Structure-Guided Hybrid GCN−Transformer Framework for Enzyme Optimal pH Prediction

Abstract

Accurate prediction of enzyme optimal pH (pHopt) is fundamental to biocatalyst engineering, pharmaceutical development, and process optimization, yet developing a reliable computational method remains challenging. To address this limitation, we present trPHopt, which is based on a structure-guided hybrid framework integrating the Graph Convolutional Network (GCN) with the Transformer. At the feature representation, trPHopt introduces new structural features to better characterize residue microenvironments relevant to pHopt. At the model architecture level, our method adopts a parallel dual-stream encoder: a structure-guided multi-head self-attention stream that incorporates geometric features as attention biases to capture long-range residue dependencies, and a Dense GCN stream that aggregates fused residue representations over distance-weighted residue graphs. The encoder-layer outputs are combined through a residual fusion gate, with the GCN contribution modulated by a learnable scalar coefficient. To alleviate the bias toward neutral predictions caused by label distribution imbalance, we further employ a two-stage calibration strategy that couples the coarse-grained acidic/neutral/alkaline classifier with a confidence-aware calibrator under a logical consistency constraint. On the independent test set, trPHopt consistently outperforms classical machine learning models and state-of-the-art methods, achieving an RMSE of 0.794 and an R2 of 0.527. These results demonstrate that trPHopt provides an accurate and practical method for enzyme pHopt prediction.