Almost 2,000 AI agents analyse a mouse brain atlas in 32 minutes

OpenAI, PerturbAI and Scripps Research used almost 2,000 AI agents to analyse a large set of mouse brain data. The agents did it in 32 minutes. OpenAI publishes the work on Wednesday. The team estimates that a person who works by hand uses months to a year for the same task. The results are in a preprint on bioRxiv.

How it works

The data is an atlas of mouse brain cells from PerturbAI. The team used CRISPR, a tool that edits genes, to stop one gene at a time in mice. Each agent had one gene change. It studies how the cells change, searches the papers and gives possible causes. The team then makes one atlas of all reports.

The numbers

Experts do not know which agent wrote a report. They select the reports of the many agents in 75% of the tests. The many agents gave more than 700 references. One agent gave 11. OpenAI reports that the agents found again the results of the first analyses by persons and also found new leads.

What is not known

No other expert examines the result. To stop Tsc2 changed many genes connected to contact between brain cells, but to stop Tsc1 changed a small number. This is possibly important for tuberous sclerosis, a gene disorder.

Insight: Agents that work apart gave wider results

In a Tsinghua University test, thousands of agents that speak to each other moved to opposite groups. Here, each agent had one gene and did the work without other agents, and the team then connected the reports. The procedure to connect agents is possibly the primary cause: isolated agents gave 700 references, and one agent gave 11. Test 2 procedures.