GPT-Rosalind agents read an entire mouse whole-brain CRISPR atlas in 32 minutes. That is the result OpenAI announced on October 21, 2026, from a collaboration with PerturbAI and Scripps Research. The atlas records how millions of brain cells responded when nearly two thousand genes were disrupted, one at a time. Reading it perturbation by perturbation and cross-checking the literature would take human analysts an estimated several months to a year; thousands of agents working in parallel compressed the analysis and literature synthesis into about half an hour.
Read it in pieces, then assemble a reasoning atlas
The approach, called distributed biological reasoning, assigns each gene disruption to one agent: examine the cellular response, search the scientific literature, and propose a possible mechanism. All the reports are then combined into a reasoning atlas that links genes, cellular effects, relevant papers, and candidate mechanisms, now publicly available for researchers to explore. The agents reproduced findings from human analyses and additionally flagged weak signals that are easy to overlook. In large experiments, the most wasteful losses are exactly those observations outside the strongest signals that nobody has time to chase.
Two leads in the atlas were rated worth following up by human experts. One concerns Tsc1 and Tsc2, two genes in the mTOR pathway that regulates cell growth. Disrupting either had similar effects in most neurons, but in the cerebellum, which coordinates movement, disrupting Tsc2 changed the activity of dozens of genes involved in neuronal communication while Tsc1 disruption produced little change, matching the link between TSC2 mutations and cerebellar abnormalities in tuberous sclerosis. The other concerns DCTN1: disrupting it raised the expression of cholesterol-making genes in neurons, leading the agents to a testable hypothesis, that disturbed cholesterol distribution inside neurons may contribute to the neurodegeneration seen in Perry syndrome, a condition with Parkinson-like symptoms.
Where a single agent loses
The team ran a controlled comparison with the model and resources held fixed. At the largest workload, a single agent cited only 11 references while the distributed agents cited more than 700, and in blinded reviews human experts preferred the distributed interpretations 75 percent of the time. The gap is not a smarter model but wider coverage: what one agent cannot finish reading, thousands reading in parallel naturally miss less of. For laboratories, the value of such a system is not a drug handed over directly, but a ranking of which experiment is worth doing next. The leads still have to be tested in the lab, and no agent can take that step for them.
