OpenAI, PerturbAI, and Scripps Research used thousands of AI agents powered by GPT-Rosalind to analyze a whole-brain CRISPR atlas, a dataset describing how gene disruptions affect different cell types across millions of mouse brain cells. The approach, called distributed biological reasoning, split the work so that each agent investigated one gene disruption, analyzed cellular responses, searched the scientific literature, and proposed an explanation. The individual reports were assembled into a publicly available reasoning atlas connecting genes, cellular effects, publications, and possible biological mechanisms. The full analysis ran in 32 minutes; the team estimates a comparable human analysis and literature review, done one perturbation at a time, would take months to a year.

Modern experiments generate far more data than researchers can thoroughly investigate, and interpreting those data for human disease requires linking complex experimental results with decades of scattered knowledge. Sarah Teichmann, chair in stem cell medicine at Cambridge and founding co-chair of the Human Cell Atlas, says cell atlases are built so others can make discoveries not yet imagined; AI could help more researchers explore these resources and turn unexpected observations into new questions to test.

The distributed results held up in human evaluation. The agents reproduced findings from earlier human analyses while uncovering additional leads, and their novelty and scientific-value assessments aligned with judgments from blinded human experts. In blinded comparisons, experts preferred distributed-agent reports over single-agent reports 75% of the time. In a scaling test with the model and resources held fixed, a single agent searched less of the literature and cited only 11 references at the largest workload, whereas the distributed agents cited more than 700. The authors argue this allows scientists to examine subtle or unexpected responses rather than focusing only on the strongest signals.

Several leads are particularly relevant to disease. Tsc1 and Tsc2 work together to control mTOR, a major regulator of cell growth, metabolism, and aging. Disrupting either gene had similar effects in several groups of neurons, but in the cerebellum Tsc2 disruption altered the activity of dozens of genes involved in neuronal communication, while Tsc1 disruption produced little change. That difference may matter for tuberous sclerosis, where cerebellar abnormalities are strongly linked to TSC2 mutations, and it gives researchers specific questions about which changes existing mTOR-reducing medicines might correct and which may need new approaches.

The agents also found a lead in DCTN1, a gene mutated in Perry syndrome, a neurodegenerative disease with Parkinson-like symptoms. DCTN1 carries cargo inside neurons, and disrupting Dctn1 in the atlas increased expression of genes involved in cholesterol synthesis. The agents connected that observation to research suggesting cholesterol transport problems can make cells produce more cholesterol, raising the testable hypothesis that DCTN1 disruption changes cholesterol distribution in neurons and contributes to neurodegeneration. Human experts considered this hypothesis novel, consequential, and worth following up.

The reasoning atlas is now available so researchers can compare related genes, investigate unexpected responses, and identify promising leads. Tomasz Nowakowski, associate professor of neurological surgery at UCSF, says clues are often already there but scattered across fields, and the possible link between neuronal function and cholesterol metabolism shows how bringing pieces together can yield a new way to think about disease biology. The authors frame the opportunity as going beyond faster analysis: combining AI with experimental biology, human expertise, and scientific literature can help turn more of an experiment’s information into testable ideas and, eventually, better medicines.