OpenAI has described a large experiment in using swarms of AI agents to read biological data. Working with the startup PerturbAI and Scripps Research, the company ran thousands of agents powered by GPT-Rosalind, its life-sciences model, across an existing whole-brain CRISPR atlas built from millions of mouse brain cells. According to OpenAI's October 21, 2026 post, the full analysis took 32 minutes. The team estimates that people working through the same material one perturbation at a time would need months to a year.
The dataset records what happens to different brain cell types when individual genes are knocked out. OpenAI calls its method distributed biological reasoning: each agent took a single gene disruption, examined how cells responded, searched the published literature and wrote up possible explanations. Those reports were then merged into what the company calls a reasoning atlas, which maps genes to cellular effects, relevant papers and candidate mechanisms. OpenAI says the atlas is open for scientists to explore.
Two findings are highlighted. The first concerns Tsc1 and Tsc2, genes that jointly regulate mTOR, a central controller of cell growth and metabolism. Knocking out either gene looked similar in several neuron populations, but in the cerebellum only Tsc2 disruption shifted dozens of genes tied to neuronal signaling. OpenAI notes that cerebellar abnormalities in tuberous sclerosis are strongly associated with TSC2 mutations, which makes the contrast a candidate for targeted lab work. The second points to DCTN1, mutated in the Parkinson-like disorder Perry syndrome: disrupting it raised expression of cholesterol-synthesis genes, and the agents linked that to research on faulty cholesterol transport, suggesting a possible route to neurodegeneration.
OpenAI also tested whether splitting the work matters. Holding the model and compute budget fixed, it compared one agent handling the whole dataset with many agents in parallel. As the workload grew, the single agent consulted less literature, citing 11 references at the largest scale against more than 700 for the distributed setup. In blinded comparisons, human experts preferred the distributed-agent reports 75% of the time, and the company says the agents reproduced findings from earlier human analyses while their novelty judgments tracked those of the blinded experts.
The post includes supportive comments from outside scientists. Sarah Teichmann, a founding co-chair of the Human Cell Atlas, said AI could help more researchers mine shared cell atlases, and Tomasz Nowakowski of UCSF said the cholesterol link offers a fresh way to think about disease biology. Neither is described as an author of the evaluation.
For research groups, the practical claim is about coverage rather than certainty. Large perturbation screens produce far more signals than scientists can chase, and parallel agents may surface weaker or unexpected effects that would otherwise be ignored. The work also extends OpenAI's push into life sciences, which this week included the acquisition of Orchestra Bio for its Rosalind suite.
The limits are significant. The results are reported by OpenAI in a blog post, not in a peer-reviewed paper, and the expert-preference and runtime figures have not been independently replicated. Every lead is a hypothesis: none of the proposed mechanisms has yet been tested in the lab, and the data come from mice rather than people.