Summary
OpenAI says it worked with PerturbAI and Scripps Research to use thousands of GPT-Rosalind agents to analyse a whole-brain mouse CRISPR atlas in 32 minutes. The agents linked gene-disruption effects with scientific literature to produce leads for follow-up research.
OpenAI, PerturbAI and Scripps Research say they used thousands of AI agents powered by GPT-Rosalind to analyse existing experimental data from millions of mouse brain cells in 32 minutes. The agents examined a whole-brain CRISPR atlas, connected gene-disruption effects with scientific literature and assembled a set of biological leads for researchers to investigate.
The work applies AI to a challenge created by large experiments: researchers can measure the effects of disrupting thousands of genes across millions of cells, but interpreting those results can require connecting them to scientific findings spread across many fields. The analysis did not generate new experimental data; it searched for patterns and possible explanations within an existing dataset.
How the distributed analysis worked
CRISPR can be used to disrupt a gene and measure how cells respond. In the atlas analysed here, the results cover different cell types across the mouse brain. The team divided the analysis among thousands of agents. Each examined a gene disruption, assessed cellular responses, searched scientific literature and proposed explanations. Their reports were combined into a “reasoning atlas” linking genes, cell effects, relevant publications and possible mechanisms.
OpenAI says the agents reproduced findings from human analyses and also surfaced additional leads. The research team estimated that comparable human analysis and literature review, performed one gene perturbation at a time, would take months to a year.
In a comparison using the same model and resources, a single agent cited 11 references at the largest workload, while the distributed agents cited more than 700. In blinded comparisons, human experts preferred the distributed-agent reports over single-agent reports 75% of the time. These results suggest that parallel agents can cover more of a large analysis and its related literature than a single agent working through the same workload.
From patterns in mouse cells to research questions
One lead concerns Tsc1 and Tsc2, genes that work together to regulate mTOR, a major controller of cell growth and metabolism. Disrupting either gene produced similar effects in several neuron groups. In the cerebellum, however, Tsc2 disruption changed the activity of dozens of genes involved in neuronal communication, while Tsc1 disruption produced little change. The team identified this difference as a specific set of cellular changes to investigate in relation to tuberous sclerosis, where cerebellar abnormalities are linked to TSC2 mutations.
Another lead involves DCTN1, a gene whose mutations cause Perry syndrome, a neurodegenerative disease with Parkinson-like symptoms. In the atlas, disrupting Dctn1 increased the expression of genes involved in making cholesterol. The agents connected this observation to research on cholesterol transport and proposed follow-up questions: whether DCTN1 disruption changes cholesterol distribution within neurons, and whether that could contribute to neurodegeneration.
These are computationally generated leads from mouse-cell data, not established disease mechanisms or treatment effects. The proposed links need experimental follow-up. Their immediate value is in helping researchers decide which patterns and hypotheses may be worth testing next.