ProtGeoNet-Pocket: A Binding Site Prediction Approach Integrating Sequence, Geometry, and Graph Structure

作者：Mingjian Jiang, Z B Zhang, Teng Ma, Huaibin Hang, Yaping Fan, Shunpeng Pang, Wei Zhou, Yuanyuan Zhang · 发表于：Journal of Chemical Information and Modeling · 年份：2025 · DOI：10.1021/acs.jcim.5c01568 · 研究领域：Protein Structure and Dynamics、Machine Learning in Bioinformatics、Computational Drug Discovery Methods

ProtGeoNet-Pocket is an innovative multimodal prediction framework designed for protein binding site recognition, effectively integrating sequence information, geometric features, and graph-based structural representations. To address the structural complexity of proteins and the diversity of binding pocket shapes, ProtGeoNet-Pocket leverages multiscale structural information for high-precision binding site prediction. It uses a PointNet module to extract geometric features from residue coordinates and enhances them using an attention mechanism. The geometric features are then fused with encoded sequence features and graph edge features. The combined features are fed into a Graph Isomorphism Network (GIN) to capture topological relationships via a message-passing mechanism. ProtGeoNet-Pocket achieved an F1 score of 72.87% on the scPDB training set and demonstrated strong predictive performance across five independent benchmark data sets: COACH420, HOLO4K, SC6K, PDBbind, and ApoHolo. Furthermore, the visualization results confirm a high spatial overlap between the predicted and actual binding sites, demonstrating the superior performance of this method compared to existing ones.