OpenAI, PerturbAI and Scripps Research split the analysis of an existing mouse brain dataset across thousands of agents powered by GPT-Rosalind. That one round of analysis, covering data from millions of brain cells and the related literature, took 32 minutes. The output is a set of research leads, not finished research: OpenAI describes the findings as directions for further experiments, and the proposed biological mechanisms still have to be tested in the lab. The research team estimated that comparable human analysis done one perturbation at a time would take months to a year.
What the agents did
The agents worked on a whole-brain CRISPR atlas, a dataset showing how gene disruptions affect different types of cells across the mouse brain. Each agent took one gene disruption, analyzed the cellular responses, searched the scientific literature and proposed explanations. Their reports were then combined into what OpenAI calls a reasoning atlas, linking genes, cellular effects, relevant publications and possible mechanisms. OpenAI calls the approach distributed biological reasoning and has made the reasoning atlas available to researchers.
OpenAI says the agents reproduced findings from earlier human analyses while turning up additional leads, and that their judgments of novelty and scientific value lined up with those of blinded human experts.
Why splitting the work helped
OpenAI tested whether a single agent given the whole dataset could do as well, keeping the model and resources fixed. As the workload grew, the single agent searched less of the literature. At the largest workload it cited 11 references, while the distributed agents cited more than 700. In blinded comparisons, human experts preferred the distributed-agent reports 75% of the time.
The explanation OpenAI gives is coverage. One agent facing the whole experiment tends to thin out its reading, much as scientists with limited time focus on the strongest signals. Giving each agent a single perturbation lets each one go deep, including on subtle or unexpected responses, and running them in parallel lets many perturbations be investigated at once. The same idea of splitting work across parallel agents has also reached consumer products, such as parallel agents in GPT-6 Pro.
The DCTN1 lead
The example OpenAI highlighted on X, at 21:32 Beijing time on October 21, involves DCTN1, a gene whose mutations cause Perry syndrome, a rare neurodegenerative disease with Parkinson-like symptoms. In the atlas, disrupting Dctn1 increased the expression of genes involved in making cholesterol. The agents connected this with research suggesting that problems with cholesterol transport may lead cells to make more cholesterol, and proposed that cholesterol may not be reaching where neurons need it.
That turns into questions for follow-up experiments: could DCTN1 disruption change how cholesterol is distributed in neurons, could that contribute to neurodegeneration, and could restoring normal distribution protect neurons? Human experts rated the hypothesis novel and worth pursuing. It remains a hypothesis from mouse data, not a demonstrated mechanism or a treatment.
The agents also flagged a difference between Tsc1 and Tsc2, two genes that work together to control mTOR. In the cerebellum, Tsc2 disruption changed dozens of genes involved in neuronal communication, while Tsc1 disruption produced little change. OpenAI frames this as a more specific set of questions for the lab.
What comes next
OpenAI says it is expanding the collaboration to identify new drug targets. The work fits OpenAI's wider push in life sciences, including its plan to fold Orchestra Bio into Rosalind. OpenAI's own framing is that scientists stay at the center: the agents surface leads, and researchers decide which to test.