Differentiable Phylogenetics by way of Hyperbolic Embeddings with Dodonaphy
Authors: Matthew Macaulay, Mathieu Fourment
Summary: Motivation: Navigating the excessive dimensional area of discrete bushes for phylogenetics presents a difficult drawback for tree optimisation. To handle this, hyperbolic embeddings of bushes supply a promising method to encoding bushes effectively in steady areas. Nonetheless, they require a differentiable tree decoder to optimise the phylogenetic chance. We current soft-NJ, a differentiable model of neighbour-joining that permits gradient-based optimisation over the area of bushes. Outcomes: We illustrate the potential for differentiable optimisation over tree area for optimum chance inference. We then carry out variational Bayesian phylogenetics by optimising embedding distributions in hyperbolic area. We examine the efficiency of this approximation method on eight benchmark datasets to state-of-art strategies. Nonetheless, geometric frustrations of the embedding areas produce native optima that pose a problem for optimisation. Availability: Dodonaphy is freely obtainable on the net at www.https://github.com/mattapow/dodonaphy. It contains an implementation of soft-NJ.
