An efficient deep learning method for amino acid substitution model selection.

Area of Science:

• Computational Biology
• Bioinformatics
• Evolutionary Biology

Background:

• Amino acid substitution models are crucial for phylogenetic analysis of protein sequences.
• Estimating these models, often with many parameters, typically requires large datasets and computationally intensive methods like maximum likelihood.
• Existing methods face theoretical and computational challenges, motivating the search for more efficient approaches.

Purpose of the Study:

• To propose an efficient deep learning-based method for selecting amino acid substitution models.
• To address the computational burden associated with traditional model selection techniques.
• To enable rapid and accurate model selection for large-scale genomic data.

Main Methods:

• Developed a deep learning network, ModelDetector, trained on millions of protein alignments.
• Utilized summary statistics derived from amino acid substitution rates for training.
• Compared ModelDetector's performance against the maximum likelihood method (ModelFinder) using simulation data.

Main Results:

• ModelDetector demonstrated accuracy comparable to the maximum likelihood method.
• The deep learning approach was orders of magnitude faster than maximum likelihood methods.
• ModelDetector efficiently analyzed large genome alignments in minutes.

Conclusions:

• Deep learning offers a promising and efficient tool for amino acid substitution model selection.
• ModelDetector significantly reduces the computational time for phylogenetic analyses.
• This method facilitates the analysis of large-scale protein and genome sequence data.