Claude AI discovers CRISPR-like gene-editing system: All you need to know

Anthropic says 950 Claude agents scanned more than 200,000 reverse transcriptases in 21 hours, identifying a novel system called ART that researchers later validated in a laboratory

Anthropic has claimed that its flagship artificial intelligence (AI) model Claude has autonomously discovered a novel enzyme system associated with an array of DNA repeats. The finding is notable because the arrangement of these repeats resembles the architecture of CRISPR, the gene-editing technology that has transformed modern medicine and biotechnology.

 

Anthropic published its findings as a pre-print on September 23. The research has not yet undergone peer review, and the company says the biological function of the newly identified system is still not known.

 

What did Claude discover?

 

The system has been named ART, short for array-associated reverse transcriptase. It is found primarily in bacteriophages, viruses that infect bacteria, and consists of three components: a reverse transcriptase (RT), a partner gene located beside it and a long array of evenly spaced DNA repeat sequences.

 

 

Reverse transcriptase is an enzyme, or a tiny biological machine, that reads RNA, a molecule carrying genetic instructions, and produces matching DNA. This is the reverse of the process cells usually follow. Some bacteria use reverse transcriptases as part of their basic immune defence against viruses.

 

The reverse transcriptase itself was not new. It had previously been identified in a jumbo phage. What Claude appears to have identified was the significance of the surrounding arrangement, particularly the repeat array and an additional accessory protein whose function remains unknown.

 

That repeat arrangement is what makes ART particularly interesting. Its structure is reminiscent of a CRISPR array, which contains a collection of RNA sequences and forms part of the system that makes CRISPR-Cas technology programmable. Anthropic's initial experiments found that the ART repeat array is transcribed into a series of distinct short RNAs. This raises the possibility that the array could play a role somewhat analogous to the RNA-containing arrays seen in CRISPR systems.

 

How did Claude find ART?

 

The discovery came from an AI-led search through huge quantities of genomic data. Researchers gave Claude a high-level instruction to search public sequence databases for interesting, uncharacterised reverse transcriptases, enzymes that convert RNA into DNA.

 

Anthropic instructed Claude to look through public DNA sequence databases for interesting and previously uncharacterised reverse transcriptases. Rather than relying on a single AI process, the company deployed a swarm of around 950 specialised Claude agents. Together, the agents examined more than 200,000 known reverse transcriptases over roughly 21 hours. The analysis involved about 210 million tokens, the units used to measure the amount of text an AI system processes.

 

From the broader search, the AI narrowed about 3,500 candidate systems to 20 that it considered particularly promising. It generated detailed analytical reports on those candidates for further examination. The process was designed to replicate a task that a biological researcher might undertake when looking for unusual relationships in genomic data, which is to identify something that does not fit existing patterns, investigate it further and check whether it has already been described.

 

During the analysis, one agent examining raw DNA sequences flagged an unusual pattern and subsequently recognised that it resembled a CRISPR-style repeat array. It counted the repeats, measured the gaps between them and searched the scientific literature to establish whether the pattern had already been identified. The analysis led Claude to a specific virus enzyme located next to a long array of evenly spaced, non-coding DNA repeats. Anthropic described this as a unique 'CRISPR-like' arrangement that had not been documented in scientific literature. It is to note that the full process, which would normally require human scientists to spend months on manual data analysis, was completed by the AI agents in under 22 hours.

 

Anthropic puts Claude's finding to test

 

Anthropic then moved the finding from computational analysis to physical testing. Its researchers synthesised the biological system in the company’s newly built molecular biology laboratory in the San Francisco Bay Area. Experiments there confirmed that the biological array produces distinct short RNAs, supporting Claude’s hypothesis.

 

Anthropic said the identification and physical laboratory validation of the system demonstrates that generative AI can act as an autonomous engine for biological exploration. Pioneering gene-editing researcher Feng Zhang of MIT and the Broad Institute praised the finding, describing it as an “exciting example of how AI agents can contribute to biological discovery”.

 

Anthropic has released a technical research paper detailing the discovery and is inviting academic and enterprise partners to work on determining ART’s exact primary function.