OpenAI, in collaboration with PerturbAI and Scripps Research, utilized nearly 2,000 GPT-Rosalind AI agents to analyze a brain atlas containing approximately 7.7 million mouse brain cells. The project aims to identify new leads for treating neurodegenerative diseases and advancing drug-target discovery.
The multi-agent system processed 1,946 genetic perturbations by linking cellular responses to existing scientific literature. The process resulted in a public reasoning atlas that documents the agents' conclusions. OpenAI announced the research on October 21, 2026, following a preprint released around October 16.
In validation tests, the system successfully recovered 52 out of 55 known scientific findings, achieving a 94.5% recovery rate. Human experts preferred the multi-agent outputs over single-agent baselines in 75% of cases, noting that the multi-agent reports cited over 700 references compared to 11 for single-agent models.
The agents identified specific biological links, such as the connection between the DCTN1 gene and cholesterol transport, relevant to Perry syndrome. They also analyzed Tsc1 and Tsc2 disruptions in the cerebellum, which aligned with established clinical patterns for tuberous sclerosis. GPT-Rosalind, a specialized model for life sciences, entered research preview in early 2026.




