Analyze a mouse brain CRISPR atlas with thousands of parallel research agents

OpenAI worked with PerturbAI and Scripps Research on an existing whole-brain CRISPR atlas: data on how knocking out thousands of genes changes cells across millions of mouse brain cells. Instead of handing one agent the whole set, they gave each gene disruption to its own agent built on the GPT-Rosalind model, thousands running in parallel. Each agent looked at the cellular response, searched the scientific literature and wrote up a possible explanation. The reports were then joined into one public reasoning atlas that links each gene to its cell effects, the papers behind them and the proposed mechanism.

The run took 32 minutes; the team guessed the same analysis done by people, one disruption at a time, would take months to a year. The agents reproduced findings from earlier human analysis and added new leads, such as different cerebellum effects for Tsc1 and Tsc2 and a possible link between DCTN1, cholesterol making and neurodegeneration. In a test with the same model and budget, one agent working the whole load cited 11 papers against more than 700 for the split agents, and blinded experts preferred the split reports 75% of the time. None of the leads has been tested in the lab yet.

This is OpenAI's own write-up, and the HN thread was sceptical: commenters said lead generation was never the bottleneck in drug discovery and that the leads still need costly experiments. The harness was built for the project, not a product you can install. Splitting the work into one agent per unit is the part that made the difference in their comparison.