Model-Derived Mechanistic Insights into Structural, Surface, and Physicochemical Determinants Underlying Enzyme pH...

The optimal pH is a core functional parameter determining the catalytic efficiency and stability of enzymes, which is crucial for the optimization of industrial biocatalysts processes. To address the challenges of time-consuming and costly traditional experimental determinations, as well as the insufficient feature representation capabilities of existing computational models, we herein propose a multi-source attention Kolmogorov-Arnold Network model, named EnzMSA-KAN, for predicting enzyme optimal pH. This model achieves deep feature fusion and efficient regression through a three-stage architectural innovation. First, a feature fusion module is constructed to integrate protein sequence evolutionary information, residue physicochemical properties, and global statistical features, adaptively and dynamically adjusting the weight contributions of different source information to pH adaptation. Second, a two-layer Graph Attention Network is employed to capture spatial topology and sequence dependencies between residues, while a multi-head attention pooling mechanism is introduced to automatically focus on key functional regions such as active sites and surface charge distributions. Finally, a regularized KAN regression head is utilized to replace traditional fully connected layers, achieving accurate fitting of the complex non-linear relationships between the enzyme microenvironment and optimal pH based on learnable univariate function bases. Experiments on a benchmark dataset of 1554 enzyme sequences demonstrate that the model achieves a test set MAE of 0.57, RMSE of 0.81, and R² of 0.52, significantly outperforming existing mainstream methods. Moreover, attention visualization results verify the model's effective capture of key functional residues, endowing the prediction results with good biological interpretability. This research not only provides an efficient tool for enzyme function prediction but also offers a new computational paradigm for the rational design and engineering of industrial enzymes.