AI Just Found a CRISPR-Like System Hiding in Virus DNA
Claude spent 21 hours combing through 200,000 enzymes and surfaced a CRISPR-like system that apparently nobody had recognized. The discovery is interesting. The method is the real story.
Almost every major breakthrough in biology started with someone noticing something odd. Restriction enzymes. The heat-loving enzyme that made PCR possible. CRISPR itself. Each one began as a strange pattern in nature's DNA that a researcher happened to stare at long enough.
Last month, the one doing the staring wasn't a researcher. It was an AI.
On September 23, Anthropic announced that Claude had identified a previously uncharacterized enzyme system hidden in the DNA of bacteriophages, the viruses that infect bacteria. They're calling it ART, short for array-associated reverse transcriptase. It pairs an enzyme with a partner gene and a long array of evenly spaced repeating DNA, an architecture that looks a lot like CRISPR.
How It Happened
The scale is what got my attention. Roughly 950 Claude agents spent 21 hours combing through genomic data. They pulled more than 200,000 reverse transcriptases, flagged about 3,500 candidate systems, and narrowed that list to 20 worth detailed reports. Anthropic says that kind of genome mining could take an expert scientist weeks or months.
One agent zeroed in on an unusual enzyme family, examined the raw DNA around it, and spotted a repeat array that apparently no one had recognized before. To be precise, the enzyme itself had shown up in earlier studies of a giant phage. What Claude appears to have been first to see was the full system: the enzyme, its neighbor gene, and the repeat array working together. It counted the repeats, measured their spacing, checked the scientific literature for prior descriptions, and flagged it for human review. Claude then proposed experiments to test the finding, and Anthropic's team ran them in its new Bay Area molecular biology lab.
The early lab result is the part worth watching. The repeat array is expressed as multiple short RNAs. CRISPR arrays do the same thing, and those RNAs are exactly what make CRISPR programmable.
The bottleneck in science was never curiosity. It was hours. AI is starting to remove the hours.
What It Is, and What It Isn't
I want to be precise about the limits, because Anthropic was. No one has shown that ART edits genes. No one yet knows what it does in nature. Dario Amodei said plainly that its function, usefulness, and significance aren't clear yet, and the work is a preprint that hasn't been through peer review.
There's another caveat worth knowing. According to one report, when Anthropic reran the same search 10 more times to test reproducibility, none of the reruns caught the repeat array. That doesn't make the finding any less real, since the lab work confirmed the array exists and produces RNAs. But it tells you the method isn't a slot machine that pays out every pull. This was one strong run, not a guaranteed repeatable result.
The most credible outside voice so far is Feng Zhang, one of the pioneers of CRISPR genome editing at MIT and the Broad Institute. He called it an exciting example of AI agents contributing to biological discovery and said the finding deserves further investigation. That's the right level of excitement: real, but earned in the lab, not in a press release.
Which is why I'd argue the discovery isn't the headline. The method is.
Why the Method Is the Real Story
For decades, scientific progress has been capped by a simple constraint: how many qualified people can look at how much data. Genomic databases have grown far faster than the number of biologists who can actually read them. Nature has been sitting on answers we never had the hours to go looking for.
Run the rough math. 950 agents for 21 hours is close to 20,000 agent-hours, which is about ten working years if you tried to staff it with people. It's not a perfect comparison, but it shows the shape of what changed. Humans picked the direction and ran the bench work. AI did the exhaustive reading, the pattern-hunting, and the hypothesis generation at a scale no lab could hire for.
From an investor's seat, that's the asymmetry I care about. If ART turns out to be nothing, the search itself cost about a day of compute, plus the lab work to test it. If it turns out to be the next programmable DNA tool, the payoff is measured in decades of therapies. And the same approach can be pointed at a different enzyme family, a different organism, or a different disease, even if no single run is guaranteed to find something.
What Comes Next
Anthropic says experiments are underway to figure out what the enzyme, its accessory protein, and those repeat-derived RNAs actually do. Amodei has floated the idea that Claude could eventually run experiments itself by controlling lab equipment, with safeguards in place. That's an aspiration, not a timeline, but it tells you where this is heading.
We used to measure AI by whether it could pass a test. Now we're measuring it by whether it can find something we didn't know was there. That's a very different scoreboard, and I think it's the one that matters.
"The next great discovery might already be sitting in a database. The question is who reads it first."
— Paul Gravette, CEO & Founder, Gravette Capital
AI & science disclaimer: This article is for general informational and educational purposes only. It summarizes public claims from Anthropic and press reports about a preprint that has not been peer reviewed, and it is not a technical or scientific audit of the findings. The function, utility, and significance of the ART system are unproven, and any potential applications in gene editing or medicine are speculative. Anthropic is a private company mentioned for context; this is not investment advice or a recommendation regarding any security. Forward-looking statements about AI capabilities and research workflows are uncertain and may change as new data emerge.