Enhancing Drug Repositioning with Interpretable Graph Neural Network (GNN) Models on Biomedical Knowledge Graphs

Abstract: Drug repositioning — finding new uses for already-approved drugs — is faster and safer than developing entirely new compounds from scratch. Biomedical Knowledge Graphs (KGs) are computational networks that map complex relationships between drugs, diseases, genes, and proteins, and can be mined to uncover hidden drug–disease connections that may point to new treatments. While Graph Neural Networks (GNNs) have shown promise in predicting these connections, they function as black boxes, providing no explanation for their predictions — making it difficult for researchers to trust or act on the results. In our work, we developed a GNN-based framework that not only predicts new drug–disease associations but also generates interpretable reasoning paths through the biomedical knowledge graph. This allows each prediction to be traced back to biologically meaningful evidence, improving the transparency and reliability of AI-driven drug repositioning.