OpenAI and startup PerturbAI assigned 1 946 AI agents one gene each to study after its inactivation in the mouse brain: a preprint reports that the agents reproduced or refined 48 of 55 assessable findings from a paper on the data withheld from them, while a single agent also proposed the cholesterol hypothesis

OpenAI and startup PerturbAI assigned 1 946 AI agents one gene each to study after its inactivation in the mouse brain: a preprint reports that the agents reproduced or refined 48 of 55 assessable findings from a paper on the data withheld from them, while a single agent also proposed the cholesterol hypothesis

On October 21, OpenAI, PerturbAI and Scripps Research announced a method that assigns each gene to a separate AI agent running GPT-Rosalind-5.5, OpenAI's model for biology. The data come from an atlas released in March, in which nearly two thousand genes were individually inactivated in mouse brains. The method is described in an October 16 preprint, a paper that has yet to undergo peer review.

Xin Jin's team at Scripps created the atlas by using a virus to deliver gene-inactivation tools into neurons, then measuring which genes were active in 7,7 million cell nuclei. In 2020, Jin tested the method on 35 autism risk genes. The atlas now covers nearly two thousand genes, and the authors write that experts have yet to interpret all the results.

Each agent receives a table showing how inactivating its assigned gene changed the activity of other genes. It selects analyses, compares its gene with others, searches the literature and writes a report with quantitative results and proposed experiments. Each gene takes about 32 minutes; the authors report that processing the entire atlas takes the same time when the agents run in parallel. Six preprint authors work at OpenAI and two at PerturbAI; Grace Zheng and Jin are PerturbAI's co-founders.

Using the same model, instructions and tools, the authors compared two approaches: a distributed approach that assigns a separate agent to each gene, and a single-agent approach in which one agent handles up to 893 genes whose inactivation produced a substantial response. For the single agent, the number of steps and tool calls stopped increasing as the workload grew, and its analysis of all 893 genes included about 11 literature references. The distributed agents provided 7 586 references for the same genes.

The agents were denied access to the atlas paper: the model's knowledge cutoff was December 2025, the paper appeared in March 2026, and a web filter blocked access to it. The authors judged that the agents reproduced or refined 48 of the paper's 55 findings that could be derived from the supplied tables.

The findings included results for Tsc1 and Tsc2, whose proteins form a complex that inhibits mTOR, a regulator of cell growth. In four groups of neurons, inactivating either gene produced similar changes in gene activity, consistent with the literature. The responses differed in cerebellar neurons that inhibit other cells: inactivating Tsc2 changed the activity of 83 genes, whereas inactivating Tsc1 changed none. The agent found no studies describing this response.

Dctn1 contributes to cargo transport within neurons, and mutations in its human counterpart cause Perry syndrome, a neurodegenerative disease. Inactivating Dctn1 increased the activity of cholesterol synthesis genes in several types of neurons. The agents proposed that disrupted cholesterol delivery prompts the cell to produce more cholesterol. In the report archive, the Dctn1 agent compared its assigned gene with others: inactivating Dync1h1, a transport partner of Dctn1, increased the activity of the same genes in the neurons that responded most strongly to Dctn1 loss, while inactivating Srebf2, a regulator of cholesterol synthesis, reduced the activity of some of those genes.

The reports underwent blinded evaluation by 30 researchers with doctoral degrees. For 20 genes, evaluators compared pairs of reports from the distributed approach and the single agent. Three researchers assessed each pair: they preferred the distributed approach in 45 of 60 judgments and the single agent in 10, with 5 ties. The Dctn1 hypothesis received a score of 4,3 out of 5 for priority for experimental testing, but the evaluation table attributes it to the single agent. The distributed agent also identified the same gene activity response. One reviewer described the response as unexpected and proposed an experiment: suppress Dctn1 in neurons and measure cholesterol.