Leveraging protein language models for accurate multiple sequence alignments

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Biology

Background:

• Multiple sequence alignment (MSA) is crucial for understanding protein sequence and function.
• Traditional MSA methods struggle with proteins exhibiting low sequence identity (the twilight zone).
• Protein language models offer a novel approach by generating contextual embeddings that capture amino acid properties.

Purpose of the Study:

• To develop a novel multiple sequence alignment (MSA) method leveraging protein language models.
• To improve alignment accuracy for proteins with low sequence identity.
• To circumvent limitations of traditional MSA algorithms.

Main Methods:

• Clustering and ordering of amino acid contextual embeddings derived from protein language models.
• Development of a novel MSA approach based on semantic consistency of protein groups.
• Avoidance of traditional MSA components like guide trees, pairwise alignments, gap penalties, and substitution matrices.

Main Results:

• The novel MSA method demonstrates higher accuracy for structurally similar proteins with low amino-acid similarity.
• The approach effectively utilizes information from contextual embeddings.
• Successful alignment of protein groups based on semantic consistency.

Conclusions:

• Protein language models provide a powerful new source of information for MSA.
• The proposed method offers a more accurate alternative for aligning challenging protein sets.
• This approach is anticipated to be a fundamental component of future MSA algorithms.