Claude Spots a Never-Before-Seen Enzyme System That Looks a Lot Like CRISPR

A biology lab inside an AI lab

Anthropic announces the creation of a life sciences research group, with a physical lab set up in the Bay Area. The goal: do fundamental biology with Claude, exploring DNA datasets to identify uncharacterized protein families, generate hypotheses at scale, and test them at the bench.

The team consists of researchers specialized in computational DNA analysis, whose prior work covered CRISPR evolution and regulation or new enzymes for cell and gene therapies. It's integrated into Anthropic's life sciences organization, alongside teams focused on drug discovery and training Claude in biology and chemistry.

The lab remains deliberately modest in terms of risk: BSL-1 and BSL-2 only, no pathogens capable of infecting humans. And all manipulations are done by humans, since robotic automation doesn't mesh well with the improvised workflows of molecular biology.

The legacy of nature's "oddities"

Anthropic places its approach within a long lineage of discoveries born from an anomaly noticed in nature. Restriction enzymes, originating from bacterial immune systems, launched the biotech industry. Taq polymerase, found in a Yellowstone hot spring, became the basis for PCR.

CRISPR, finally, was first noticed as an unusual repeated sequence in the DNA of certain bacteria before becoming the foundation for gene-editing therapies. The question Anthropic is asking: can a general-purpose model systematize and speed up this kind of discovery?

How Claude found ART

The starting point is a simple prompt asking to search for interesting new examples of reverse transcriptases, enzymes that copy RNA into DNA. About 950 agents worked for 21 hours, scanning RT families and judging for themselves which candidates deserved attention.

One of them, while reading the raw sequence near an atypical RT found in a jumbo phage, spotted a tandem repeat array. It then proceeded like a researcher would: counting repeats, measuring spacing, comparing with known RT systems, doing a literature search, then submitting a report to humans.

The RT itself had already been identified in prior studies. What Claude appears to have been the first to note are the system's distinctive features: the associated array of non-coding sequences and an accessory protein of unknown function.

A still-mysterious but promising system

ART, found mostly in bacteriophages, brings together three elements: the RT, a partner gene, and a long series of regularly spaced repeats. This combination has only been observed in a handful of other systems, all of them programmable and capable of cutting, copying, or pasting DNA.

Early experiments show the array is transcribed into distinct short RNAs, similar to the guide RNA libraries that make CRISPR-Cas programmable. But ART's main function remains unknown, and Anthropic is choosing to publish early, via a pre-print, to showcase Claude's capabilities and open the topic up to the community.

Hypotheses become an object of study

The typical workflow: Claude reads the literature, reproduces established results to validate its methods, searches for proteins that don't match any described system, writes a report per candidate, then critiques its own evidence. Most leads get eliminated; a single campaign can end with just one candidate, or none at all.

Since Claude produces hundreds or thousands of reports per campaign, the team studies what distinguishes the proposals deemed worth testing. These lessons are fed back into the prompts to teach Claude to mimic researchers' "scientific taste." All of this runs inside Claude Science and Claude Code, sometimes with an in-house harness orchestrating parallel sessions.