Predicting Enzyme pH Optima from Structure Using Equivariant Graph Neural Networks

Abstract

Enzyme activity and stability are strongly modulated by pH, making the catalytic pH optimum (pHopt) a key parameter in enzyme development and biotechnological applications. Experimental determination of pHopt is, however, labor-intensive and time-consuming, motivating the development of accurate computational prediction methods. Here, we introduce pHoptNN, an E(n)-equivariant graph neural network designed to predict enzyme pHopt directly from three-dimensional protein structures. pHoptNN was trained on a curated data set comprising nearly 12,000 enzymes with experimentally determined pHopt values and high-confidence structural models obtained from the Protein Data Bank and AlphaFold. The model represents enzymes as atomic-level molecular graphs, integrating structural, chemical, and electrostatic features. Model development was assisted by hyperparameter optimization using genetic and Bayesian search strategies. On a held-out test set, pHoptNN achieved a root-mean-square error (RMSE) of 0.588 pH units, substantially outperforming the sequence-based method EpHod (RMSE = 0.879). The model also showed strong out-of-distribution generalization, achieving RMSE values of 0.594 and 0.611 on test sets comprising held-out EC class 4 enzymes and enzymes sharing <20% sequence identity with the training set, respectively, compared with RMSE values of 0.878 and 0.883 for EpHod. Moreover, pHoptNN maintains robust predictive performance across different enzyme classes and pH ranges. These results demonstrate the utility of structure-based equivariant deep learning for enzyme pHopt prediction and highlight the potential of pHoptNN to accelerate enzyme discovery and engineering workflows.