OpenAI, PerturbAI and Scripps run thousands of agents over a mouse brain CRISPR atlas

OpenAI, the startup PerturbAI and Scripps Research ran thousands of AI agents in parallel over existing data from millions of mouse brain cells. According to OpenAI's account of the work, the agents ran on GPT-Rosalind, OpenAI's model for biology. They finished the analysis in 32 minutes. The research team estimates that people doing comparable analysis and literature review, one perturbation at a time, would need months to a year.

The point is not only speed. Large biology experiments now produce more data than researchers can examine closely. OpenAI says the agents reproduced findings from earlier human analyses, added leads that had not been reported, and pointed to specific experiments worth running next.

How the work was divided

The dataset is a whole-brain CRISPR atlas. It records how disrupting individual genes changes different types of cells across the mouse brain. Researchers can now alter thousands of genes and measure the effects across millions of cells. Working out what each result means for human disease still requires connecting it to decades of published research spread across many fields.

OpenAI calls its approach "distributed biological reasoning". Each agent took one gene disruption. It analysed the cellular response, searched the scientific literature and proposed explanations. The reports were then combined into what the team calls a reasoning atlas, which links genes, cellular effects, relevant papers and possible biological mechanisms. OpenAI says the atlas is publicly available for other scientists to explore.

The team also tested whether one agent given the whole dataset could do the same job. The model and the resources were kept the same. As the workload grew, the single agent searched less of the literature. At the largest workload it cited 11 references. The distributed agents cited more than 700. In blinded comparisons, human experts preferred the distributed-agent reports 75% of the time. OpenAI also says the agents' own judgments of how novel and valuable a finding was matched the judgments of blinded human experts.

Two leads from the data

OpenAI describes two findings in detail.

The first concerns Tsc1 and Tsc2, two genes that work together to control mTOR, a major regulator of cell growth, metabolism and aging. In several groups of neurons, disrupting either gene had similar effects. In the cerebellum, which helps coordinate movement, the two genes behaved differently. Disrupting Tsc2 changed the activity of dozens of genes involved in communication between neurons. Disrupting Tsc1 changed very little. OpenAI notes that in people with tuberous sclerosis, cerebellar abnormalities are strongly linked to TSC2 mutations. Drugs that reduce mTOR activity can treat some brain tumours and seizures in the disease. Learning and behavioural difficulties remain hard to treat. The agents identified a specific group of neurons and a set of cellular changes. That gives researchers narrower questions to test: which changes existing drugs might correct, and which might call for new approaches.

The second concerns DCTN1. Mutations in this gene cause Perry syndrome, a neurodegenerative disease with Parkinson-like symptoms. DCTN1 helps move cargo inside neurons. In the atlas, disrupting Dctn1 raised the expression of genes that make cholesterol. The agents linked this to research suggesting that when cholesterol transport fails, cells may respond by making more of it. That led to a hypothesis for follow-up work. DCTN1 disruption might change how cholesterol is distributed inside neurons, and that could contribute to neurodegeneration. If so, restoring normal distribution might protect neurons. OpenAI says human experts judged the hypothesis novel, consequential and worth pursuing.

Both findings are hypotheses. None has been tested in the lab yet. OpenAI presents them as a way to decide which experiments come next, not as results.

What the claims rest on

The figures come from OpenAI's own description of the project. These include the 32-minute run time, the months-to-a-year estimate for human work, the 11 versus more than 700 references and the 75% expert preference. The source does not say how many experts took part or how the blinded comparisons were set up. A citation count measures how widely the agents searched, not whether their reasoning was correct. The value of the Tsc2 and DCTN1 leads will depend on whether laboratory experiments confirm them.

The data itself is also a limit. The atlas comes from mouse cells. Any link to human disease is an inference that needs separate validation.

What it means for work

The work being automated here is the slow first stage of interpretation. That means going through each perturbation in a large screen, checking the literature and drafting a possible mechanism. In many labs, postdocs and analysts do this by hand, and they usually start with the strongest signals because time is short. With agents running in parallel, the weaker and unexpected responses can be examined too, so the human role moves towards judging, ranking and testing what the agents propose. OpenAI says scientists "remain at the center of discovery" and that the goal is to help them choose ideas, design experiments and interpret results. Some questions remain open. It is not yet clear how many of these leads will survive the bench. It is also unclear how labs will check agent reasoning at this scale, and what training early-career researchers will get if the reading and synthesis they once did is done for them.