PrioriTree: a utility for improving phylodynamic analyses in BEAST

Jiansi Gao, Michael R. May, Bruce Rannala, Brian R. MoorebioRxiv 2022 · DOI

TL;DR

This paper introduces PrioriTree, an R package designed to help researchers understand and manage the sensitivity of phylodynamic analyses to prior assumptions. By providing tools for robust Bayesian inference, data cloning, and model fit assessment, PrioriTree aims to improve the reliability of inferences about disease outbreak histories.

What Problem Does This Paper Address?

Phylodynamic methods are crucial for reconstructing the geographic and demographic history of disease outbreaks. These methods often rely on discrete-geographic models, which can involve a large number of parameters, especially when dealing with many distinct geographic regions. Inferring these parameters from limited data (like the sampling location of pathogen isolates) makes the analysis highly sensitive to the initial "prior beliefs" about these parameters. The default priors in common software like BEAST are often biologically unrealistic, potentially leading to misleading conclusions. Researchers need a way to systematically assess how much their results depend on these prior choices and to select more appropriate priors.

Methods

PrioriTree offers a suite of functions to address prior sensitivity in phylodynamic analyses within the BEAST framework. Its core features include:

  • Robust Bayesian Inference: This involves running multiple analyses on the same dataset, each time systematically altering one or more prior distributions. By comparing the resulting posterior distributions, researchers can identify parameters whose estimates are heavily influenced by prior choices. PrioriTree aids this by dynamically generating plots of prior distributions and their implications.
  • Data Cloning: This technique assesses the impact of a specific prior on the posterior distribution. It works by running analyses on iteratively increasing numbers of "clones" (replicate copies) of the original dataset. As the number of clones increases, the data's influence on the posterior grows, and the posterior should converge to the maximum likelihood estimate (MLE). PrioriTree helps visualize this convergence rate, indicating how informative a prior is relative to the data.
  • Model Adequacy and Fit Assessment: PrioriTree implements posterior-predictive simulation to evaluate how well a chosen model (including its priors) describes the observed data. If simulated data from the model resembles the real data, the model is considered adequate. It also supports power-posterior analyses for computing marginal likelihoods to compare the relative fit of different candidate models using Bayes factors.

Datasets

The paper uses a SARS-CoV-2 dataset for demonstration purposes in its figures. No new benchmark datasets are introduced.

Key Insights and Ideas

  • Prior Sensitivity is a Major Hurdle: The paper strongly emphasizes that phylodynamic inferences, particularly for complex geographic models, are inherently sensitive to prior parameter choices. This sensitivity can lead to unreliable conclusions if not properly addressed.
  • Default Priors Can Be Problematic: The default priors in widely used software like BEAST are often biologically unrealistic, highlighting the need for tools that help researchers move beyond these defaults.
  • Transparency is Key: PrioriTree's interactive visualization of prior distributions and their implications (e.g., showing the number of dispersal events corresponding to a given prior on the dispersal rate) helps researchers develop more informed prior choices.
  • Multiple Approaches for Prior Assessment: The utility offers distinct but complementary methods (robust Bayesian inference and data cloning) to tackle prior sensitivity from different angles.

Key Results

The paper demonstrates PrioriTree's utility using a SARS-CoV-2 dataset.

  • Figure 1 shows how different prior choices for the average dispersal rate (μ) can lead to substantially different inferred posterior distributions. It also illustrates how data cloning helps assess the convergence of these posteriors towards the MLE.
  • Figure 2 demonstrates the use of posterior-predictive simulation for model adequacy. It shows that a hierarchical exponential prior for μ provides an adequate fit to the SARS-CoV-2 data, while the default CTMC-rate reference prior does not.

Limitations and Future Work

  • Sequential Inference Assumption: PrioriTree assumes that the phylogeny and geographic history are inferred sequentially. While it offers ways to account for phylogenetic uncertainty, this sequential approach might be a limitation for some complex scenarios.
  • Focus on Discrete-Geographic Models: The utility is specifically designed for discrete-geographic phylodynamic models. Its applicability to other types of phylodynamic models or different inference problems is not discussed.
  • Future Work: The paper implicitly suggests that further development could involve expanding the range of models and priors supported, as well as potentially integrating more advanced methods for prior selection and model averaging.

Summary updated

Abstract

Abstract Summary Phylodynamic methods are central to studies of the geographic and demographic history of disease outbreaks. Inference under discrete-geographic phylodynamic models—which involve many parameters that must be inferred from minimal information—is inherently sensitive to our prior beliefs about the model parameters. We present an interactive utility, PrioriTree , to help researchers identify and accommodate prior sensitivity in discrete-geographic inferences. Specifically, PrioriTree provides a suite of functions to generate input files for—and summarize output from— BEAST analyses for performing robust Bayesian inference, data-cloning analyses, and assessing the relative and absolute fit of candidate discrete-geographic (prior) models to empirical datasets. Availability and Implementation PrioriTree is distributed as an R package available at https://github.com/jsigao/prioritree , with a comprehensive user manual provided at https://bookdown.org/jsigao/prioritree_manual/ . Contact [email protected]
BibTeX
@misc{gao2022prioritree,
  author = {Jiansi Gao and Michael R. May and Bruce Rannala and Brian R. Moore},
  title = {{PrioriTree: a utility for improving phylodynamic analyses in BEAST}},
  year = {2022},
  doi = {10.1101/2022.08.24.505196},
  howpublished = {bioRxiv},
}
Figure