RealKcat

Summary

RealKcat is a gradient-boosted enzyme kinetics predictor built to fix a specific failure mode of prior models: existing kcat/Km predictors (including DLKcat, TurNuP, UniKP, and CatPred) tend to give near-identical predictions for alanine substitutions at catalytic residues, i.e. they don't detect when a mutation destroys the active site. RealKcat addresses this via a manually re-curated training set and an explicit synthetic "inactive" class. It was validated on held-out mutations in the alkaline phosphatase PafA dataset.[1]

Data and Method

• KinHub-27k dataset: 27,176 paired kcat/Km entries curated from BRENDA and SABIO-RK, cross-checked against 2,158 primary-literature articles to resolve database inconsistencies (988 discrepancies in kcat, Km, mutant identity, and substrate identity corrected; ~1,072 new entries added).[1]
• Synthetic inactive class: catalytic residues (per UniProt/InterPro active-site annotations) were computationally substituted with alanine and assigned to a dedicated "Cluster 0" (kcat≈0), expanding an initial ~5,000 negative points to ~17,000 via InterPro-based motif matching, to teach the model that disrupting the catalytic apparatus abolishes activity.[1]
• Features: enzyme sequences are embedded with ESM-2 (1,280-dim) and substrates with ChemBERTa (768-dim) from isomeric SMILES, concatenated into a 2,048-dimensional feature vector.[1]
• Model: kcat and Km are framed as multi-class classification (order-of-magnitude bins, with dedicated clusters for extreme values) rather than regression, and predicted with XGBoost, selected from a grid search over ten model architectures and 25,671 hyperparameter combinations.[1]

Performance

• Held-out test accuracy: >88% (kcat) and >85% (Km) exact-bin accuracy; e-accuracy (within one order of magnitude) of 94.7% (kcat) and 97.2% (Km).[1]
• Alkaline phosphatase (PafA) mutant validation (1,016 single-site mutants from Markin et al.; R164A/R164G catalytic mutations held out of training): exact-bin accuracy of 53% (kcat) and 93% (Km), with e-accuracy of 96% (kcat) and 100% (Km) — the R164A/R164G shifts in particular confirmed the model is sensitive to disruption of the catalytic apparatus rather than relying on sequence similarity alone.[1]

Availability

Code and the KinHub-27k dataset are intended for release at chowdhurylab.github.io/downloads.html and github.com/TKAI-LAB-Mali/RealKcat upon publication. This is a preprint, not yet peer-reviewed; confidence is set to medium pending publication of the dataset.

See Also

• DLKcat and DLKcat Generalization Failure and Train/Test Leakage — the mutation-insensitivity problem RealKcat targets directly.
• CatPred — a comparable framework addressing generalization to dissimilar sequences (rather than catalytic-site mutation sensitivity specifically) via out-of-distribution test splitting.
• KcatNet — a contemporary structure-aware kcat predictor targeting genome-wide coverage rather than mutation sensitivity.
• Enzyme Kinetic Parameter Prediction — field overview.

Citations

[1] Chowdhury Lab et al. (2025), "Robust Prediction of Enzyme Variant Kinetics with RealKcat", bioRxiv Supports: dataset curation, method, all performance figures above. Location: Abstract; Results (dataset construction; model architecture; PafA mutation validation).