Hierarchical Deep Learning Framework Integrating Structural Interaction Potentials and Evolutionary Information for Protein-Protein Interaction Affinity Prediction

Abstract

Protein-protein interactions (PPIs) are fundamental to cellular biology, making accurate prediction of protein-protein interaction affinity (PPIA) crucial for drug discovery and disease mechanism elucidation. Existing computational methods often lack explicit physical constraints or operate as "black boxes," hindering interpretability and generalization. To address these limitations, we introduce StructFuncNet, a novel hierarchical deep learning framework designed for highly accurate, generalizable, and interpretable PPIA prediction. Our approach uniquely integrates multi-level Structural Interaction Potentials (SIP) with rich evolutionary information and advanced graph neural networks (GNNs). This dual paradigm leverages SIP for physical and geometric realism, while GNNs and Transformer-like components learn synergistic dynamics and contextual corrections. StructFuncNet incorporates comprehensive input feature categories, including pre-trained protein residue features, SIP interface energies, evolutionary conservation scores, global molecular descriptors, and interface interaction counts. We extensively evaluated StructFuncNet on diverse and challenging benchmarks, including those for mutation affinity, multi-domain complexes, and intrinsically disordered proteins, as well as complexes derived from predicted structures. Our framework consistently achieves state-of-the-art performance, demonstrating high correlations across all tested scenarios. Furthermore, StructFuncNet proves robust on complexes derived from predicted structures. Ablation studies confirm the synergistic contributions of our multi-modal features and architectural components, underscoring StructFuncNets capacity to provide a robust, interpretable, and efficient solution for complex PPIA prediction challenges.