Anthropic AI unearths novel CRISPR-like system

In an impressive demonstration of computational biology capabilities, artificial intelligence has expanded its reach beyond data crunching directly into the realm of primary scientific discovery. Anthropic recently disclosed that its flagship AI architecture, Claude Mythos 5, successfully unveiled a previously unmapped enzymatic process concealed within the genetic framework of bacteriophages—viruses that infect bacteria. Marking the first validated breakthrough from the company’s newly launched molecular biology branch, the finding underscores the accelerating convergence between cutting-edge software engineering and life sciences following significant strategic investments made earlier in 2026.

Dubbed ART, an acronym for array-associated reverse transcriptases, the system operates on mechanisms centered around reverse transcriptase enzymes, which synthesize DNA directly from RNA templates. While microbial organisms regularly deploy these enzymes as protective shields to counter viral intrusions, the architecture of ART presents a unique structural layout. It pairs the primary enzyme with a neighboring partner gene alongside a distinct series of uniformly spaced repeating DNA units.

This structural design bears an uncanny resemblance to the widely studied CRISPR system, where repetitive arrays act as biological libraries storing RNA guides to direct site-specific genome targeting. According to the research team's preprint report, these newly identified ART arrays feature anywhere from 3 to 21 instances of a short, repeating sequence. However, a crucial distinction sets ART apart: it lacks cas genes, which traditionally serve as the functional molecular scissors in standard CRISPR pathways. Subsequent laboratory validation confirmed that these arrays transcribe into compact, individual RNA molecules. In tests using a staphylococcal phage, these specific RNA strands surged to account for nearly 8% of total viral RNA within just 15 minutes of initial infection.

The mechanism behind the discovery offers a compelling look at the power and quirks of modern autonomous agents. Researchers provided Claude with a straightforward prompt: hunt for uncharacterized reverse transcriptase systems across a massive dataset encompassing 1.9 billion protein sequences. Operating independently, a swarm of Claude Mythos 5 agents took charge of the entire workflow. One agent mapped out and executed granular subtasks, while a secondary agent acted as an auditor, scrutinizing outputs and opening new execution branches based on real-time findings.

Executing seamlessly for 21.5 hours without any human intervention, the automated process consumed resources equivalent to 949 agent sessions and 215.6 million tokens. The pipeline systematically pinpointed roughly 200,000 enzyme clusters, scored 3,564 families of candidate partner genes, and assembled 19 detailed briefs for human evaluation. Notably, the ART mechanism emerged from a serendipitous side path. While attempting to read raw DNA sequences adjacent to an unfamiliar enzyme, an agent noted in its execution logs that it could visually spot a repeating sequence. It then proceeded on its own initiative to tally the repetitions, compare them against known biological databases, review existing scientific literature, and compile its final findings.

Despite this remarkable outcome, the experiment shed light on persistent challenges surrounding the reliability and reproducibility of current large language models. When researchers attempted to replicate the exact search campaign ten consecutive times, the autonomous agents consistently failed to inspect the specific DNA regions adjacent to the enzyme, leaving the array completely undetected in every rerun. The study authors attribute this variance to the vastness of the search space and the inherent randomness of multi-agent dynamics. Under tightly controlled testing environments, the company's top four models accurately identified the array over 90% of the time when provided the DNA segment directly. However, success rates dropped sharply to 32% when the models were required to navigate external tools and files autonomously.

The facility enabling these discoveries reflects a broader shift toward AI-driven biological research. Situated in the San Francisco Bay Area, Anthropic’s laboratory operates strictly under Biosafety Levels 1 and 2, ensuring that research remains isolated from pathogens harmful to humans. This development follows the company’s $400 million acquisition of Coefficient Bio in April 2026, solidifying its intent to lead the next frontier of biological innovation.