learnMSA2: deep protein multiple alignments with large language and hidden Markov models.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Traditional protein sequence alignment tools struggle with accuracy as sequence identity decreases or dataset size increases.
• Existing methods often use limited prior knowledge and data, impacting their performance on challenging alignments.

Purpose of the Study:

• To develop an improved protein sequence alignment method that leverages deep learning embeddings.
• To enhance the accuracy and scalability of multiple sequence alignment for large and diverse protein families.

Main Methods:

• Extended traditional profile hidden Markov models (HMMs) to incorporate protein sequence embeddings from deep learning models.
• Utilized gradient descent and a differentiable HMM layer for model fitting.
• Jointly aligned unaligned protein sequences and their embeddings to a protein family model.

Main Results:

• The upgraded HMM-based aligner, learnMSA2, combined with the ProtT5-XL protein language model, achieved an average of nearly 6% higher accuracy in correctly aligned columns compared to state-of-the-art competitors.
• learnMSA2 demonstrated superior performance with lower sequence identity and larger numbers of sequences.
• The method scales well with increasing sequence numbers.

Conclusions:

• Protein language models' embeddings contain rich evolutionary, structural, and biophysical information valuable for bioinformatics tasks.
• learnMSA2 represents a significant advancement in multiple sequence alignment, particularly for challenging datasets.
• The findings highlight the potential of deep learning embeddings to drive downstream applications in bioinformatics.