Scripps and OpenAI run 1,946 parallel agents to interpret a whole-brain Perturb-seq atlas
jiqizhixin · x · 2026-11-02
A preprint from Scripps Research, OpenAI, and PerturbAI demonstrates distributed biological reasoning on a whole-brain in vivo Perturb-seq atlas with 1,946 perturbation profiles.
- Architecture: one agent per perturbation, each receiving cell-type-resolved expression changes, experimental context, analysis tools, and literature access — with the source paper's conclusions withheld.
- Results: parallel analysis finishes in 32 minutes; in blinded expert comparison (3 assessors), distributed interpretations won 45 of 60 judgments vs. 10 for single-agent and 5 ties.
- Scientific boundary: authors stress runtime covers computational analysis only; the system aids interpretation without turning every proposed explanation into an established finding.
- Validation: of 71 findings, 35 matched the source study (Shi et al.), 13 refined, 9 partial, 14 uncovered; agents also surfaced insights beyond the original manuscript (e.g., Psmb4's three-part phenotype).
The figure shows representative report structure (Psmb4 coupling proteostasis and autophagy to graded withdrawal of neuronal modules) plus validation charts.
More from AGI Musings
- AI bird sticker project swarmed by artists on Xiaohongshu over 'poisoning' claims — AlchainHust · 2026-11-02
- IBM's STAIR retriever cuts invalid references 65× and hits 82.6% accuracy by using document structure — solyarisoftware · 2026-11-02
- Why Anthropic's 6-to-1 agent-hour ratio isn't a staffing forecast — krishnan · 2026-11-02
- Managing agents is babysitting: every ping arrives equally loud — krishnan · 2026-11-02
- Anthropic report: agent hours already outnumber employee hours 6-to-1 — krishnan · 2026-11-02
- The moving goalposts of Reddit anti-AI sentiment, from poetry to cures — GreatConsideration72 · 2026-11-02