An embedding-based framework enables statistical testing of gene-set function hypotheses inferred by large language models.
Emerging large language models (LLMs) can infer gene functions directly from gene lists, enabling hypothesis generation without predefined gene sets. However, these LLM-derived predictions are qualitative, and principled statistical validation is lacking. Here, we develop an embedding-based statistical framework that transforms gene and function descriptions into vector representations, enabling statistical testing of gene-gene and gene-function relationships and quantitative prioritization of de novo functional hypotheses inferred by LLMs. We benchmark seven state-of-the-art embedding models using curated and retrieval-augmented literature-derived gene descriptions across diverse biological contexts. OpenAI's text-embedding-3-large and Google's gemini-embedding-001 perform best, capturing gene-gene functional relationships in 88.7-92.5% of Gene Ontology biological processes and approximately 98.6% of canonical pathways. In gene-function association analyses, these models achieve high sensitivity (95.2-98.4%) and specificity (72.7-84.3%). Through contamination analysis and evaluation using experimentally informed protein assembly gene sets, our framework distinguishes biologically meaningful LLM-inferred hypotheses from noise, outperforming confidence-based inference and conventional enrichment analysis. We further develop the open-source R package DEGEmbedR and demonstrate its utility for interpreting a drug perturbation-derived differentially expressed gene (DEG) signature lacking significant conventional enrichment results. Together, these results establish LLM-derived embeddings as a quantitative foundation for functional genomics and the statistical validation of LLM-based gene function inference.