A Pattern Beside an Enzyme
The search began with reverse transcriptases, enzymes known for copying RNA into DNA. Anthropic asked Claude agents to look for unusual examples across a database containing about 1.9 billion protein clusters. The agents identified nearly 200,000 reverse transcriptase clusters and investigated genes found near them.
One candidate drew attention because of DNA next to its reverse transcriptase gene. An agent examining the sequence noticed a series of regularly spaced repeats that had not been described as part of that enzyme system. The research team subsequently named the family array-associated reverse transcriptases, or ART. Its members are found chiefly in bacteriophages, viruses that infect bacteria.
The arrangement is striking because CRISPR systems also contain repeated DNA sequences. Similarity in layout, however, does not establish similarity in function. Anthropic’s researchers found no CRISPR-associated cas genes near the ART sites they examined, and the spaces between ART repeats differ from those in typical CRISPR arrays. The comparison points to a feature worth investigating, rather than evidence that ART can edit genes.
From a Computer Search to the Lab
Claude’s work went beyond sorting a list of candidate enzymes. During the campaign, agents proposed follow-up tasks, checked possible associations and produced reports for human review. The technical report records 949 agent sessions across 119 tasks, completed over 21.5 hours of elapsed time. That account gives a measure of the search, though the number of agents alone says little about how often the process would yield a finding of similar value.
Human scientists then examined whether the pattern extended beyond the sequence that first caught the agent’s attention. They identified 95 distinct ART reverse transcriptase clusters in their searches of genetic databases and public genomes. Detectable repeat arrays appeared beside 28 of them. The researchers also found a neighboring gene that appears to be associated with the enzyme family.
The team examined RNA evidence as well. In data from a previously published study of a Staphylococcus bacteriophage infection, RNA produced from an ART repeat array was abundant. Anthropic’s scientists separately expressed an ART system in E. coli and observed short RNAs from the array. Those experiments support the finding that the repeated sequences produce RNA. They do not yet show what the RNAs do, or whether the reverse transcriptase uses them.
The Evidence and the Open Questions
The unknowns are central to the story. In their report, the researchers say they have not shown that the ART reverse transcriptase is active or that the short RNAs are its substrates. They also have not established whether the enzyme interacts with its neighboring protein or what advantage the system might give the phage. Claims that ART is a new gene-editing tool, or that it works like CRISPR, would go beyond the published results.
Anthropic frames the project as a demonstration of how AI agents can spot an unexpected feature in raw biological data and pursue it far enough to guide experiments. The company’s scientists set the research goal, built the system used for the search, reviewed its findings and performed the lab work. The observation of the repeat pattern arose during the agents’ analysis.
The researchers also tested how repeatable that discovery was within their workflow. They ran the campaign 10 more times; some runs examined the same enzyme lineage, but none independently read the neighboring DNA and recognized the defining repeat array in the way the initial run did. That result leaves an important question open: how reliably can the approach make a useful observation when the next unusual pattern looks different?
An Early Result for Anthropic’s Biology Team
The finding arrives as Anthropic expands its work in life sciences. Its new research group aims to use Claude to search genetic datasets, develop hypotheses and bring selected candidates into the laboratory. In this case, the path from a computer-generated lead to experimental evidence produced a specific subject for further study: an enzyme family associated with repeat arrays and short RNAs.
The work is available as an Anthropic preprint, rather than a peer-reviewed journal article. Other researchers can inspect the methods and findings, while further experiments are needed to establish ART’s biological role. Anthropic says that work is underway.
A practical use may emerge, or the system may prove interesting for what it reveals about bacteriophages themselves. At present, the result is more precise than either prediction: an AI agent noticed a repeat pattern, scientists found related systems and RNA evidence, and the underlying mechanism remains unresolved.
