Anthropic's Claude Found a CRISPR-Like Enzyme System — ART in Phages

Anthropic says Claude spotted a novel enzyme system with CRISPR-like DNA repeats — array-associated reverse transcriptases (ART) — in bacteriophage genomes, with only high-level direction from scientists. Primary source: Anthropic news — Claude discovers a novel enzyme system (Sep 23, 2026). A company pre-print is linked from that page.

What shipped (claimed)

• New life sciences research group and Bay Area lab; focus on fundamental biology with Claude — DNA datasets, uncharacterized protein families, hypotheses at scale, wet-lab experiments by humans
• Early result: Claude discovered ART — CRISPR-like repeat arrays next to an unusual reverse transcriptase (RT)
• Underlying RT in a jumbo phage was previously identified; Anthropic claims Claude was first to notice the associated non-coding repeat array plus an accessory protein of unknown function
• Pattern of traits found together in only a handful of other systems that are programmable (cut/copy/paste DNA); ART function still unknown; work ongoing
• Feng Zhang (MIT/Broad, CRISPR pioneer) quoted calling AI agents contributing to discovery intriguing; RNA-repeat arrays with RTs worth further study
• Soft CTA: Anthropic invites other scientists to collaborate on proposals

Caveat: Discovery narrative, scale figures, and biological interpretation are company claims from Anthropic’s news post and linked pre-print — not independent bake-offs.

How they ran it

• Prompt: search a massive DNA database for interesting RTs; scientists set the initial prompt and did all lab work; Claude agents searched
• Workflow (claimed): survey a protein family, reproduce literature, find undescribed neighbors, write human-readable reports, eliminate most, test survivors in lab
• Stack: Claude Science + Claude Code; parallel harness coordinating many sessions
• Scope: bacteriophages only; BSL-1 / BSL-2; no human pathogens; all wet-lab work by humans

Scale and signal (claimed)

• ~21 hours, ~950 agents, ~210M tokens; one agent spotted repeating DNA next to an odd RT
• Campaign funnel: 200k+ RTs → ~3,500 candidate systems → 20 most compelling reports (weeks–months of expert work claimed for that pass)
• ART anatomy: RT + partner gene + long evenly spaced DNA repeat array (CRISPR-like); early experiments show the array expressed as distinct short RNAs

Operator takeaway

Read this as a task bundle, not a magic discovery button: search at scale → anomaly detect → structured report → human gate → wet-lab. Same pattern fits operator stacks — leave / augment / automate per step, with humans owning judgment and biosafety. Soft next step if useful: an Ada conversation to map which research or ops search tasks in your stack should stay human, which to augment with agent swarms, and which to automate once evals and gates hold.

Source: Anthropic — Claude discovers a novel enzyme system (Sep 23, 2026); company pre-print linked from that page (PDF).