Anthropic’s life‑sciences team and laboratory reported on September 23, 2026 that their general AI model Claude autonomously identified a previously uncharacterized enzyme system, which they name array‑associated reverse transcriptases (ART). The group says the finding emerged from an automated genomic search followed by laboratory follow‑up.
What happened and how it was found
The team tasked Claude with scanning a large database of DNA sequences to find novel reverse transcriptase (RT) families and associated systems. The automated search ran for about 21 hours, involving roughly 950 Claude agents and processing approximately 210 million tokens. From the data the agents collected over 200,000 RT sequences, flagged about 3,500 candidate systems, and narrowed these to 20 of the most compelling candidates for human‑readable reports.
One agent detected an unusual pattern: a tandem array of regularly spaced, non‑coding DNA repeats adjacent to an atypical RT gene. The repeat arrangement resembled CRISPR arrays—long banks of repeats that in CRISPR systems hold programmable RNA guides. In this case the repeat array co‑occurs with an RT and an additional accessory protein of unknown function, a combination the team says had not previously been recognized. They labeled the system ART (array‑associated reverse transcriptases).
What is known about ART so far
Anthropic reports that ARTs are primarily found in bacteriophages (viruses that infect bacteria) and comprise three main components: the reverse transcriptase, a neighboring partner gene, and a long array of evenly spaced repeat sequences. Initial laboratory experiments indicate the ART arrays are expressed as distinct short RNAs, analogous to CRISPR arrays, but the precise biological function and mechanism of ART remain unresolved.
The particular RT observed had been reported before in a jumbo phage, but Claude’s agents were, according to Anthropic, the first to notice the defining combination of the RT, the repeat array, and the accessory protein and to raise it as a distinct system. Further experiments to determine ART’s function are ongoing.
Lab setup and how the team works
The laboratory is based in the Bay Area and is equipped as a conventional molecular biology lab. Work is conducted at biosafety levels BSL‑1 and BSL‑2 only; Anthropic states they do not handle pathogens capable of infecting humans. Human scientists carry out all lab procedures. Claude is used to search large genomic datasets, generate hypotheses, and help interpret experimental data, while the human team selects which model‑generated candidates to validate experimentally.
A typical workflow begins with Claude surveying literature and public data for a protein family, reproducing established results to verify methods, and then searching for uncharacterized family members or genomic neighbors. Claude generates concise, human‑readable reports proposing functions and evidence. Humans then critically evaluate and often eliminate most candidates; a few survive to laboratory testing, where expression, biochemical, and structural characterization are performed with Claude’s assistance for interpretation.
Because Claude produces large numbers of hypotheses, the team studies what distinguishes candidates worth testing; the findings inform refinements to the instructions given to Claude and improve its selection process.
External reaction and documentation
Feng Zhang (Massachusetts Institute of Technology and Broad Institute), a pioneer in CRISPR genome editing who reviewed the preprint, commented: “This is an exciting example of how AI agents can contribute to biological discovery. The identification of RNA‑repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation. I hope this work encourages more scientists to explore how AI can support their research.”
Anthropic has released a technical report and a preprint detailing methods and the initial results.
Why this matters
The announcement illustrates how automated, large‑scale AI searches can accelerate genomic discovery by surfacing patterns that are time‑consuming to detect manually. The characteristics of ART—an RT associated with a repeat array and an accessory protein—occur together only rarely in systems that perform programmable DNA or RNA operations, which makes ART interesting as a potential molecular tool. However, based on the available data, the team stresses that ART’s function and biotechnological applicability are still to be established through further experiments.
Collaboration invitation
Anthropic says it is open to collaborating with other researchers to apply this AI‑driven hypothesis‑generation approach to broader problems in genomics and other fields and invites proposals from interested scientists.
