OpenAI, PerturbAI and Scripps Research split a whole-brain CRISPR atlas among 1,946 agents. Each agent analyzed gene disruptions and proposed research leads. The agents completed the work in 32 minutes. A human team would need months to a year for the same task. The agents retained 7,586 references, 711 times more than a single agent. Thirty biologists reviewed the agents’ reports in a blinded test. The agents matched or refined 48 of 55 original findings without seeing the study. The team envisions a cycle of agent interpretation and predictive modeling. Scientists would still make the final decisions in that loop.
What changed
Single-agent runs averaged about 11 references at largest workload.
What it unlocks
Running thousands of agents in parallel to analyze biological data and generate hypotheses.
- 1,946 GPT-Rosalind agents
- 32 minutes to propose leads
- 7,586 references retained
- 711 times more references than single-agent
Sources