Hybrid Local-Global GNN for Protein Binding Interface Prediction via Ensemble Learning

Predicting protein-protein binding interface is essential for understanding biological mechanisms and developing novel therapeutic interventions, yet it remains a challenging task due to computational complexities and label imbalance issues. To address these limitations, we propose a Hybrid Local-Global Graph Neural Network (HLG-GNN) framework that integrates message passing neural network (MPNN) with state space models (SSMs) to predict protein binding interface. By incorporating SSMs - efficient in processing long-range dependencies - our model captures both global dependencies and local interaction patterns within protein structures. Additionally, we employ ensemble learning techniques to address the prevalent label imbalance problem in protein binding interface prediction. Our model outperforms conventional MPNN, achieving an AUC of 0.967 on the DIPS docking benchmark dataset. These findings highlight the effectiveness of integrating local and global features through our HLG-GNN framework to achieve more accurate protein binding interface predictions.