OpenAI runs thousands of GPT-Rosalind agents over a mouse whole-brain CRISPR atlas in 32 minutes: 700+ citations vs 11 for one agent, and new Tsc2 and Dctn1–cholesterol leads

Why it matters

【Key points】

OpenAI, PerturbAI and Scripps Research ran thousands of GPT-Rosalind agents in parallel over a whole-mouse-brain CRISPR atlas in 32 minutes, versus an estimated months to a year by hand.

1. One agent per gene perturbation analyzes responses, searches literature and proposes mechanisms; outputs form a public "reasoning atlas".

2. With fixed resources, a single agent cited 11 references at the largest workload versus 700+ for distributed agents; blinded experts preferred distributed reports 75% of the time.

3. Leads: Tsc2 vs Tsc1 differences in cerebellar neurons; Dctn1 disruption raising cholesterol-synthesis genes, suggesting a cholesterol–neurodegeneration hypothesis.

【Verifiable】

https://openai.com/index/scaling-ai-research-for-complex-biological-data/

【Our read】

1. Background: follows the Orchestra Bio acquisition and Rosalind Academic Plan — life sciences is a core OpenAI vertical.

2. Comparison: Anthropic's Claude binder work included wet-lab tests; this is hypothesis generation only, but gives a clean single- vs multi-agent comparison.

3. Impact: a cheap lead-screening pattern for labs with large perturbation datasets.

4. Caveats: hypotheses untested; compute/cost undisclosed; mouse data.