Robust enzyme kinetics prediction through pairwise relative learning

Authors:

Abstract
Enzyme kinetics prediction is a fundamental challenge, and numerous computational methods have been developed. However, most methods rely on absolute regression across heterogeneous datasets, which are noisy and prone to regression-to-the-mean bias, limiting accurate identification of highly active enzymes. Here we present DeltaKcat, a siamese neural network framework that reformulates kinetics prediction as a pairwise relative learning problem. By learning pairwise differences within consistent experimental contexts, DeltaKcat improves accuracy, robustness, and generalization over state-of-the-art methods. This framework enables DeltaKcat to (i) mitigate inter-study variability, (ii) generalize to unseen enzymes and substrates, and (iii) prioritize highly active enzymes. We demonstrate these capabilities through systematic benchmarking and external adenylate kinase dataset validation. We anticipate that this approach will be broadly useful for enzyme discovery and engineering, and more generally for learning from noisy biological data.