E(Q)AGNN-PPIS: Attention Enhanced Equivariant Graph Neural Network for Protein-Protein Interaction Site Prediction

Identifying protein binding sites, the specific regions on a protein’s surface where interactions with other molecules occur, is crucial for understanding disease mechanisms and facilitating drug discovery. Although numerous computational techniques have been developed to identify protein binding sites, serving as a valuable screening tool that reduces the time and cost associated with conventional experimental approaches, achieving significant improvements in prediction accuracy remains a formidable challenge. Recent advancements in protein structure prediction, notably through tools like AlphaFold, have made vast numbers of 3-D protein structures available, presenting an opportunity to enhance binding site prediction methods. The availability of detailed 3-D structures has led to the development of Equivariant Graph Neural Networks (GNNs), which can analyze complex spatial relationships in protein structures while maintaining invariance to rotations and translations. However, current equivariant GNN methods still face limitations in fully exploiting the geometric features of protein structures. To address this, we introduce E(Q)AGNN-PPIS ¹ , an Equivariant Attention-Enhanced Graph Neural Network designed for predicting protein binding sites by leveraging 3-D protein structure. Our method augments the Equivariant GNN framework by integrating an attention mechanism. This attention component allows the model to focus on the most relevant structural features for binding site prediction, significantly enhancing its ability to capture complex spatial patterns and interactions within the protein structure. Our experimental findings underscore the enhanced performance of E(Q)AGNN-PPIS compared to current state-of-the-art approaches, exhibiting gains of 8.33% in the Area Under the Precision-Recall Curve (AUPRC) and 10% in the Matthews Correlation Coefficient (MCC) across benchmark datasets. Additionally, our method demonstrates robust generalization across proteins with varying sequence lengths, outperforming baseline methods.