Differentiable Phylogenetics by means of Hyperbolic Embeddings with Dodonaphy
Authors: Matthew Macaulay, Mathieu Fourment
Abstract: Motivation: Navigating the extreme dimensional space of discrete bushes for phylogenetics presents a troublesome downside for tree optimisation. To deal with this, hyperbolic embeddings of bushes provide a promising methodology to encoding bushes successfully in regular areas. Nonetheless, they require a differentiable tree decoder to optimise the phylogenetic likelihood. We present soft-NJ, a differentiable mannequin of neighbour-joining that allows gradient-based optimisation over the world of bushes. Outcomes: We illustrate the potential for differentiable optimisation over tree space for optimum likelihood inference. We then perform variational Bayesian phylogenetics by optimising embedding distributions in hyperbolic space. We study the effectivity of this approximation methodology on eight benchmark datasets to state-of-art methods. Nonetheless, geometric frustrations of the embedding areas produce native optima that pose an issue for optimisation. Availability: Dodonaphy is freely obtainable on the web at www.https://github.com/mattapow/dodonaphy. It comprises an implementation of soft-NJ.
