phyddle: Software for Exploring Phylogenetic Models with Deep Learning.

Area of Science:

• Evolutionary Biology
• Computational Biology
• Bioinformatics

Background:

• Phylogenetic models are crucial for understanding life's diversity and evolutionary history.
• Standard inference methods struggle with realistic phylogenetic models that lack tractable likelihood functions.
• There is a need for computational tools that can handle complex phylogenetic models.

Purpose of the Study:

• To introduce phyddle, a novel software pipeline for phylogenetic modeling using likelihood-free deep learning.
• To demonstrate phyddle's capability in performing phylogenetic modeling tasks on trees.
• To provide a flexible tool for integrating deep learning into evolutionary research workflows.

Main Methods:

• Phyddle utilizes a pipeline approach with five steps: Simulate, Format, Train, Estimate, and Plot.
• The software employs likelihood-free deep learning methods to analyze phylogenetic data.
• Phylogenetic datasets are transformed from raw input into numerical and visual model-based output.

Main Results:

• Phyddle accurately performs phylogenetic inference tasks, including macroevolutionary parameter estimation and model selection for continuous trait evolution.
• The software successfully passes coverage tests for epidemiological models, even those with intractable likelihoods.
• Benchmarks demonstrate comparable accuracy between phyddle's deep learning-based inferences and traditional likelihood-based methods.

Conclusions:

• Phyddle offers a powerful and flexible solution for phylogenetic modeling, particularly for models with complex or intractable likelihood functions.
• The software facilitates the integration of advanced deep learning techniques into evolutionary biology research.
• Phyddle enhances the ability to extract evolutionary insights from phylogenetic trees, even in challenging modeling scenarios.