OpenAI's GPT-Rosalind Agents Mine a CRISPR Brain Atlas

A single AI agent, given a whole experiment to interpret, cited 11 scientific papers. Split the same job across hundreds of agents, each handed one gene, and the citations ran past 700. That comparison sits at the centre of work OpenAI published on Wednesday with PerturbAI and Scripps Research. The three say that 1,946 agents running OpenAI’s GPT-Rosalind life-sciences model worked through a CRISPR atlas of the mouse brain “in just 32 minutes”.

The real claim is about who interprets biology’s ever-larger experiments. Labs can now switch off thousands of genes and read the effects across millions of cells, but making sense of the results still depends on people reading the literature one gene at a time. The team estimated that comparable human analysis “would take months to a year”. What the agents produced are leads, not discoveries: hypotheses that still have to be tested at the bench.

One agent per gene, nearly 2,000 at once

The data came from PerturbAI, a San Francisco biotech start-up. In March it released a brain-wide CRISPR atlas covering 8 million cells, which it says has been downloaded more than 40,000 times, according to PerturbAI’s announcement. Each agent took one gene disruption, studied how cells responded, searched the published research and proposed an explanation. Their reports were then stitched into what OpenAI calls a “reasoning atlas”, which researchers can now browse. OpenAI calls the approach “distributed biological reasoning”.

The details are in a preprint on bioRxiv, posted on 16 October and not yet peer reviewed. Its authors, from Scripps, OpenAI and PerturbAI, say they withheld the original study and its conclusions from the agents. Of 55 findings in that study drawn directly from the data, the agents at least partly recovered 52. Across all the runs they kept 16,445 citations. The preprint describes the “months to a year” comparison as an “illustrative” estimate rather than a measured one.

Joy Jiao, OpenAI’s head of life sciences, framed it as a matter of scale. “Running nearly 2,000 agents in parallel shows how we can scale biological interpretation alongside high-throughput experimentation,” she said in the release.

The leads point at mTOR genes and cholesterol

Two examples stand out in OpenAI’s account. Tsc1 and Tsc2 work together to control mTOR, a key regulator of cell growth. Disrupting either gene had similar effects in several groups of neurons. In the cerebellum, though, Tsc2 disruption changed dozens of genes linked to communication between neurons, while Tsc1 disruption barely registered. OpenAI notes that cerebellum abnormalities in people with tuberous sclerosis are strongly linked to TSC2 mutations.

The second lead concerns DCTN1, a gene whose mutations cause Perry syndrome, a neurodegenerative disease with Parkinson-like symptoms. Disrupting it raised the activity of genes that make cholesterol. The agents tied that to research suggesting faulty cholesterol transport can push cells to make more. That raises a question for the lab, not an answer: does losing DCTN1 change where cholesterol sits inside neurons, and does that contribute to neurodegeneration?

Sarah Teichmann of the University of Cambridge, a founding co-chair of the Human Cell Atlas, said in OpenAI’s post that AI “could help more researchers explore these resources and turn unexpected observations into new questions to test”.

Blinded experts preferred the split approach

The team also asked biology experts to rate reports without knowing which system wrote them. They preferred the distributed agents’ interpretations over a single agent’s in 45 of 60 judgments, or 75%, and their views of novelty and follow-up priority broadly matched the agents’ own rankings. The preprint’s authors say that tests on independent datasets and new experiments are needed to show how well the method carries over, and that lab experiments remain the test of any proposed mechanism.

That stops short of efforts such as Google Research’s wet-lab agents, which close the loop with robots. It also fits OpenAI’s wider push into biology, from the Rosalind Academic Plan for US university labs to buying Orchestra Bio for its Rosalind suite.

PerturbAI says the year-long collaboration will now move into joint drug-target discovery using its proprietary data. Whether the DCTN1 and Tsc2 leads survive their first experiments will say more about the method than the 32 minutes did.