Abstract
MOTIVATION: Gene set overrepresentation analysis (ORA) is widely used to interpret high-throughput transcriptomics and proteomics data, but traditional methods rely on human-curated gene set databases that lack flexibility. RESULTS: We introduce llm2geneset, a framework that leverages large language models (LLMs) to dynamically generate gene set databases tailored to input query genes, such as differentially expressed genes and a biological context specified in natural language. These databases integrate with methods, such as ORA, to assign biological functions to input genes. Benchmarking against human-curated gene sets demonstrates that LLMs generate gene sets comparable in quality to those curated by humans. llm2geneset also identifies biological processes represented by input gene sets, outperforming traditional ORA and direct LLM prompting. Applying the framework to RNA-seq data from iPSC-derived microglia treated with a TREM2 agonist highlights its potential for flexible, context-aware gene set generation and improved interpretation of high-throughput biological data. AVAILABILITY AND IMPLEMENTATION: llm2geneset is available as open source at https://github.com/Alector-BIO/llm2geneset and via a web interface at https://llm2geneset.streamlit.app.