Ancestral state reconstruction with discrete characters using deep learning.

Area of Science:

• Phylogenetics
• Computational Biology
• Machine Learning

Background:

• Ancestral state reconstruction is crucial in phylogenetics.
• Likelihood-based methods are limited by tractable likelihood functions.
• Complex, biologically realistic models often have intractable likelihoods.

Purpose of the Study:

• To adapt the deep learning software phyddle for ancestral state reconstruction.
• To evaluate phyddle's performance on various models and tree sizes.
• To compare phyddle with Bayesian inference for ancestral state reconstruction.

Main Methods:

• Modification of the phyddle software for ancestral state reconstruction.
• Performance evaluation under diverse methodological and modeling conditions.
• Comparison with Bayesian inference where feasible.

Main Results:

• Phyddle performance mirrors Bayesian inference for simple models and small trees.
• Performance degrades as phylogenetic tree size increases.
• Phyddle adequately handles complex models (e.g., speciation/extinction) but shows greater divergence from Bayesian inference.

Conclusions:

• Phyddle provides a viable deep learning approach for ancestral state reconstruction, especially for models with intractable likelihoods.
• The method shows promise for empirical datasets, including genus Liolaemus and Ebola virus evolution.
• Further research may be needed to optimize performance for large phylogenetic trees.