Hyperbolic Wavelet Interaction Network (HyWinNet): Multi-Scale Geometric Learning for Interpretable Protein-Protein Interaction Prediction

Abstract

Understanding the multi-scale organization of protein-protein interactions (PPIs) is fundamental to deciphering cellular signaling, allosteric regulation, and disease mechanisms, yet existing computational approaches fail to simultaneously resolve atomic-scale binding interfaces and pathway-level coordination. We present HyWinNet, a geometric deep learning framework that unifies Lorentzian hyperbolic graph neural networks with spectral graph wavelet transforms to intrinsically model the hierarchical architecture of biological interactions. The core innovation lies in three synergistic components: (1) Hyperbolic embeddings that preserve evolutionary-conserved topological relationships through negatively curved space projections, capturing scale-free properties of protein networks; (2) Multi-scale wavelet decomposition operating at biologically grounded resolutions to jointly analyze local residue contacts and global pathway dynamics; (3) Contrastive learning with stochastic feature dropout that mimics natural binding interface plasticity. Benchmarked against state-of-the-art methods, HyWinNet demonstrates superior performance in recovering known interactions while predicting previously unannotated functional sites validated through independent structural and biochemical studies. By bridging hyperbolic geometry with graph signal processing, this work establishes a new paradigm for analyzing multi-scale biological networks, offering both interpretable computational predictions and testable hypotheses for experimental validation. The frameworks ability to map hierarchical relationships from atomic details to system-level modules provides a transformative tool for drug discovery and mechanistic studies of complex diseases.