The Pulse
OpenAI Deploys Thousands of Agents on Mouse Brain Data
OpenAI, PerturbAI and Scripps Research used thousands of GPT-Rosalind agents to analyze data from millions of mouse brain cells in 32 minutes. The teams say the system surfaced research leads involving Tsc1, Tsc2 and DCTN1, but the proposed
AI.info Team · Written with AI · fact-checked ·
“Blinded human experts preferred distributed-agent over single-agent interpretations and supported their discovery rankings, establishing a scalable approach turning large-scale experiments into grounded insights and follow-up priorities.”
Alexander Sun et al., “Distributed AI agents enable comprehensive discovery of biological insights at scale”
OpenAI, PerturbAI and Scripps Research say thousands of AI agents analyzed data from millions of mouse brain cells in 32 minutes, turning an existing gene-disruption atlas into a set of research leads. OpenAI described the collaboration on October 21, while the study appeared as a bioRxiv preprint on October 16. The teams estimate that comparable human analysis, done one gene perturbation at a time, would take months to a year.
The method, which OpenAI calls “distributed biological reasoning,” runs agents powered by GPT-Rosalind in parallel. Each agent examines a gene disruption, assesses how cells respond, searches scientific literature and proposes an explanation. The team then combines the individual reports into a “reasoning atlas” connecting genes, cellular effects, papers and possible mechanisms.
Agents examine different effects of Tsc1 and Tsc2
One lead concerns Tsc1 and Tsc2, genes that work together to regulate mTOR, a pathway involved in cell growth and metabolism. The agents found that disrupting either gene produced similar effects in several neuron groups, but the patterns diverged in the cerebellum. There, disruption of Tsc2 altered dozens of genes tied to neuronal communication, while disruption of Tsc1 produced little change.
The distinction could help researchers formulate more specific questions about tuberous sclerosis, a disease associated with mutations in these genes. OpenAI’s account says cerebellar abnormalities in people with the condition are strongly linked to TSC2 mutations, while existing medicines that reduce mTOR activity address some symptoms but not all learning and behavioral difficulties. The agent analysis points to cellular changes to investigate; it does not establish a treatment or a cause of those difficulties.
The DCTN1 result is a hypothesis, not a treatment lead
A second line of inquiry involves DCTN1, whose mutations cause Perry syndrome, a neurodegenerative disease with Parkinson-like symptoms. In the atlas, disrupting Dctn1 increased the expression of genes involved in cholesterol production. The agents connected that result to literature suggesting that problems with cholesterol transport can prompt cells to make more cholesterol, and proposed testing whether DCTN1 disruption changes cholesterol distribution inside neurons.
That proposed connection remains a question for experiments, not a demonstrated mechanism of neurodegeneration. OpenAI quotes Tomasz Nowakowski, associate professor of neurological surgery at UCSF, describing the cholesterol link as a “possible link” that could suggest a new way to think about disease biology. The team says the resulting atlas is publicly available for researchers to explore.
Parallel reports cite more literature
In a comparison using the same model and resources, the team tested distributed agents against a single agent given the broader workload. At the largest workload, the single agent cited 11 references, while the distributed reports cited more than 700. In blinded comparisons, human experts preferred the distributed-agent reports 75% of the time.
Those figures measure literature coverage and expert preference, not whether the proposed biology is correct. The authors say the agents reproduced findings from earlier human analyses and that their assessments of novelty and scientific value aligned with expert judgments. The next test is experimental: the Tsc1/Tsc2 differences and the DCTN1-cholesterol hypothesis need laboratory follow-up before they can be treated as confirmed biological mechanisms.