OpenAI agents analyze mouse brain CRISPR atlas in 32 minutes
OpenAI teamed up with PerturbAI and Scripps Research to deploy thousands of AI agents that analyzed biological data from millions of mouse cells. According to OpenAI, the agents ran on GPT-Rosalind and worked through a whole-brain CRISPR atlas, which maps how gene changes affect different cell types in the mouse brain.
The full analysis took 32 minutes. Each agent took on a single gene perturbation, analyzed how the cells responded, and searched the scientific literature for explanations. The team estimates that doing the same analysis by hand would take anywhere from several months to a year.
Among other findings, the agents turned up unexpected differences between two disease-related genes and a possible link between cholesterol regulation and neurodegeneration. In blind comparisons, human experts preferred the distributed agents' reports over those from a single agent 75 percent of the time.
The resulting "Reasoning Atlas" is available on GitHub, and the accompanying paper is out on bioRxiv. OpenAI says these leads should help researchers decide which mechanisms to test in the lab first.
Key facts
- OpenAI, PerturbAI and Scripps Research deployed thousands of agents, running on GPT-Rosalind according to OpenAI, over data from millions of mouse cells.
- The full analysis of the whole-brain CRISPR atlas took 32 minutes; the team estimates the same work by hand would take several months to a year.
- Each agent handled one gene perturbation, analyzed how the cells responded and searched the scientific literature for explanations.
- Human experts preferred the distributed agents' reports over a single agent's 75 percent of the time in blind comparisons.
- Findings include unexpected differences between two disease-related genes and a possible link between cholesterol regulation and neurodegeneration; the "Reasoning Atlas" is on GitHub and the paper is on bioRxiv.
Why it matters
The headline claim is about speed: 32 minutes for an analysis the team estimates would take several months to a year by hand. The approach splits the work so that each agent covers one gene perturbation, then reads the cell response data and the literature for it. The comparison with a single agent is also part of the claim: experts preferred the distributed setup's reports 75 percent of the time.
Who it affects
The most direct audience is researchers working with large perturbation datasets such as CRISPR atlases, who could use the published Reasoning Atlas to pick mechanisms to test in the lab first. The work involved OpenAI, PerturbAI and Scripps Research.
How to use it
The "Reasoning Atlas" is available on GitHub and the accompanying paper is on bioRxiv, so researchers can inspect the outputs and the method. OpenAI says the leads should help researchers decide which mechanisms to test in the lab first.
How solid is it
The speed and expert-preference figures come from OpenAI and the team. The by-hand estimate is the team's own, with no method given. The 75 percent figure comes from blind comparisons by human experts; the source does not say who the experts were or how many comparisons were run. The source does not say that the paper has been peer reviewed.
Risks and caveats
The findings are described as unexpected or possible, not confirmed, and no lab validation is reported. The two disease-related genes are not named. No exact agent or cell counts are given, only "thousands" and "millions". No hardware, compute cost or token cost for the 32-minute run is given.