Anthropic Unveils ART: A CRISPR-Like Enzyme System Discovered by Large Language Models

Anthropic's latest announcement claims its large language model Claude has identified an enzyme system exhibiting characteristics remarkably similar to CRISPR, the Nobel Prize-winning gene-editing technology. In a September 23 release, the company reported that approximately 950 Claude agents operated simultaneously to locate intriguing genetic sequences across vast genomic databases within just 21.5 hours.

The system, dubbed ART (array-associated reverse transcriptases), belongs to a family of reverse transcriptases found primarily in jumbo phages—large bacteriophages that infect bacteria. These enzymes typically function in the opposite direction of standard cellular processes, where they copy RNA into DNA rather than transcribing DNA to produce RNA. After screening over 200,000 potential candidates and narrowing the list to several thousand novel variants, researchers isolated an "unusual" family featuring extensive repeat sequences that echo CRISPR patterns.

While the initial results are exciting, leading scientists caution that the discovery is still in early stages. Le Cong, a Stanford professor specializing in AI-driven genome engineering, notes that "the experiments are still in the queue" and that peer review has not yet occurred. He uses an apt analogy comparing the effort to scanning endless sand for a diamond: AI may spot something shiny, but lab validation is essential to confirm whether it is truly valuable or merely a false positive.

This development sits at the intersection of AI infrastructure and biotechnology, highlighting how advanced language models can now perform tasks traditionally requiring human expertise. For organizations building GPU clusters for AI workloads, understanding such cross-disciplinary advances helps guide investment in compute resources capable of sustaining iterative biological discovery pipelines.