Scoring alignments by embedding vector similarity

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Sequence similarity is vital for understanding protein function and evolutionary relationships.
• Existing scoring matrices (e.g., PAM, BLOSUM) are context-independent, limiting their accuracy.
• Deep learning offers a way to create context-dependent representations.

Purpose of the Study:

• To develop a novel, context-dependent scoring method for amino acid similarity.
• To improve the accuracy of biological sequence alignment.
• To leverage deep learning embeddings for protein sequence analysis.

Main Methods:

• Utilized deep learning architectures with self-supervised learning on large unlabeled protein sequence datasets.
• Generated contextual embedding vectors for individual amino acid residues.
• Defined the E-score as the cosine similarity between residue embedding vectors.

Main Results:

• Alignments generated using the E-score method, particularly ProtT5-score, showed significant improvement over BLOSUM-based alignments.
• The new method demonstrated superior performance across various reference multiple sequence alignments.
• The E-score effectively captures context-dependent amino acid similarity.

Conclusions:

• The E-score offers a more accurate and context-aware approach to sequence similarity scoring.
• This method has the potential to revolutionize sequence alignment and related bioinformatics tasks.
• The developed tool is accessible via a web server and open-source code.