Non-negative matrix factorization for learning alignment-specific models of protein evolution.

Area of Science:

• Computational Biology
• Molecular Evolution
• Bioinformatics

Background:

• Protein evolution models are typically generalist (one-size-fits-all) or specialist (data-intensive).
• Specialist models offer better performance but require extensive data, which is often unavailable.
• Generalist models are widely applicable but less accurate for specific datasets.

Purpose of the Study:

• To develop a novel method for creating alignment-specific protein evolution models.
• To adapt model complexity based on the available data richness for a given alignment.
• To improve the accuracy and performance of protein evolution modeling.

Main Methods:

• Utilized non-negative matrix factorization (NNMF) to learn basis matrices from diverse protein alignments.
• Developed a weighting system to create alignment-specific models as a weighted sum of basis matrices.
• Constrained model complexity to the dimensions justified by the data, reducing parameters.

Main Results:

• The NNMF-based alignment-specific models outperformed existing methods on 49 out of 50 test alignments.
• Learned basis matrices confirmed conserved amino acid properties and quantified variation in conservation strength.
• Application to phylogeny inference yielded different and improved likelihood phylogenies compared to standard models.

Conclusions:

• The proposed NNMF method effectively generates accurate alignment-specific protein evolution models.
• This approach balances model complexity with data availability, offering superior performance.
• The method enhances phylogenetic inference accuracy and provides insights into amino acid property conservation.