AI and the World's Hard Problems · Life sciences · arXiv · 2026-10-21
GenoMAS: Guided Multi-Agent Programming for Scientific Discovery from Genomic Data
Gene expression analysis underpins many biomedical discoveries, but extracting insights from raw transcriptomic data is difficult because it involves multiple large semi-structured files and extensive domain expertise. GenoMAS is a team of LLM-based agents that combines structured workflows with autonomous adaptability: in a guided-planning framework, programming agents decompose task guidelines into Action Units and choose to advance, revise, bypass, or backtrack as the data demands. On the GenoTEX benchmark it achieves 89.13% CSC for data preprocessing and 60.48% F1 for identifying trait-associated genes, surpassing prior art by 10.61% and 16.85%, while also outperforming strong agent baselines on ScienceAgentBench and DiscoveryBench and surfacing gene-phenotype associations corroborated by the literature.