Vector-clustering Multiple Sequence Alignment

Vector-clustering Multiple Sequence Alignment applies protein language model-derived contextual embeddings to perform multiple sequence alignment of protein sequences, improving alignment accuracy for structurally similar proteins with low amino acid sequence identity ("twilight zone").

Key Features:

• Protein language model embeddings: Generates high-dimensional contextual embeddings for each amino acid using protein language models trained on massive protein sequence datasets.
• Clustering of contextual embeddings: Clusters embeddings to group semantically consistent residues or sequence regions across proteins.
• Ordering embeddings for alignment: Orders clustered embeddings to produce alignments of semantically consistent groups of proteins.
• Guide-tree and pairwise alignment avoidance: Circumvents construction of initial guide trees and intermediate pairwise alignments.
• Gap penalty and substitution matrix independence: Eliminates dependence on gap penalties and substitution matrices for scoring alignments.
• Captures structural and functional signals: Embeddings encode physicochemical, higher-order structural, and functional attributes beyond raw sequence similarity.
• Improved twilight-zone performance: Targets improved MSA accuracy for proteins with low amino acid identity but potential structural or functional similarity.

Scientific Applications:

• Aligning proteins in the twilight zone: Produces MSAs for protein sets with low amino acid sequence identity to reveal structural or functional relationships.
• Studying elusive sequence similarity: Enables analysis of protein groups where traditional sequence-based methods fail to detect relationships.
• Investigating protein evolution and function: Facilitates comparative studies to infer evolutionary relationships and functional signals when sequence similarity is weak.

Methodology:

Uses protein language models trained on massive protein sequence datasets to generate per-amino-acid contextual embeddings, then clusters and orders those embeddings to produce alignments while avoiding initial guide trees, intermediate pairwise alignments, gap penalties, and substitution matrices.