I Trained Knowledge Graph Embeddings to Find a Cure for My Disease. The Model Found One That Causes It.

A software engineer tested knowledge graph embeddings on DRKG to predict ulcerative colitis treatments.

A small training run looked promising, ranking known UC drugs such as Infliximab, Cyclosporine, Prednisone, Azathioprine, and Methotrexate highly.

But a broader evaluation also surfaced a drug linked to causing or worsening the disease, suggesting the model may be learning network patterns rather than biology.

The case raises questions about how reliably biomedical graph models can support drug repurposing claims and how they should be evaluated.