Abstract
ABSTRACT: Understanding drug–gene relationships is essential for advancing targeted cancer therapies and drug repurposing strategies. However, the vast volume of biomedical literature poses significant challenges in efficiently extracting relevant insights. In this study, we developed an automated pipeline that leverages retrieval-augmented large language models (LLM) to infer drug–gene interactions using the most up-to-date biomedical literature. By integrating PubMed and state-of-the-art LLMs, our pipeline generates accurate, evidence-based inferences while addressing the limitations of static LLMs, such as outdated knowledge and the risk of producing misleading results. We systematically validated the pipeline’s performance using curated databases and demonstrated its ability to accurately identify both well-established and emerging drug targets. Using our pipeline, we constructed a pan-cancer drug–gene interaction network among hundreds of FDA-approved drugs and key oncogenes. In a case study on liver cancer, we identified and validated an association between CTNNB1 mutations and enhanced sensitivity to sorafenib, highlighting a potential therapeutic strategy for this challenging mutation. To facilitate broad accessibility, we developed GeneRxGPT, a user-friendly web application that enables cancer researchers to utilize the pipeline without programming expertise or extensive computational resources. It provides intuitive modules for drug–gene inference and network visualization, streamlining the exploration and interpretation of drug–gene relationships. We anticipate that GeneRxGPT will empower researchers to accelerate drug discovery and development, making it a valuable resource for the cancer research community. SIGNIFICANCE: This study presents a novel approach that integrates LLMs with real-time biomedical literature to uncover drug-gene relationships, transforming how cancer researchers identify therapeutic targets, repurpose drugs, and interpret complex molecular interactions. GeneRxGPT, our user-friendly tool, enables researchers to leverage this approach without requiring computational expertise.