Using thousands of AI agents to uncover new leads in biological data

OpenAI, PerturbAI and Scripps Research ran thousands of agents built on GPT-Rosalind over an existing whole-brain CRISPR atlas covering millions of mouse brain cells. Each agent took one gene disruption, analyzed how cells responded, searched the literature and proposed an explanation. The reports were merged into a public reasoning atlas. The full run took 32 minutes; the team estimates comparable human work would take months to a year.

The outputs are concrete. The agents flagged a cerebellum-specific difference between Tsc1 and Tsc2, two genes that normally act together and are tied to tuberous sclerosis, and a possible link between DCTN1, a gene behind the Parkinson-like Perry syndrome, and cholesterol production in neurons.

This matters because modern biology generates far more data than anyone reads closely. Systematically mining existing experiments for overlooked leads could stretch every research dollar further.

We are genuinely encouraged by this, and the reason is the method more than the speed. Splitting the work so that each agent goes deep on one perturbation beat a single agent handed everything. With the model and resources held fixed, the lone agent cited just 11 references at the largest workload, versus more than 700 across the distributed agents, and blinded human experts preferred the distributed reports 75 percent of the time. That is a useful, transferable lesson: agents do better research when the problem is cut into pieces they can actually read through.

The validation choices are also sensible. The agents reproduced findings from earlier human analyses, and their judgments about novelty lined up with blinded experts. The reasoning atlas is public on GitHub, so other labs can check the work rather than take it on faith.

Now the limits, stated clearly. Every lead here is a hypothesis. Nothing in the announcement says a single one has been confirmed at the bench. The data comes from mice, and the jump to human disease runs through a long, expensive chain of experiments. And literature-connected reasoning has a known failure mode: it can stitch together a plausible story out of weak or mismatched citations. Expert agreement is reassuring, but it is not replication.

There is also a scaling risk worth naming. If one run can generate hundreds of credible-sounding mechanisms, the bottleneck shifts to wet labs and reviewers. A flood of AI-generated leads could crowd out slower, human-driven questions, or tempt groups to publish hypotheses as if they were results. The researchers themselves frame scientists as remaining at the center of discovery. We hope that framing holds when the volume rises.

Our read: this is what useful AI in science looks like at its best right now, a tireless reader that points to where to look next. The proof will be whether the cholesterol and Tsc2 leads survive contact with the lab.

The underlying dataset is a whole-brain CRISPR atlas showing how disrupting individual genes changes different cell types across the mouse brain. The work is described in a bioRxiv preprint, which has not yet gone through peer review.