(Accepted & Published: 16 September 2026)
Predicting enzyme kinetics directly from protein sequences remains a major challenge in computational biology, particularly in understanding how mutations at catalytically essential residues influence enzyme activity. Existing machine learning models often fail to account for the functional consequences of such mutations, sometimes predicting wild-type-like kinetic properties even when substantial losses in catalytic activity occur. In this work, we introduce CatRange, a machine learning framework designed to robustly predict enzyme kinetic regimes for both wild-type and mutant enzymes. The framework is trained on CatLog-27k, a carefully curated dataset containing 27,176 in vitro enzyme–substrate kinetic records, developed through systematic auditing and correction of data from established biochemical databases, including BRENDA and SABIO-RK. Mutant records were manually verified against 2,158 source articles to improve data reliability. Rather than predicting exact numerical values, CatRange classifies catalytic turnover (kcat) and substrate affinity (KM) into logarithmically spaced ranges, providing biologically meaningful predictions that accommodate experimental variability while preserving distinctions between functional states. By integrating enzyme and substrate representations with gradient-boosted classification models, the framework demonstrates robust prediction across standard held-out, out-of-distribution, and few-shot mutation evaluation settings. Importantly, CatRange captures mutation-induced transitions between kinetic regimes, including substantial activity losses resulting from alterations to annotated catalytic residues. The framework also identifies non-enzyme sequence inputs, helping prevent biologically inappropriate kinetic predictions. Overall, this study establishes a mutation-sensitive and biologically informed approach to enzyme kinetic prediction, highlighting the importance of trustworthy data curation, transparent dataset development, and balanced model evaluation for advancing enzyme engineering, biocatalysis, and kinetic metabolic modelling.