Tree-Guided Graph Neural Networks with Multilevel Optimization for Protein–Protein Interaction Prediction

Predicting human-virus protein-protein interactions (PPI) is crucial for revealing viral infection and disease development. Graph neural networks serve as a powerful tool to identify potential PPIs. However, their performance is often compromised by topological disturbances caused by noise, which will reduce their ability to capture global and local pattern information. To address this limitation, inspired by the distinct distribution of nutrients and energy in the trunk and leaves within natural trees, we propose Tree-Guided Graph Neural Networks with Multi-Level Optimization (TGGNN), aiming to capture fine-grained structures and improve their applicability in PPI prediction. Specifically, TGGNN introduces a hierarchically decoupled mechanism that decomposes the original graph into multilevel semantic subgraphs, including the trunk graph and leaf graph, achieving information aggregation across different levels. The trunk graph learns the core topology, while the leaf graph captures the fine-grained interaction patterns. To ensure semantic consistency across hierarchical levels, we design an attention-based routing, which serves as a cross-level semantic fusion interface to integrate information effectively. Extensive experiments demonstrate that TGGNN achieves superior prediction performance compared with state-of-the-art methods on four benchmark datasets. Additionally, case studies validate the ability of TGGNN to accurately identify human-viral PPIs, highlighting its practical utility in biomedical research.