Evaluating Knowledge-Graph Drug Repurposing for MASH Against Clinical Evidence

Biomedical knowledge graph (KG) methods for drug repurposing are commonly evaluated by their ability to recover known drug–disease associations. However, this evaluation does not determine whether models prioritize clinically effective drugs. This study addresses this limitation in the context of metabolic dysfunction-associated steatohepatitis (MASH), a condition with few approved treatments and frequent late-stage failures. Degree-matched network proximity, random walk with restart (RWR), ComplEx, R-GCN, and TxGNN checkpoint transfer were compared using the Final Biomedical KG, which integrates PrimeKG and Open Targets 26.03. For general drug repurposing, RWR achieved the highest AUROC (0.731), while network proximity provided the best early retrieval (filtered MRR 0.0295, Hits@50 0.1286). In MASH drug repurposing, R-GCN achieved the highest POS–NEG AUROC (0.800), although all confidence intervals included 0.5. Among 5,241 candidates, proximity achieved the highest filtered MRR (0.0165), whereas R-GCN had a filtered MRR of 0.0009 and ranked positive drugs substantially lower (median rank 1,576). Incorporating curated MASH mechanism genes improved RWR retrieval but did not enhance clinical separation. General drug repurposing performance, MASH clinical separation, and early retrieval across the full candidate library each evaluate distinct capabilities and should be reported independently.