Has Claude Found the Next CRISPR? Not Yet

Has Claude found the next CRISPR? Anthropic says its AI agents spotted a CRISPR-like enzyme system in bacteriophage DNA, but nobody yet knows what it does. Pharma R&D teams should read the finding as a promising lead, not a proven platform.

Image

What Claude Actually Found in Phage DNA

On 23 September 2026, Anthropic reported that Claude agents had found an uncharacterised enzyme system in bacteriophages, the viruses that infect bacteria.

Anthropic calls it array-associated reverse transcriptases, or ART. It released the work as a preprint, so independent reviewers have not yet tested it.

Some background helps though.

CRISPR arrays store short genetic snippets that act as a guide library. Cas enzymes use those guides to find matching DNA and incise it. That pairing of a guide bank with an enzyme made CRISPR programmable, and programmability made it a tool.

ART has an array and an enzyme, but its enzyme copies RNA into DNA rather than cutting it. That difference is not small.

How the search works

A scientist gave Claude one high-level brief to find unusual reverse transcriptases, which are enzymes that copy RNA into DNA.

The agents gathered more than 200,000 of them and flagged 3,500 possible systems. They then narrowed the list to 20 and wrote reports for human review.

Anthropic says an expert could need weeks or months for that analysis.

The campaign ran for about 21.5 hours across 949 agent sessions and consumed 215.6 million tokens, according to the preprint. Human scientists did all of the bench work.

The three parts of ART

• A reverse transcriptase: Earlier work on a jumbo phage had already identified this enzyme.
• A partner gene: A neighbouring gene whose protein has no known role.
• A repeat array: A long, evenly spaced run of three to 21 copies of a short DNA sequence.

Image

Claude appears to be the first to link these parts into one system. That link is the real novelty.

Early lab tests show the array produces short, distinct RNAs. Yet the research team has not shown that the enzyme is active or that it acts on those RNAs.

Why hunt in phages? These viruses carry a vast, poorly explored reservoir of genes shaped by a long arms race with bacteria. That history has already yielded compact editors such as CasΦ, so phage DNA is a rational place to search. The strategy is sound. The open question is what this particular find delivers.

View interview How Should Pharma Companies Audit AI Vendors?

How Should Pharma Companies Audit AI Vendors?

“This is one of the riskiest industries there is because you're dealing with human lives, you're dealing with experimental protocols, and you're dealing with regulatory bodies where you might not get another shot at that clinical trial."

Image

Why the Next CRISPR Claim Runs Ahead of the Evidence

Spotting a pattern is not the same as explaining it.

Scientists first noticed CRISPR as an odd repeat pattern in bacterial DNA. Years of experiments then revealed what those repeats do.

Repeat arrays beside enzymes are a familiar sight in genomics, and genome mining across microbial and phage DNA is a mature field. A strange gene cluster is a useful starting point. It is only the first step.

Two well-known cases show what the harder step looks like.

A 2024 Nature paper showed that bridge RNAs guide a recombinase, and that two RNA loops can be reprogrammed to insert, excise, and invert DNA.

Additionally, a 2020 Science study described CasΦ, a single compact protein paired with a CRISPR array in huge phages, which worked in human and plant cells.

Both teams explained how their systems function. Both also showed that users can steer them. Anthropic has not reached that stage with ART.

The gap between explaining pattern and tool can be long.

Decades passed between the first sightings of CRISPR repeats and the editing tools that labs and clinics use today. Each step needed experiments that ruled out plausible but wrong explanations.

AI can speed the first stage. It cannot skip the later ones.

Image

Related biology already exists

Reverse transcriptases that work with non-coding RNA are not new.

A study of the DRT2 defence system found a reverse transcriptase bound to a neighbouring RNA. After phage infection, it copies part of that RNA into a tandem repeat array. The array then yields a toxic protein that halts the infection. That mechanism looks nothing like CRISPR editing.

The lesson is simple. Similar-looking genetic layouts can hide very different biology. A repeat array alone says little about purpose.

ART could turn out to be a defence system, a regulator, or a new editing tool. The current data cannot choose between those options.

What Expert Reaction Means for Pharma Decision-Makers

Reaction has been split.

Anthropic quotes Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute, who called the link intriguing and worth pursuing. He also praised the work as an example of AI agents contributing to discovery.

However, microbiologists quoted by Gizmodo took a firmer line. They said ART resembles CRISPR in layout, with no evidence yet of a shared function.

Reproducibility deserves scrutiny

The preprint reports that 10 reruns of the original task all missed the array, because none read the DNA upstream of the enzyme.

When given that DNA directly, Anthropic’s four most capable models described the array in at least 90% of attempts. The models read the pattern well once pointed at it. They did not reliably decide where to look.

Read the announcement as both science and strategy

Anthropic’s Life Sciences Research Group sits beside teams that work on drug discovery, and the company invites outside scientists to submit research proposals.

The announcement is therefore a science update and a business signal. Pharmatica’s earlier look at Anthropic’s AI biology lab explains that wider context.

Framing shapes expectations

Headlines set expectations. When an early finding arrives framed as a major breakthrough, partners, investors, and the public recalibrate at once.

If the science then moves slowly, trust erodes.

Careful framing protects credibility for the stronger results that follow. Anthropic did state plainly that it does not know what ART does, so much of the debate concerns emphasis.

Where ART could matter if it works

Suppose later experiments show that ART is programmable.

Novel enzymes can widen the toolbox for gene and cell therapy developers, especially if they are compact or work in ways existing editors cannot. Intellectual property and licensing would then become live issues.

Even in that best case, years of characterisation, safety work, and delivery optimisation would sit between the enzyme and any therapy. Leaders should hold both facts at once: real upside, and a long road.

Five questions to ask before acting

• Function: Does ART edit, defend, regulate, or do something else?
• Activity: Has anyone shown the enzyme works on its RNAs?
• Replication: Do independent labs confirm the array and its expression?
• Programmability: Can users direct the system to chosen targets?
• Peer review: Has a journal tested the claims?

Until those answers arrive, treat ART as a research lead. Do not build a pipeline thesis on it.

Image

Practical fit matters too. The CasΦ team stressed a protein half the size of Cas9 because compact editors ease delivery into cells. Any ART-derived tool would face the same test.

The bigger signal is speed. If agents can screen 200,000 enzymes in under a day, discovery teams will receive candidate lists faster.

Human validation remains the slow and costly step. Pharma should plan resources for that gap, and Pharmatica’s AI in drug discovery coverage tracks how others are doing so.

Business development teams face a similar choice. AI-discovery announcements will keep arriving from frontier labs, and they will vary widely in maturity.

A simple screen helps. Separate the pattern found, the function shown, and the result replicated. ART currently clears only the first of those three.

View interview AI in Pharma: Why the Future of Healthcare Starts With Patients, Not Tech

AI in Pharma: Why the Future of Healthcare Starts With Patients, Not Tech

Kate O’Reilly, President & Chair of the Healthcare Businesswomen’s Association (HBA) Dublin-Ireland Chapter & Healthcare Transformation Partner at Roche, discusses AI in pharma, patient engagement, and the future of healthcare innovation.

Image

Next CRISPR or Next Hypothesis? The Pharma Verdict

Claude found something real in an unusual three-part system that earlier work had missed. But, it has not found a new gene editor.

ART has a CRISPR-like layout, an unknown function, and no peer review. The comparison stays premature until labs show what the enzyme, the partner protein, and the RNAs do.

This current research also shows where AI adds value today.

Broad, fast screening that surfaces overlooked patterns. It does not replace the experiments that turn a pattern into knowledge.

Anthropic frames the result as early and asks outside scientists for proposals, which gives the community a clear route to test the claims.

For R&D teams, they should track ART as a lead, not a platform. Watch for function data, replication, and peer review.

Additionally, ask AI-discovery partners how they validate hits. Keep budget for wet-lab confirmation, because the Pharmatica Insights library shows that validation decides which discoveries reach patients.

At Pharmatica, we separate genuine breakthroughs from early signals. We track the science, the sources, and the strategic stakes behind each AI announcement, so pharmaceutical leaders can act with confidence when evidence, not hype, points the way.

Pharmatica: Insight. Connection. Impact.