Anthropic: AI model Claude autonomously discovers novel enzyme system ART in phage DNA

By Zhang Mengran, Science and Technology Daily

According to an official announcement from Anthropic, the company published findings on March 23 from its Life Sciences Research Group and laboratory: using only high-level search directions provided by human researchers, the AI model Claude autonomously analyzed genomic data through parallel agent-based processing and discovered a previously uncharacterized enzyme system in bacteriophages. The system has been named Array-associated Reverse Transcriptase (ART).

Key Findings

The ART system consists of three components:

• Reverse transcriptase: An enzyme that synthesizes DNA from an RNA template
• An adjacent companion protein gene: Function currently unknown
• A DNA array with regularly spaced repeats: Resembling CRISPR-like repetitive sequences

While the system may offer new tools for nucleic acid manipulation, its actual biological function remains to be confirmed experimentally.

How the Discovery Was Made

The research began with a single prompt instructing Claude to search large-scale DNA sequences for novel reverse transcriptases of interest. Approximately 950 Claude agents worked in parallel for 21 hours, consuming roughly 210 million tokens. The process involved:

1. Screening nearly 200,000 reverse transcriptases to narrow down to approximately 3,500 candidate systems
2. Further focusing on 20 high-priority targets
3. One agent, while examining raw sequences adjacent to an unusual phage reverse transcriptase, identified a row of regularly spaced repetitive DNA
4. The agent then compared the sequence against known families, counted repeat units, and searched relevant literature before submitting the system for human review

Laboratory Validation

The human research team conducted further wet-lab experiments and confirmed that the repetitive array is expressed and produces a set of distinct short RNAs, indicating active transcriptional activity.

Expert Commentary

Dr. Feng Zhang, a professor at the Massachusetts Institute of Technology and a pioneer in the CRISPR field, reviewed the materials and stated that this is "an exciting example of an AI agent contributing to biological discovery." He noted that such RNA repeat arrays associated with reverse transcriptases warrant further investigation.

Anthropic's Perspective

Anthropic emphasized that AI was responsible only for data retrieval, candidate screening, and hypothesis organization, while all wet-lab experiments were conducted by human scientists. The company noted that if this type of genomic mining can be scaled, it could compress analyses that would typically take experts weeks or months down to just hours.

However, Anthropic also stressed that whether ART can ultimately be developed into a biotechnology tool comparable to CRISPR remains dependent on future experimental validation of short RNA targets, companion protein functions, and overall regulatory mechanisms.

Editor's Note

Many discoveries that have fundamentally transformed biology and medicine began with scientists noticing peculiarities in the diversity of molecular systems in nature. Yet, faced with massive biological datasets, traditional screening methods are often time-consuming, labor-intensive, and costly. With its powerful data processing and pattern recognition capabilities, AI can rapidly extract valuable information from complex data—and, as demonstrated here, identify novel systems worthy of experimental validation. Looking ahead, AI is expected to further enhance experimental prediction, providing directional guidance for biological research, thereby significantly reducing trial-and-error processes and accelerating the pace of scientific discovery and translation.