Anthropic announced that its AI model Claude Mythos 5 independently scanned 1,94 miliardy protein sequence clusters across twenty-one hours, discovering a previously unknown enzymatic system in bacteriophage DNA that resembles CRISPR.
Autonomous Discovery Across 1,94 miliardy Sequences
The California-based AI lab deployed nearly 950 autonomous Claude Code agents to comb through an immense genetic database. Running for 21 hours straight, the agents consumed over 210 milionů tokens to pinpoint novel reverse transcriptase systems. Human scientists provided only the initial prompt and subsequent work in the laboratory. Out of roughly 3 500 candidates, the system flagged twenty promising targets for deep analysis.
During the run, one oversight agent directed a peer to examine an unusual DNA region located directly adjacent to a group of reverse transcriptase genes. Reading the sequence straight as strings of Adenine, Cytosine, Guanine, and Thymine letters, the agent flagged the segment as spectacular. It spotted repeating blocks separated by consistent gaps. The system didn’t receive any instructions regarding repetitive sequences beforehand. It calculated the spacing independently, compared it against known literature, and compiled a report for human review.
Anatomy of the ART System
Researchers named the newly found architecture ART, short for array-associated reverse transcriptases. The system consists of three distinct components: the enzyme itself, an adjacent partner gene, and an array of 3 to 21 regularly spaced repeats. While scientists previously knew the enzyme from jumbo phages—extraordinarily large bacteriophages with extraordinarily long DNA—neither the flanking repeats nor the partner protein had been documented prior to this AI sweep.
Follow-ups directed by human researchers revealed that cells produce a suite of short RNA working copies from these repeat arrays. This matches how CRISPR operates, where repeat arrays transcribe into short RNA that function as navigation, directing the enzyme to specific genetic targets. Anthropic notes that ART could operate via a similar mechanism. Measurements from the Staphylococcus-targeting phage SA1 show these repeat transcripts spike during infection, accounting for up to 8 percent of all phage-derived RNA in the cell just 15 minutes after invasion.
Structural Distinctions and Experimental Verification
Despite functional parallels, critical structural gaps separate ART from classic CRISPR mechanisms. The surrounding DNA lacks cas genes, which bacteria typically use to build Cas proteins like the well-known molecular scissors Cas9. The spacers between ART's repeats span 120 to 220 letters of DNA, dwarfing the roughly 30-letter intervals found in CRISPR. Researchers haven’t yet proven whether the core enzyme is functionally active, though Anthropic points out that a handful of known systems with a similar combination of properties are programmable, allowing scientists to direct them to cut, copy, or paste genetic material.
Feng Zhang of the MIT and Broad Institute, a pioneer in genome editing, reviewed the findings prior to publication and called the discovery genuinely interesting and worthy of further study. At the same time, the technical report highlights the fragility of automated discovery. When researchers re-run the search ten times, agents spot the core enzyme family reliably, but miss the preceding repeat arrays entirely unless the sequence is fed explicitly into their context.
Anthropic pushed out the early finding to demonstrate Claude’s raw capability while inviting broader academic collaboration. The release follows the company’s confidential filing for an initial public offering with the US Securities and Exchange Commission in June. Rival labs have similarly highlighted AI-driven breakthroughs, such as OpenAI’s unreviewed solution to one of the seven millennium math problems. Whether ART proves as transformative as CRISPR or remains a biological anomaly depends entirely on ongoing laboratory experiments.
