Anthropic's Claude Discovers Enzyme System With CRISPR-Like Repeats in Phage DNA

A swarm of roughly 950 Claude agents spent 21 hours scanning public DNA databases on September 23, 2026, and flagged a pattern in bacteriophage genomes that human researchers had previously overlooked: an orderly run of repeating DNA sequences sitting beside a known reverse transcriptase gene. Anthropic has named the system array-associated reverse transcriptases, or ART, and published its first scientific finding from its new Bay Area molecular biology laboratory in a blog post and an accompanying preprint. The preprint has not been peer reviewed.

What ART Is: Three Parts, One Unknown Function

ART is a three-component system found primarily in bacteriophages — the viruses that infect bacteria. Its first part is a reverse transcriptase (RT), an enzyme that copies RNA into DNA. Bacteria use diverse RT families as components of their immune systems; ART's RT had been identified in earlier studies of jumbo phages, bacteriophages with unusually large genomes. The second part is a neighboring partner gene of unknown function. The third is the element Claude actually spotted: a long, evenly spaced array of DNA repeats, ranging from 3 to 21 copies, arranged in a pattern that resembles a CRISPR array.

The structural resemblance to CRISPR is why researchers paid attention. In CRISPR-Cas systems, the repeat array holds a bank of distinct RNA sequences that make the system programmable — guide RNAs help Cas proteins locate and cut specific DNA targets. That programmability is what transformed CRISPR from a bacterial immune mechanism into the gene-editing technology now underlying approved medicines. Whether ART's repeat array serves an analogous role is exactly what Anthropic's lab has not yet determined.

How Claude Found What Human Researchers Missed

Anthropic's biology team gave Claude a single high-level prompt: search a massive database of DNA sequences — spanning roughly 1.9 billion protein clusters — for interesting new examples of reverse transcriptases. The agents worked autonomously from there. They gathered more than 200,000 RTs, narrowed those to 3,500 novel candidate systems, and produced detailed human-readable reports on the 20 most compelling candidates. The type of analysis that expert scientists estimate can take weeks to months of focused work ran in 21 hours across the parallel agent cluster, consuming approximately 210 million tokens.

The discovery itself came from a single agent that went further than its peers. While examining the raw DNA sequence surrounding an unusual RT family, the agent documented its own realization in its output log: "The DNA next to the RT is spectacular: I can see by eye a tandem repeat array … that's a CRISPR-like … repeat array?!" The agent then counted the repeats, measured their spacing, compared the layout against known RT systems, and searched the literature for any prior report of the pattern — before filing its report for human review.

That chain of behavior — noticing anomaly, characterizing it quantitatively, cross-referencing prior literature, and writing a structured scientific report — is what Anthropic is calling the proof-of-concept for agentic AI-assisted biological discovery. The underlying RT was not new; the contribution Claude claims is spotting the system's defining features: the associated non-coding repeat array and an additional accessory protein of unknown function.

What Wet Lab Experiments Have Confirmed So Far

Anthropic's Bay Area lab — a conventional BSL-1 and BSL-2 facility that does not handle pathogens capable of infecting humans — ran experiments to test whether the repeat array does anything measurable. Human scientists at the lab found that the ART array is expressed as a set of distinct short RNAs. In published genomic data from a Staphylococcus phage, those RNAs made up as much as 8% of total phage RNA fifteen minutes after infection, suggesting they accumulate rapidly during phage activity.

That laboratory observation matters because it elevates ART from a purely computational finding to one with a detectable biochemical signal. But the team has not yet shown that the RT enzyme is active, and it has not shown that the enzyme works on these RNAs. "We don't yet know what this system does," Anthropic writes in its blog post. Further experiments are underway. The preprint, titled "Autonomous AI agents discover reverse transcriptases with tandem repeat arrays," describes the computational search, the structural findings, and the early wet lab results — but stops well short of characterizing ART's biological role.

Feng Zhang, a CRISPR pioneer and professor at MIT and the Broad Institute who reviewed the preprint before publication, offered a measured assessment: "The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation." Zhang's phrasing is deliberate — "merits further investigation" is not a claim that ART will become a gene-editing tool.

The Reproducibility Problem That Didn't Make the Headlines

One detail Anthropic disclosed that most coverage downplayed: after the initial discovery, the company ran the identical agent campaign ten additional times. In none of the reruns did any agent read the DNA upstream of the enzyme, and all ten missed the repeat array entirely.

This finding cuts in two directions. It demonstrates that the discovery was real — the array exists in the data and in wet lab experiments. But it also reveals that the workflow that produced the discovery was not reliably systematic. The finding appears to have emerged from a combination of agent judgment and circumstance in a single run, not from a repeatable pipeline that would have surfaced it again. For researchers evaluating whether to adopt similar agentic workflows, the honest implication is that AI agents can surface unexpected patterns — but that the conditions under which they do so may not yet be fully understood or controllable.

Dario Amodei, writing on X on September 23, calibrated the claim carefully: "Its precise function, biotechnological utility (if any), or level of significance is not yet clear, but at minimum it is work I would have been proud to do as a PhD student." He also acknowledged that a team at Stanford independently described a novel RT system with an associated non-coding array — similar in some respects to ART, though Amodei characterized the two as distinct systems that evolved independently.

Anthropic's Biology Stack and Where It Sits Competitively

The ART announcement represents the first published scientific output from an Anthropic life sciences effort that has been assembling for roughly a year. In April 2026, Anthropic acquired Coefficient Bio, a stealth biotech startup of fewer than ten people, mostly former Genentech computational biologists, for approximately $400 million in stock. In June 2026, it launched Claude Science, a research workbench with connections to more than 60 scientific databases covering genomics, proteomics, structural biology, and cheminformatics. In August 2026, it previewed the Model Hardware Standard, a software interface that lets AI agents control laboratory instruments — liquid handlers, robotic arms, plate readers — through a standardized driver layer. The Bay Area wet lab, confirmed to Reuters by life sciences head Eric Kauderer-Abrams on September 18, is the physical execution layer connecting all of the above.

The combination — a domain-specific research platform, a hardware interface, and physical laboratory capacity — is a vertically integrated biology stack that no other frontier AI lab has fully assembled at the same time. OpenAI launched GPT-Rosalind, a biology-focused model, in April 2026, and Google offers Gemini for Science integrating AlphaFold and AlphaGenome. Neither has confirmed physical laboratory operations or published a preprint from an internally run biological discovery program. Isomorphic Labs, the DeepMind-founded AI drug discovery company that has the most advanced publicly disclosed program, is targeting its first clinical trials by the end of 2026 — a deadline already pushed back one year from its original timeline.

Read more: Anthropic's Bay Area Biology Lab Confirmed as Claude Moves Into Physical Drug Research

Biosecurity Context: How Anthropic Is Managing the Tension

The same AI capabilities that can find patterns in genomic data could be pointed at harmful biological questions. Anthropic has been transparent about the tension. According to its published safety reports, the company disclosed multiple biology-related incidents between late 2025 and mid-2026 in which users attempted to use Claude to access sensitive biological research through bypassed controls. No case produced a synthesized pathogen or real-world harm. The wet lab operates at BSL-1 and BSL-2 only and handles no pathogens capable of infecting humans — constraints that meaningfully limit the scope of experiments the lab can run. The biology team's focus on bacteriophages, combined with BSL-1/2 restrictions, reflects a deliberate decision to pursue discovery at arm's length from pathogens while biosecurity governance catches up to AI capability.

What Comes Next for ART and Agentic Biology

ART's placement in the broader landscape of reverse transcriptase biology is important for evaluating its potential. The past several years have seen a wave of RT family discoveries in bacteria and bacteriophages. Defense-associated reverse transcriptases (DRTs), diversity-generating retroelements (DGRs), and retrons — a bacterial immunity mechanism — have all been characterized recently using computational genome mining of the kind Claude performed. Each began as a sequence-level finding before wet lab work established what the system does. ART follows the same trajectory, at a much earlier stage.

The CRISPR comparison is instructive in scale as much as in structure. The progression from CRISPR's discovery as an odd repeat pattern in bacterial genomes to its use in approved medicines took roughly two decades and required connecting sequence observations to immunity, engineering the system as a programmable tool, and adapting it for human cells — contributions from multiple research groups across multiple continents. The ART announcement sits at the very beginning of an analogous process, with no guarantee the endpoint resembles CRISPR's.

Anthropic says it is now running further experiments to characterize what ART actually does. The questions that matter are whether the RT enzyme is catalytically active, whether it acts on the short RNAs the array produces, and whether the overall system performs anything analogous to the programmable operations that make CRISPR commercially and scientifically valuable. Each question requires additional wet lab work. The preprint describes early structural and expression findings; the functional story is months of experiments away at minimum.

The broader implication of the ART case is less about this specific enzyme and more about what the workflow demonstrates. A single agent — without human guidance between the initial prompt and the final report — read raw sequence data, identified a structural anomaly, connected it to prior literature, and generated a hypothesis that survived initial wet lab testing. The failure to replicate in ten subsequent runs is a methodological signal, not a dismissal of the result. It suggests that the conditions enabling this kind of discovery are not yet fully understood — and that improving the reliability of agentic genome mining is a tractable engineering problem worth pursuing. For the genomics and structural biology communities considering whether to add AI agents to their own genome-mining workflows, the ART preprint is now a concrete reference point. The peer-reviewed version, when it appears, will tell the scientific community how much weight the structural and expression evidence actually carries — and whether ART joins the small catalog of programmable biological systems or remains an interesting footnote to a demonstration of what AI agents can find when pointed at the right data.