Close to 2,000 AI agents analysed a whole-brain mouse CRISPR atlas in 32 minutes, work the team estimates would take researchers months to a year by hand. The project was carried out by OpenAI with PerturbAI and Scripps Research and published on Wednesday with a bioRxiv preprint. For companies running research and analytics operations, this matters as a test of parallel agents on a real scientific dataset.
How the 32-minute parallel analysis worked
The dataset came from a whole-brain CRISPR atlas built by PerturbAI, in which single genes were switched off in mice to show effects across brain cell types. Each agent took one gene perturbation, studied the cell response, searched scientific literature, and proposed possible explanations. The agents ran on GPT-Rosalind, a model built for life sciences research. The preprint describing the results has not yet been peer-reviewed.
The team then merged the individual agent reports into a single reasoning atlas linking genes, cell effects, papers, and likely mechanisms. That structure turns thousands of separate outputs into one connected resource rather than a pile of files. According to OpenAI, the agents reproduced findings from earlier human analyses and also produced new leads. The reasoning atlas is now available on GitHub, while the underlying dataset sits on Hugging Face.
Blinded tests favored the split approach over a single agent in 75% of expert comparisons. The difference in literature coverage was wide: at the largest workload the many agents together cited more than 700 references, against 11 for a single agent. Sarah Teichmann, chair in stem cell medicine at the University of Cambridge, said cell atlases are built so other scientists can make discoveries the creators have not imagined. The comparison suggests scale changes both speed and breadth of evidence gathering.
What parallel agents mean for research operations
Two biological leads show the practical output. Switching off Tsc2 altered dozens of genes tied to nerve-cell communication in the cerebellum, while switching off Tsc1 changed little, a gap potentially relevant to tuberous sclerosis. Switching off Dctn1, linked to Perry syndrome, raised activity of cholesterol-making genes, with proposed lab tests on nerve-cell damage. For R and D teams, the gain is faster triage of hypotheses before costly bench work.
Small teams gain access to literature review and hypothesis drafting that once required large specialist groups, while large organizations can run many perturbations in parallel across portfolios. The constraint is validation: preprint status, need for lab follow-up, and dependence on GPT-Rosalind plus the PerturbAI atlas limit direct transfer. Buyers should ask what data formats are supported, how references are checked, and how the merged atlas is kept current.
The marker to watch is whether independent labs reproduce the Tsc1 and Tsc2 and Dctn1 leads through the proposed experiments and cite the GitHub reasoning atlas in follow-up work. Peer review of the bioRxiv preprint will be a second signal. If both advance, parallel agents will look like a standard layer in discovery pipelines.
