torchtree: Flexible Phylogenetic Model Development and Inference Using PyTorch.

Area of Science:

• Computational Biology
• Evolutionary Biology
• Bioinformatics

Background:

• Markov chain Monte Carlo (MCMC) is the traditional method for Bayesian inference in phylogenetics but is computationally intensive for complex models.
• The computational bottleneck of MCMC hinders scalability for large phylogenetic datasets and time tree models.
• Variational Bayes methods offer a scalable alternative for Bayesian inference, particularly for large datasets.

Purpose of the Study:

• To introduce torchtree, a flexible Python framework for developing phylogenetic models and algorithms.
• To enable efficient Bayesian inference using variational Bayes with PyTorch.
• To explore alternative optimization criteria for variational inference in phylogenetics.

Main Methods:

• Development of torchtree, a Python framework utilizing PyTorch for phylogenetic model implementation.
• Implementation of variational inference within torchtree, supporting automatic differentiation and analytical gradients.
• Evaluation of the forward KL divergence as an optimization criterion for variational inference, comparing it with the evidence lower bound (ELBO).

Main Results:

• Torchtree demonstrates comparable speed to existing tools like BEAST for phylogenetic inference.
• Variational inference with torchtree yields promising approximation accuracy, though performance varies across different scenarios.
• Forward KL divergence inference shows faster iterations than ELBO, while ELBO may offer faster convergence in some cases.

Conclusions:

• Torchtree provides a flexible and efficient framework for phylogenetic model development and Bayesian inference.
• The framework facilitates the application of variational Bayes methods to complex phylogenetic problems.
• Exploration of forward KL divergence offers a viable alternative optimization strategy for variational inference in phylogenetics.