OpenAI, PerturbAI and Scripps Research used thousands of AI agents running on GPT-Rosalind to analyze a whole-brain mouse CRISPR atlas in 32 minutes, according to research OpenAI published Wednesday. The agents worked through existing experimental data from millions of mouse brain cells, linked the effects of individual gene disruptions to the published literature, and proposed biological mechanisms that researchers can now test in the lab.
The researchers estimated that comparable human analysis and literature review, conducted one perturbation at a time, would take months to a year. The work is described in a bioRxiv preprint, and the resulting "reasoning atlas" is available on GitHub so scientists can compare related genes.
The underlying dataset comes from PerturbAI, which emerged from stealth in March with what it called the world's largest in vivo CRISPR atlas: roughly 8 million cells across the mouse brain, covering thousands of genetic perturbations and hundreds of neuronal cell types. Researchers can now alter thousands of genes and measure the effects across millions of cells, but interpreting each result still takes expert time. That bottleneck is what the agents were pointed at.
“We build cell atlases so that other scientists can make discoveries we haven't imagined yet.”— Sarah Teichmann, Chair in Stem Cell Medicine, University of Cambridge, and founding co-chair of the Human Cell Atlas
Key facts
- 01OpenAI, PerturbAI and Scripps Research used thousands of GPT-Rosalind agents to analyze a mouse brain CRISPR atlas in 32 minutes.
- 02The researchers estimate comparable human analysis, one perturbation at a time, would take months to a year.
- 03At the largest workload, distributed agents cited more than 700 references versus 11 for a single agent.
- 04Blinded human experts preferred the distributed-agent reports over single-agent reports 75% of the time.
- 05The work is a bioRxiv preprint, and the resulting reasoning atlas is public on GitHub.
OpenAI calls the approach "distributed biological reasoning." Instead of handing one model the entire dataset, the team split the work so that each agent investigated a single gene disruption. Each one analyzed the cellular response, searched the literature and proposed an explanation. The individual reports were then combined into the atlas, which links genes, cellular effects, publications and possible mechanisms.
The split made a measurable difference. With the model and resources held constant, a single agent given the full dataset searched less literature as the workload grew. At the largest workload, the single agent cited 11 references, while the distributed agents cited more than 700. In blinded comparisons, human experts preferred the distributed-agent reports 75% of the time.
The agents also reproduced findings from earlier human analyses and surfaced additional leads, and their judgments of novelty and scientific value lined up with those of blinded human experts. One lead involves Tsc1 and Tsc2, two genes that both regulate mTOR. Disrupting either one produced similar effects in several neuron groups, but in the cerebellum, disrupting Tsc2 changed the activity of dozens of genes involved in neuronal communication while disrupting Tsc1 caused little change. That difference may be relevant to tuberous sclerosis, where cerebellar abnormalities are strongly linked to TSC2 mutations.
A second lead centers on Dctn1, a gene whose mutations cause Perry syndrome. Disrupting it increased expression of cholesterol-making genes, and the agents connected that to research suggesting cholesterol transport problems can push cells to make more cholesterol. The proposed follow-up question is whether DCTN1 disruption changes how cholesterol is distributed inside neurons and contributes to neurodegeneration.
“Sometimes the clues are already there, scattered across different fields—we just haven't connected them.”— Tomasz Nowakowski, Associate professor of neurological surgery, UCSF
OpenAI is explicit about the limits. The Tsc1 and Tsc2 difference and the cholesterol link are leads that need experimental testing, not established findings. The stated value of the agents is in prioritizing which experiments to run, while scientists remain responsible for designing those experiments and interpreting the results.
Sarah Teichmann, who chairs stem cell medicine at the University of Cambridge and co-founded the Human Cell Atlas, said AI could help more researchers explore resources like this one and turn unexpected observations into new questions to test. Tomasz Nowakowski of UCSF said the possible link between neuronal function and cholesterol metabolism shows how scattered pieces, once brought together, can give researchers a new way to think about disease biology.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
