Abstract
The turnover number (k(cat)) is a key parameter in enzyme kinetics that quantifies catalytic efficiency and underpins mechanistic understanding of enzyme activity. However, existing computational approaches for k(cat) prediction often suffer from limited accuracy and generalization, largely because most models rely only on substrate information and neglect the role of the full reaction context. Although TurNuP partially addresses this limitation by encoding entire reactions using binary fingerprints, its two-stage design separates enzyme and reaction modeling, limiting both predictive performance and interpretability. Here, we present PMAK, a deep learning framework that integrates pre-trained representations of enzyme sequences and reaction SMILES with a residue-aware attention mechanism. By jointly modeling enzymes and reactions, PMAK captures their interactions and highlights key residues contributing to catalytic activity. Comprehensive evaluations demonstrate that PMAK consistently outperforms existing methods, achieving average R(2) improvements of approximately 16.9% and 10.0% over TurNuP under new-reaction and new-enzyme settings, respectively, in five-fold cross-validation. Beyond improved accuracy, PMAK provides interpretable insights into residues associated with enzyme catalysis, offering a robust and informative tool for enzyme kinetic prediction and related applications.