AI Agents Spotted a New Reverse Transcriptase Family, ART, in Jumbo Phages (Preprint Explained)

This video explains a preprint from Anthropic, "Autonomous AI agents discover reverse transcriptases with tandem repeat arrays" (Yoon, Athukoralage, Ameisen, Kauderer-Abrams, Perry, Durrant). The authors had Claude agents survey reverse transcriptases (RTs) across 1.9 billion metagenomic protein clusters. One agent noticed an unannotated array of tandem repeats beside a jumbo-phage RT. That became ART, array-associated reverse transcriptases: a repeat array of about 200-nucleotide units, an RT with a long N-terminal extension, and a dedicated partner gene. The video asks one question: can an agent notice what a pipeline was never told to look for? It follows the biology first (retrons, the ART locus, how it differs from a CRISPR array, the expressed array RNAs, three unrelated partner proteins, and the authors' hypothesis), then how the agents found it and how the authors tested why. What the paper does and does not show. The function of ART is untested: the retron-like model is a hypothesis built on expression data and structure predictions. The discovery came from one session, and none of ten reruns found it again. The benchmark asks a model to describe a known finding, and no comparison with a conventional pipeline or a human expert is reported. The authors work for the company that built the models, one of the tested models also served as the judge, and the paper is a preprint that has not been peer reviewed. Traditional Chinese subtitles are available (turn on CC). Chinese terms follow the National Academy for Educational Research (NAER) terminology database; terms without an NAER entry are kept in English or described. CHAPTERS 0:00 Hook: an enzyme next to a strange repeat array 0:40 Reverse transcriptases and retrons 1:41 Too many sequences to read 2:44 The agent campaign 4:19 The detour: an anomaly upstream 6:27 ART: a new family of reverse transcriptases 8:04 Not a CRISPR array 9:02 The array is expressed 10:05 Three unrelated partner proteins 11:13 A hypothesis: a retron with a bank of RNAs 12:13 Why did the AI see it? The benchmark 14:01 Repeat signals inside the model 15:16 Appraisal 17:00 Recap SOURCES Yoon PH, Athukoralage JS, Ameisen E, Kauderer-Abrams E, Perry NT, Durrant MG (Anthropic). Autonomous AI agents discover reverse transcriptases with tandem repeat arrays. Preprint, not peer reviewed, 2026. https://www.anthropic.com/news/claude... Structure shown: Cryo-EM structure of E. coli retron-Ec86 in complex with its effector. RCSB PDB 7V9X: https://www.rcsb.org/structure/7V9X