AI-Driven Drug Repurposing: A Knowledge Graph-based Approach for Rare Diseases

Abstract

Drug repurposing has emerged as a powerful strategy for accelerating the discovery of new therapeutic uses for existing drugs, with particular value in rare diseases where treatment options are limited. This work presents a knowledge graph-based framework developed within the SIMPATHIC project to extract, integrate, and analyze biomedical literature to identify potential drug candidates for nine rare neurological, neurometabolic, and neuromuscular syndromes.

Our approach begins with retrieving disease-specific literature from PubMed and PubMed Central using semantic search with Medical Subject Headings (MeSH). Entity recognition and relation extraction tools, SemRep and MetaMap, identify biomedical entities and relationships, converting raw text into structured knowledge. The extracted data is integrated into a knowledge graph enriched with curated information from DrugBank and Disease Ontology, resulting in 215,000 nodes and 5.5 million relations.

To identify novel drug repurposing candidates, we conduct link prediction experiments using our knowledge graph and three ground-truth drug-indication datasets collected from public repositories. We focus on three tasks: drug-disease, drug-gene, and drug-phenotype interactions. Using semantic path-based features, we train machine learning models to predict unknown links, while we also explore Graph Neural Networks.

Our experiments on the nine studied rare diseases show promising results, demonstrating that literature-based concept co-occurrence and article-topic relationships significantly enhance predictive performance. However, limited literature data results in missing drug-disease interactions, highlighting the need to incorporate gene and phenotype associations for low-resource diseases. Concept granularity remains a key challenge requiring careful handling. These findings represent an encouraging step toward effective drug repurposing for rare diseases, pending validation through wet-lab experiments.