From Thousands to a Handful: Streamlining Breast Cancer Pathway Enrichment with LLMs

Over-representation of pathways (ORA) analysis on large gene lists from RNA-seq typically generates hundreds of terms, making it difficult to draw conclusions. In our work, we compare classical ORA (GO + KEGG), with LLM approach (llm2geneset) on 16 gene clusters generated from TCGA-BRCA data. Using GPT-4-turbo and GPT-4.1-mini models, the aforementioned package generates short, biologically significant pathways whose overrepresentation is tested. With the parameter npaths = 100 the average number of results drops to 5-15 terms per cluster, each of which appears in at least 100 publications with the phrase "breast cancer."Instead of the percentage reduction in "noise"or Dice ratio, we consider the number of citations in the literature as the key indicator, because it is relatively independent of the tool parameters. The entire experiment was conducted on GPT-4.1-mini at a cost of about $2.50. Our results show that LLMs can effectively filter ORA results, simplifying the interpretation of transcriptomic data and speeding up the formulation of biological hypotheses.