Enzyme kinetics prediction learns pairwise differences: kcat R² of 0.71 under cold start, zero-shot PCC of 0.40 on external adenylate kinase data
Synopsis
Enzyme kinetics prediction has relied on absolute regression over heterogeneous data, where the same enzyme-substrate system can differ by orders of magnitude across studies and models drift toward the mean. DeltaKcat recasts the task as pairwise relative learning: within one study it fixes either the enzyme or the substrate and a siamese network predicts the kinetic difference between two enzyme-substrate complexes, after which an experimentally characterized reference complex restores absolute values. Under a cold-start split, absolute kcat prediction reaches R² 0.71 and PCC 0.85, against R² 0.14 to 0.48 for DLKcat, CataPro and UniKP; on external adenylate kinase data excluded from training, zero-shot PCC is 0.40.
Interpretation
The prediction target moves from estimating each enzyme-substrate complex independently to estimating the kinetic difference between two complexes measured under the same experimental context. Earlier approaches regressed absolute values across heterogeneous datasets, where reported values for the same enzyme-substrate system can differ by orders of magnitude between studies, so models captured dataset-specific bias and systematically underestimated extreme values. Under the cold-start split, absolute kcat prediction reaches R² 0.71 and PCC 0.85, while DLKcat, CataPro and UniKP fall between R² 0.14 and 0.48; in publication-level evaluation, median PCC for kcat is 0.78 for DeltaKcat against 0.11 to 0.36 for baselines.
Relative differences can be converted back into absolute kcat, Km and kcat/Km through experimentally characterized reference complexes, making the model usable where absolute values are required. Relative predictions carry no absolute scale on their own, so a route from differences back to absolute values is needed. Under the random split, absolute prediction reaches R² 0.79 for kcat, 0.80 for Km and 0.84 for kcat/Km; repeated estimates for the same test complex across different reference anchors show standard deviations centered around 0.25.
The model still ranks highly active enzymes in an enzyme family absent from training and adapts quickly with a small amount of family-specific data. The external adenylate kinase dataset was completely excluded from training, and its enzyme family does not appear in the training data. Zero-shot PCC is 0.40 and SCC 0.33, against UniKP at PCC 0.22 and SCC 0.21; taking the top 5% of true kcat as highly active, DeltaKcat retrieves one highly active enzyme within its top 5 candidates while baselines retrieve none within their top 10, and across the top 30 candidates DeltaKcat recovers 56% against 22% for the best baseline UniKP; fine-tuning on 20 to 140 samples raises PCC to 0.69.
Perspective
The demonstrated scope covers relative and absolute kinetics prediction under random and cold-start splits, publication-level evaluation, evaluation stratified by sequence and substrate similarity, and zero-shot plus few-sample fine-tuning on an external adenylate kinase dataset excluded from training; the authors propose the approach for enzyme discovery, enzyme engineering and functional annotation.
The authors note that kinetic measurements remain limited overall, especially for rare reaction types and underrepresented enzyme classes; key environmental factors such as pH, temperature and assay conditions are not explicitly incorporated; and protein structural information and enzyme-substrate complex structures are not used. In fine-tuning experiments, performance rises steadily with training set size, with a substantial improvement beyond 60 samples.