Claude just discovered a brand-new CRISPR-like enzyme system

Most breakthrough tools in modern medicine started when a sharp-eyed scientist spotted an anomaly in nature. PCR came from an enzyme in a Yellowstone hot spring, while CRISPR began as an odd pattern of repeating DNA in bacterial immune systems.

Anthropic is now trying to automate that intuition. On September 23, the AI company announced its newly established life sciences research group and wet lab in the Bay Area, revealing an early discovery driven almost entirely by Claude: a novel enzyme system dubbed array-associated reverse transcriptases, or ART.

A 21-hour digital treasure hunt

To find it, Anthropic researchers simply gave Claude a broad prompt to search sequence databases for unusual reverse transcriptases, enzymes that copy RNA back into DNA. The model handled the heavy lifting.

Anthropic deployed roughly 950 Claude agents in parallel. Over 21 hours and 210 million tokens, the swarm evaluated more than 200,000 enzymes, narrowed the field to 3,500 candidates, and generated human-readable dossiers on the top 20 prospects. For a human specialist, this genomic mining would typically take weeks or months.

During the run, one Claude agent even displayed a remarkably human reaction while examining a raw genomic stretch, noting in its log that the flanking sequence was spectacular and asking if it was looking at a CRISPR-like repeat array.

CRISPR’s viral cousin

What Claude spotted was an overlooked architecture inside giant bacteriophages, viruses that prey on bacteria. While the underlying enzyme was already cataloged, scientists had missed the machinery sitting right beside it.

ART consists of three parts: the enzyme, an uncharacterized accessory protein, and an evenly spaced array of non-coding repeat sequences. In molecular biology, this repeating architecture is famous for directing programmable machines to cut, copy, or paste genetic material.

Anthropic scientists verified the prediction in their BSL-1 and BSL-2 facility, confirming that the array produces distinct small RNA molecules. MIT professor and CRISPR pioneer Feng Zhang reviewed the preprint, calling the identification of RNA-repeat arrays beside reverse transcriptases genuinely intriguing.

Anthropic CEO Dario Amodei noted the effort was mostly, though not entirely, Claude's work, since human scientists still executed the physical lab experiments. What the system actually does in nature remains an open mystery.

Why it matters

We think this marks a critical shift in AI's role in the sciences, moving past summarization and into generating novel empirical hypotheses.

For genomics researchers, public databases are overflowing with uncharacterized sequences that humans simply lack the bandwidth to inspect. Claude proved that a fleet of agents can filter hundreds of thousands of entries down to actionable lab candidates in less than a day, effectively handing scientists the needle from the haystack.

Across the AI sector, the battle lines are expanding into fundamental scientific discovery. Just days after OpenAI touted automated breakthroughs in advanced mathematics, Anthropic demonstrated that general-purpose models can systematically surface biological novelties.

Anthropic's hybrid setup looks clever. By letting AI handle data exploration while seasoned human biologists run targeted bench tests, the company avoids the brittle overhead of full laboratory automation.

Yet the bottleneck has simply moved downstream. AI can dream up high-confidence hypotheses in 21 hours, but validating biological mechanisms in living cells still takes months of tedious benchwork. Whether ART evolves into a versatile biotechnology tool or remains a genomic curiosity will ultimately be decided by biology, not tokens.