AI models have read every paper, summarized every forum thread, and answered every homework question. Now they're doing actual science.
This week Anthropic announced something genuinely startling: its Claude AI autonomously discovered a brand-new enzyme system hidden inside viral DNA — one that no human researcher had ever spotted. The find looks structurally like CRISPR, the gene-editing breakthrough that won a Nobel Prize. Anthropic is calling the new system ART, short for array-associated reverse transcriptases, and it published the results as a preprint on September 23.
Here's how it happened. Earlier this year Anthropic quietly built a life sciences research group — including a small wet lab in the Bay Area — to test a bold idea: that AI could systematize scientific discovery, not just answer questions about it. For this first big test, the team pointed roughly 950 Claude agent sessions at a huge database of reverse transcriptase enzyme sequences and gave them a broad instruction to look for something unusual.
About 21 hours and 210 million tokens later, the search had narrowed more than 200,000 sequences down to 3,500 candidates, then 20 genuinely compelling reports, then one standout. One agent flagged a long, evenly spaced run of repeating DNA sequences next to an unusual reverse transcriptase gene — a pattern that looked architecturally like a CRISPR array. As one of the agents reportedly wrote, in language closer to a lab notebook than a chatbot log: "I can see by eye a tandem repeat array... that's a CRISPR-like... repeat array?!"
What is ART, exactly? It has three parts: a reverse transcriptase enzyme (which copies RNA back into DNA), a neighboring partner gene whose function is currently unknown, and that long run of evenly spaced DNA repeats. It lives in bacteriophages — viruses that infect bacteria, not humans. Early lab experiments, performed entirely by human scientists, confirmed the system is real and showed the repeat array producing short RNAs — exactly what you'd expect if the system uses a CRISPR-like guide mechanism. Only a handful of known systems share this feature set, and all of them are programmable tools that can cut, copy, and paste DNA.
The reactions are what make this hard to dismiss as hype. CRISPR pioneer Feng Zhang, whose lab helped turn CRISPR into a gene-editing tool, reviewed the preprint and called the discovery "genuinely intriguing," saying it "merits further investigation." Anthropic CEO Dario Amodei suggested on X that ART "may constitute a new gene editing mechanism that could have applications in gene therapy" — though he himself called the finding "preliminary."
And the honest caveats matter here: nobody knows what ART actually does yet. It hasn't been peer-reviewed. There is no evidence it edits genes, and at least one microbiologist is openly skeptical it could ever rival CRISPR. Anthropic says all its lab work runs at biosafety levels 1 and 2, with no human-infecting pathogens, and it has invited outside scientists to propose research into the system.
The pattern is what makes this a bigger story than one enzyme. Restriction enzymes, Taq polymerase, CRISPR — biology's greatest molecular tools all began with someone noticing something odd in natural DNA. Now the noticing can be done by AI. Claude spotted in 21 hours something that every human who ever looked at this data had missed.
That doesn't mean scientists are being replaced — the lab experiments, the judgment about what it all means, the skepticism that keeps science honest: all human. But the discovery funnel just gained a very fast new member. If this works once, labs will start pointing agents at every corner of biology where there's too much data and too few eyes. The next discovery might take 21 hours too.
Comments
Post a Comment