PhyloCNN: Improving Tree Representation and Neural Network Architecture for Deep Learning from Trees in Phylodynamics...

Area of Science:

• Evolutionary biology
• Computational biology
• Genomics

Background:

• Phylodynamic and diversification studies often employ complex evolutionary models.
• Traditional likelihood-based methods struggle with these complex models.
• Likelihood-free simulation-based approaches offer an alternative for incorporating complex scenarios.

Purpose of the Study:

• To develop a novel deep learning (DL) method for phylodynamics and diversification.
• To enable accurate selection and parameter estimation of birth-death models.
• To improve computational efficiency and accuracy compared to existing methods.

Main Methods:

• A simulation-based deep learning (DL) approach using a convolutional neural network architecture named PhyloCNN.
• Encoding phylogenetic trees using the neighborhood context of nodes and leaves.
• Comparing PhyloCNN's accuracy with varying neighborhood sizes and training set sizes.

Main Results:

• PhyloCNN accurately selects birth-death models and estimates parameters.
• Broader neighborhood context improves accuracy, especially for complex models and smaller training sets.
• PhyloCNN achieves high accuracy with significantly smaller training datasets (10,000-100,000 trees) compared to other DL methods.
• PhyloCNN performs comparably to state-of-the-art likelihood-based methods.

Conclusions:

• PhyloCNN offers a computationally efficient and accurate alternative for phylodynamics and diversification studies.
• The method successfully handles complex evolutionary models and specific sampling scenarios.
• PhyloCNN demonstrates potential for analyzing real-world datasets, such as HIV evolution and primate seed dispersal.