Differentiable Phylogenetics via Hyperbolic Embeddings with Dodonaphy

TL;DR

This paper introduces soft-NJ, a differentiable neighbor-joining algorithm that enables gradient-based optimization of phylogenetic trees in hyperbolic space, facilitating Bayesian inference and improving tree optimization methods.

Contribution

It presents a novel differentiable tree decoder, soft-NJ, allowing gradient-based optimization in hyperbolic embeddings for phylogenetics, and demonstrates its application in Bayesian inference.

Findings

Differentiable optimization over tree space is feasible with soft-NJ.

Hyperbolic embeddings improve encoding of phylogenetic trees.

Local optima due to geometric frustrations pose challenges.

Abstract

Motivation: Navigating the high dimensional space of discrete trees for phylogenetics presents a challenging problem for tree optimisation. To address this, hyperbolic embeddings of trees offer a promising approach to encoding trees efficiently in continuous spaces. However, they require a differentiable tree decoder to optimise the phylogenetic likelihood. We present soft-NJ, a differentiable version of neighbour-joining that enables gradient-based optimisation over the space of trees. Results: We illustrate the potential for differentiable optimisation over tree space for maximum likelihood inference. We then perform variational Bayesian phylogenetics by optimising embedding distributions in hyperbolic space. We compare the performance of this approximation technique on eight benchmark datasets to state-of-art methods. However, geometric frustrations of the embedding locations produce local optima that pose a challenge for optimisation. Availability: Dodonaphy is freely available on the web at www.https://github.com/mattapow/dodonaphy. It includes an implementation of soft-NJ.