OTalign: optimal transport alignment for remote protein homologs using protein language model embeddings

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Protein sequence alignment is vital in bioinformatics.
• Aligning remote homologs with low sequence identity is challenging due to difficulties in gap handling.

Purpose of the Study:

• Introduce a novel method for protein sequence alignment using Optimal Transport (OT) theory.
• Provide a mathematically principled framework for modeling residue matches and gaps.

Main Methods:

• Formulate sequence alignment as an entropy-regularized unbalanced optimal transport (UOT) problem.
• Utilize embeddings derived from protein language models (PLMs).
• Introduce position-specific, adaptive gap penalties.

Main Results:

• OTalign consistently outperforms traditional and recent PLM-based methods on remote-homolog benchmarks (SABmark, MALIDUP, MALISAM).
• Achieved high F1 scores: 0.594 on SABmark Superfamily and 0.358 on SABmark Twilight.
• Provides a quantitative metric for PLM embedding effectiveness and enables end-to-end PLM fine-tuning.

Conclusions:

• OTalign offers a superior approach to protein sequence alignment, especially for challenging remote homologs.
• The method provides interpretable insights into sequence similarity and facilitates the development of alignment-optimized embeddings.
• The code is publicly available, promoting further research and application.