Abstract
Duplication is an essential mechanism of molecular evolution, which operates across biological scales, from whole genomes to single basepairs. Its study is central to understanding protein evolution, but the detection of duplication events often becomes challenging over evolutionary time, due to the accumulating sequence divergence. The most sensitive sequence-based protein repeat detection method, HHrepID, relies on the construction of multiple sequence alignments (MSAs) to enhance statistical signals of internal similarity and thus facilitate the detection of ancient duplications. However, such an alignment-based approach comes at the expense of speed, severely limiting its applicability to large-scale scans. Recent advances in protein representation learning have introduced sequence embeddings extracted from protein language models (pLMs) as a powerful and faster alternative to MSAs. Such representations have been shown to be effective in detecting distant sequence similarity, as exemplified by the pLM-BLAST software developed in our group. In this study, we describe pLM-Repeat, a pipeline built on top of pLM-BLAST to identify repeat patterns encoded in sequence representations. pLM-Repeat achieves comparable sensitivity to HHrepID in detecting the presence of repeats, while identifying many more repeat units and providing shorter runtimes, allowing us to detect novel repeat proteins in the AlphaFold Protein Structure Database with the aid of a pre-filtering model trained on repeat protein representations. pLM-Repeat is available as an open-source tool at https://github.com/KYQiu21/plmrepeat.