Array-associated reverse transcriptases (ART): what Claude found scanning DNA at scale
Anthropic's new life sciences research group has used its Claude models to comb through DNA datasets and surface a genetic system nobody had named before: array-associated reverse transcriptases, or ART. The finding shows what happens when a language model is set loose on raw sequence data with instructions to look for patterns rather than confirm known ones.
Anthropic's new life sciences research group has used its Claude models to comb through DNA datasets and surface a genetic system nobody had named before: array-associated reverse transcriptases, or ART. The finding shows what happens when a language model is set loose on raw sequence data with instructions to look for patterns rather than confirm known ones.
A group built to point Claude at raw genomic data
Anthropic formed its life sciences research group in spring 2026. The group uses Claude to study DNA datasets and identify protein families that have not been characterized before. ART is the system this work surfaced: a previously uncharacterized enzyme system associated with DNA repeat arrays.
What an ART system actually looks like
An ART system has three parts sitting next to each other in the genome: a reverse transcriptase gene, a neighboring partner gene, and a long array of evenly spaced DNA repeats. That repeat layout resembles the structure of a CRISPR array. ARTs are found mainly in bacteriophages. Initial experiments found that ART arrays are expressed as distinct short RNAs. The primary function of ART systems remains unknown; Anthropic has not attributed a role to them beyond that open question.
The scale of the search that found it
Finding ART meant searching an enormous amount of sequence data. Roughly 950 agents ran for 21 hours, consuming 210 million tokens across the search. That run gathered more than 200,000 reverse transcriptases and identified 3,500 candidate systems worth a second look. That set of 3,500 was narrowed down to 20 leading systems for detailed analysis. The available facts do not specify how that narrowing was carried out, only that ART was one of the systems that emerged from it and became the one described publicly.
Where the physical lab work fits in
The computational search is only part of the process. Anthropic's laboratory operates at BSL-1 and BSL-2 biosafety levels and does not handle pathogens capable of infecting humans. All laboratory work is performed by human scientists, not by Claude itself. The model's role was to search and sort candidates out of a dataset far larger than a human team could review directly, while the physical follow-up work sits with people trained to run it.
An open question, published as one
The unresolved function of ART is the honest ending here, not a gap to explain away. Anthropic has shared the finding through its own site, so anyone wanting to follow how the function of ART gets resolved, or to see the underlying account of the search itself, can go directly to that source. For readers who work with AI-assisted research or who train teams to use language models for large-scale pattern search rather than simple text generation, this kind of workflow, hundreds of agents run against raw data for hundreds of millions of tokens, is a concrete reference point worth studying alongside their own applied AI practice.
Specs
- Search duration: 21 hours
- Agents used: approximately 950
- Tokens used: 210 million
- Reverse transcriptases gathered: more than 200,000
- Candidate systems identified: 3,500
- Leading candidates analyzed: 20
- Biosafety levels: BSL-1 and BSL-2
Sources
- Anthropic — Anthropic