On the utility of Deep Learning for model classification and parameter estimation on complex diversification...

Area of Science:

• Evolutionary biology
• Computational phylogenetics
• Machine learning in macroevolution

Background:

• Birth-Death models are crucial for studying diversification dynamics using phylogenetic trees.
• Traditional likelihood-based inference methods face computational challenges with complex models.
• Deep Learning offers a computationally tractable approach to increase model complexity.

Purpose of the Study:

• To evaluate the effectiveness of Convolutional Neural Networks (CNNs) for classification and regression tasks in phylogenetic diversification analysis.
• To compare CNN performance against Maximum Likelihood Estimation (MLE) for extant-only phylogenies under various diversification scenarios.
• To apply trained CNN models for parameter estimation on empirical phylogenies.

Main Methods:

• Simulated 10,000 phylogenetic trees for six diversification scenarios (e.g., Constant Birth-Death, Mass Extinction).
• Encoded phylogenetic trees using the CDV vectorization procedure to capture branch length information.
• Trained CNN models and compared their performance with Maximum Likelihood Estimation (MLE).

Main Results:

• CNNs achieved 80-93% classification accuracy, significantly outperforming MLE (70-74%).
• CNNs showed slightly lower mean average errors in regression tasks compared to MLE.
• Both methods struggled with estimating ratio parameters like mass-extinction survival.

Conclusions:

• CNNs provide a powerful and accurate tool for inferring diversification dynamics from phylogenetic data.
• Further research should focus on incorporating rate-variable and lineage-heterogeneous models into deep learning frameworks.
• Deep learning advances computational phylogenetics for macroevolutionary studies.