THE CRUNCH
OpenAI, PerturbAI and Scripps Research deployed thousands of AI agents to comb through biological data from millions of mouse brain cells, and the whole analysis took 32 minutes. Each agent handled a single gene perturbation, examined how the cells responded and searched the scientific literature for explanations. OpenAI estimates the same job done by hand would take anywhere from several months to a year.
The agents, which OpenAI says ran on its GPT-Rosalind model, worked through a whole-brain CRISPR atlas: a map of how gene changes affect different cell types in the mouse brain. Among the findings were unexpected differences between two disease-related genes and a possible link between cholesterol regulation and neurodegeneration.
In blind comparisons, OpenAI reports, human experts preferred the reports from the distributed agents over those from a single agent 75 percent of the time. OpenAI positions the results as research leads rather than settled findings, suggesting which mechanisms scientists should test in the lab first. The resulting Reasoning Atlas is available on GitHub, and the accompanying paper is on bioRxiv.