OpenAI, PerturbAI and Scripps Research used thousands of AI agents to analyze biological data from millions of mouse cells in 32 minutes. The agents worked through a whole-brain CRISPR atlas, which maps how gene changes affect different cell types in the mouse brain.
According to OpenAI, the agents ran on GPT-Rosalind. Each agent examined one gene perturbation, analyzed the cells’ responses and searched scientific literature for possible explanations.
The team estimates that people doing the same analysis by hand would need several months to a year. That estimate describes the time required for this analysis, while the 32-minute figure is the time reported for the agents’ work.
The agents identified unexpected gene differences
The agents identified unexpected differences between two disease-related genes. They also found a possible link between cholesterol regulation and neurodegeneration. These are leads for further research, not confirmed mechanisms.
In blind comparisons, human experts preferred reports from the distributed agents over reports from a single agent 75 percent of the time. The resulting “Reasoning Atlas” is available on GitHub, and an accompanying paper is available on bioRxiv.
OpenAI says the leads could help researchers choose which mechanisms to test in the lab first.
