Exploiting protein language model sequence representations for repeat detection

Abstract

Duplication is an essential evolutionary mechanism that operates at the scale of chromosomes, large chunks of DNA sequences, genes, protein domains, and shorter motifs. The study of duplication is central to understanding protein evolution, but the detection of repetitive sequence patterns is often challenging due to decreasing similarity between internal repeats resulting from long-term divergence. The most sensitive sequence-based repeat detection method, HHrepID, relies on the construction of multiple sequence alignments (MSAs) to enhance homology signals and thus facilitate the detection of very ancient duplications. However, such an alignment-based approach is slow and limits the ability to perform large-scale scans. Recent advances in protein representation learning have introduced sequence embeddings extracted from protein language models as a powerful and much faster alternative to MSAs. Protein sequence representations have been shown to be effective in homology detection, as exemplified by software such as our recently developed pLM-BLAST. In this study, we implement pLM-Repeat, a pipeline built upon pLM-BLAST, to identify repeats encoded in sequence embeddings. pLM-Repeat achieves comparable sensitivity to HHrepID in detecting the presence of repeats, while predicting many more repeat units and providing significantly better run times. We also trained an auxiliary neural network, DeepRepeat, to detect domains with patterns similar to well-characterized repeat folds to support rapid filtering. Using our newly developed tools, we scanned the AFDB90v4 database and identified a collection of novel and undescribed repeat domains.