MOTIVATION: Protein sequence alignment is a crucial task in bioinformatics, yet aligning remote homologs with low sequence identity remains a longstanding challenge, particularly due to the difficulty of handling gaps. We introduce a new method that applies Optimal Transport (OT) theory to sequence alignment, providing a mathematically principled framework for modeling residue matches and gaps. RESULTS: OTalign formulates sequence alignment as an entropy-regularized unbalanced optimal transport (UOT) problem over embeddings derived from protein language models (PLMs). Unlike traditional methods, it introduces position-specific gap penalties that adapt to each sequence pair. On challenging remote-homolog benchmarks (SABmark, MALIDUP, MALISAM), OTalign consistently outperforms baselines (Needleman-Wunsch, HHalign) and recent PLM-based methods (PLMAlign, DeepBLAST), achieving F1 scores of 0.594 on SABmark Superfamily and 0.358 on SABmark Twilight. Furthermore, OTalign provides a quantitative and interpretable metric of how effectively PLM embeddings represent sequence similarity relationships. Finally, its differentiable nature enables end-to-end fine-tuning of PLMs, establishing a framework for learning embeddings explicitly optimized for alignment tasks. AVAILABILITY AND IMPLEMENTATION: This code is available at https://github.com/DeepFoldProtein/OTalign.

MOTIVATION: Protein structure prediction has been revolutionized and generalized with the advent of cutting-edge AI methods such as AlphaFold, but reliance on computationally intensive multiple sequence alignments (MSA) remains a major limitation. RESULTS: We introduce DeepFold-PLM, a novel framework that integrates advanced protein language models with vector embedding databases to enhance ultra-fast MSA construction, remote homology detection, and protein structure prediction. DeepFold-PLM utilizes high-dimensional embeddings and contrastive learning, significantly accelerate MSA generation, achieving 47 times faster than standard methods, while maintaining prediction accuracy comparable to AlphaFold. In addition, it enhances structure prediction by extending modeling capabilities to multimeric protein complexes, provides a scalable PyTorch-based implementation for efficient large-scale prediction. Our method also effectively increases sequence diversity (Neff = 8.65 versus 4.83 with JackHMMER) enriching coevolutionary information critical for accurate structure prediction. DeepFold-PLM thus represents a versatile and practical resource that enables high-throughput applications in computational structural biology. AVAILABILITY AND IMPLEMENTATION: Source codes and user-friendly Python API of all modules of DeepFold-PLM publicly available at https://github.com/DeepFoldProtein/DeepFold-PLM.