM features for kcat/Km and Substrate+Seq-Attn+EGNN for Ki.
Performance¶
- Held-out test sets: R²=0.607 (kcat), 0.648 (Km), 0.637 (Ki) using the best Seq-Attn+pLM feature combination — matching or exceeding prior published models for each parameter.
- Out-of-distribution test sets: R²=0.390 (kcat), 0.536 (Km), 0.409 (Ki) — a much smaller relative drop than reported for DLKcat, whose R² becomes negative (R²=-0.61) on comparably dissimilar sequences. CatPred's out-of-distribution kcat accuracy (R²=0.40 comparable evaluation) matches TurNuP's reported R²=0.40 in a similar setting.
- Robustness down to 40% sequence identity: Km R²=0.48 and kcat R²=0.33 are maintained even for out-of-distribution enzymes below 40% maximal training-set sequence identity when pLM features are enabled.
- Comparison to UniKP and DLKcat: CatPred outperforms both baselines across held-out and out-of-distribution sets for all three parameters; the improvement over UniKP is most pronounced out-of-distribution (kcat R²=0.365 vs. 0.260 for CatPred vs. UniKP at ≤40% sequence identity; Km and Ki improved by ~8% and ~5% on average). DLKcat underperforms even a simple similarity-based baseline for kcat prediction specifically.
- Order-of-magnitude accuracy: averaged across ten replicate models, 76% (kcat), 84% (Km), and 68% (Ki) of held-out predictions fall within one order of magnitude of the true value, dropping to 61%, 79%, and 60% respectively out-of-distribution.
Availability¶
Code, pretrained models, and a Google Colab interface: github.com/maranasgroup/CatPred. Reproducibility capsule: CodeOcean. Data: Zenodo 10.5281/zenodo.14775002.
See Also¶
- DLKcat and DLKcat Generalization Failure and Train/Test Leakage — the leakage problem CatPred's evaluation design was built to expose.
- TurNuP — comparable out-of-distribution kcat accuracy via a different (reaction-fingerprint) architecture.
- UniKP — directly benchmarked against CatPred on the same CatPred-DB datasets.
- Enzyme Kinetic Parameter Prediction — field overview.