OpenAI, PerturbAI, and Scripps Research jointly demonstrated an approach in which thousands of GPT‑Rosalind–powered AI agents ran in parallel to analyze a whole‑brain CRISPR atlas from mice. According to the team, the agents processed and interpreted experimental data from millions of cells in 32 minutes, linking perturbation effects to scientific literature and revealing unexpected biological connections that suggest new research directions.
Why this approach matters
Modern biological experiments generate datasets far larger than researchers can manually inspect: thousands of gene perturbations measured across millions of cells create complex patterns that must be interpreted in the context of decades of scattered literature. AI can help surface promising leads that human analysts might miss, by connecting experimental signals with relevant prior findings across fields.
How the agents worked
The research team distributed the workload among thousands of AI agents, each built on the GPT‑Rosalind model. Each agent focused on a single gene perturbation: analyzing cellular responses, searching the scientific literature, and proposing explanations for observed effects. Individual reports were combined into a reasoning atlas that links genes, cellular effects, pertinent publications, and plausible biological mechanisms.
The team estimates that comparable human analysis done one perturbation at a time would take months to a year.
Key findings
The agents reproduced results known from prior human analyses and also generated additional leads. Two notable examples cited by the researchers:
-
Tsc1 vs. Tsc2: Tsc1 and Tsc2 jointly regulate mTOR, a central controller of cell growth, metabolism, and aging. While perturbing either gene produced similar effects across many neuron groups, in the cerebellum Tsc2 disruption changed the activity of dozens of genes linked to neuronal communication whereas Tsc1 disruption produced little change. This distinction is relevant because cerebellar abnormalities are strongly associated with TSC2 mutations in tuberous sclerosis.
-
DCTN1 and cholesterol: Mutations in DCTN1 cause Perry syndrome, a neurodegenerative disorder with Parkinson‑like symptoms. In the whole‑brain CRISPR atlas, disrupting Dctn1 increased expression of genes involved in cholesterol biosynthesis. The agents connected this to prior research suggesting that defects in cholesterol transport can trigger compensatory cholesterol production. That raises a testable hypothesis: does DCTN1 disruption alter intracellular cholesterol distribution in neurons in a way that contributes to neurodegeneration, and could restoring normal cholesterol distribution protect neurons? Human experts judged this hypothesis novel and worth following.
Distributed agents vs. a single agent
The team compared the distributed‑agent approach with a single agent running under the same model and resources. As workload increased, the single agent searched far less literature — citing only 11 references at the largest workload — while the distributed agents collectively cited over 700 references. In blinded comparisons, human experts preferred the distributed‑agent reports 75% of the time.
The researchers argue that parallel agents enable a broader survey of experimental signals, including subtle or unexpected responses, and link them more thoroughly to existing literature.
Next steps and implications
The combined reasoning atlas is available for researchers to explore, allowing comparisons across related genes, investigation of surprising responses, and identification of promising leads to pursue experimentally. The authors emphasize that AI is a tool to expand the set of hypotheses scientists can test — not a replacement for human expertise and laboratory validation.
This work illustrates that accelerating data analysis is only part of the opportunity: greater value comes from connecting large experimental datasets with dispersed scientific knowledge to generate testable ideas that can guide future experiments and, ultimately, therapeutic development.
