A Biophysical Prior-Based Hierarchical Graph Pooling Strategy for Protein Function Prediction

Accurate prediction of protein function is fundamental to understanding biological systems. Graph Neural Networks (GNNs) have demonstrated promise in this field; however, effectively aggregating residue-level features into discriminative global representations remains a core challenge. Existing graph pooling methods often rely on data-driven clustering, neglecting the inherent hierarchical organization of proteins, which may lead to the loss of critical structural information and limit their interpretability. To address this, we propose BioPhysPool, a biophysical prior-based hierarchical pooling strategy. Unlike traditional approaches, BioPhysPool utilizes secondary structure elements as a deterministic blueprint for graph coarsening, transforming fine-grained residue interaction networks into coarse-grained secondary structure networks without requiring computationally intensive assignment matrices. We further develop a hierarchical GNN architecture equipped with a cross-granularity attention fusion mechanism to adaptively integrate local chemical features with macro-topological patterns. Evaluations on rigorous benchmarks demonstrate that BioPhysPool outperforms different types of classical pooling methods in both binary and multi-class classification tasks. Furthermore, interpretability analysis reveals that attention allocation is primarily driven by structural context rather than amino acid type, with observed patterns consistent with known residue functional importance. These findings validate the effectiveness of embedding biophysical priors into the graph learning framework, offering a robust and interpretable solution for protein function prediction.