BioMedReasoner: Towards Multi-Hop Reasoning using Path-based Relational Learning on Biomedical Knowledge Graphs

Scientific discovery of novel drug assets requires rigorous reasoning over the complex interplay of biomedical entities. By integrating multi-modal entities, biomedical knowledge graphs (KGs) have emerged as powerful resources for tasks such as identifying targets and uncovering mechanisms of drug action. However, effective reasoning over those KGs is often limited by their incompleteness and noisy information, hindering reliable downstream tasks. To address this challenge, we propose BioMedReasoner, a modular reasoning framework that performs KG completion as a precursor to interpretable multi-hop reasoning. Our approach builds on Neural Bellman-Ford Networks (NBFNet) that formulates link prediction as a path-based relational learning problem, enabling interpretable multi-hop predictions for biomedical applications. We evaluated BioMedReasoner on PrimeKG, a publicly available large-scale multi-modal biomedical KG, and demonstrated its effectiveness in KG completion and downstream reasoning. As a proof of concept, we focus on genes associated with inflammatory bowel disease (IBD), including the JAK family, TNFSF15, IL17A, and IL17RA. By extracting paths connecting the genes to IBD and to their biological processes and molecular functions, and supplying these paths to a large language model (LLM) while withholding gene names, we show that interpretable graph paths enable biological reasoning and assessment of gene suitability as therapeutic targets. These results highlight the potential of path-based, interpretable biomedical KG reasoning to guide early-stage drug discovery.