GPT-Rosalind agents and PerturbAI's brain atlas

OpenAI, PerturbAI, and Scripps Research deployed 1,946 GPT-Rosalind-5.5 agents in parallel to analyze a whole-brain mouse CRISPR atlas, completing the analysis in 32 minutes. Each agent analyzed a single gene knockout and independently searched scientific literature. The agents matched or refined 48 of 55 published findings from the atlas paper, even though that paper appeared after the model's December 2025 knowledge cutoff. Running the analysis in parallel produced 711 times more literature citations than routing all 893 genes to a single agent, and blinded PhD reviewers preferred the multi-agent reports 75% of the time. Among the outputs, the agents identified a potential drug discovery lead tied to the DCTN1 gene and Perry syndrome: GPT-Rosalind found that disrupting DCTN1 increases expression of cholesterol-synthesis genes and hypothesized that cholesterol may not be reaching neurons where it is needed. The collaboration, which ran for a year, produced an AI reasoning framework that connects experimental results with existing research and generates testable hypotheses. Findings were published in a joint preprint and blog post. The CRISPR atlas underlying the experiment was launched by PerturbAI in March 2026 and has been downloaded more than 40,000 times; PerturbAI described it as 'the largest known causal genomics dataset produced in living tissue'. OpenAI and its collaborators said they are expanding the partnership to identify additional drug targets.