Beyond profiles: supervised repeat annotation using protein embeddings

TREAD, a supervised repeat-annotation framework using protein language-model embeddings, matches or outperforms HMMER on RepeatsDB and Pfam, especially in low-data and high-divergence settings.

• Why it matters: Repeat units can diverge in sequence while preserving structure, making alignment/profile-based detection inflexible and less scalable.
• What they did: TREAD reformulates repeat detection as residue-level annotation and learns repeat-specific features directly from protein embeddings, avoiding explicit probabilistic profile construction.
• The result: It yields residue-resolved scoring with robust thresholding and better generalization to independent test sets, and it enabled a census of {beta}-propeller proteins in the AlphaFold Database and representative proteomes.