The KGC Newsletter highlights a presentation that I believe will resonate with a larger audience. This blog describes how a carefully curated biomedical knowledge graph fuels drug repurposing. In her KGC 2025 presentation, Carolina González Cavazos, Staff Scientist at Scripps Research, showed step-by-step how her team finds new uses for existing drugs.
Developing new drugs is risky and expensive. New drugs fail clinical trials 90% of the time. Existing drugs, tested and known to be safe, often work for seemingly unrelated diseases.
Carolina worked with SMEs at Scripps Research to build a biomedical knowledge graph. DrugMechDB integrates 5,000 paths from drug to gene to disease, tracing their mechanism of action. Domain experts curate this knowledge graph.
Running a GNN over the knowledge graph identifies diseases existing drugs may address. It's a black box that tells you what but not why.
GNNs use embeddings, vector representations of the graph. “What we expect to see is that drugs that treat a disease with the same mechanistic path should end up close together in the embedding space,” said Carolina.
Proximity signals similarity. "What they are missing -- and what is really important for domain experts to make better informed decisions -- is to understand the biological rationale, the sequence of nodes and edges that explain the prediction coming from these GNN models,” Carolina explained.
To fill this gap, the team created Drug-Based Reasoning Explainer (DBR-X), which uses case-based reasoning, link prediction, and path identification to find new uses for existing drugs. Given a drug, DBR-X retrieves similar pathways, typically about 100. Then the team narrows them down to the most important pathway.
“The important paths module basically takes the set of paths that we identified with the link prediction module and ranks them based on importance,” explains Carolina. This is done through a process of elimination called perturbation. If a path is removed, and a probability score that was similar to the GNN’s score drops, that path shouldn’t be removed. Other paths can be omitted with little or no impact on the probability score.
DBR-X provides a framework for finding new uses for existing drugs. This presentation serves a broad audience, including anyone in the biomedical field as well as those interested in explainability using KGs.
To learn more:
KGC 2025 session on YouTube: DBR-X: Drug-Based Reasoning Explainer for Interpretable Drug Repositioning by Carolina González Cavazos
Papers
Drug Repositioning using Consilience of Knowledge Graph Completion Methods
A Case-Based Explainable Graph Neural Network Framework for Mechanistic Drug Repositioning
DrugMechDB: A Curated Database of Drug Mechanisms
Github
DBR-X
DrugMechDB