DrugRepur AI: An Explainable Knowledge Graph Embedding Platform for Drug Repurposing with Multi-Level Biological Validation

Abstract

Background: Most computational drug repurposing systems generate predictions without providing mechanistic justification, biological context, or clinical evidence. This disconnect between prediction and understanding limits practical adoption by pharmaceutical researchers. Objective: To design and validate an integrated platform combining knowledge graph embedding-based prediction with multi-level biological validation for drug repurposing hypothesis generation. Methods: A RotatE knowledge graph embedding model was trained on a DrugBank-derived knowledge graph containing 16,698 entities and 2.94 million relational triples using PyKEEN. The prediction engine was wrapped in a validation architecture comprising dual-tiered explainability (path-based and embedding-based reasoning), disease pathway analysis, drug target identification with druggability scoring, biomarker discovery via transcriptomic reversal analysis, chemical similarity search, ClinicalTrials.gov integration, literature mining, pharmacovigilance-based safety profiling, and novelty assessment. Results: Internal validation against RepoDB yielded an MRR of 0.422, Hits@10 of 65.4%, and AUC-ROC of 0.847. External validation against the independent PREDICT dataset produced Hits@10 of 40%, confirming cross-dataset generalization. Case studies on Metformin and Hydroxychloroquine demonstrated the system across known associations, novel embedding-based predictions, and path-supported hypotheses with convergent biological evidence from pathway, transcriptomic, and chemical similarity analyses. Conclusion: Knowledge graph embeddings, when integrated within structured biological validation, can produce drug repurposing hypotheses that are scientifically defensible and clinically contextualized. The multi-level evidence architecture transforms numerical predictions into testable scientific hypotheses suitable for guiding early-stage experimental investigation.