Anthropic says Claude discovered a novel CRISPR-like enzyme system

Anthropic's life sciences research group says Claude autonomously discovered a previously uncharacterized enzyme system while searching a large DNA sequence database, identifying a repeating pattern of non-coding DNA associated with an unusual reverse transcriptase found in bacteriophages. The system, which Anthropic calls array-associated reverse transcriptases (ART), resembles CRISPR in structure even though nothing pointed Claude toward that comparison in advance.

What's new

The discovery came out of a large-scale automated search rather than a single directed query. "After 21 hours spent searching this data by roughly 950 agents using 210 million tokens, one of the agents spotted something remarkable: a repeating pattern of DNA sequences," Anthropic wrote. That pattern turned out to be an array of non-coding sequences sitting next to a reverse transcriptase gene, plus an accessory protein whose function is still unknown.

Reverse transcriptases are enzymes that copy RNA into DNA; they're well studied in isolation, but Anthropic says this particular arrangement, an array flanking the RT gene, had gone unnoticed in the existing literature despite being present in the database Claude searched. "Claude appears to be the first to notice the system's defining features, an associated array of non-coding DNA sequences and an additional accessory protein of unknown function," the company said.

Context

This follows Anthropic's earlier work showing Claude could design working protein binders against 14 of 15 targets in a separate research push, part of a broader pattern of the company running large agent swarms against biological datasets rather than treating Claude purely as a coding or chat assistant. The scale here is notable on its own terms: nearly a thousand agents running in parallel for the better part of a day, burning through 210 million tokens, to surface one pattern that a human researcher scanning the same database manually would have had to get lucky to spot.

CRISPR itself was discovered by researchers noticing an odd repeating DNA pattern in bacteria decades before anyone understood what it did or that it could be turned into a gene-editing tool. Anthropic's framing leans on that parallel deliberately: repeating array structures next to enzymes with unclear function are exactly the kind of thing that historically turned out to matter far more than it looked like at first glance.

Why it matters

The scientific value of ART itself is still unproven. Anthropic has identified the pattern and the associated genes, not demonstrated a function or a use case, and the accessory protein's role remains unknown. What this establishes more concretely is a working method: point a large number of Claude agents at a big, messy biological dataset and let them search for anomalies without a specific hypothesis to test.

That's a different mode of use than most current AI-in-science work, which tends to apply models to a narrow, pre-defined question (predict this protein's structure, screen these compounds). An open-ended anomaly hunt across a sequence database, run by agents rather than a human curator, is closer to how genuinely new biology tends to get found. If the pattern holds up under further study and turns into something with a defined function, it would be a concrete instance of an AI system generating a novel research lead rather than just accelerating one a human already had in mind.