OpenAI, PerturbAI and Scripps Research have reported a way to divide the interpretation of a large biological dataset among thousands of AI agents. The work produced research leads from existing mouse-brain data, rather than a new treatment or a completed experiment.
OpenAI presented the collaboration on October 21. The system used agents powered by GPT-Rosalind to examine the effects of gene disruptions, connect those effects with scientific literature and propose explanations that scientists could investigate.
From experimental results to research questions
Large experiments can leave researchers with more observations than they can investigate in detail. Here, individual agents examined different gene disruptions. Their reports were assembled into a linked collection of genes, cellular responses, publications and possible mechanisms, described as a reasoning atlas.
The reported 32-minute runtime refers to analysis and synthesis of data already collected. It excludes the biological experiments and follow-up work needed to establish whether the explanations hold. OpenAI’s comparison with months to a year of human work is an estimate for sequential interpretation and literature review, rather than a timed contest with a human team.
What the evaluation measured
OpenAI says blinded experts preferred the distributed reports over single-agent reports in 75% of comparisons. That is a preference measure between reports, not a score showing that 75% of the biological claims were correct.
One research lead concerned Dctn1, a gene involved in transport within neurons. The agents connected changes following its disruption with literature about cholesterol regulation. The proposed mechanism still requires experiments; its relevance to human disease cannot be settled by interpretation of mouse-cell data alone.
The next step is experimental
For laboratories, the potential value is broader investigation of an existing dataset: more observations can be connected with prior work and turned into specific questions. That benefit depends on the quality of the evidence behind each connection and the ability to test it.
The collaboration does not establish that ordinary ChatGPT subscribers can run the same system. The announcement does not describe a general subscription entitlement, operating costs or a Hebrew-language workflow. Researchers considering the approach should distinguish access to published results from access to the full analysis system.