Surfacing new leads in biological data with thousands of AI agents | OpenAI
October 21, 2026
How OpenAI, PerturbAI, and Scripps Research used distributed AI agents to help scientists identify promising research leads from data on millions of brain cells
00:10 01:23
Learn how OpenAI, PerturbAI, and Scripps Research deployed thousands of AI agents in parallel to identify potential biological connections from complex data across millions of brain cells.
Video credit: PerturbAI.
What if scientists could uncover new leads for Alzheimer’s and Parkinson’s in minutes rather than months? In a collaboration between OpenAI, PerturbAI, and Scripps Research, thousands of AI agents powered by GPT-Rosalind analyzed and interpreted existing experimental data from millions of mouse brain cells in just 32 minutes. By connecting the effects of gene disruptions with findings in scientific literature, these agents revealed unexpected biological associations, helping scientists understand disease and pointing toward directions for new therapies. This work illustrates how AI can help researchers dig deeper into what large-scale biological experiments have to teach us.
Modern experiments generate far more data than scientists can study in full detail. Researchers can now perturb thousands of genes and measure their effects across millions of cells. However, understanding what these results mean for human disease requires bridging complex experimental findings with decades of fragmented scientific literature. This is where AI can excel: helping scientists discover promising leads that might otherwise be overlooked.
“We build cell atlases so other scientists can make discoveries we haven’t yet imagined. AI can help more researchers explore these resources and turn unexpected observations into new questions to test.”
— Dr. Sarah Teichmann, Head of Stem Cell Medicine at the University of Cambridge, Co-founder and Co-chair of the Human Cell Atlas
The team distributed the analysis across thousands of AI agents powered by GPT-Rosalind, an approach we call distributed biological reasoning. These agents reviewed a whole-brain CRISPR atlas, a dataset showing how gene disruptions affect different cell types across the mouse brain. Each agent was tasked with investigating a single gene disruption, analyzing cellular responses, searching the scientific literature, and proposing potential explanations for the results. Their reports were then synthesized into a reasoning atlas that connects genes, cellular effects, relevant papers, and underlying biological mechanisms.
The team estimated that conducting comparable analysis and literature review by human researchers, perturbation by perturbation, would take months to a year. The agents not only reproduced the results of human analyses but also uncovered additional leads. Their evaluations of novelty and scientific merit aligned closely with those of blinded human experts. These findings help prioritize follow-up experiments to test the proposed biological mechanisms.
Several of these insights pointed to new leads for understanding brain diseases. The agents uncovered unexpected differences in how two disease-linked genes affect brain cells, as well as a potential link between cholesterol regulation and neurodegeneration. These findings help scientists investigate disease mechanisms and identify where future therapies might intervene.
“Sometimes the clues are already out there, scattered across different fields—we just haven’t connected them yet. The potential link between neuronal function and cholesterol metabolism shows how bringing these pieces together can give us fresh ways to think about disease biology.”
— Dr. Tomasz Nowakowski, Associate Professor of Neurological Surgery, University of California, San Francisco (UCSF)
Researchers can begin exploring these findings today. The reasoning atlas is now publicly available, enabling scientists to compare related genes, investigate unexpected responses, and pinpoint promising leads to pursue. Breakthroughs often happen when researchers from different fields combine their expertise, linking familiar concepts in unexpected ways. AI can significantly amplify these connections by bridging new data with knowledge scattered across disciplines. The opportunity is not just about analyzing data faster. By uniting AI with experimental biology, human expertise, and the scientific literature, we can help scientists turn more data into testable hypotheses.
Massive experiments, unexpected leads
Some of the most intriguing leads in an experiment lie outside what scientists originally set out to study. In a dataset spanning thousands of perturbations, these leads can easily be missed.
The agents identified an unexpected divergence between Tsc1 and Tsc2, two genes that work together to regulate mTOR—a master regulator of cell growth, metabolism, and aging. Across several neuronal populations, disrupting either gene produced similar effects. But in the cerebellum, which coordinates movement, Tsc2 disruption altered the activity of dozens of genes involved in signaling between neurons, whereas Tsc1 disruption caused almost no change.
This difference may be directly relevant to human disease. In patients with tuberous sclerosis, cerebellar abnormalities are strongly associated with TSC2 mutations. While drugs that reduce mTOR activity can treat certain brain tumors and seizures, the learning and behavioral challenges associated with the condition remain difficult to address. The agents pinpointed a specific group of neurons and cellular changes worth investigating: which changes can be corrected with existing drugs, and which might point to new therapeutic approaches? This provides researchers with a more focused set of questions to test in the lab.
Dividing the work to deepen reasoning
Could a single agent handle the entire dataset as effectively as thousands of agents working in parallel? We tested this while keeping the model and compute resources fixed. As the workload increased, the single agent retrieved less literature and cited…