Discovery

Anthropic uses autonomous AI agents to uncover novel bacteriophage enzyme system

Anthropic uses autonomous AI agents to uncover novel bacteriophage enzyme system

Researchers at San Francisco-based Anthropic have used autonomous AI agents to identify a previously uncharacterized enzyme system in bacteriophages, according to a preprint posted September 23, 2026. The computational search was conducted autonomously after a high-level prompt, with human scientists handling laboratory validation. The system, designated array-associated reverse transcriptases (ART), was identified from roughly 1.94 billion protein clusters, completed in 21.5 hours of wall-clock time by approximately 950 autonomous agents consuming 210 million tokens. The reverse transcriptase itself had been observed previously; the novel finding was its association with a distinctive repeat array and downstream partner protein, defining a previously uncharacterized system.

A general-purpose large language model (LLM) agent system autonomously detected a structural anomaly — a CRISPR-like repeat array adjacent to an unusual reverse transcriptase (RT) gene — that fell outside predefined pipeline features, then flagged it for laboratory follow-up. The agents gathered nearly 200,000 RT sequence clusters, narrowed these to 3,564 candidate partner families across 10,983 loci, promoted 17 for deeper investigation, and identified three previously unreported RT associations, with ART emerging as the principal finding.

ART loci consist of three co-occurring components: a repeat array of 3–21 copies of a 15–49 nucleotide repeat with a palindromic core separated by unique spacers, an RT gene with an unusually long N-terminal domain, and a partner protein of unknown function encoded directly downstream. The architecture has some superficial similarity to other reverse transcriptase-associated defense systems such as retrons, but ART’s biological function and the role of its array-derived RNAs remain unknown.

Laboratory validation in Escherichia coli confirmed that the ART locus is transcribed and produces discrete array-derived RNA species with reproducible boundaries. Separately, reanalysis of previously published RNA-seq data from Staphylococcus lentus infected with phage SA1 showed that ART array-derived transcripts accounted for up to 8% of phage RNA at 15 minutes post-infection.

Critical aspects of ART biology remain uncharacterized: whether the RT is enzymatically active has not been demonstrated, whether array-derived RNAs serve as RT substrates is unconfirmed, the effector function of the partner protein is unknown, and whether the system benefits the phage or the bacterium is unclear. Anthropic said work to understand ART's primary function is ongoing.

The AllSci BriefFree, systematic R&D and deal news. Daily.

In benchmarking across approximately 3,500 attempts using seven Claude model versions, Mythos 5 performed competitively with or better than purpose-built genomic language models Evo 2 and gLM2 on selected repeat-detection tasks, despite not being trained specifically on genomic data.

The autonomous discovery approach places the work alongside recently published agentic biology platforms, including FutureHouse's Robin system, published in Nature in May 2026, and the Virtual Biotech framework published in Science in September 2026, both of which apply multi-agent architectures to therapeutic candidate discovery rather than enzyme sequence-space exploration from metagenomics.

Feng Zhang, a professor at MIT and the Broad Institute and a pioneer of CRISPR genome editing, said after reviewing the preprint: "The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation."


Access the AllSci platform to explore the science behind the news.


Spot something wrong? Report an issue with this article