Protein embedding based alignment

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Sequence Analysis

Background:

• Traditional protein sequence alignment methods struggle with sequences below 35% identity.
• Substitution matrices, developed in the 1970s, are suboptimal for scoring alignments in the low-identity 'twilight zone'.

Purpose of the Study:

• To develop a novel algorithm, Protein Embedding based Alignments (PEbA), for improved protein sequence alignment.
• To leverage protein language models for enhanced alignment accuracy in the low-identity twilight zone.

Main Methods:

• PEbA utilizes a dynamic programming approach similar to Smith-Waterman.
• Amino acid matching scores in PEbA are derived from the similarity of embeddings generated by protein language models.
• The algorithm was evaluated on over 12,000 benchmark pairwise alignments from BAliBASE.

Main Results:

• PEbA significantly outperformed traditional BLOSUM substitution matrix-based alignments, especially for sequences with <10% identity (over fourfold improvement).
• Comparisons with different protein language models showed ProtT5-XL-U50 embeddings yielded the best alignment performance.
• PEbA demonstrated superior performance compared to other recent embedding-based alignment methods like DEDAL and vcMSA.

Conclusions:

• General-purpose protein language models provide valuable contextual information for protein sequence alignment.
• PEbA offers a more accurate alignment method than traditional approaches, particularly for divergent protein sequences.