Thousands of AI agents analyzed biological data from millions of mouse cells in 32 minutes, a task the team estimates could take people several months to a year. The work, conducted by OpenAI with PerturbAI and Scripps Research, explored how gene changes affect different cell types in the mouse brain.
Agents examined one gene perturbation at a time
The analysis used GPT-Rosalind agents to work through a whole-brain CRISPR atlas. The atlas maps the effects of gene changes across cell types in the mouse brain, giving the agents a large set of cellular responses to examine.
Each agent was assigned a single gene perturbation. It analyzed how cells responded to that change and searched scientific literature for possible explanations. In combination, the agents covered the atlas as a distributed analysis rather than having one agent handle every perturbation.
The full process took 32 minutes. OpenAI estimates that a manual analysis of the same scope would take anywhere from several months to a year. That comparison highlights the speed of dividing a large dataset into many focused investigations, though the findings still point researchers toward questions to examine further.
Reports surfaced leads for further study
Among the findings were unexpected differences between two disease-related genes and a possible link between cholesterol regulation and neurodegeneration. These are leads from the analysis, not results described as confirmed through laboratory testing.
OpenAI says the leads could help researchers decide which mechanisms to test in the lab first. That is a practical role for the analysis: it can help prioritize where researchers direct experimental attention, while the source describes lab work as the next step for testing mechanisms.
The resulting collection of findings is called the “Reasoning Atlas.” It brings together the agents’ analyses and explanations, connecting observed cellular responses with possible interpretations drawn from scientific literature.
Experts compared distributed and single-agent reports
In blind comparisons, human experts preferred reports from the distributed agents over reports from a single agent 75 percent of the time. The comparison offers one measure of how the multi-agent approach performed in this exercise.
The result does not, by itself, establish that every report is correct or that every biological interpretation is validated. It shows that experts favored the distributed reports in the blind comparison described by the team, while the study’s suggested mechanisms remain candidates for researchers to investigate.
Atlas and paper are available
The Reasoning Atlas is available on GitHub, and the accompanying paper is out on bioRxiv. Together, the atlas and paper provide access to the reported analysis and its underlying research account.
The work brings together biological data, literature searching, and many specialized AI agents in a single analysis. Its immediate contribution is a set of findings that may help scientists choose what to test next, alongside a large-scale demonstration completed in a fraction of the estimated time for manual analysis.