Differentiable phylogenetics via hyperbolic embeddings with Dodonaphy.

Area of Science:

• Computational Phylogenetics
• Machine Learning
• Bioinformatics

Background:

• Optimizing phylogenetic trees in high-dimensional discrete spaces is computationally challenging.
• Hyperbolic embeddings offer efficient continuous space encoding for trees.
• Differentiable tree decoders are necessary for optimizing phylogenetic likelihoods.

Purpose of the Study:

• To present soft-NJ, a differentiable version of the Neighbor-Joining algorithm.
• To enable gradient-based optimization directly over the space of phylogenetic trees.
• To apply differentiable tree embeddings for variational Bayesian phylogenetics.

Main Methods:

• Developed soft-NJ, a differentiable tree decoder for phylogenetic analysis.
• Utilized hyperbolic embeddings for efficient tree representation in continuous space.
• Performed variational Bayesian inference by optimizing embedding distributions.

Main Results:

• Demonstrated the potential of differentiable optimization for maximum likelihood inference.
• Evaluated soft-NJ's performance on eight benchmark datasets against state-of-the-art methods.
• Showcased soft-NJ as a powerful and efficient approach for phylogenetics via tree embeddings, despite potential local optima.

Conclusions:

• Soft-NJ facilitates gradient-based optimization in the space of phylogenetic trees.
• Hyperbolic embeddings combined with soft-NJ offer a parametrically efficient approach to phylogenetics.
• The Dodonaphy software package implements soft-NJ, making it accessible for research.