OpenAI and PerturbAI's Agent Fleet Analyzes Brain Atlas in 32 Minutes

A distributed fleet of nearly 2,000 AI reasoning agents completed a full analysis of PerturbAI's whole-brain perturbation atlas in 32 minutes, recovering the source manuscript's conclusions without ever seeing the paper, according to a bioRxiv preprint published this week by PerturbAI, OpenAI and their collaborators.

The preprint, titled "Distributed AI agents enable comprehensive discovery of biological insights at scale," caps a year-long collaboration between the two companies and academic scientists at Scripps Research Institute, the University of California, San Diego, and the Howard Hughes Medical Institute. For R&D managers watching whether multi-agent architectures can genuinely compress analysis timelines, the reported numbers deserve scrutiny — and the authors provide enough detail for that scrutiny.

How did the system perform on real data?

The team applied its framework to data from PerturbAI's whole-brain atlas, published earlier this year. Thousands of agents worked in parallel to analyze cellular responses, search the scientific literature, and connect experimental results with existing biological knowledge. The system then synthesized its output into a reasoning atlas linking genes, cell responses, literature details, and candidate biological mechanisms.

The headline measurement is blunt: "thousands of parallel agents thoroughly examined the dataset in minutes," the scientists wrote, and — without access to the source manuscript — the system "recovered its conclusions, predicted held-out perturbation effects, and identified perturbation relationships with shared or opposing outcomes across effect strengths."

The agents also proved scalable in literature work. Distributed agents retrieved hundreds of times more references than a single-agent mode, scaling coverage with workload. That figure comes from the authors' own comparison, and the preprint reports it without independent replication.

What did human biologists prefer?

To test interpretation quality, the team had biologists review agent-generated reports and score them for novelty, evidential support, follow-up priority, and confidence level. In blinded comparisons against reports from single AI agents, human experts preferred the multi-agent reasoning interpretations in 45 of 60 judgements — 75% of the time. Single-agent interpretations won 10 judgements, or 16.7%, and five comparisons, 8.3%, ended in ties.

A 60-judgement sample is modest, and the evaluation was designed and reported by the same organizations building the system. Even so, a three-to-one preference margin among domain experts is a data point that portfolio leads evaluating AI-assisted analysis tools cannot easily ignore.

What happens next?

The collaboration is now expanding from methodology into drug-target discovery. PerturbAI and OpenAI will apply their AI reasoning frameworks to PerturbAI's proprietary causal in vivo biological data — measurements generated directly in living systems — to identify new drug targets.

Grace Zheng, PhD, CEO and co-founder of PerturbAI, framed the commercial logic: "Biomedical experiments are generating data faster than scientists can interpret it, meaning promising new discoveries are likely to remain hidden. By combining large-scale causal data generated directly in living systems with frontier AI, our partnership with OpenAI is making it possible to investigate more biological data, faster and with greater depth, and to make target discovery more systematic."

Joy Jiao, head of life sciences at OpenAI, pointed to the throughput argument: "By combining PerturbAI's experimental data with OpenAI's reasoning capabilities, we can help scientists uncover new biological insights and identify promising leads for drug discovery." She added that "the scale, quality, and speed of PerturbAI's discovery engine make them an ideal partner, and this is just the beginning of what we can achieve together."

Jiao's most concrete claim concerns operational capacity: "Running nearly 2,000 agents in parallel shows how we can scale biological interpretation alongside high-throughput experimentation, helping researchers turn those findings into better medicines."

What should R&D managers take from it?

Three measured results stand apart from the projections: the 32-minute analysis, the 75% expert preference rate across 60 blinded judgements, and the hundredfold-plus improvement in literature retrieval under distributed workload. The drug-target discovery program, by contrast, remains a projection — a stated plan with proprietary data behind it and no reported targets yet.

Both quotes from the partnering companies carry the usual commercial incentives, and the evaluation involved no external validation. Still, the work demonstrates a concrete workflow shift: analysis capacity that scales with compute rather than headcount, applied to a published dataset with checkable conclusions. The next test will come when the same agent fleets surface a drug target that survives experimental validation.