Generating explainable hypotheses for drug repurposing with graph neural networks

Biomedical knowledge discovery increasingly relies on computational tools to uncover patterns in complex datasets, yet generating explainable, evidence-based hypotheses about biological interactions remains challenging. This study introduces XAIPath, an interpretable pipeline that leverages biomedical knowledge graphs and Graph Neural Networks (GNNs) to uncover and explain novel drug-disease relationships, aiding applications such as drug repurposing. The pipeline combines GNN-based predictions with a post-hoc interpretability layer that extracts simple paths connecting drug and disease nodes in a biomedical knowledge graph and compares them using MinHash-based similarity. Similar paths are grouped via K-means clustering to build interpretable clusters that represent alternative mechanistic hypotheses. The method was applied to the NeDRex knowledge graph for drug indication prediction, with performance evaluated using AUROC, AUPRC, precision, sensitivity, and specificity. XAIPath achieved strong predictive performance, with AUROC exceeding 95% and AUPRC over 90% across training, validation, and test sets, while precision, sensitivity, and specificity all surpassed 85%. Most high-scoring predictions were supported by existing literature, and the extracted path clusters closely aligned with DrugMechDB annotations, supporting the plausibility of the generated hypotheses. Overall, XAIPath offers a scalable and explainable approach for identifying drug-disease associations, facilitating hypothesis generation and biological validation, and highlighting the value of explainable AI in biomedical research and drug repurposing.