OpenAI accelerates distributed biological analysis on mouse brain
OpenAI, PerturbAI, and Scripps Research used thousands of AI agents to analyze mouse brain data in 32 minutes. The approach produced a public reasoning atlas and biological leads to test.
OpenAI, PerturbAI, and Scripps Research report having used thousands of GPT-Rosalind-powered AI agents to analyze experimental data on millions of mouse brain cells in 32 minutes. The work, conducted on a whole-brain-scale CRISPR atlas, resulted in a public reasoning atlas linking genes, cellular effects, and scientific publications. The goal is clear: to accelerate a research phase that would take several months or even a year if done manually, according to them.
Des milliers d’agents pour fractionner le travail
The approach, described by OpenAI as distributed biological reasoning, distributes the analysis among specialized agents. Each examined a genetic perturbation, observed the cellular response, searched the scientific literature, and proposed a possible explanation. The reports were then aggregated into an atlas that cross-references genes, relevant publications, and plausible biological mechanisms.
According to the announcement, this method reproduced results already obtained by human analyses while uncovering new leads. Human experts, blind to the source, assessed the novelty and scientific interest of the proposals in alignment with the agents’ evaluations. The process was also compared to a single agent handling the entire dataset: as the load increases, the isolated model reads less, down to only 11 references at the highest tested scale, versus more than 700 for the distributed agents. Not quite the same appetite for bibliography.
Deux pistes biologiques ont retenu l’attention
Among the results, the researchers report an unexpected difference between Tsc1 and Tsc2, two genes that jointly regulate mTOR, a major driver of cell growth, metabolism, and aging. In the cerebellum, invalidation of Tsc2 altered the activity of dozens of genes related to neuronal communication, whereas invalidation of Tsc1 had little effect. The announcement links this divergence to forms of tuberous sclerosis associated with TSC2, where cerebellar abnormalities are strongly involved.
Another lead concerns DCTN1, a gene whose certain mutations cause Perry syndrome, a neurodegenerative disease with symptoms similar to Parkinson’s. In the atlas, perturbation of Dctn1 increased the expression of genes involved in cholesterol production. The authors see this as a possible entry point to study cholesterol transport in neurons and its potential role in neurodegeneration. Tomasz Nowakowski, associate professor of neurosurgery at UCSF, believes this kind of cross-disciplinary signal shows how clues scattered across several fields can change the understanding of a disease.
Une base publique, des données massives et un papier encore en prépublication
The reasoning atlas and agent framework are published on GitHub under an MIT license. The repository includes the execution engine, corpus tools, and notebooks used for analyses, figures, and supplementary tables. The analyzed data come from PerturbAI and are hosted on Hugging Face: 7.7 million cells, 205 GB of data, 19,070 mRNAs, and 8,588 sgRNAs, with unique cell profiles on nearly 2,000 disease-related genes. All this relies on the Whole Mouse Brain taxonomy from the Allen Institute.
The companion paper, Distributed AI agents enable comprehensive discovery of biological insights at scale, was posted on bioRxiv on October 16, 2026, and has not yet been peer-reviewed. The authors state that the distributed system greatly expands literature coverage and helps prioritize experiments to launch next. In this context, AI does not replace the laboratory; it mainly helps decide where to look first.
By the numbers
- 32 minutes — duration of the analysis announced by OpenAI.
- 7.7 million cells — size of the dataset on Hugging Face.
- 75% — experts’ preference for distributed reports in blind comparison.
- More than 700 references — citations found by distributed agents at peak load.
- 11 references — citations by a single agent at the maximum tested load.