Improving the Scalability of Bayesian Phylodynamic Inference through Efficient MCMC Proposals
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
In Bayesian phylodynamics, we jointly reconstruct the posterior distribution of timed phylogenetic trees, evolutionary and population dynamic parameters. Most approaches in the field use Markov chain Monte Carlo (MCMC) for inference. Inferring timed phylogenetic trees using MCMC relies on a set of proposal distributions -- operators -- to explore different phylogenetic tree topologies, node ranks, and heights. However, inefficiencies in these proposal distributions limit the speed and scale with which phylodynamic analyses can be performed. Here, we introduce two classes of operators that allow scaling of phylodynamic analyses by reducing the time needed to reach convergence. We propose an improved proposal distribution to scale trees based on the intervals between consecutive nodes in the tree, and a set of operators informed by a pseudo parsimony score to explore different tree topologies. Contrary to already existing operators that make large changes to areas of the tree with no mutations at all, our new parsimony-focused operators target areas of the tree with many mutations. We first demonstrate the correctness of these proposal distributions using a well-calibrated simulation study. We then demonstrate the improvements of using the updated proposal distributions using a number of large real-world datasets, showing a substantial increase in effective sample size per hour across viral datasets as well as a reduction in the number of samples required for the MCMC chain to reach stationarity. Finally, we apply the operators to a 10,000-sequence H3N2 dataset including phylogeography and finish the analysis within two weeks. The operators are implemented in the TargetedBeast package, which is available as an open-source package for BEAST2 and a user-friendly graphical user interface. The speed improvements in the TargetedBeast package are beneficial to any phylodynamics method implemented in BEAST2 that relies on standard phylogenetic trees, such as coalescent-skyline, birth-death skyline, and structured phylodynamic methods.