Enzyme Kinetic Parameter Prediction via Catalytic Pocket-Augmented Machine Learning

Predicting enzyme kinetic parameters is vital for enzyme engineering, yet modeling these relationships remains challenging due to catalytic complexity. While Large Language Model (LLM)-based methods perform well by encoding sequences, their accuracy is often limited by neglecting explicit substrate-binding interactions. Here, we present GraphKcat, a deep learning framework integrating 3D enzyme−substrate conformations for kinetic prediction. GraphKcat demonstrates strong predictive performance across multiple sequence-identity-based evaluation settings, outperforming existing models. Using GraphKcat, we identified an L-glutaminase and a raspberry zingerone synthase with significantly enhanced catalytic efficiencies. Model interpretation analyses and in silico mutation studies suggest that the model captures functionally relevant residue−substrate interactions and catalytic regions. Further molecular dynamics and electric field analyses indicate that these enhancements are associated with strengthened substrate binding and a more favorable electrostatic environment. Overall, our results suggest that GraphKcat captures biologically relevant structure−function relationships, providing a structure-informed framework for rational enzyme engineering and industrial biocatalysis.