OpenAI sent thousands of agents through a mouse brain atlas. In 32 minutes they produced hypotheses, and the lab still has to confirm them

OpenAI, PerturbAI and Scripps Research had thousands of agents running GPT-Rosalind work through existing data from millions of mouse brain cells. In 32 minutes they produced a map of hypotheses that, by the team's estimate, would take humans months to a year. The interesting part is the way the work was divided, because no lab has yet confirmed the new leads.

Each agent took one genetic perturbation and chased the literature behind it

The data comes from a whole-brain CRISPR atlas, which shows how knocking out individual genes changes different cell types across the mouse brain. The team gave each agent one genetic perturbation. The agent reviewed the cellular response, searched the scientific literature and proposed an explanation. The individual reports were then combined into a reasoning atlas that links genes, cellular effects, publications and possible biological mechanisms. OpenAI calls the approach distributed biological reasoning.

According to OpenAI, the agents reproduced conclusions humans had already reached and added further leads. One concerns Tsc1 and Tsc2, two genes that jointly control the mTOR pathway. In the cerebellum, disrupting Tsc2 changed the activity of dozens of genes involved in neuronal communication, while Tsc1 barely changed anything. A second lead points to DCTN1, whose mutations cause Perry syndrome, a disease with Parkinson-like symptoms. Disrupting it raised the activity of cholesterol-synthesis genes, and the agents turned that into a question: could faulty cholesterol transport contribute to neurodegeneration?

The study appeared on October 16 as a bioRxiv preprint and has not been peer reviewed. Six authors work for OpenAI and two for PerturbAI. The reasoning atlas is public on GitHub and the dataset is on Hugging Face.

Many agents read more literature than one overloaded agent

The most useful comparison in the paper pits one agent against thousands. The team gave the same model, with the same resources, the whole dataset at once. The more work it got, the less literature it searched. At the largest workload it cited 11 references, while the distributed agents cited more than 700. In blinded comparisons, experts preferred the distributed-agent reports 75% of the time.

For labs sitting on large screening datasets, that is a practical lesson. A single agent with a giant brief starts cutting corners. Small, narrow tasks can run in parallel and be merged afterwards. Engineering teams know the same pattern from subagents, and it will likely spread anywhere data grows faster than the number of people reading it.

There is also a commercial layer. Over the past weeks OpenAI has been building a portfolio around the Rosalind line: an academic programme with selective access and the acquisition of the startup Orchestra Bio. The brain atlas study shows what the model is meant to do in practice.

The 32 minutes are measured against an estimate, not against a human team

The numbers need context. The months-to-a-year figure is the team's estimate for a human analysis going one perturbation at a time. No human team actually ran alongside the agents. The 75% preference compares two configurations of the same model, not the model against scientists. And agreement with expert ratings shows the hypotheses look plausible and novel. It does not show they are true.

The announcement also does not say how much compute thousands of parallel agents consumed. GPT-Rosalind is not a generally available model either: OpenAI grants access only to selected organisations and labs. Anyone hoping to repeat the approach on their own data currently needs an invitation.

A cholesterol experiment in neurons will settle it

The method will be confirmed or sunk by the first lab that takes a specific hypothesis and tests it. The DCTN1 lead is the obvious candidate. If cholesterol distribution in neurons really shifts, the agent-built map gets its first hard confirmation.

The second signal will be peer review, and whether groups with no ties to OpenAI adopt the method. The public atlas and data make that possible. The closed model, for now, less so.

Lilith's verdict

In half an hour, thousands of agents collected more than seven hundred citations and a list of suspicious genes for the mouse brain. Now comes the postdoc with a pipette, who has to find out whether cholesterol really wanders off in neurons or the agents just read about it.