Anthropic has opened a life sciences lab where its Claude models comb through DNA databases looking for enzyme families nobody has named yet. The lab’s first public result: Claude flagged a repeating stretch of DNA sitting beside an unusual gene, checked it against known systems on its own, and concluded the pattern marked an undiscovered piece of biology.
Anthropic calls the system ART, short for array-associated reverse transcriptase. A reverse transcriptase is an enzyme that turns RNA into DNA, and this one sat next to a long, evenly spaced array of DNA repeats plus an extra protein whose job is still unknown. That layout echoes CRISPR, the bacterial defense system scientists turned into a programmable gene-editing tool by exploiting its own repeat array, which stores a bank of RNA snippets the system uses to target specific DNA sequences.
Anthropic’s scientists gave Claude a single instruction: hunt through a massive DNA sequence library for reverse transcriptase variants nobody had documented. Roughly 950 agents took it from there, spending 21 hours and 210 million tokens on the search, working through more than 200,000 candidate enzymes before narrowing the field to 3,500 unfamiliar systems and finally 20 worth a closer look. One agent, reading raw sequence data, is quoted in Anthropic’s post noting a “tandem repeat array” that looked “CRISPR-like.” Human scientists then took over, expressing the protein in the lab and running the biochemical tests that confirmed the pattern was real, with Claude helping interpret the results.
That division of labor matters. The search and the pattern recognition came from the model. The wet lab work and the confirmation came from people. Anthropic has not established what ART actually does. Early experiments show its DNA array gets copied into short RNA fragments, a hint at a CRISPR-like function, but the company is explicit that the system’s primary purpose remains under investigation.
Feng Zhang, a CRISPR pioneer who splits his time between MIT and the Broad Institute, reviewed Anthropic’s pre-print and called the RNA-repeat pattern “genuinely intriguing” and worth further study. That is an endorsement of the finding’s interest, not a confirmation that ART functions as a gene-editing tool. Anthropic released the work as a pre-print, meaning it has not gone through external peer review or journal publication.
The history here is instructive. Restriction enzymes, Taq polymerase, and CRISPR itself were all first spotted by human scientists noticing something odd in raw genetic data, then spending years working out what the oddity meant. Anthropic’s pitch is that Claude can compress that noticing step, the part that used to take a career of pattern recognition, into hours of automated search. The claim comes almost entirely from Anthropic’s own blog post and pre-print, with one outside quote for context, so it should be read as the company’s account of its own model’s output.
Biotech researchers evaluating AI-assisted genomic screening should treat ART as an early lead generated by a vendor’s internal tool, not a validated discovery, until the pre-print clears peer review and the enzyme’s function is characterized.
Anthropic, published on the company’s official blog.
