Anthropic says a team of Claude agents found a previously uncharacterized enzyme system while searching bacteriophage DNA, the first result it has disclosed from a new in-house biology lab. The finding concerns a pattern of repeated DNA next to a reverse-transcriptase gene. Researchers do not yet know what the system does.
The company calls it array-associated reverse transcriptases, or ART. It consists of a reverse transcriptase, a nearby partner gene and an array of evenly spaced DNA repeats. Reverse transcriptases copy RNA into DNA. The enzyme in one large bacteriophage had been identified before; Anthropic says the new observation is the combination of the enzyme, the adjacent gene and the repeat array as a possible system.
Anthropic’s scientists gave Claude a high-level prompt to look for interesting reverse transcriptases in a large sequence database. According to the company, agents examined more than 200,000 enzymes, identified about 3,500 candidate systems and reduced those to 20 for deeper analysis. Roughly 950 agent sessions ran over 21 hours and used 210 million tokens. One agent noticed the repeat pattern beside an unusual enzyme and prepared a report for human review.
Human researchers then tested the candidate in Anthropic’s Bay Area laboratory. Their early experiments found that the DNA array produces a set of short RNAs. That gives the repeated sequences a measurable biological output, but it does not establish what the enzyme or the RNAs do. The resemblance to CRISPR lies in the repeat-array structure; the researchers have not shown that ART can edit genes or serve as a programmable biotechnology tool.
Anthropic published a technical preprint with the announcement. The work has not been peer reviewed, and its proposed biological role remains an open question. Feng Zhang, a CRISPR researcher at MIT and the Broad Institute who reviewed the preprint, described the identification of the repeat arrays as intriguing and said it warranted further investigation.
The company formed the research group this spring to combine genome searches with laboratory experiments. It says human scientists conduct all physical lab work and that the facility handles only lower-biosafety-level material, not pathogens that infect people. Claude Science, Claude Code and a system coordinating parallel agent sessions are used in the computational work.
The result is narrower than the phrase AI-discovered enzyme might suggest. Claude identified a candidate pattern that researchers had not characterized as a system; scientists still had to examine it and test the RNA output. Anthropic has not reported ART’s biological function, a therapeutic application or evidence that the method consistently finds useful systems. The disclosed result shows how an agent search can direct laboratory attention, while the scientific claim remains provisional.
Sources: Anthropic
–
By the Control Plane Editorial Team