XNews

Claude Discovers Novel Enzyme System ART in Bacteriophages

Anthropic announced Claude discovered ART, a previously unnoticed CRISPR-like enzyme system in bacteriophage DNA. About 950 Claude agents screened 200,000+ reverse transcriptases over 21 hours. MIT professor Feng Zhang reviewed the preprint, calling it a strong example of AI agents in biological discovery.

Key insights

ART: A New CRISPR-Like Enzyme System

Anthropic named the newly identified system ART (array-associated reverse transcriptase). It consists of a reverse transcriptase, a partner gene of unknown function, and a long array of evenly spaced repeated DNA sequences — structurally similar to CRISPR. The array is transcribed into a set of distinct short RNAs, suggesting a potentially programmable mechanism. Previously, only a handful of systems with these features had been found, all capable of cutting, copying, or pasting DNA. ART's specific function is not yet clear, and Anthropic acknowledged its biotechnological utility is unknown.

Claude-Driven Discovery Process

The discovery was nearly entirely completed by Claude. Researchers gave an initial prompt to search for interesting new reverse transcriptases in DNA databases. About 950 Claude agents ran for 21 hours, consuming 210M tokens. They screened over 200,000 reverse transcriptases, narrowed to 3,500 candidates, and selected 20 for human-readable reports. One agent directly identified the repeated array while reading raw sequences, counted repeats, measured spacing, compared with known systems, and confirmed it was novel before submitting for human review.

Dario Amodei: Biology AI on Exponential Trend

Anthropic CEO Dario Amodei emphasized that AI performance in new intellectual domains has gone from weak to superhuman within a few years: 2023 models struggled with high-school math, 2024 they performed well on national math competitions, 2025 they solved minor open problems, and in late 2026 they are beginning to solve top open problems in mathematics. He believes AI for biology follows a similar exponential trend. The main difference is that biology requires experimentation, but humans can collaborate with AI to validate key results in weeks.

Wet Lab Collaboration Model

The work came from Anthropic's life sciences team formed in spring, with a BSL-1/BSL-2 wet lab in the Bay Area that does not handle materials dangerous to humans. All experiments were performed by human scientists using publicly released Claude Science and Claude Code tools. This means other labs could replicate the 'AI proposes hypotheses, humans verify' workflow. Anthropic is soliciting external research collaboration proposals and aiming to grow its life sciences team.

Potential Gene Editing Implications

A system with all these features — reverse transcriptase, partner gene, and repeated arrays — has previously been found only a few times, and all such systems performed DNA cutting, copying, or pasting. CRISPR has already produced gene-editing drugs, and the other systems are being developed into new tools. Even though ART's precise function remains unknown, a Stanford team independently described a similar but distinct RT system that evolved separately. Amodei said he believes we are at the very beginning of finding such systems and developing them into powerful biotechnology tools.

Source posts