DeepMSA: constructing deep multiple sequence alignment to improve contact prediction and fold-recognition for...

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Bioinformatics

Background:

• Genome sequencing has led to a vast increase in protein sequences.
• Accurate multiple sequence alignment (MSA) is crucial for modeling protein structure and function.
• Existing pipelines lack efficiency and sensitivity, especially for large-scale genomic and metagenomic data.

Purpose of the Study:

• To develop a novel, sensitive, and efficient pipeline for multiple sequence alignment (MSA) construction.
• To improve the accuracy of protein structure and function prediction methods.
• To provide a robust tool for protein structural bioinformatics applications.

Main Methods:

• Developed DeepMSA, an open-source method for sensitive MSA construction.
• Utilized complementary hidden Markov model algorithms with multi-source databases.
• Evaluated DeepMSA in three large-scale benchmark experiments using 614 non-redundant proteins.

Main Results:

• DeepMSA improved residue-level contact prediction accuracy by up to 24.4% for long-range contacts.
• Average TM-score for homologous structure identification increased by over 7.5%.
• Achieved statistically significant improvements in secondary structure prediction (Q3 accuracy).

Conclusions:

• DeepMSA offers robust and general usefulness in protein structural bioinformatics.
• Improvements were achieved without re-training existing models, demonstrating method's versatility.
• Particularly beneficial for proteins without homologous templates in the Protein Data Bank (PDB).