Anthropic's Claude Flags Uncharacterized Enzyme System in Bacteriophage DNA

Anthropic announced on September 23 that its Claude model had identified a previously uncharacterized enzyme system tucked inside the DNA of bacteriophages, viruses that infect bacteria. The company is calling the system ART, short for array-associated reverse transcriptases, and it posted the result simultaneously as a blog entry and an unreviewed preprint. The structure, the discovery workflow, and the limits of what's actually been proven all deserve equal attention.

The mechanics of the search are worth spelling out. According to the Anthropic blog post reviewed by Reuters, researchers gave Claude a single prompt: find interesting reverse transcriptases in a large DNA database. From there, the agents worked largely without human guidance. As reported by phys.org, roughly 950 Claude-powered software agents combed the database for 21 hours, processing 210 million tokens. They surfaced more than 200,000 genes for one class of enzyme, identified 3,500 candidate systems, and narrowed the list to 20 for human review. The company says that kind of triage can take an expert scientist weeks to months.

What caught the agents' attention was structural. ART has three components: a reverse transcriptase, a partner gene sitting next to it, and a long array of evenly spaced DNA repeats. As The Next Web reported, citing the preprint directly, those arrays hold 3 to 21 copies of a short repeat, and none of the ART systems carry the cas genes that define CRISPR. The layout is reminiscent of CRISPR's repeat architecture, where an array stores RNA guides, but the similarities are structural, not functional. Anthropic has not shown ART does what CRISPR does. According to the Anthropic blog post, the company has not yet determined the function of the system at all.

The replication detail buried in The Next Web's reporting is the most important caveat in the paper. Anthropic ran the same discovery campaign ten additional times. None of the reruns read far enough upstream of the enzyme to catch the repeat array, and all ten missed it. The authors attribute this to the size of the search and the agents' unpredictable behavior. In controlled tests where the DNA was handed to the model directly, Anthropic's four best models described the array correctly at least 90 percent of the time, but when agents had to navigate files and tools autonomously, that rate fell as low as 32 percent. That's not a footnote. It means the original find may have been as much a product of scale and stochasticity as it was of genuine analytical reliability.

The peer-review gap here is substantial. As phys.org noted, the results are a preprint and have not been peer reviewed. Anthropic's wet lab, operating at biosafety levels 1 and 2 in the San Francisco Bay Area, did confirm experimentally that the ART array is transcribed into a set of distinct short RNAs. In published Staphylococcus phage data, those RNAs accounted for a meaningful share of the phage's total RNA output at a specific timepoint. That's a real experimental signal. Whether it points to a programmable gene-editing mechanism, a phage defense evasion strategy, or something else entirely is open.

Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute, reviewed the preprint at Anthropic's request and called the identification of the repeat arrays "genuinely intriguing" and said it "merits further investigation," according to remarks released by Anthropic and confirmed in reporting by the Malay Mail. Anthropic CEO Dario Amodei also acknowledged on X that a Stanford team had independently described a comparable but distinct system. The company says Claude's contribution was specifically identifying the features surrounding the previously known reverse transcriptase enzyme that appeared to define a system no one had characterized as a unit.

The underlying reverse transcriptase had been identified in earlier work on a jumbo phage. What Claude appears to have added was context, pattern recognition across a large sequence space, and the flagging of a structural neighborhood that human researchers had not assembled into a coherent system. That's a meaningful contribution to a hypothesis. It is not, yet, a characterized biological mechanism. The peer-review pipeline will tell us how much of it holds.

Sources cited:
- Anthropic blog post (anthropic.com) (https://www.anthropic.com/news/claude-discovers-novel-enzyme-system)
- phys.org (https://phys.org/news/2026-09-anthropic-touts-ai-biology-discovery.html)
- The Next Web (https://thenextweb.com/news/anthropic-claude-enzyme-system-crispr-like-repeats)
- Reuters via WHBL (https://whbl.com/2026/09/23/anthropic-says-claude-ai-helped-discover-novel-enzyme-system/)
- Malay Mail (https://www.malaymail.com/news/tech-gadgets/2026/09/24/anthropic-says-its-claude-ai-helped-uncover-previously-unknown-enzyme-system-that-may-enable-new-gene-editing/236311)