Anthropic launches biology lab after AI identifies new enzyme system

Anthropic has established a life sciences research group and laboratory aimed at using Claude to investigate fundamental biology. The company says one of the group’s first programmes has identified a previously uncharacterized enzyme system linked to repeated DNA sequences and bearing some similarities to CRISPR.

The system, which Anthropic calls array-associated reverse transcriptases, or ARTs, has not yet been assigned a primary function. Its underlying reverse transcriptase was already known from research into a jumbo bacteriophage, but Anthropic says Claude was the first to connect that enzyme with the surrounding repeat array and an additional protein whose role remains unclear.

The work was reported by Anthropic, which has also released a preprint. The company describes the finding as an early result rather than a completed biological discovery, with laboratory studies continuing to determine what ARTs do.

From sequence searches to laboratory tests

Anthropic formed the research group in spring 2026 to examine whether general-purpose AI models could make the search for novel biological systems faster and more systematic. Its scientists gave Claude a broad task: examine a large database of DNA sequences for unusual examples of reverse transcriptases, enzymes that use RNA as a template to produce DNA.

According to Anthropic’s account, roughly 950 Claude agents worked for 21 hours and processed 210 million tokens. Together, they collected more than 200,000 reverse transcriptases, selected 3,500 possible systems for closer consideration and reduced those to 20 leading candidates. The company says that comparable analysis can take an expert researcher weeks or months.

The agents were not simply asked to retrieve matching sequences. They surveyed protein families, read relevant scientific literature, reproduced established findings from public data and looked for genes or nearby DNA that did not fit known descriptions. For each promising case, Claude produced a report proposing a possible function and setting out the evidence. Further rounds of criticism usually removed most candidates before any experimental work began.

In the ART case, Claude noticed a distinctive run of repeated DNA next to an unusual reverse transcriptase. It then examined the sequence in more detail, counted and measured the repeats and investigated the nearby genetic material. Anthropic’s scientists reviewed the result and tested the candidate in the laboratory.

Why unusual enzymes matter

Modern biotechnology has repeatedly benefited from biological features first noticed in unexpected places. Restriction enzymes, discovered in bacterial defence systems, cut DNA at particular sequences and became tools for assembling genetic material. Taq polymerase, taken from a heat-tolerant bacterium found in a Yellowstone hot spring, enabled the polymerase chain reaction, a technique used widely in diagnostics. CRISPR began as an odd pattern in bacterial DNA before becoming the basis for gene-editing treatments.

These examples help explain the significance Anthropic attaches to genome mining. Researchers can search the rapidly expanding catalogue of DNA for proteins and molecular arrangements that have not yet been understood, but the scale of the data makes manual inspection difficult. Reverse transcriptases provide a useful example: many newly identified families occur in bacteria, where they can form part of immune systems, and most have been found through computational analysis rather than conventional experiments.

ARTs share a combination of features seen in only a small number of other systems, according to Anthropic. Those systems include tools that can be programmed to cut, copy or paste DNA. The resemblance to CRISPR is therefore suggestive, but it does not establish that ARTs perform the same tasks or will become useful biotechnology.

A human-led AI research model

Anthropic says the project is designed around collaboration between scientists and AI agents. Claude conducts large-scale sequence searches, proposes candidates and helps interpret experimental results, while human researchers decide which possibilities merit testing and perform all laboratory work.

The company’s Bay Area facility operates at biosafety levels 1 and 2. It does not handle pathogens capable of infecting humans, and its experiments involve expressing candidate proteins in standard laboratory strains before biochemical and structural analysis.

Anthropic says its scientists use Claude Science and Claude Code, sometimes coordinating many sessions through an internal system. The volume of proposals has also created a second research question: what separates ideas worth testing from those that should be rejected? The team is studying those judgments and feeding the lessons back into the instructions given to Claude.

The group includes scientists whose earlier work covered CRISPR evolution and regulation, enzymes for cell and gene therapies, and computational methods for detecting potentially harmful DNA variants. It sits within Anthropic’s wider life sciences organisation, alongside teams working on drug discovery and on Claude’s biology and chemistry capabilities.

For now, the ART finding is best understood as a demonstration of a discovery process, not a new genetic technology. Its biological role remains unresolved, and Anthropic says the next stage is to establish whether the system has useful or programmable activity.