Ankh-Score Produces Better Sequence Alignments Than AlphaFold3.

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Biology

Background:

• Protein sequence alignment is crucial for bioinformatics with numerous applications.
• Advancements in protein structure prediction (AlphaFold) and protein language models (Ankh, ProtT5, ESM-C) offer new avenues for alignment improvement.

Purpose of the Study:

• To evaluate the efficacy of AlphaFold-predicted structures and protein language model embeddings for enhancing protein sequence alignment.
• To compare the Ankh-score method against structure-based alignment and traditional methods.

Main Methods:

• Utilized AlphaFold3 for protein structure prediction.
• Employed protein language models, specifically Ankh, to generate protein embeddings.
• Performed sequence alignment using the Ankh-score method and compared it with structure-based alignment (US-align) and traditional BLOSUM matrices.
• Validated performance on BAliBASE and CDD datasets.

Main Results:

• The Ankh-score method achieved superior sequence alignments compared to structure-based alignments of AlphaFold3-predicted structures.
• Both Ankh-score and AlphaFold3-based alignments outperformed traditional methods using BLOSUM matrices.
• Demonstrated that Ankh embeddings capture information potentially absent in AlphaFold3 structures.

Conclusions:

• Protein language model embeddings, particularly from Ankh, show significant promise for improving protein sequence alignment.
• Ankh-score represents a novel and effective approach to sequence alignment, outperforming current state-of-the-art methods.
• The Ankh-score software is publicly available as a web server and open-source code.