Analysis
Anthropic says its Claude model has identified a novel enzyme system hidden in bacteriophage DNA that the company believes could represent a new gene-editing mechanism with properties reminiscent of CRISPR, according to Anthropic's own announcement and TechCrunch. Anthropic calls the system 'array-associated reverse transcriptases,' or ART.
The system has three parts: a reverse-transcriptase enzyme, a partner gene of unknown function, and a long array of repeating DNA sequences -- structurally similar in outline to the CRISPR systems bacteria use as an immune defense against viruses. Anthropic's researchers found it in the DNA of bacteriophages, the viruses that infect and replicate inside bacteria, using Claude inside the company's recently established biology research lab in San Francisco.
Claude made the identification in about 21 hours of what Anthropic calls 'concerted effort,' deploying roughly 950 agents that together processed 210 million tokens scanning genomic data for the pattern. All of the actual wet-lab experiments confirming the finding were still carried out by human scientists -- Claude's role was pattern identification across a search space too large for a small human team to comb through manually, not running the biology itself.
Anthropic's own disclosure includes a significant caveat: when researchers reran the identical search ten separate times, not one rerun surfaced the ART pattern again, meaning the discovery depended in part on how the AI agents happened to explore the search space that specific run -- a reproducibility gap the company states plainly rather than glossing over. Further experiments still need to confirm whether ART can actually cut or edit DNA the way CRISPR-Cas systems do; right now this is a computationally identified candidate, not a validated editing tool.
The result lands alongside a broader push by AI labs into scientific discovery -- Google DeepMind's AlphaFold reshaped structural biology and Isomorphic Labs is running its own AI-driven drug-discovery pipeline -- and Anthropic's version is notable mainly for how transparent the company was about the method's limits: a zero-for-ten reproducibility rate is not the framing of a company trying to oversell its own result.
