Biology now produces data faster than people can read it. A whole-brain CRISPR atlas, which records how disrupting thousands of genes changes cells across the mouse brain, holds far more leads than any lab could chase one gene at a time. OpenAI, PerturbAI and Scripps Research tried a different approach: split the analysis across thousands of AI agents running GPT-Rosalind, OpenAI's life sciences model, and let each one take a single gene.
Each agent looked at one gene disruption, analysed how cells responded, searched the published literature and proposed an explanation. Their reports were merged into what the team calls a reasoning atlas, linking genes, cellular effects, papers and possible mechanisms, which is now available for scientists to explore. The full run over data from millions of mouse brain cells took 32 minutes; the team estimates that comparable human analysis would take months to a year. The agents reproduced findings from earlier human analyses, and their judgements of which leads were novel agreed with those of blinded human experts.
Two leads stand out. Tsc1 and Tsc2 normally work together to control mTOR, a central regulator of cell growth, but in the cerebellum disrupting Tsc2 changed dozens of genes involved in neuronal communication while disrupting Tsc1 did little, which may matter for tuberous sclerosis, where cerebellar problems are linked to TSC2 mutations. And disrupting Dctn1, a gene whose mutations cause the Parkinson-like Perry syndrome, raised the activity of cholesterol-making genes, suggesting a testable link between cholesterol handling and neurodegeneration. A comparison also favoured the split design: one agent given the whole dataset cited only 11 references at the largest workload, against more than 700 for the distributed agents.
The approach builds on OpenAI's push into biology with GPT-Rosalind, which also includes free seats for academic labs and the purchase of Orchestra Bio. For researchers sitting on large perturbation datasets, the method is the takeaway: parallel agents can read the long tail of results that people skip.
The caution is the usual one, and it matters here. The work is a bioRxiv preprint and has not been peer reviewed. Generating hypotheses quickly has never been the main bottleneck in biology; testing them in the lab is, and none of these leads has been tested yet. Agreement with human experts on novelty is not the same as being right. The reasoning atlas is a list of good questions, not a set of findings.