Anthropic: Claude Agents Independently Discovered ART, a Novel Enzyme System Architecturally Similar to CRISPR, in Bacteriophage Genomes
Anthropic: Claude Agents Independently Discovered ART, a Novel Enzyme System Architecturally Similar to CRISPR, in Bacteriophage Genomes
On September 23 Anthropic reported that the company assigned roughly 950 instances of Claude a single task: search genomic databases for novel reverse transcriptases, enzymes that copy RNA into DNA. Within 21 hours one of the agents noticed a DNA pattern resembling CRISPR adjacent to an unusual enzyme and developed the observation into a report for scientists. The laboratory confirmed that the identified DNA region is transcribed into RNA. The exact function of the system has not yet been established.
The agents were given a routine task: collect reverse transcriptases and identify their neighboring genes. One agent discarded its initial hypothesis about a neighbor and set itself a next step on its own, checking the DNA region upstream of the gene for a short RNA, as seen in retrons, another family of bacterial defense systems. After reading the raw sequence rather than its annotation, the agent wrote in its log: 'this is striking: I can see a tandem repeat array with my own eyes… CRISPR-like?!' It then counted the repeats on its own, measured their spacing, and checked the literature.
Scientists had already seen this region of the genome: another researcher, before Claude's finding, described the reverse transcriptase itself in a report and even proposed a nearby short RNA, but noticed neither the repeat array nor the partner gene. The preprint authors explain this simply: genomic data has accumulated faster than scientists can track. Reading the same region again, the agent noticed what a human had missed.
The system was named ART: a reverse transcriptase adjacent to an array of 3–21 noncoding DNA repeats and a partner gene of unknown function. CRISPR has a similar architecture (repeats next to a programmable enzyme) but a different origin: ART lacks cas genes, the 'scissor' proteins of CRISPR, and its repeats are conserved across related phages rather than swapped from strain to strain. Structurally, ART is closer to retrons, which protect bacteria by causing the infected cell to kill itself along with the virus. The laboratory confirmed part of the hypothesis experimentally: upon phage infection the ART array is transcribed into short RNAs and by the 15th minute reaches 8% of all phage transcripts; the same result was confirmed when the system was inserted into E. coli. Feng Zhang, an MIT professor and one of the pioneers of CRISPR, called the finding 'genuinely intriguing and worthy of further study.'
A detail from the preprint that the press release omitted: the campaign was rerun ten more times, and the array was not found in any of them. The reason was identified in a separate test: when the models were given files and analysis tools instead of raw sequence, accuracy dropped to 32% because the model did not read regions of 200 nucleotides or more in full. With direct reading, as in the original discovery, accuracy rose above 90%. Internal signals were found within the model that respond specifically to repetitive DNA. Previously only models trained on genomes, such as Evo 2, exhibited this capability, not general-purpose language models.
Anthropic CEO Dario Amodei wrote on X: 'the work was done mostly, but not entirely, by Claude' and 'at minimum, it is work I would have been proud of as a graduate student.' He also acknowledged that a similar system was recently and independently discovered by a Stanford team. He placed the campaign alongside CRISPR and the bridge recombinase DNA editor discovered at Arc Institute in 2024, noting that interest in reverse transcriptase-based systems is now growing across the field. The result was presented by the same Anthropic laboratory whose launch the company had announced three days earlier.