saved
Using thousands of AI agents to uncover new leads in biological data
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
OpenAI, PerturbAI and Scripps report 1,946 GPT-Rosalind agents that read a whole-brain mouse CRISPR atlas in 32 minutes. Agents recovered 94.5% of directly testable source findings, logged 39,185 paper reads, and at full workload kept about 711× more citations than one agent (vs 11); blinded experts preferred distributed reports 75% of the time.
ideas
- One agent per gene disruption, then a shared atlas. Each agent inspected cell-type responses, searched literature and proposed mechanisms; reports were merged into a reasoning atlas of genes, effects, papers and hypotheses for scientists to browse.
- Distributed agents keep literature depth as the screen grows. Holding model and resources fixed, a single agent’s citations collapsed to 11 at the largest workload while parallel agents retained hundreds of times more references and broader coverage of subtle responses.
- Recovered known science and ranked new leads. Without the source manuscript, agents recovered 52 DEG-derived findings (94.5%), including Tsc1 versus Tsc2 differences in cerebellar neurons and a DCTN1–cholesterol lead tied to Perry syndrome / Parkinson-like disease.
- Human judgment still selects what to test. Novelty and value rankings aligned with blinded experts; the stated goal is more testable follow-ups, not autonomous discovery, with human scientists choosing experiments and interpreting results.
quotes
“We build cell atlases so that other scientists can make discoveries we haven’t imagined yet.”
“At the largest workload, distributed agents cited more than 700 references, compared with 11 for a single agent.”
“In blinded comparisons, human experts preferred distributed-agent reports over single-agent reports 75% of the time.”