Attention-Driven Graph Neural Networks for Knowledge Graph Completion in Biomedical Domains

Knowledge graphs (KGs) serve as critical infrastructure for biomedical knowledge representation, yet their inherent incompleteness limits downstream applications. We introduce ADGNN-Bio, an attention-driven graph neural network framework specifically designed for biomedical knowledge graph completion. Our model employs a hierarchical attention mechanism that captures both local structural patterns and global semantic relationships, incorporating domain-specific biomedical ontologies as inductive biases. Evaluated on three benchmark datasets—FB15k-237-Bio, UMLS, and DrugBank-KG—ADGNN-Bio achieves state-of-the-art performance with MRR scores of 0.524, 0.891, and 0.743 respectively, representing improvements of 8.3%, 5.1%, and 11.7% over the strongest baseline. Ablation studies confirm the critical contributions of both the hierarchical attention mechanism and ontology-guided initialization. Our framework enables more accurate prediction of novel drug-target interactions and disease-gene associations.