OpenAI pointed nearly 2,000 AI agents at a massive mouse brain dataset. They finished their analysis in 32 minutes.
Human experts were projected to need anywhere from several months to a year for the same job. The agents also flagged a potential new drug target for neurodegenerative disease.
What the agents actually did
The project, announced on October 21, 2026, was a collaboration between OpenAI, PerturbAI and Scripps Research. The agents ran on OpenAI’s GPT-Rosalind system.
Their target was PerturbAI’s CRISPR atlas. It covers 8 million mouse brain cells and nearly 2,000 genes linked to disease.
The agents examined how cells responded to each gene disruption, then cross-referenced those responses against the existing scientific literature.
The result is what the team calls a reasoning atlas. It ties gene effects and cellular responses to established research, essentially layering an annotated explanation over raw data.
The reasoning atlas has been made publicly available.
Many agents beat one
The distributed agents cited more than 700 references. The single agent cited 11.
In blinded reviews, specialists preferred the multi-agent outputs 75% of the time over the single-agent versions.
The findings that stood out
The first involved two genes called Tsc1 and Tsc2. The agents found unexpected differences between how disrupting each gene affected movement-related regions of the brain, including the cerebellum. Those genes are relevant to tuberous sclerosis.
The second finding is tied to drug discovery. The agents generated a new hypothesis linking disruptions in a gene called DCTN1 to altered cholesterol regulation in neurons.
That connection matters for Perry syndrome. It is a neurodegenerative condition that resembles Parkinson’s disease, and cholesterol handling in neurons is considered a critical factor in it.
How we got here
The project grew out of a year-long collaboration between OpenAI and PerturbAI.
PerturbAI first released its CRISPR atlas publicly in March 2026. Since then, it has been downloaded more than 40,000 times.
The teams have also shared a preprint on bioRxiv.
What this means for drug discovery and AI
The results open the door to potential joint drug-target discovery efforts using PerturbAI’s in vivo datasets.
The findings still need experimental validation, the preprint still needs peer review, and a 75% expert preference rate also means reviewers favored the single agent a quarter of the time.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.