Abstract 3640: Mapping drug-gene interactions in cancer with large language models

Abstract Understanding drug-gene relationships is critical for advancing targeted cancer therapies, identifying opportunities for drug repurposing, and improving patient outcomes. However, the vast and growing volume of biomedical literature presents significant challenges for researchers attempting to efficiently extract relevant insights. To address this, we developed an automated pipeline that integrates state-of-the-art large language models (LLMs) with up-to-date literature to reliably infer drug-gene interactions. By seamlessly interfacing PubMed with LLMs, the pipeline generates efficient, evidence-based inferences, effectively overcoming the limitations of static LLMs, including outdated knowledge and the potential for inaccurate conclusions. We systematically validated the pipeline using curated collections of drug-gene pairs from established knowledgebases. Comparative performance evaluations of various LLMs, including GPT-4o (OpenAI), Gemini 1.5 Pro (Google), and several open-source models, revealed that GPT-4o slightly outperformed Gemini 1.5 Pro, with both achieving an area under the ROC curve (AUROC) exceeding 90%. These proprietary models demonstrated substantial performance advantages over open-source alternatives. Using this pipeline, we analyzed hundreds of key oncogenes and FDA-approved drugs to construct a comprehensive pan-cancer drug-gene interaction network. This network identifies therapeutic targets and drug repurposing opportunities. The highest density of drug-gene pairs was observed in breast, colorectal, and lung cancers, reflecting significant therapeutic development and research interest in these prevalent cancer types. Additionally, our analysis of liver cancer, the third leading cause of cancer-related deaths worldwide, uncovered intricate multi-target effects of approved drugs such as sorafenib and lenvatinib. Beyond cancer-type-specific findings, the analysis also revealed numerous drug-gene associations spanning multiple cancer types. In summary, our pipeline integrates real-time literature retrieval with advanced LLMs and offers a robust and accessible resource for anti-cancer drug development. Designed with user accessibility in mind, the pipeline has the potential to enable cancer researchers to analyze drug-gene relationships without requiring programming expertise or extensive computational resources. This work provides a foundation for advancing therapeutic exploration and supporting innovation in oncology. Citation Format: Ying-Ju Lai, Li-Ju Wang, Yufei Huang, Yu-Chiao Chiu. Mapping drug-gene interactions in cancer with large language models [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 3640.