ModelRevelator: Fast phylogenetic model estimation via deep learning.

Area of Science:

• Computational Biology
• Phylogenetics
• Machine Learning

Background:

• Selecting appropriate evolutionary models is crucial for accurate phylogenetic tree reconstruction.
• Traditional methods like maximum likelihood (ML) are computationally expensive due to tree reconstruction and parameter optimization.
• Existing model selection tools may not always be computationally feasible for large datasets.

Purpose of the Study:

• To introduce ModelRevelator, a novel tool for selecting evolutionary models using deep neural networks.
• To demonstrate that neural networks can perform model selection without reconstructing trees or calculating likelihoods.
• To provide a computationally efficient alternative to traditional model selection methods.

Main Methods:

• Development of two deep neural networks: NNmodelfind for model recommendation and NNalphafind for rate heterogeneity (Γ-distribution) and its shape parameter (ɑ).
• Inputting a multiple-sequence alignment (MSA) into ModelRevelator for swift model and parameter output.
• Comparative analysis against likelihood-based methods and the machine learning tool ModelTeller.

Main Results:

• ModelRevelator successfully recommends evolutionary models, including rate heterogeneity and the ɑ parameter.
• Performance is comparable to established likelihood-based methods and ModelTeller across various parameter settings.
• The tool demonstrates robust performance on both training and unseen empirical data.

Conclusions:

• ModelRevelator offers a computationally efficient and accurate method for evolutionary model selection in phylogenetics.
• It is particularly valuable for phylogeneticists facing computationally prohibitive traditional methods.
• The neural network-based approach has significant potential for streamlining phylogenetic analyses.