1. GPT-Rosalind agents finish a whole-brain analysis in 32 minutes
OpenAI, working with the biotechnology company PerturbAI and the nonprofit Scripps Research, deployed nearly 2,000 GPT-Rosalind agents against a whole-brain CRISPR perturbation atlas covering roughly 7.7 million mouse brain cells. The full analysis took 32 minutes. OpenAI announced the work on October 21, 2026, following a companion preprint dated October 16.
The purpose is practical rather than a benchmark exercise. The collaborators want the system to surface leads that could help scientists understand and eventually treat neurodegenerative diseases, and they describe the project as a step toward drug-target discovery carried out jointly by human researchers and AI systems.
Instead of handing a single model one enormous task, the team split the atlas into many smaller pieces and assigned each piece to its own agent. Each agent examined what happened in the brain when a particular gene was switched off, tied those observations to published research, and proposed fresh hypotheses where the existing literature fell short. The combined output was released as a public resource that the partners call a reasoning atlas.
2. What is a CRISPR perturbation atlas?
A CRISPR perturbation atlas is a map of how cells change when specific genes are deliberately disrupted. CRISPR is the gene-editing technique that lets researchers target a chosen stretch of DNA with high precision. In a perturbation experiment, scientists use it to knock out one gene at a time and then measure how the affected cells respond, typically by reading which other genes become more or less active.
Repeat that process across many genes and many cells, and the result is a large table linking each disruption to its downstream consequences. The atlas used in this project covers the entire mouse brain, which means the measured responses span different regions and many types of neurons and supporting cells rather than a single dish of cultured tissue.
Datasets on this scale are valuable precisely because they record so many interactions at once. They are also hard to use. A human team can study a handful of genes in depth, but working through every disruption in a whole-brain atlas, checking each result against decades of published papers and writing up the possible explanations would normally take months or years of expert time.
3. How the multi-agent system divided the work
The approach rests on parallel effort. Each GPT-Rosalind agent was responsible for interpreting a slice of the atlas, focused on individual genetic perturbations. For every disruption it handled, an agent looked at the observed cellular responses, searched for prior research that could explain them, and then laid out new hypotheses where the evidence pointed somewhere the literature had not yet gone.
Because the agents worked at the same time rather than one after another, the whole job finished in about half an hour. That speed is the headline figure, but the partners place equal weight on transparency. The reasoning atlas documents how the system arrived at its conclusions, not merely the conclusions themselves, so researchers can inspect the chain of logic behind each proposed link between a gene and a cellular effect.
That design matters in biology, where a plausible-sounding explanation is worth little unless a scientist can trace the evidence behind it. By publishing the reasoning alongside the results, the collaborators are inviting outside researchers to judge which leads look strong enough to test in a laboratory and which ones fall apart under scrutiny.
The structure also makes the work easier to extend. Because every agent handled a defined unit of the atlas and recorded its own line of reasoning, individual entries can be revisited, compared or challenged without rerunning the entire analysis. For researchers who care about only one gene or one brain region, that means the relevant reasoning can be read on its own, with its supporting references attached, rather than buried inside a single sprawling report.
4. How well did the agents perform?
Human experts preferred the reports produced by the multi-agent system over those from single-agent baselines 75% of the time. In those comparisons, specialists reviewed the outputs and judged which analysis was more useful, and the distributed approach came out ahead in three out of four cases.
The gap in research depth was even more striking. Reports from the multi-agent setup cited more than 700 references, while the single-agent model cited only 11. Reviewers reportedly valued that comprehensiveness, which makes sense in a field where the context for a single gene may be scattered across hundreds of papers in genetics, neuroscience and clinical medicine.
The difference reflects a basic limit of asking one model to do everything. A single agent facing a very large workload has to ration its attention, and in practice that means searching less of the literature for each question. Splitting the task gives every agent room to dig into its own narrow problem, which is how the combined system ended up drawing on a far broader base of published evidence.
5. The DCTN1 hypothesis and Perry syndrome
Among the hypotheses the agents generated, one centers on a gene called DCTN1. The system proposed that disrupting DCTN1 is connected to changes in how neurons transport and synthesize cholesterol.
DCTN1 is associated with Perry syndrome, a rare inherited neurodegenerative disorder. The gene carries instructions for part of the dynactin complex, a piece of cellular machinery that helps move cargo along the internal tracks of cells, a job that is especially demanding in neurons because of their long projections.
A proposed link to cholesterol handling is the kind of lead the project was built to produce. It is a hypothesis, not a finding: it would need to be tested in laboratory experiments before anyone could say whether it explains part of the disease. Still, it shows how the system can connect a gene with a known clinical association to a specific biological process that researchers can then examine directly.
6. Tsc1 and Tsc2 behave differently in cerebellar neurons
A second result involves two closely related genes, Tsc1 and Tsc2. In the atlas, disrupting Tsc2 significantly altered gene expression in neurons of the cerebellum, the brain region that plays a central role in coordinating movement. Disrupting Tsc1, by contrast, did not produce a comparable effect.
Both genes are linked to tuberous sclerosis, a genetic condition that causes noncancerous growths in the brain and other organs and is often accompanied by seizures and developmental differences. The two genes produce proteins that work together, so a clear divergence in their effects is noteworthy.
The collaborators say the pattern matches what is known from clinical experience with tuberous sclerosis. That alignment serves two purposes. It suggests the agents can pick out biologically meaningful differences rather than noise, and it gives researchers a concrete, testable observation about how the two genes behave in a particular cell population.
7. What is GPT-Rosalind?
GPT-Rosalind is a specialized reasoning model from OpenAI built for the life sciences. It was first released as a research preview earlier in 2026. Unlike general-purpose chat models, it is aimed at scientific work such as interpreting biological data, connecting findings to existing literature and proposing directions for further study.
The brain atlas project is a showcase of that intent. Rather than answering isolated questions, the model was run as a large team of agents, each handling a defined scientific subtask, with the results assembled into a single structured resource.
Each partner brought a distinct strength to the collaboration. PerturbAI contributed its expertise in perturbation biology, the experimental approach behind the atlas itself. Scripps Research, one of the best-known biomedical research institutions in the United States, contributed scientific credibility and domain knowledge in disease biology. OpenAI supplied the model and the agent framework that coordinated the work.
8. Why speed matters for neurodegenerative disease research
Neurodegenerative diseases, a group that includes conditions such as Alzheimer's disease and Parkinson's disease alongside rarer disorders like Perry syndrome, remain among the hardest problems in medicine. Drug development in this area has a long record of setbacks, partly because the underlying biology involves many genes and cell types interacting in ways that are difficult to untangle.
Large perturbation datasets are one of the tools researchers use to tackle that complexity, but their value depends on how quickly and thoroughly they can be interpreted. A system that can work through an entire atlas in 32 minutes, cite hundreds of relevant studies and produce inspectable reasoning could change where scientists spend their time, shifting effort away from initial data review and toward designing experiments that test the most promising ideas.
The collaborators frame the work in exactly those terms: as a way to generate leads, not to replace laboratory science. Every hypothesis in the reasoning atlas still has to be confirmed or rejected through experiments, and the preprint format means the work is still open to review by the wider research community.
9. What comes next for the reasoning atlas
With the reasoning atlas now public, the next phase depends on how researchers use it. Scientists studying specific genes or brain regions can examine the system's proposed explanations, check its citations and decide which leads merit laboratory follow-up.
The DCTN1 and Tsc1/Tsc2 observations offer early test cases. If experiments support the proposed cholesterol link or confirm the divergent effects in cerebellar neurons, that would strengthen the case for multi-agent AI as a practical partner in drug-target discovery. If they do not, the transparent reasoning should make it easier to see where the system went wrong.
For OpenAI, PerturbAI and Scripps Research, the project is presented as a step rather than an endpoint: a demonstration that a coordinated team of GPT-Rosalind agents can turn one of the largest brain datasets available into a structured set of ideas in about half an hour.
