OpenAI set 1,946 AI agents loose on a mouse-brain gene atlas and got months of analysis in 32 minutes, plus a few leads worth testing

OpenAI, PerturbAI and Scripps Research said on Wednesday, Oct. 21, 2026, that 1,946 AI agents running OpenAI's life-sciences model GPT-Rosalind in parallel, one per gene disruption, analyzed an existing whole-brain CRISPR dataset from millions of mouse brain cells in 32 minutes and produced testable research leads.

Biology's bottleneck has quietly moved. Labs can now break thousands of genes in one experiment, but nobody has time to read what all of that means, so most of the data sits there. Throwing an agent at every gene is a brute-force fix that actually seems to work: it rediscovered what the human authors found and handed back a short list of things worth testing. The skeptic's notes matter, though. The "months to a year" figure is an estimate, the leads are mouse hypotheses, and the paper isn't peer reviewed yet. And OpenAI is clearly building GPT-Rosalind into something it can sell to drugmakers. The proof will be whether any of these leads survive a wet lab.

OpenAI, PerturbAI, and Scripps Research call the method distributed biological reasoning. Each of 1,946 agents received one gene disruption, a perturbation, from a previously published whole-brain Perturb-seq atlas of the mouse brain. Perturb-seq pairs CRISPR, a tool that breaks one chosen gene, with a readout of which other genes then change their activity, and it does that cell type by cell type. In vivo means the disruptions were made in living mice, not in a dish. Each agent read those cell-type gene-expression changes, searched the scientific literature, and proposed an explanation. The reports were merged into a searchable reasoning atlas. The agents read data that already existed. They did not run experiments. Those lines are OpenAI's and the preprint's.

The run took 32 minutes. The team estimated that a comparable human analysis and literature review, one perturbation at a time, would take months to a year. R&D World notes that this figure is an illustrative estimate, built from assumed times per gene and per paper, not a clocked race against a lab team. The 32 minutes are measured. The months-to-a-year line is the estimate, with that caveat.

The agents did not get the source study's manuscript. From the data they were given, they matched or refined 48 of the 55 findings they could test, R&D World reported. At the largest workload, a single agent cited about 11 references. The distributed agents cited 700-plus, and 7,586 of those references were kept. In blinded comparisons, where the reviewers did not know which method produced a report, human experts preferred the distributed-agent reports 75% of the time. Thirty doctoral-level biologists reviewed reports and broadly supported how the agents ranked the leads. Those comparisons are the study's, as reported with the result.

Two of the leads are hypotheses. They still need experiments at a lab bench. The data are from mice. The posts do not present a treatment, and they do not say the findings apply to people. First, the genes Tsc1 and Tsc2, which together regulate mTOR, a pathway that helps control how cells grow, behaved similarly in many neurons but differently in the cerebellum, the part of the brain that coordinates movement. There, disrupting Tsc2 changed dozens of genes involved in how neurons communicate. In people with tuberous sclerosis, a genetic condition, abnormalities in the cerebellum are strongly linked to mutations in TSC2. Second, disrupting Dctn1 raised the activity of genes that make cholesterol. Mutations in the human gene DCTN1 cause Perry syndrome, a neurodegenerative disease that looks like Parkinson's. The suggested link is between how neurons move cholesterol and neurodegeneration, the progressive loss of nerve cells. Both ideas are for a lab to test. Those lines are the research post's.

The reasoning atlas is available now for researchers to explore, OpenAI says. The written account is a bioRxiv preprint by Sun and colleagues, posted Oct. 16, 2026, under the title “Distributed AI agents enable comprehensive discovery of biological insights at scale.” A preprint is a paper shared before a journal has peer reviewed it. This one has not been peer reviewed. A day earlier, OpenAI said it had acquired Orchestra Bio and was folding that team into Rosalind.

Sarah Teichmann of the University of Cambridge, a founding co-chair of the Human Cell Atlas, said, "We build cell atlases so that other scientists can make discoveries we haven't imagined yet." Tomasz Nowakowski of UCSF said, "Sometimes the clues are already there, scattered across different fields—we just haven't connected them." Those quotations are theirs.

In plain terms, OpenAI, PerturbAI, and Scripps said on Wednesday that 1,946 copies of GPT-Rosalind read an existing mouse-brain gene atlas in 32 minutes and wrote a lead for each disrupted gene. Without seeing the original manuscript, they recovered most of the findings the human authors had already drawn from that data, and they handed back two ideas worth testing in a lab: a cerebellum-specific effect of Tsc2, and a cholesterol clue around Dctn1. Neither is a treatment. The paper has not been peer reviewed.