ClusterEmbed: Efficient Protein Structure Prediction with Clustering and Embeddings

Protein structure prediction has been revolutionized by methods like AlphaFold2, which rely on large-scale multiple sequence alignments (MSAs) to achieve near-experimental accuracy. However, the computational cost and data demands of such approaches limit their accessibility, particularly for proteins with few homologs. To address this, we introduce ClusterEmbed, a novel, lightweight method for generating training data for protein structure prediction. ClusterEmbed combines rapid clustering-based MSA generation using MMSeqs2 with small-scale sequence sets and embedding extraction via pretrained protein language models, bypassing the need for extensive MSA datasets like those in OpenProteinSet. We evaluated ClusterEmbed across five experiments, testing variables such as sequence set size, clustering sensitivity, and embedding model type against metrics including RMSD, TM-score, GDT-TS, and computational efficiency.