Effective enzyme engineering through modeling kinetic parameter changes upon mutations with geometric deep learning

Quantifying enzyme kinetic parameter changes upon mutations is essential for accelerating enzyme evolution. Current methods focus on predicting absolute values of kinetic parameters (turnover numbers kcat and Michaelis constants Km), yet they fail to accurately capture mutation effects. To our knowledge, no method directly models mutation effects (Δkcat and ΔKm) due to the scarcity of such training data. Moreover, these methods haven’t fully exploited protein structural geometry. Here, we establish a high-quality enzyme mutation effect dataset by pairing wild-type and mutant kinetic parameters measured under identical experimental conditions via a semi-automatic procedure combining language model-assisted parsing and manual verification. Leveraging this dataset, we develop DeltaCata, a Siamese geometric graph neural network to directly predict the changes in kinetic parameters induced by residue substitutions. DeltaCata demonstrates substantial performance improvements over existing methods, especially for enzymes unseen in the training sets. Its utility is further validated through in silico beneficial mutation screening experiments using deep mutational scanning datasets. Finally, we apply DeltaCata to engineer the methyltransferase sGmOMT2, achieving up to a 7.0-fold improvement in catalytic efficiency for in vitro icaritin synthesis. These results suggest the high potential of DeltaCata for industrial enzyme engineering. Predicting how mutations affect enzyme catalytic efficiency remains challenging. Here the authors build a benchmark dataset and develop DeltaCata, a geometric deep learning model that predicts changes in enzyme kinetics and guides enzyme engineering.