After searching more than 200,000 reverse transcriptases, Anthropic's Claude identified a previously uncharacterized enzyme system in a bacteriophage, demonstrating how AI agents can help scientists discover hidden biological systems in vast DNA datasets. The newly identified system is called array-associated reverse transcriptase (ART), which combines a reverse transcriptase with adjacent partner genes and a long array of evenly spaced DNA repeats. This arrangement has similarities to CRISPR systems, but researchers do not yet know ART's function.

Claude discovered the system while analyzing DNA sequences as part of Anthropic's new life sciences research program. About 950 agents spent 21 hours using roughly 210 million tokens to search genomic data, producing thousands of potential biological systems for further analysis. The agents collected more than 200,000 reverse transcriptases, identified about 3,500 candidate systems, and ultimately narrowed the field to 20 candidates for detailed reporting. Anthropic says this kind of genome mining work could take expert scientists weeks or months.

One Claude agent focused on an unusual reverse transcriptase family and began examining the raw DNA around it. It detected a tandem repeat array next to the RT gene, a pattern that had apparently not been recognized before. The agent then counted the repeats, measured their spacing, compared the arrangement with known reverse transcriptase systems, and searched the scientific literature for prior descriptions. It ultimately flagged the sequence as a potential new biological system for human researchers to review.

Reverse transcriptases themselves are not new. Previous research had already identified the enzyme in a giant bacteriophage. Anthropic researchers say the newly identified feature is the broader system around it, including the noncoding repeat array and an additional protein whose function remains unknown. Preliminary laboratory experiments found that the ART repeat array is expressed as multiple distinct short RNAs. This is particularly interesting because CRISPR arrays also produce RNAs that help make CRISPR-Cas systems programmable.

The researchers caution that ART has not been shown to work like CRISPR. Its biological role remains unknown, and further experiments are underway to determine the actual functions of the reverse transcriptase, accessory protein, and repeat-derived RNAs. However, Anthropic says the combination of these features resembles a small group of previously discovered biological systems capable of programmable operations involving DNA, including cutting, copying, or inserting genetic material.

After reviewing the work, Feng Zhang, a CRISPR genome-editing pioneer and a professor at MIT and the Broad Institute, said the discovery warrants further investigation. "This is an exciting example of how AI agents can contribute to biological discovery," he said. "Identifying an RNA repeat array associated with a reverse transcriptase is truly fascinating and worth further study."

Anthropic has also established a Bay Area molecular biology laboratory to test candidates generated by computational searches. All laboratory experiments are conducted by human scientists, while Claude is used to search genomic datasets, generate hypotheses, analyze candidates, and help interpret experimental results. The ART work is still at an early stage, and researchers are investigating its biological function and whether its unusual structure might eventually have biotechnology applications.

From the perspective of AI in scientific discovery, this case demonstrates the potential of agents in genome mining. Traditionally, finding systems with specific structures among millions of gene sequences requires researchers to write search rules, filter results, and manually inspect candidate sequences, a process that is time-consuming and prone to missing things. Claude's agent approach integrates search, pattern recognition, literature comparison, and hypothesis generation into an automated workflow, allowing weeks or months of work to be completed in 21 hours. However, there is an essential difference between discovering a candidate system and proving its function. The structural similarity between ART and CRISPR suggests it may have programmable DNA manipulation capabilities, but the expression of the RNA repeat array and the function of the accessory protein still require laboratory validation. Anthropic's establishment of a molecular biology laboratory and its use of human scientists to conduct experiments indicate that AI's role is to generate hypotheses and narrow the range of candidates, rather than replace experimental validation. Zhang's assessment provides recognition of the discovery within the field, but he likewise emphasizes the need for further research. The biological function and potential biotechnology applications of ART remain uncertain, and if subsequent experiments confirm that it has programmable DNA manipulation capabilities, it could add a new candidate system to gene-editing tools.