AI agents discover novel enzyme system with CRISPR features in 200K enzyme search

This discovery provides concrete support for the argument that AI has value in life sciences, an area previously driven more by potential than outcomes. A tangible demonstration of AI agents shortening early-stage research timelines could improve sentiment towards AI-powered drug discovery, genomics tools, and gene-editing platforms. Because Anthropic is not publicly traded and ART's commercial potential is unconfirmed, any impact will probably be felt across the sector rather than tied to a single company. The news also balances the wider discussion about AI agents at scale, following a recent security breach involving an AI agent.

AI agents have detected something in a dataset of 200,000 enzymes that human researchers had overlooked. Although the function of this system remains unknown, the rapid pace of the search could be as significant as the discovery.

Summary:

• Approximately 950 Claude agents spent 21 hours searching through genomic data, collecting over 200,000 reverse transcriptases and filtering roughly 3,500 candidate systems down to 20.
• The most notable discovery, called array-associated reverse transcriptase (ART), links an enzyme with a companion gene and a lengthy array of evenly spaced DNA repeats that resemble CRISPR.
• Initial laboratory experiments indicate the repeat array is transcribed into multiple short RNAs, a characteristic that enables CRISPR systems to be programmed.
• ART has not yet been demonstrated to function like CRISPR, and its biological purpose remains unclear.
• Feng Zhang, a CRISPR pioneer, stated that the finding warrants further investigation.
• A new laboratory in the Bay Area employs human scientists for all experiments, while Claude handles data searches, hypothesis generation, and analysis.

According to Interesting Engineering, Anthropic's Claude AI model discovered an enzyme system in bacteriophages that was not previously characterized, sharing certain features with CRISPR, following a search of over 200,000 reverse transcriptases by AI agents. This finding provides an early instance of AI agents directly aiding biological discovery, though researchers emphasize that the system's role is still unknown.

This research was part of Anthropic's newly established life sciences program. Using approximately 210 million tokens, around 950 Claude agents searched genomic data for 21 hours. They collected over 200,000 reverse transcriptases, identified about 3,500 candidate systems, and selected 20 for in-depth reporting. Anthropic noted that mining the genome at this scale would require an expert scientist weeks or months to complete.

The leading candidate, termed array-associated reverse transcriptase (ART), couples a reverse transcriptase with an adjacent partner gene and a lengthy array of DNA repeats that are evenly spaced. While scanning raw DNA around an unusual enzyme family, one agent detected the repeat array. It then counted the repeats, measured their distances, compared the pattern against known systems, and reviewed scientific literature before alerting human researchers. Although the enzyme itself was previously known, the surrounding system—including the repeat array and an accessory protein whose function is unknown—had not been described.

Preliminary lab studies showed that the ART repeat array generates several different short RNAs. This is notable because CRISPR arrays similarly produce RNAs that allow CRISPR-Cas systems to be programmable. According to Anthropic, the set of features is similar to a handful of known systems that can perform programmable actions on DNA, including cutting, copying, or inserting genetic material.

Scientists warn that ART has not yet been proven to function like CRISPR, and experiments are ongoing to establish the actual roles of its components. Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute, reviewed the work and said it was an exciting demonstration of AI agents assisting biological discovery, warranting further research.

Anthropic has also established a molecular biology lab in the Bay Area to test candidates produced by computational searches. Human scientists perform all experiments, while Claude is used for data searching, hypothesis generation, and result interpretation. Regardless of whether ART becomes a new biotechnology tool or remains a biological oddity, the method indicates that AI agents can shorten the early discovery phase of life sciences research.