Claude finds a new bacterial enzyme system by itself

Anthropic's new life sciences group gave Claude a single instruction: comb a large DNA sequence database for something interesting. What came back was a class of enzymes nobody had described before, which the team calls array-associated reverse transcriptases. The system has three parts, a reverse transcriptase, a partner gene, and a repeated stretch of DNA that looks a lot like the arrays in CRISPR, and it appears in bacteriophages, the viruses that infect bacteria. The find was strong enough that Feng Zhang, one of the scientists who built CRISPR into a gene-editing tool, called it a real example of AI agents contributing to biological discovery.

The numbers give a sense of the scale. Claude agents worked through more than 200,000 reverse transcriptases, proposed roughly 3,500 candidate systems, then narrowed those to 20 worth a close look. About 950 agents ran for 21 hours and spent 210 million tokens, and Anthropic says the search covered ground that would normally take human scientists weeks or months. The researchers still did the wet-lab validation themselves at their Bay Area facility, but the hypotheses came from the model. You can read the full writeup for how the group set up the run.

The honest caveat is that this is one result, endorsed by one prominent scientist, and the broader hit rate of agent-driven discovery is still unknown. But the shape of the work is what stands out: a database search, a set of ranked hypotheses, and a lab confirmation, with the model doing the tedious middle.

Why it matters

If you work in computational biology, this is a concrete template you can copy: point agents at a large sequence database, let them rank candidates, and reserve your bench time for validating the top handful. The result also means model providers are now competing on scientific discovery, not just chat and code, so it is worth watching whether the hit rate holds up beyond a single flagship example.