Nearly 2,000 AI agents analyze mouse brain gene data in 32 minutes
OpenAI and its research partners say the approach reproduced earlier findings and generated new leads, but the results have not been peer-reviewed.
A collaboration involving OpenAI, PerturbAI and Scripps Research used close to 2,000 AI agents to examine a mouse brain dataset, completing the analysis in 32 minutes. The team estimates that doing the equivalent work manually would take researchers months to a year.
The agents used GPT-Rosalind, OpenAI’s life sciences research model. Their results appear in a preprint on bioRxiv and have not yet undergone peer review.
PerturbAI created the dataset using CRISPR gene editing to deactivate individual genes in mice and observe the effects on different brain cell types. Each agent examined one gene alteration, reviewed scientific literature and proposed explanations for the cellular responses. The team combined those reports into a reasoning atlas connecting genes with cell effects, research papers and potential mechanisms.
OpenAI said the agents recovered findings from previous human analyses and identified additional research leads. Experts conducting blinded evaluations favored reports produced by multiple agents over a single agent’s work 75% of the time. Under the largest workload, the multi-agent approach cited more than 700 references; one agent cited 11.
One lead concerned Dctn1, a gene associated with Perry syndrome, a rare brain disease. Deactivating it increased the activity of cholesterol-producing genes. The agents proposed laboratory tests to investigate whether that response contributes to nerve cell damage.
The reasoning atlas is available through GitHub, while Hugging Face hosts the underlying dataset.
