Beyond profiles: supervised repeat annotation using protein embeddings

Protein repeats give rise to diverse protein structures and functions, yet their sequence-based detection is often challenging because individual repeat units can diverge substantially in sequence while retaining similar structures. Most sensitive approaches rely on sequence information directly represented by alignments or derived profiles, thereby limiting flexibility and scalability. Here, we introduce TREAD (Transfer learning-based REpeat Annotation using Protein EmbeDdings), a supervised framework that reformulates repeat detection as a residue-level annotation problem and operates directly on embeddings from protein language models. Instead of constructing explicit probabilistic profiles, TREAD learns repeat-specific features implicitly, enabling residue-resolved scoring and flexible repeat segment extraction. Across complementary benchmarks based on RepeatsDB and Pfam, TREAD consistently matches or outperforms the widely used profile-based tool HMMER, particularly in low-data and high-divergence settings. The model exhibits robustness to score thresholding and demonstrates better generalization to independent test sets, including those containing remotely related families. To illustrate the practical utility of TREAD, we applied it to survey {beta}-propeller proteins across the AlphaFold Database and representative proteomes to generate a comprehensive census of this fold. This analysis highlights extensive propeller diversity, identifies lineage-specific expansion patterns across the tree of life, and suggests previously unrecognized relationships between propellers and other repeat folds. Together, TREAD provides a flexible and scalable alternative to profile-based repeat annotation and establishes a general motif-centric framework for protein sequence annotation.