OpenAI Uses Thousands of AI Agents to Analyze Mouse Brain Cell Data in 32 Minutes
Working with PerturbAI and Scripps Research, OpenAI analyzed existing experimental data from millions of mouse brain cells to surface leads for follow-up work. The results are now public as a "reasoning atlas."
On October 21, OpenAI published a study, conducted with PerturbAI and Scripps Research, in which thousands of AI agents analyzed existing experimental data from millions of mouse brain cells. The team said it divided the analysis among thousands of GPT‑Rosalind-powered agents and ran them all at once, finishing in 32 minutes. The goal was to mine the changes observed in the cells for leads worth checking in follow-up experiments.
What the thousands of AI agents divided up
The researchers assigned each agent, a system that carries out tasks on its own, a single experimental case in which the function of one gene had been altered. Each agent examined how the cells responded when that gene was perturbed. The agents also reviewed the relevant literature and proposed biological mechanisms that could account for the changes they saw.
Pulling all of those results together produced what the team calls a “reasoning atlas”: a kind of research map that links genes, the changes seen in cells, the related literature, and possible explanations. Researchers can explore the atlas starting October 21, 2026, according to the team.
How it compared with a single AI doing the work
The team estimated that having humans run the same analysis and literature review for every experimental case would take anywhere from several months to a year. The 32 minutes reported here covers the time needed to analyze the existing experimental data. Confirming the hypotheses through follow-up experiments still lies ahead.
The team also compared the result with what a single AI produced on the same task. With the model and resources held constant, the report written collectively by many agents cited more than 700 references at the largest workload tested. The report written by a single AI cited 11. In a blinded comparison, where reviewers did not know which approach produced which report, human experts chose the multi-agent report in 75 percent of the paired comparisons, the team said.
Which follow-up experiments this points to
In the cerebellum data, a difference between the genes Tsc2 and Tsc1 emerged as one lead. Altering the function of Tsc2 shifted the activity of dozens of genes involved in communication between neurons. Altering Tsc1, by contrast, produced only small changes. The differing cellular responses to perturbing the two genes are captured in the analysis.
After the function of Dctn1 was altered, the activity of genes tied to cholesterol synthesis increased. On that basis, the team proposed the relationship between how cholesterol is distributed inside neurons and neurodegeneration as a target for follow-up experiments.
What this study demonstrates is a way of working: scanning vast amounts of cell data alongside the literature to surface the next question worth testing. For researchers, the publicly available atlas is a starting point for examining those questions. The thing to watch from here is whether the explanations the AI stitched together hold up in actual experiments.