IT Home On October 23, OpenAI, in collaboration with PerturbAI and Scripps Research, used thousands of AI agents to analyze biological data from millions of mouse brain cells. The entire analysis process took only 32 minutes.
The research team estimates that if human researchers analyze each experiment result individually and search for relevant literature, completing the same work would take months to a year.
According to the introduction released by OpenAI on October 21, this research used the GPT-Rosalind model designed specifically for life science research to analyze a CRISPR map of the entire mouse brain. The map records the changes in various brain cells after different genes are intervened, which helps researchers understand gene functions and potential connections with diseases.
The research team adopted a distributed AI agent architecture, splitting numerous analysis tasks and executing them in parallel. Each agent was responsible for studying a genetic intervention, analyzing the responses of related cells, and searching scientific literature to find possible biological explanations for the observed changes.
Later, the researchers integrated the reports of various agents into a reasoning map called “Reasoning Atlas,” linking genes, cellular changes, references, and potential mechanisms of action.
Researchers have found that distributed agents can not only reproduce some of the conclusions from existing studies, but also provide some clues that are worth further verification. For example, Tsc1 and Tsc2 are two genes that jointly regulate processes such as cell growth and metabolism. However, studies have shown that in certain neurons in the cerebellum, interfering with Tsc2 alters the expression of various genes related to neuronal communication, while interference with Tsc1 does not produce the same level of change. This difference may help researchers better understand the brain abnormalities associated with tuberous sclerosis.
Another clue involves the DCTN1 gene related to neurodegenerative diseases. Studies have found that after interfering with this gene, the expression levels of some genes involved in cholesterol synthesis increase. The AI agent links this phenomenon to research on abnormal cholesterol transport in existing literature, and raises a question for verification: whether abnormal DCTN1 function may affect the cholesterol distribution within neurons and further contribute to neurodegenerative changes.
The research team also conducted a blind test comparison between distributed agents and individual agents. In 20 paired reports, biological experts preferred the reports generated by distributed agents in 75% of the evaluations; reports from individual agents received 16.7% of the evaluation preferences, while the remaining 8.3% resulted in tie outcomes.
Researchers say that as the scale of analysis tasks increases, distributed architectures can retrieve and reference more scientific literature. In the largest test, the distributed agent referenced over 700 papers, while a single agent only referenced 11.
The related research paper was published on the bioRxiv preprint platform on October 16, 2026, and has not yet been peer-reviewed. The research team emphasized that the biological connections and mechanisms proposed by AI still belong to research hypotheses that need verification, and further confirmation through experiments is still required.
The relevant reasoning graphs, analysis processes, and code have been made available to researchers, which can be used to explore gene intervention results, compare responses of different cell types, and identify research directions that warrant priority experimentation.
IT Home paper address:
https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1
Related Reading: