Deep Learning from Phylogenies for Diversification Analyses.

Area of Science:

• Evolutionary biology
• Computational phylogenetics
• Bioinformatics

Background:

• Birth-death (BD) models are crucial for studying species diversification using phylogenies.
• Current likelihood-based inference methods lack generalizability and can be computationally intractable for complex models.
• Deep learning offers a potential solution by learning relationships between simulations and model parameters.

Purpose of the Study:

• To adapt deep learning for phylogenetic diversification inference.
• To extend deep learning to state-dependent diversification models using trait data.
• To evaluate the accuracy and efficiency of deep learning inference compared to traditional methods.

Main Methods:

• Adapted a deep learning method from pathogen phylodynamics for diversification inference.
• Trained deep neural networks to perform regression between simulated phylogenies and model parameters.
• Applied the method to time-constant homogeneous BD models and Binary-State Speciation and Extinction (BiSSE) models.
• Reanalyzed a primate phylogeny with ecological trait data (seed dispersal).

Main Results:

• Deep learning inference demonstrated accuracy comparable to likelihood-based methods.
• The deep learning approach was orders of magnitude faster than traditional methods.
• Successfully inferred parameters for both homogeneous BD and BiSSE models.
• Validated the approach on a real-world primate phylogeny dataset.

Conclusions:

• Deep learning provides an accurate and highly efficient alternative for phylogenetic diversification inference.
• This method broadens the applicability of inference to complex, state-dependent models.
• The speed and generalizability of deep learning will facilitate the development and deployment of novel evolutionary models.