OpenAI, PerturbAI and Scripps Research have used close to 2,000 AI agents to analyse a large mouse brain dataset in 32 minutes.
The team estimates that researchers working by hand would need months to a year for the same analysis, which OpenAI published on Wednesday.
The agents ran on OpenAI’s GPT-Rosalind, a model built for life sciences research. The results are set out in a bioRxiv preprint, which has not yet been peer-reviewed.
The data comes from a whole-brain CRISPR atlas built by PerturbAI. CRISPR is a gene-editing tool, and here it was used to switch off single genes in mice, showing how each change affects different types of brain cells.
Each agent took one of these gene changes. It studied how the cells responded, searched the scientific literature, and suggested possible explanations.
The team then merged all the reports into a single reasoning atlas that links genes, cell effects, papers, and likely mechanisms.
According to OpenAI, the agents reproduced findings from earlier human analyses and also found new leads. In blinded tests, experts preferred the reports from this split approach over those from a single agent 75% of the time.
The split approach also read more widely. At the largest workload, the many agents together cited more than 700 references, compared with 11 for a single agent.
One lead involves two genes, Tsc1 and Tsc2, which both help control cell growth. Switching off Tsc2 changed dozens of genes linked to communication between nerve cells in the cerebellum, while switching off Tsc1 changed little. That gap may matter for tuberous sclerosis, a rare genetic disorder.
In another case, switching off Dctn1, a gene linked to the rare brain disease Perry syndrome, raised the activity of genes that make cholesterol. The agents suggested lab experiments to test whether this plays a part in nerve cell damage.
“We build cell atlases so that other scientists can make discoveries we haven’t imagined yet,” said Sarah Teichmann, chair in stem cell medicine at the University of Cambridge.
The reasoning atlas is now available on GitHub, and the underlying dataset is on Hugging Face.