VorRel-Net: Topology-Driven Multirelational Learning with Voronoi Contact Area for Protein–Ligand Interaction Site Prediction

Accurate identification of protein–ligand binding sites is a core task in computational biology and drug discovery. Existing protein graph construction methods generally define graph edges using only a single topological distance threshold, ignoring the complementary roles of various noncovalent interactions (such as hydrogen bonds, hydrophobic interactions, and long-range salt bridges) in stabilizing protein structures and mediating specific ligand recognition. Furthermore, prevailing models suffer from feature degradation and oversmoothing issues during deep-layer stacking, alongside insufficient capacity for adaptive fusion of multimodal features. To address these bottlenecks, we proposed VorRel-Net (Topology-Driven Multirelational Learning with Voronoi Contact Area for Protein–Ligand Interaction Site Prediction), which constructs a hierarchical multirelational graph to jointly model multimodal residue representations and diverse structural interactions. An adaptive feature evolution mechanism is designed to effectively capture high-order structural dependencies and fuse complementary biological information. In addition, a semantic feedback strategy is introduced to refine low-level representations using high-level semantic guidance, thereby alleviating information loss during deep feature propagation. Experimental results verify that VorRel-Net resolves dimensional mismatch and information redundancy stemming from conventional feature fusion strategies. The proposed method achieves state-of-the-art predictive accuracy and biological interpretability on the protein–ligand binding site prediction benchmark.