A Large-Scale Pharmacogenomic Knowledge Graph for Drug-Gene-Variant-Disease Discovery

Precision therapeutics depends on the ability to reason jointly over genes, variants, drugs, diseases, adverse drug reactions (ADRs), and molecular pathways without contaminating evaluation with future knowledge. I present a large-scale pharmacogenomic knowledge graph (PGx-KG) that integrates PharmGKB, ClinVar, SIDER, and Reactome—harmonized to HGNC, RxNorm, MeSH, and ChEBI identifiers—yielding 3,744,727 nodes and 9,645,367 edges across six major relation families. A leakage-free processing pipeline enforces version-aware chronological splits, publication-date audits, symmetric and transitive consistency checks, and cross-database de-duplication, eliminating temporal violations in held-out audits. As a first benchmark, a bilinear link-prediction model implemented in PyTorch Geometric achieves mean reciprocal rank (MRR) 0.347 (95% bootstrap CI [0.321, 0.368]), Hits@1/3/10 of 0.234/0.417/0.589, and AUROC 0.823 on validation data, with five-fold temporal cross-validation yielding 0.341 ± 0.018 MRR and a 2024 hold-out achieving MRR 0.329. Ranked candidate lists surface clinically relevant hypotheses, including CYP2D6–codeine dosing and HLA-B*15:02–carbamazepine risk, while also proposing pathway-level drug repurposing opportunities for expert review.