Loading page…
A new preprint describes how AI agents spotted a pattern in DNA data and human scientists tested it. The system’s function remains unknown.
Listen to this story
Anthropic’s biology group is using Claude to surface candidates for lab research, with ART offering an early example of both the approach’s promise and its limits. The agents found a repeat pattern beside a known reverse transcriptase; scientists then confirmed that the array produces distinct short RNAs. ART’s main biological function remains unknown, and the CRISPR-like appearance does not show that it can edit genes.
The agents screened more than 200,000 reverse transcriptases, selected 3,500 candidate systems, then narrowed them to 20 for detailed reports.
Anthropic reports that roughly 950 agents searched for 21 hours, using 210 million tokens.
ART combines a reverse transcriptase, a neighboring partner gene, and evenly spaced DNA repeats; it occurs mainly in bacteriophages.
Claude agents searching DNA data identified what Anthropic calls a previously uncharacterized enzyme system in viruses that infect bacteria. In a new preprint, the company describes the pattern that drew the agents’ attention and the first lab result supporting further study. The finding puts AI-led searching in the spotlight, but the system’s biological function is still unknown.
Anthropic calls the system array-associated reverse transcriptases, or ART. A reverse transcriptase copies RNA into DNA. Researchers had identified the underlying enzyme in an earlier study of a jumbo phage, a virus that infects bacteria. Anthropic says Claude’s contribution was spotting features around that enzyme that appeared to mark a system no one had characterized.
The scientists started the search with a prompt asking Claude to find interesting reverse transcriptases in a large DNA database. Anthropic says agents gathered more than 200,000 of the enzymes, selected 3,500 candidate systems and narrowed those to 20 for detailed reports. One agent examined DNA beside an unusual enzyme and noticed a long, orderly run of repeats.
Anthropic says roughly 950 Claude agents worked on the DNA-data search.
Anthropic says the agents spent 21 hours searching the data.
Anthropic reports that the search used 210 million tokens.
The agent counted the repeats, checked their spacing, compared the layout with known systems and searched for earlier descriptions before submitting a report for human review. Anthropic says scientists supplied the initial prompt and performed all laboratory work; the agents handled the database search and proposed which leads deserved attention.
ART has three parts, according to Anthropic: the reverse transcriptase, a neighboring partner gene and an array of evenly spaced DNA repeats. It occurs mainly in bacteriophages. The repeats resemble a feature of CRISPR systems, whose arrays hold sequences that help make those systems programmable. A similar-looking layout is a reason to investigate ART, not evidence that it can edit genes.
The first experiments found that the ART array produces distinct short RNAs. That adds a laboratory observation to the sequence-based finding, but it does not settle what those RNAs do or what the enzyme system is for. Anthropic says experiments to determine ART’s primary function are continuing.
The result comes from a biology research group Anthropic formed in spring 2026, with its own laboratory for testing AI-generated ideas. In the workflow Anthropic describes, Claude surveys scientific literature and DNA data, writes reports on possible systems, then reassesses its evidence. Scientists review promising candidates before testing proteins in the lab. Many proposed leads fall away before that stage.
Anthropic says a single campaign can yield hundreds or thousands of candidate reports. Its researchers study which proposals they consider worth testing, then use those judgments to refine Claude’s instructions. ART shows why that filter matters: an agent can notice a pattern humans want to pursue, while experiments still have to establish whether the pattern points to a useful biological mechanism.
Story updates
Anthropic CEO Dario Amodei has renewed his forecast that AI could help cure most diseases within five to 10 years, pointing to a discovery made with Claude. The model found a distinctive pattern around an enzyme in genetic data, and Anthropic’s scientists began testing it. They still do not know what the system does, let alone whether it could help treat disease.
Amodei made the case Wednesday while discussing Anthropic’s newly disclosed research. His forecast is not a claim that Claude found a cure: he described curing most diseases in five to 10 years as a goal he had set out before, and said it would require AI to help at every stage of the biomedical pipeline. The reported discovery is an early test of a narrower idea—that AI can help scientists find biological systems worth investigating.
The enzyme itself was not new. Earlier researchers had identified it in a virus that infects bacteria. Anthropic says Claude spotted something that had escaped notice nearby: an associated gene and a long array of evenly spaced DNA repeats. The company calls the resulting system array-associated reverse transcriptases, or ART. A reverse transcriptase is an enzyme that copies RNA into DNA.
Scientists gave Claude a broad prompt to search DNA-sequence data for interesting reverse transcriptases. Anthropic says roughly 950 agents spent 21 hours on the search, gathering more than 200,000 enzymes and narrowing thousands of candidate systems for closer examination. One agent flagged the repeat pattern; further analysis led to a report for human review. That division of labor matters: Claude identified a candidate system, rather than independently proving its function.
Human scientists carried out the lab experiments, with Claude helping propose what to test. Their initial work found that ART’s repeat array produces distinct short RNA molecules. Anthropic sees a possible resemblance to CRISPR, a naturally occurring system that has been turned into a gene-editing tool. But resemblance is not the same as a working editing mechanism. Anthropic says experiments to determine ART’s primary function are still underway.
a goal that sounds impossible, but one I believe is just barely possible if AI is applied to every stage of the pipeline.
The early result has drawn both interest and restraint. Stanford bioengineering associate professor Stanley Qi told Al Jazeera that Claude’s pursuit of a hard-to-spot pattern was exciting; he said AI could help researchers explore complex biological systems faster. Microbiologist Kevin Blake warned against a much bigger leap. A CRISPR-like pattern in nature does not show that ART rivals CRISPR as a technology, he said, or that it can become a practical tool or therapy.
Amodei’s argument is that faster fundamental discoveries could give researchers more promising drug targets, therapies and tools to investigate. He does not expect AI to shorten clinical trials. Any treatment emerging from this work would still need standard testing and regulatory review. For ART, even those downstream questions are premature: its biological role and possible use remain unestablished.
Anthropic’s research setup also draws a present-day boundary around the word “autonomous.” Its scientists supplied the initial direction, reviewed Claude’s work and performed the physical experiments. Amodei has suggested Claude might someday control lab equipment with safeguards, but said Anthropic is not doing that today. The next test of this finding is less sweeping than the disease-cure forecast: establishing what ART actually does.
anthropic.comLoading discussion...
Join the conversation
What part of the work earns that credit?
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.