Anthropic has discovered a previously unknown enzyme system in a bacterial virus with nearly 1,000 Claude agents in its new biology lab. The structure resembles the DNA repeats of CRISPR but combines them with a so-called reverse transcriptase. The function of the system referred to as ART in the genetic material is still unclear.
Nearly 1,000 agents review 200,000 enzyme candidates
According to Anthropic, 949 instances of Claude searched public genome databases for unusual enzyme families in a roughly 21-hour, largely autonomous run. The agents consumed about 215.6 million tokens and initially collected more than 200,000 so-called reverse transcriptases – enzymes that transcribe RNA into DNA and often serve as a defense against viruses in bacteria. From this mass, the agents filtered out 3,564 candidate families and ultimately condensed them into 19 detailed reports for human researchers. A single agent instance recognized the characteristic repeat pattern that later made the structure the actual find. The research took place in the company’s own biology lab in the Bay Area, the existence of which Anthropic had only publicly confirmed a few days earlier. Human scientists conducted all subsequent laboratory experiments themselves, while Claude assisted solely with data analysis, hypothesis generation, and evaluation.
Enzyme combines reverse transcriptase with CRISPR-like pattern
The team calls the new system “array-associated reverse transcriptase” (ART). It consists of three parts: the namesake reverse transcriptase, a neighboring partner gene with an unknown function, and a series of evenly distributed DNA repeats. This arrangement resembles the repeat arrays of CRISPR systems, in which bacteria store various RNA guide sequences – the underlying enzyme in ART, however, is a reverse transcriptase and not a CRISPR-typical nuclease. The system has so far only been found in bacteriophages, i.e., viruses that infect bacteria themselves. Initial laboratory experiments suggest, according to Anthropic, that the repeat region produces independent, short RNA pieces. What role these play in cellular processes remains open.
CRISPR co-developer Feng Zhang from MIT and the Broad Institute has reviewed the unpublished preprint. The RNA repeat arrangements are remarkable and deserve further investigation, according to his assessment – independent confirmation of the actual function is still pending.
Anthropic expands biology program around Claude
The find is the first published result of a new research group for fundamental biology that Anthropic has recently established. The company operates its lab at biosafety levels BSL-1 and BSL-2 and works exclusively with pathogens that are not transmissible to humans. In parallel, Anthropic offers a commercial working environment for pharmaceutical companies with Claude Science, whose clients, according to previous reports, include Novo Nordisk, Genentech, and Bristol Myers Squibb. CEO Dario Amodei wrote on the platform X that the company suspects ART to be a potential new mechanism for gene editing. However, the function, utility, and significance of the system have not yet been clarified. In addition to the blog post, Anthropic has published a technical preprint that describes the methodology and all 19 tested candidate systems in detail. An independent, peer-reviewed evaluation is still pending.
It will be crucial whether external laboratories can independently verify the existence and actual function of ART, as the preprint has not yet undergone peer review. If Amodei’s suspicion of a new gene editing tool is confirmed, Anthropic’s biomedical arm could grow into a second, scientifically driven business field alongside data centers and cloud contracts.
