Thousands of GPT-Rosalind Agents Mapped Millions of Mouse Brain Cells in 32 Minutes
Thousands of AI agents powered by GPT-Rosalind processed a whole-brain CRISPR atlas — millions of mouse brain cells across thousands of gene disruptions — in 32 minutes. Human researchers doing the same work one perturbation at a time would take months to a year.
The work, a collaboration between OpenAI, PerturbAI, and Scripps Research, is published as a preprint on bioRxiv. The dataset itself — the PerturbAI Whole Brain CRISPR Atlas — covers how gene disruptions affect cell types across the full mouse brain, and is hosted on Hugging Face.
Distributed Biological Reasoning
Each agent in the swarm took a single gene disruption. It analysed the cellular response, searched scientific literature, and proposed an explanation linking observed effects to known biological mechanisms. Those individual reports were aggregated into what the team calls a reasoning atlas — a structured document mapping genes, cellular effects, relevant publications, and proposed mechanisms.
The approach is called distributed biological reasoning. The practical advantage is obvious: parallelisation at a scale no human team can match. But the more interesting result is qualitative — the agents’ novelty and scientific value assessments aligned with those of blinded human experts reviewing the same findings.
What the Agents Found
Two findings stand out. First, the agents identified unexpected differences in how two disease-linked genes affect distinct brain cell populations — a result that human analysts confirmed was consistent with prior literature but had not been explicitly connected.
Second, agents found a potential link between cholesterol regulation and neurodegeneration. Cholesterol metabolism’s role in brain health is an active research area, but the specific connection the agents surfaced points to a mechanism worth testing experimentally.
Neither finding is a cure. Both are leads — specifically the kind of unexpected cross-domain connection that structured literature review misses because it requires holding simultaneous context across dozens of gene-function relationships.
What It Means for Scale
Modern genomics experiments generate data faster than science can interpret them. A CRISPR screen across millions of cells produces results that, if fully analysed, would consume research careers. The standard approach is to pick the top hits and ignore the rest.
GPT-Rosalind’s application here is a direct attempt to close that gap. The agents reproduced what human analysis would have found, while also surfacing connections the manual review would have missed — at 32 minutes instead of a year.
The next step, as the team notes, is designing experiments to test whether the proposed mechanisms hold. AI found the leads. Human scientists run the validation.