TL;DR

Anthropic has introduced a life sciences research group with its own laboratory. Its first reported result is a previously undescribed enzyme system in bacteriophages, viruses that infect bacteria, which the company says Claude agents identified with only high-level direction from scientists. The system has a repeat pattern reminiscent of CRISPR, but its function is not yet known. A pre-print has been released.

What Claude found

Scientists asked Claude to search a large database of DNA sequences for new examples of reverse transcriptases (RTs), enzymes that turn RNA into DNA. According to Anthropic, roughly 950 agents worked for 21 hours, using 210 million tokens. From a haul of more than 200,000 RTs, they flagged 3,500 possible new systems, then wrote up the 20 strongest. The company says that kind of survey can take an expert weeks or months.

One agent noticed a regular array of repeated DNA next to an unusual RT gene. The RT itself had appeared in earlier studies. What Anthropic credits Claude with is spotting, apparently before anyone else, the two features that define the system: the repeat array and an extra protein of unknown purpose. The team calls it array-associated reverse transcriptase, or ART.

Early experiments show the array is transcribed into several separate short RNA molecules. That is suggestive, because the RNAs stored in a CRISPR array are what let CRISPR-Cas tools be reprogrammed, but Anthropic is clear it does not yet know what ART does.

Feng Zhang, a CRISPR pioneer who holds posts at the Broad Institute and MIT, reviewed the pre-print. He called the finding “genuinely intriguing” and said it “merits further investigation”.

How the lab works

All bench work is done by human scientists, in a Bay Area lab restricted to the two lowest biosafety levels and with no human pathogens. Claude reads the literature, reproduces known results to check its own methods, proposes functions for unexplained proteins and then critiques its own candidates. Most candidates are rejected at that stage. The team works in Claude Code and Claude Science, both of which are commercially available.

This follows Reuters’ report last week that Anthropic had quietly set up a biology lab in the Bay Area. Anthropic’s post confirms a Bay Area lab and puts a first research result to it.

Looking forward

The result is a company announcement backed by a pre-print, not peer-reviewed science. Its value depends on what ART turns out to do. The more durable claim is about method: agents surveying huge sequence databases and ranking candidates for humans to test. For UK genomics groups and biotech start-ups, the question is whether that search-then-test loop can be reproduced outside a frontier lab, and at what cost in compute.