Gene Set Function Discovery with LLM-Based Agents and Knowledge Retrieval

Daniela Pinto Veizaga, Aécio Santos, Juliana Freire, Wenke Liu, Sarah Keegan, David Fenyo

Track: Tiny paper track (up to 4 pages)

Abstract: Advancements in high-throughput technologies have generated complex biomedical datasets, posing significant challenges for knowledge discovery. Traditional tools like Gene Set Enrichment Analysis (GSEA) and over-representation analysis (ORA) map gene sets to known pathways but are limited in their ability to uncover novel biological-mechanisms, often relying on manual interpretation to synthesize insights. While large language models (LLMs) aid in summarization, they lack transparency, adaptability to new knowledge, and integration with computational tools. To address these challenges, we introduce $\texttt{Discovera}$, an agentic system that combines LLMs with established computational bioinformatics pipelines, and retrieval-augmented generation (RAG) to support mechanistic discovery. $\texttt{Discovera}$ bridges the gap between computation and interpretation, enabling users to explore hypotheses grounded in data and literature. We demonstrate the utility of $\texttt{Discovera}$ in the context of endometrial carcinoma research, where it supports functional enrichment analysis and the summarization of potential mechanisms of action for gene sets associated with an observed phenotype.

Submission Number: 90