OpenAI and research partners use thousands of agents to find biological leads

OpenAI said on October 21 that a collaboration with PerturbAI and Scripps Research used thousands of AI agents to examine biological data in 32 minutes. The analysis covered existing measurements from millions of mouse brain cells.

The agents ran on GPT-Rosalind and divided up a dataset showing the effects of gene disruptions. They examined cellular changes, searched research literature and assembled possible explanations. In blinded evaluations, human experts favored their reports over those from a single agent in 75% of comparisons.

The approach addresses a bottleneck in biology: experiments can produce more results than researchers have time to investigate closely. One lead connected disruption of Dctn1, a gene involved in transporting material inside neurons, with increased activity in genes responsible for cholesterol production.

That connection remains a hypothesis requiring laboratory tests. Researchers can investigate whether Dctn1 disruption changes cholesterol distribution inside neurons and whether those changes contribute to degeneration. The findings identify possible experiments rather than establish a treatment.