A broad search produced one unusual lead

Anthropic says a group of Claude agents found a previously undescribed family of phage enzymes while searching public genome and metagenome data. The system combined many Claude Code sessions in a staged workflow: gather data, classify reverse transcriptases, inspect their genetic neighborhoods and investigate the most unusual families.

The accompanying preprint says the campaign surveyed about 1.9 billion protein clusters. It narrowed nearly 200,000 reverse-transcriptase clusters into a set of deeper investigations. One agent noticed a repeating stretch of DNA beside an enzyme in a jumbo-phage genome and treated the pattern as worth pursuing. The team calls the family array-associated reverse transcriptases, or ARTs.

What the researchers actually found

The paper describes 95 distinct ART enzyme clusters in cultured jumbo phages and predicted viral sequences. Twenty-eight had a detectable repeat array upstream. A dedicated partner gene sat downstream, giving the loci a three-part architecture: repeat array, reverse transcriptase and partner protein.

That arrangement resembles some known bacterial defense systems, but it is not CRISPR. The repeats are longer and organized differently, and the researchers found no nearby cas genes. Public RNA data from a Staphylococcus phage infection showed that one ART array was heavily transcribed. Human scientists then expressed that locus in E. coli and observed discrete short RNAs from the array. Those experiments support the existence of a coherent system. They do not reveal its job.

The boundary between agent work and lab work

The computational campaign was unusually autonomous. The paper reports 119 tasks, with worker, supervisor, curator and editor agents sharing a versioned research record. The agents used sequence databases, literature search, structure prediction and standard bioinformatics tools. They filed reports for human review rather than operating laboratory equipment.

People made the experimental decisions and performed the bench work. That distinction matters because a pattern in sequence data is a hypothesis generator. Physical validation can show that an RNA exists or that a predicted component is expressed. It still takes more experiments to establish mechanism, function and biological importance.

A discovery claim with open ends

Anthropic presents ART as evidence that general-purpose agents can notice anomalies in raw biological data, connect them to the literature and carry an investigation far enough to hand scientists a useful lead. The scale is real, and the reported result is more substantial than a benchmark score. It also comes from Anthropic researchers using Anthropic models, in a preprint that has not yet passed peer review.

The largest questions are still biological. The authors have not shown that the reverse transcriptase is active, that the repeat RNAs are its substrates, that the partner proteins interact with it, or what advantage the system gives a phage. ART may become a useful example of AI-assisted discovery. For now, it is a well-documented candidate system with a long experimental road ahead.

Sources

  1. Anthropic: Claude discovers a novel enzyme system with CRISPR-like repeatsPrimary September 23 announcement describing the agent campaign, human validation and the company's interpretation of the result.
  2. Anthropic preprint: Autonomous AI agents discover reverse transcriptases with tandem repeat arraysPrimary 40-page methods and results paper supporting the search scale, ART family description, experimental evidence and stated limitations.
  3. NCBI BioProject PRJNA836150Public primary RNA-sequencing dataset from the phage infection study that the ART paper reanalysed for evidence of array transcription.