OpenAI and Others Use Thousands of AI Agents to Uncover New Clues in Biological Data

OpenAI, PerturbAI, and the Scripps Research Institute announced a collaborative study in which the team used thousands of AI agents powered by GPT-Rosalind to analyze whole-brain CRISPR atlases in parallel, characterizing changes in millions of mouse brain cells following gene perturbation and retrieving relevant literature. OpenAI stated that the analysis took 32 minutes; researchers estimated that performing similar manual analysis and literature review on an item-by-item basis could take several months to a year. The reports from individual agents were aggregated into a “reasoning map” linking genes, cellular responses, publications, and potential biological mechanisms. Clues generated by the study include differing effects of Tsc1 and Tsc2 perturbation in cerebellar neurons, and increased expression of cholesterol synthesis-related genes following Dctn1 perturbation. In a set of blind comparative evaluations, human experts preferred reports generated by the parallel agents over those from a single agent for 75% of the evaluable responses. The team has made the atlas available for exploration by other researchers. The value of these results lies in helping to prioritize next-step experimental questions; the underlying mechanisms still require experimental validation, and data from mouse cells do not directly demonstrate therapeutic benefits for patients with Alzheimer’s or Parkinson’s disease.