On September 23 local time, Anthropic announced the establishment of a life science research team and its own molecular biology laboratory, and revealed an early achievement in biological research involving Claude: without any specific instructions from human scientists, Claude independently discovered a previously uncharacterized enzyme system named Array-associated Reverse Transcriptases (ART). This discovery came from its San Francisco Bay Area laboratory and is the first public result from the life science team since its formation earlier this year. The preprint paper has been released on the same day and is awaiting peer review.
950 Agents Working in Parallel, Identifying a New Enzyme System in 21 Hours
According to Anthropic, the research team instructed Claude to search for new reverse transcriptases in a large DNA sequence database. Human scientists were only involved in the initial prompt design and subsequent laboratory validation, while the rest of the work was completed by approximately 950 Claude agents working in parallel. These agents consumed about 210 million tokens in 21 hours, screening out around 3,500 new candidate systems from over 200,000 reverse transcriptases, narrowing it down to 20 most promising candidates, and generating human-readable analysis reports for each one.
Ultimately, one of the agents noticed an unusual reverse transcriptase gene adjacent to a series of tandem repeats while processing a bacteriophage DNA sequence. The agent then calculated the length and spacing of the repeat sequences, compared the structure of known reverse transcriptase systems, and checked the literature to confirm if there had been any prior reports. Anthropic describes ART as consisting of three parts: a reverse transcriptase, a neighboring partner gene, and a long series of uniformly spaced non-coding DNA repeat sequences. This repeat array is structurally similar to CRISPR arrays, and the paper shows that it contains 3 to 21 short repeated copies. Preliminary experiments suggest that it can be expressed as a set of different short RNAs with biological activity. However, the exact function of ART remains unclear, and a significant amount of follow-up experiments are needed to verify whether it can be stably and controllably used for gene editing.
Zhang Feng Praises, Amodei Acknowledges the Need for Further Verification
Zhang Feng, a pioneer in CRISPR gene editing and professor at MIT and the Broad Institute, reviewed the preprint and stated that this is an exciting case of AI agents driving biological discoveries. He found the identification of RNA repeat arrays "extremely intriguing" and said it is "completely worth further in-depth research," and hoped to encourage more scientists to explore how AI can support research. Dario Amodei, CEO of Anthropic, acknowledged on X that this discovery builds upon previous work—specifically, a team from Stanford had described a system that is "similar in some aspects." He also emphasized that "most but not all" of the discovery was made by Claude, and that the system's exact function and biotechnology practicality remain "unclear" at this stage.
Anthropic stated that the laboratory only handles BSL-1/BSL-2 research and does not involve pathogens that can infect humans. All wet experiments are conducted by human scientists, while Claude is responsible for data mining, hypothesis generation, and data analysis. The company believes that if such genome mining is scaled up, it could compress the analysis typically requiring weeks or months into just a few hours, and is willing to collaborate with scientists to expand the method to genomics and other fields.