Protein embeddings and local alignments.

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Sequence Analysis

Background:

• Protein embeddings provide contextual representations crucial for bioinformatics, complementing traditional sequence alignments.
• While embedding-based improvements exist for global alignments, local alignment optimization remains underexplored.
• Accurate local alignment is vital for understanding protein function and evolution.

Purpose of the Study:

• To identify the most accurate local alignment algorithm for protein sequences.
• To introduce and validate a novel scoring function for protein local alignment using Ankh embeddings.

Main Methods:

• Developed a new algorithm incorporating Ankh embeddings into the E-score framework.
• Created a comprehensive evaluation framework with a new algorithm for local alignment extraction, localization, and quality assessment.
• Utilized five distance metrics and multiple datasets (CDD, BAliBASE, GPCRdb) for rigorous testing, performing over 2.5 million comparisons.

Main Results:

• The Ankh-score-based algorithm demonstrates superior accuracy in protein local alignment compared to existing methods, including BLOSUM, GPCRtm matrices, PEbA, DEDAL, vcMSA, and pLM-BLAST.
• Analysis revealed differential performance of protein language models on natural versus artificial sequences.
• Ankh embeddings showed limited benefit when combined with other embedding types.

Conclusions:

• The Ankh-score-based program represents a significant advancement, outperforming all current local alignment methods.
• Findings offer new insights into protein embeddings, guiding future research and development.
• The method and protocol are publicly available via a web server and source code for broader accessibility.