DeepMSA

DeepMSA constructs deep, high-quality multiple sequence alignments by integrating sequences from Uniclust30, UniRef90, Metaclust and whole-genome and metagenome databases to support protein structure and function prediction for distant homologs.

Key Features:

• Comprehensive Sequence Integration: Merges sequences from large-scale databases including Uniclust30, UniRef90 and Metaclust to increase MSA depth and diversity.
• Hidden Markov Model Algorithms: Employs complementary hidden Markov model (HMM) algorithms to create sensitive MSAs and enhance homolog detection across varied data sources.
• Robust Performance without Re-training: Improves outcomes in protein structural bioinformatics applications without requiring re-training of existing parameters or neural-network models.

Scientific Applications:

• Contact Prediction: DeepMSA profiles improve residue-level contact prediction accuracy, reporting up to a 24.4% improvement for long-range contacts compared with default HHblits and PSI-BLAST methods.
• Threading and Structure Identification: Enhances homologous structure identification across multiple threading programs, yielding over a 7.5% increase in the average TM-score of template alignments.
• Secondary Structure Prediction: Contributes to statistically significant improvements in secondary structure (Q3) prediction accuracy.

Methodology:

Constructs MSAs by integrating sequences from Uniclust30, UniRef90, Metaclust and whole-genome/metagenome collections using complementary HMM algorithms within a multi-source pipeline, validated on large-scale benchmarks of 614 non-redundant proteins.