Claude agents uncover unfamiliar DNA enzyme system
Anthropic said on 23 September that Claude agents found an uncharacterized enzyme system in bacteriophage DNA, and human scientists at its new biology laboratory followed up with experiments. Roughly 950 agents spent 21 hours and 210 million tokens searching and analyzing sequence data before one noticed a repeating DNA pattern near a reverse-transcriptase gene. The team named the candidate array-associated reverse transcriptases, or ART, and published a preprint. The pattern resembles one feature of CRISPR systems, but the team has not established ART's biological function or a gene-editing use. The result is an early, concrete test of whether a general AI model can contribute a useful hypothesis to laboratory science, rather than merely summarize existing papers.
Anthropic says the researchers supplied a high-level prompt and handled laboratory work, while Claude agents searched a large sequence database, investigated families of reverse transcriptases and proposed candidates. The agents gathered more than 200,000 enzymes, identified 3,500 possible new systems and narrowed them to 20 for detailed reports. One agent noticed a tandem repeat array beside a gene that had been studied in another context, a combination the team says had escaped earlier description. Scientists then analyzed and tested the candidate in their lab. The report is specific about the division of labor: Claude generated and refined the lead; human researchers validated aspects of the finding and continue to determine what the system actually does.
Feng Zhang, an MIT and Broad Institute scientist associated with CRISPR research, called the repeat arrays ‘genuinely intriguing’ after reviewing the preprint and encouraged further investigation. His reaction supports the scientific interest of the observation without certifying a practical application. Anthropic's lab describes working with biosafety level 1 and 2 material and says humans perform all physical experiments. Its broader life-sciences group works on model training for biology and chemistry and on drug discovery. The company used Claude Science and Claude Code, along with a custom harness for parallel sessions, in the research workflow. The finding gives it a demonstrable example for customers who want AI to search broad scientific spaces and hand back testable leads.
The claim should be read at the level of discovery actually made. A repeated sequence near a reverse transcriptase and a neighboring accessory protein define a previously uncharacterized system; they do not yet reveal the enzyme's natural role, programmability or therapeutic value. Anthropic itself says the function is still under study. The 210 million token figure is a measure of the search effort, not a cost comparison with an equivalent human research campaign. The team argues that a scientist might spend weeks or months on a similar family-level survey, but a useful economic comparison would also count laboratory validation and the rate of false leads across many campaigns. The preprint creates an avenue for outside scientists to test the result.
Anthropic's move into a wet lab changes the evidence it can offer for Claude's research value. Instead of relying on model benchmarks or hypothetical biological tasks, it can couple computational search with experiments and report a tangible candidate. The announcement came as the company expanded a Life Sciences Verification Program for qualified research users and courted enterprise customers in specialized fields. The new ART system may eventually prove useful or remain an interesting biological anomaly; either outcome will help calibrate the method. The more durable asset for Anthropic could be a repeatable pipeline in which Claude proposes unusual, well-documented candidates and scientists spend their scarce bench time on the best of them.
Analysis
The commercial mechanism is a reduction in the search cost per credible scientific lead, not revenue from ART itself. Anthropic spent 210 million tokens and substantial parallel orchestration to produce a candidate that still required human experiments; a lab buying this workflow would compare that total cost with expert time saved and the number of leads that survive validation. If Claude can repeatedly narrow hundreds of thousands of sequences to a manageable bench-testing set, Anthropic can sell high-value usage and deepen life-sciences adoption. A single discovery cannot establish the hit rate, but the published trail from sequence search to experimental follow-up is more informative than a generic claim that AI accelerates science.