Obtaining extremely large and accurate protein multiple sequence alignments from curated hierarchical alignments

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Machine learning for protein analysis requires large multiple sequence alignments (MSAs).
• Existing automated MSA generation methods (e.g., PSI-BLAST, JackHMMER) struggle with distant homologs and evolutionary divergence, necessitating manual curation.
• Manually curated MSAs are often too small for statistical methods.

Purpose of the Study:

• To address limitations in automated MSA generation for machine learning.
• To provide a robust method for creating large, accurate MSAs from curated hierarchical MSAs (hiMSAs).
• To enable improved protein sequence and structural analysis using deep learning and big data approaches.

Main Methods:

• Development and release of 252 curated hiMSAs comprising over 26 million sequences.
• Introduction of the MAPGAPS search program for querying hiMSAs to align vast numbers of database sequences.
• Hierarchical MSA construction with subgroup and template alignments for accurate sequence relationships.

Main Results:

• Demonstrated accurate alignment of database sequences using hiMSAs and MAPGAPS, comparable to curated MSA quality.
• Successfully applied the method to the exonuclease-endonuclease-phosphatase superfamily and pleckstrin homology domains.
• Generated extremely large MSAs suitable for deep learning and big data analyses.

Conclusions:

• The developed hiMSA framework and MAPGAPS program significantly enhance the accuracy and scale of MSAs for computational biology.
• This resource facilitates advanced machine learning applications in protein sequence and structural analysis.
• Public availability of hiMSAs and associated software promotes further research in the field.