Thousands of AI agents worked through a mouse brain atlas in 32 minutes, a job researchers say takes months

OpenAI, PerturbAI and Scripps Research reported that 1,946 AI agents running on OpenAI's GPT-Rosalind-5.5 model analyzed a whole-brain CRISPR atlas of the mouse brain in 32 minutes, work the team estimates would take human researchers months to a year. Each agent handled one gene disruption and produced a report, which further agents merged into a searchable "reasoning atlas" of 3,397 findings released on GitHub; the agents recovered the Scripps lab's main conclusions despite a December 2025 knowledge cutoff preceding the lab's March paper. In a blinded evaluation, 30 doctoral-level experts preferred the distributed agents' interpretations in 45 of 60 head-to-head judgments (75%), and the agents surfaced new leads on Tsc1/Tsc2 in tuberous sclerosis and Dctn1 in Perry syndrome.

OpenAI’s artwork for its distributed biological reasoning research. Image: OpenAI Thousands of AI agents working side by side picked through a vast atlas of gene experiments on the mouse brain in 32 minutes, work their creators estimate would take human researchers months to a year. OpenAI, the start-up PerturbAI and Scripps Research described the project on Wednesday https://openai.com/index/scaling-ai-research-for-complex-biological-data , alongside a preprint on bioRxiv https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1 that has not yet been peer reviewed. The agents ran on GPT‑Rosalind, OpenAI’s model for life sciences, and the team says they reproduced what human scientists had found in the same data while turning up new leads on brain disease. What the agents actually did The data is a whole-brain CRISPR atlas built by Xin Jin’s lab at Scripps Research and published on Hugging Face https://huggingface.co/datasets/perturbai/wholebrain crispr atlas : nearly 2,000 genes switched off one at a time in living mice, with the effects measured across millions of brain cells. The lab’s own analysis needed “a team of experts in statistical analysis, genomics, and neuroscience”, the preprint says, and still left much of the biology “under-interpreted”. The new approach, which OpenAI calls distributed biological reasoning, handed each of 1,946 gene disruptions to its own agent. Each one looked at how cells responded, searched the scientific literature and wrote a report proposing what was going on. Further agents then merged the reports into a searchable “reasoning atlas” of 3,397 findings, now released on GitHub https://github.com/jinlabneurogenomics/distributed agent . To keep the test honest, the team says the agents never saw the lab’s original paper, and the model version used, GPT‑Rosalind‑5.5, has a stated knowledge cutoff of December 2025, before that paper came out in March. The agents still recovered its main conclusions. New leads on brain disease Two leads stand out. Tsc1 and Tsc2 are genes that work together to control mTOR, a key regulator of cell growth, and switching off either one had similar effects in several kinds of neuron. In the cerebellum, which coordinates movement, the agents found Tsc2 disruption changed dozens of genes involved in nerve signalling while Tsc1 disruption barely registered. That could matter for tuberous sclerosis, where cerebellum problems are strongly linked to TSC2 mutations. The second involves Dctn1, a gene behind Perry syndrome, a rare disease with Parkinson’s-like symptoms. Switching it off raised the activity of genes that make cholesterol, and the agents linked that to research suggesting faulty cholesterol transport can push cells to make more. That raises a question for the lab: could disrupted cholesterol inside neurons feed neurodegeneration? Tomasz Nowakowski, a neuroscientist at the University of California, San Francisco, who is not among the authors, said in OpenAI’s announcement: Sometimes the clues are already there, scattered across different fields, we just haven’t connected them. Tomasz Nowakowski, associate professor of neurological surgery, UCSF One agent versus thousands The team also gave a single agent the whole job, with the same model and resources. As the workload grew, it read less: at 893 gene disruptions it cited 11 references, against more than 700 for the distributed agents. To check quality, 30 doctoral-level experts scored 60 reports without knowing which system wrote them. In 20 head-to-head pairs they preferred the distributed agents’ interpretation in 45 of 60 judgments 75% . The agents’ own novelty ratings tracked the experts’ moderately well a correlation of 0.61 . Sarah Teichmann of the University of Cambridge, a founding co-chair of the Human Cell Atlas, said AI “could help more researchers explore these resources and turn unexpected observations into new questions to test.” The agents were less sure-footed at prediction. Asked how strongly 100 held-out gene disruptions would change cells, they got 59 right; they correctly spotted 142 of 172 that had no effect. The work also follows Claude designing antivenom proteins https://madrobot.blog/2026/10/20/anthropic-claude-scorpion-antivenom-protein-binders-citrini/ and Claude spotting a new CRISPR-like system https://madrobot.blog/2026/09/23/claude-discovers-crispr-like-enzyme-system-art/ as AI labs race to show their models can do real biology. The caveats Everything here is a hypothesis. The preprint says lab experiments remain “the test of the proposed mechanisms”, and none of the new leads has been tested yet. The figures come from the team itself: six authors work at OpenAI, two at PerturbAI, and Jin, the senior author, co-founded PerturbAI. The months-to-a-year comparison is the team’s own illustrative estimate. GPT‑Rosalind itself is not a public chatbot. OpenAI describes it https://openai.com/index/introducing-gpt-rosalind/ as a reasoning model for biology, drug discovery and translational medicine, and opened it to eligible organisations worldwide through a trusted-access programme in September, in ChatGPT, Codex and its API. Bigger datasets are coming: tech giants are pooling $1.8 billion to build an AI “virtual cell” https://madrobot.blog/2026/10/07/biohub-1-8-billion-virtual-cell-ai-biology-data-google-deepmind-meta-doe-nih/ . Why it matters Labs can now generate far more data than their scientists have time to read, and this is a credible sign that swarms of AI agents can do the first pass and hand back a ranked list of ideas. Whether any of those ideas cures anything depends on the slow part, the lab work, which no agent has done yet. Sources: Sun et al., bioRxiv preprint https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1 ; OpenAI announcement https://openai.com/index/scaling-ai-research-for-complex-biological-data Teichmann and Nowakowski quotes ; OpenAI, GPT‑Rosalind https://openai.com/index/introducing-gpt-rosalind/ ; reasoning atlas on GitHub https://github.com/jinlabneurogenomics/distributed agent ; whole-brain CRISPR atlas on Hugging Face https://huggingface.co/datasets/perturbai/wholebrain crispr atlas .