Anthropic’s September 23 enzyme discovery announcement now faces an attribution question: computational biologist Mario Rodríguez Mestre says his team had studied related enzymes for four years and shared key findings with Anthropic models, an account reported September 27. Anthropic says Claude agents helped identify the array-associated reverse transcriptase system it calls ART.
Mestre’s account raises an attribution question
Mestre says his team studied the enzymes and associated molecules for four years and shared key findings with Anthropic while regularly using its models for other work. His account puts earlier research alongside Anthropic’s description of Claude-assisted discovery.
What Anthropic says Claude did
Anthropic says it tasked Claude agents with searching DNA-sequence data for interesting reverse transcriptases. The agents flagged a repeat pattern beside a reverse-transcriptase gene and prepared a report for human review; Anthropic says human scientists performed the laboratory work.
Anthropic reports that about 950 Claude agents searched for 21 hours and used 210 million tokens. In the described process, agents gathered more than 200,000 reverse transcriptases, identified 3,500 candidate systems and selected 20 for detailed, human-readable analysis reports.
What ART is—and what its function is
Anthropic describes ARTs—array-associated reverse transcriptases—as a system made up of a reverse transcriptase, a neighboring partner gene and an array of evenly spaced DNA repeats. A reverse transcriptase is an enzyme that copies RNA into DNA. Anthropic says the systems are found mainly in bacteriophages, viruses that infect bacteria.
Anthropic reports that initial experiments detected distinct short RNAs produced from the repeat array. The company says ART’s biological function remains unknown.
The earlier MarsHill research
A 2021 study described the MarsHill gp206 reverse transcriptase and a 1,200-base-pair noncoding region upstream of it. The study’s authors hypothesized that the region contained a noncoding RNA associated with a retroelement. Anthropic says the reverse transcriptase itself had appeared in earlier studies.
The attribution question
Anthropic attributes the candidate search and pattern recognition to Claude agents, with human researchers reviewing the reports and carrying out the laboratory work. Mestre says his team had already studied related enzymes and shared key findings with Anthropic models. He also voiced concern that information from his unpublished research could be used to train future model versions.
