Close to 2,000 AI Agents Analyse a Mouse-Brain Gene Atlas in 32 Minutes

OpenAI, PerturbAI and Scripps Research have used close to 2,000 AI agents to analyse a large mouse-brain dataset in just 32 minutes, with findings published on Wednesday in a bioRxiv preprint that has not yet been peer-reviewed. The agents ran on GPT-Rosalind, a model built specifically for life sciences research.

The underlying data comes from a whole-brain CRISPR atlas built by PerturbAI, in which individual genes in mice were switched off one at a time to show how each change affects different brain cell types. One agent was assigned to each gene disruption. Each agent examined how the cells responded, searched the scientific literature, and suggested possible mechanistic explanations. All reports were then merged into a unified reasoning atlas linking genes, cellular effects, research papers, and likely biological mechanisms.

Had human researchers tackled the same task manually, the team estimates it would have taken months to a year. In blinded evaluations, independent experts preferred the multi-agent reports over those produced by a single agent 75% of the time. The multi-agent system also cited more than 700 references at its largest workload, compared with just 11 for a single agent.

Among the new leads identified: disrupting Tsc2 altered dozens of genes tied to nerve-cell communication in the cerebellum, while disrupting Tsc1 caused little change — a gap potentially relevant to tuberous sclerosis. In a separate finding, switching off Dctn1, a gene linked to the rare brain disease Perry syndrome, raised the activity of genes involved in cholesterol production, prompting suggested lab experiments to test its role in nerve-cell damage.

The reasoning atlas is publicly available on GitHub, and the underlying dataset is on Hugging Face.