Anthropic’s AI-assisted enzyme lead is promising, but its function remains unknown

Anthropic says Claude helped researchers identify a previously uncharacterized bacterial system involving reverse transcriptase and repeated DNA sequences. Early experiments support that the system is expressed. They do not yet show that it works like CRISPR or what biological job it performs.

A research team using Anthropic’s Claude models reports a new lead in phage-associated molecular biology: an array-associated reverse transcriptase, or ART, system. The company says researchers used an agentic workflow to search literature, analyze sequence data and prioritize candidates before conducting laboratory experiments. The work is described in a bioRxiv preprint and in Anthropic’s account. Anthropic says roughly 950 agent sessions ran over 21 hours and used about 210 million tokens; all laboratory work was performed by human researchers.

The reported system combines a reverse-transcriptase gene, a nearby partner gene and a repeated sequence array. The enzyme itself was previously known; the proposed combination is the uncharacterized system. Initial laboratory work suggests that the array is expressed as distinct short RNAs. That gives researchers evidence that the genomic pattern is biologically active, rather than merely an interesting arrangement in a sequence database.

It does not settle what the system does. Anthropic’s account and independent reporting both describe its function as unknown. In particular, a resemblance to CRISPR in the presence of repeated sequences should not be read as evidence that the system cuts DNA, edits genomes or provides an equivalent defense mechanism. Nature’s report emphasizes that the evidence for a CRISPR-like role is not yet established.

This is an important distinction in discovery science. Computational analysis can narrow a huge search space, connect clues across literature and help identify an experiment worth running. But a candidate is not a mechanism, and a mechanism is not a demonstrated application. The lab work is what starts to test whether the computational lead corresponds to a real biological system; additional experiments must establish its function, partners and behavior under relevant conditions.

Anthropic says the effort involved hundreds of model sessions over roughly a day and a large volume of generated tokens. Those details illustrate how the company organized the work, but they are not a measure of scientific validity. A large number of agent interactions does not establish that a conclusion is reproducible, and the company has a commercial interest in showing its models contributing to research.

The preprint is therefore the central document for readers who want to assess the claim: it contains the methods and experiments that should be evaluated by other specialists. Peer review, independent reproduction and functional follow-up will determine whether ART is a substantial discovery or an intriguing genomic clue. The materials available now support the narrower claim that AI-assisted analysis helped identify a candidate system and that initial lab experiments detected its RNA expression.

The story is less “AI invents a new CRISPR” than a case study in the handoff between pattern recognition and experimental biology. Models may be effective research partners when they help scientists find and prioritize hypotheses. The decisive work remains showing what those hypotheses mean in living systems.