Knowledge-graph embeddings for osteoarthritis candidate prediction

Osteoarthritis (OA) is a prevalent, disabling joint disease with no approved disease modifying treatments. We present a knowledge-graph based approach to discover candidate treatments for OA by integrating large-scale biomedical data. We introduce the Osteoarthritis Knowledge-graph (OKG), a comprehensive network derived from the Drug Repurposing Knowledge-graph (DRKG) and enriched with causal genetic associations from OA genome-wide association study (GWAS) involving nearly 2 million individuals. We propose CausalPathKG, a knowledge-graph embedding model built upon RotatE that integrates domain specific innovations: (i) weighted gene OA edges reflecting GWAS significance, (ii) a path based regularization term to encourage drug gene OA causal connectivity, (iii) multi hop graph attention to prioritize informative paths, and (iv) self adversarial negative sampling with type consistent corruptions for robust training. CausalPathKG was trained to predict missing links, while withholding known OA-related edges for testing. In experiments, CausalPathKG outperformed TransE and RotatE baselines in predicting held out OA treatments, achieving higher link prediction accuracy and classification performance. Case studies highlight that top ranked repurposed drugs engage key OA-associated genes and pathways identified in human genetics. These results demonstrate that incorporating genetic evidence into knowledge-graph models can improve the discovery of therapeutics, offering a computational strategy to bridge human genomic data with drug repurposing.