Gene expression analysis is central to biomedical discovery, but raw transcriptomic data is messy: multiple large, semi-structured files demand both computational flexibility and deep domain knowledge. According to an arXiv preprint accepted to EMNLP 2026 Findings, existing approaches fall short because they are either rigid workflows that fail on edge cases or fully autonomous agents that lack scientific precision.
GenoMAS, developed by Haoyang Liu, Yijiang Li, and Haohan Wang, takes a middle path. It orchestrates a team of LLM-based agents that communicate through typed message-passing protocols. A guided-planning framework lets programming agents decompose task guidelines into Action Units and then dynamically choose to advance, revise, bypass, or backtrack, keeping the analysis logically coherent while adapting to the idiosyncrasies of genomic data.
The system shows strong results on GenoTEX, achieving 89.13% CSC for data preprocessing and 60.48% F1 for identifying trait-associated genes, surpassing prior art by 10.61% and 16.85%, respectively. It also beats strong agent baselines on ScienceAgentBench and DiscoveryBench. Qualitative analysis further suggests that the gene-phenotype associations GenoMAS finds are biologically plausible and supported by the literature.