Drug Repurposing with a Graph-of-Thoughts Inspired Reasoning Framework

ICLR 2026 Conference Submission20991 Authors

Keywords: Drug, Repurposing, Reasoning, LLM, Graph, Grounding, Discovery, Prompt, Reasoning, Theraputic, Biological, Biomedical

Abstract: Drug repurposing is a promising strategy to accelerate therapeutic development, and provides a viable and effective way to treat diseases, especially rare diseases, that otherwise do not have established and approved treatment options. As the magnitude of available biological data continues to increase, computational methods have become vital for extracting meaningful insights and identifying candidates for repurposing. Although large language model (LLM)-driven agentic workflows show potential, their computational cost and latency often make them impractical. We present a novel, efficient platform that addresses this challenge by grounding LLM reasoning in a biomedical knowledge graph (PrimeKG). This uses known relationships between biological entities to deduce relevant drug repurposing information. Our method first identifies multiple diverse paths between a given drug-disease pair from a natural language query. A Graph-of-Thoughts (GoT)-inspired module then constrains the LLM to reason over these structured paths and synthesize the information into a coherent biological hypothesis. Our platform presents key novelties such as Graph-of-Thoughts-inspired reasoning and diverse KG path-finding that seek to ground reasoning in biological knowledge and mitigate hallucinations. Our evaluation demonstrates that this constrained approach performs comparably in accuracy to unconstrained agentic workflows. Our platform achieves this with significantly fewer LLM calls, 55.3% lower token consumption, and 40.4% less time. This GoT-inspired framework, grounded on knowledge graph data, presents an efficient and powerful system for LLM-driven drug repurposing, effectively balancing reasoning with computational efficiency.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 20991