Anthropic said on 23 September that Claude had identified a previously uncharacterised enzyme system in bacteriophages, viruses that infect bacteria. In its announcement of the result, the company described a search directed by scientists and carried out largely by AI agents, followed by analysis and experiments in its own laboratory. The system has a pattern of DNA repeats reminiscent of CRISPR, though Anthropic says it does not yet know what the system does.
Key points
- Anthropic says Claude spotted DNA repeats beside a previously identified enzyme.
- Roughly 950 agents searched for 21 hours, according to the company.
- Ten further searches failed to find the repeat array, Quartz reported.
- The system’s biological function remains unknown.
Claude spotted repeats beside a known enzyme
Anthropic calls the system array-associated reverse transcriptases, or ART. A reverse transcriptase copies RNA into DNA. Researchers had identified the underlying enzyme in earlier studies of a jumbo phage, Anthropic says, but Claude appears to have been the first to notice an associated stretch of repeating DNA and another protein whose function is unknown. The company describes the work as a discovery of a system, rather than of an entirely new enzyme.
The layout helps explain the comparison with CRISPR. In that system, repeated sequences sit alongside genetic material that can guide molecular machinery to a target. ART also has an array of repeats, but a similar arrangement of parts does not give it a known function: Anthropic says experiments to determine ART’s primary role are ongoing. The company says some other systems with a comparable combination of characteristics can be programmed to cut, copy or paste DNA.
Anthropic says its scientists set Claude the task of looking through DNA data for unusual reverse transcriptases. Agents then examined enzyme families and selected candidates for closer attention. One agent noticed repeats beside an unusual enzyme gene, prompting further analysis and laboratory testing by human scientists. The company characterises the result as autonomous because its scientists’ involvement in the search was limited to the initial prompt, while they performed the subsequent lab work.
Getting a medical test could eventually depend on an enzyme system identified through a search of genetic data, if experiments established a useful role for it. Anthropic points to an earlier enzyme discovery that became part of a DNA-copying method used in diagnostics. For ART, the company says the basic question of what the system does remains open.
A 21-hour search produced one lead
The search took 21 hours, used 210 million tokens and involved roughly 950 Claude agents, according to Anthropic. Tokens are the pieces of text a model processes as it reads and generates responses. Here, many agent sessions worked through genetic data, with one ultimately flagging a pattern for researchers to inspect. Quartz reported that the database contained roughly 1.9 billion protein clusters; it described a preprint in which the agents narrowed around 200,000 enzyme clusters to 3,564 candidate partner families and produced 19 reports for human review.
That funnel matters to the method. The agents were searching stored sequences and selecting leads, while the physical work belonged to the scientists. Anthropic says subsequent analysis and testing in its lab led it to recognise ART as a previously uncharacterised system in bacteriophages. Its account places the unusual observation in the neighbouring DNA, rather than in the earlier identification of the reverse transcriptase itself.
The search was hard to repeat. Anthropic ran it ten more times and none of those runs found the repeat array, Quartz reported. According to Quartz, the company attributed that outcome to the size of the search and variation in the routes its agents took through it. A developer building a research workflow around such agents would therefore have to distinguish a lead worth testing from a search that reliably returns the same lead.
Anthropic’s Bay Area lab tests the finding
Anthropic formed its life sciences research group in spring 2026 to investigate whether general AI models could speed up biological discovery. Its Bay Area laboratory works at biosafety levels 1 and 2, does not handle pathogens capable of infecting humans and leaves all laboratory work to human scientists, the company says. This result therefore joins a computational search to physical experiments within the same research group.
Feng Zhang, a professor at MIT and the Broad Institute who has worked on CRISPR genome editing, reviewed the preprint and called the identification of repeat arrays associated with reverse transcriptases “genuinely intriguing”, in a statement published by Anthropic. Lucas Harrington, who earned his PhD under CRISPR co-discoverer Jennifer Doudna, took a narrower view: genome mining has been used for decades, and the difficult step is finding out what a system does, he said, according to The Decoder.
The result also arrives as Anthropic prepares to go public and seeks to attract scientists to its laboratory, The Verge reported. Anthropic says its life sciences organisation also includes work on drug discovery and on training Claude in biology and chemistry. For ART, its scientists’ experiments to establish the system’s primary function are continuing.
