OpenAI says agents analysed mouse brain atlas in 32 minutes

OpenAI, PerturbAI and Scripps Research said thousands of GPT-Rosalind agents processed a mouse-brain CRISPR atlas in 32 minutes. They said human researchers would have needed months to a year, and the work appeared in a bioRxiv preprint that had not yet been peer reviewed.

The atlas came from Xin Jin’s lab at Scripps Research and was published on Hugging Face. It covered nearly 2,000 genes that were switched off one by one in living mice, and the effects were measured across millions of brain cells.

The team said it gave each gene disruption to its own agent, which examined cell responses, searched the literature and wrote a report. Later agents combined those reports into a searchable atlas of 3,397 findings, and the authors said the model still recovered the paper’s main conclusions.

The paper highlighted possible leads involving Tsc1, Tsc2 and Dctn1, but said they remained hypotheses and had not yet been tested experimentally. The team also compared the distributed setup with a single agent using the same model and resources, and said 30 doctoral-level experts preferred the distributed interpretation in 45 of 60 judgments.