OpenAI, working with PerturbAI and Scripps Research, has demonstrated a new way to speed up biological research using thousands of AI agents working in parallel. The team used nearly 2,000 agents powered by GPT-Rosalind to analyze an 8-million-cell mouse brain atlas in 32 minutes - work that researchers estimated would take humans anywhere from several months to a year to complete using conventional methods.
The project examined PerturbAI's whole-brain CRISPR atlas, which shows how disrupting nearly 2,000 different genes affects various cell types across the mouse brain. Each AI agent investigated one gene disruption, analyzed how cells responded, searched scientific literature for relevant findings, and proposed explanations for the results.
The agents' reports were combined into what the team calls a reasoning atlas - a resource linking genes, cellular effects, related publications, and possible biological mechanisms. This atlas is now publicly available for other researchers to explore.
The work matters because modern biological experiments generate far more data than scientists can thoroughly investigate. By dividing analysis among many agents, researchers can examine subtle or unexpected responses that might otherwise be missed, and connect these findings with relevant scientific literature across different fields.
New leads for brain disease research
The agents identified two findings that may point toward new research directions. Tsc1 and Tsc2 are genes that work together to control mTOR, a major regulator of cell growth and metabolism. While disrupting either gene had similar effects in most neurons, in the cerebellum Tsc2 disruption altered dozens of genes involved in neuronal communication while Tsc1 disruption produced little change. This is relevant to tuberous sclerosis.
The agents also found a lead in DCTN1, a gene involved in moving materials inside neurons. Disrupting Dctn1 increased expression of genes that make cholesterol. The agents connected this to research suggesting that problems with cholesterol transport can cause cells to produce more cholesterol. This could matter for Perry syndrome, a neurodegenerative condition with Parkinson-like symptoms.
These findings illustrate how AI can help researchers connect unexpected experimental results with existing knowledge scattered across disciplines, turning clues into testable hypotheses. While the findings require experimental validation, they demonstrate how distributed AI reasoning can help scientists prioritize which research to pursue next.