Assessing the Validity of the Fixed Tree Topology Assumption in Phylodynamic Inference

Mathieu Fourment, Jiansi Gao, Marc A. Suchard, Frederick A. MatsenarXiv 2025 · DOI · arXiv

Abstract: Fixed tree topologies are widely used in phylodynamic analyses to reduce computational burden, yet the consequences of this assumption remain insufficiently understood. Here, we systematically assess the impact of various fixed-topology strategies on phylogenetic and phylodynamic parameter estimates across a diverse set of viral datasets. We compare fully Bayesian joint inference with fixed-topology strategies, including conditioning on maximum likelihood trees subsequently dated with LSD or TreeTime. Our analyses show that global parameters of the substitution and site models are largely robust to the fixed-topology assumption, whereas parameters that depend on the temporal structure of the tree, such as molecular clock rates, node ages, and demographic histories, can exhibit substantial biases. We do treat unconstrained Bayesian analyses as the reference, although we recognize that these too are model-based approximations. Nevertheless, our results highlight serious discordance associated with fixing the topology and underscore the need for faster, time-aware methods that simultaneously integrate topology and parameter estimation. These findings raise important questions about the balance between computational efficiency and inferential accuracy in phylodynamic studies.

Introduction

Phylodynamics has become a cornerstone of infectious disease research, allowing scientists to infer the hidden dynamics of epidemics from pathogen genomes. Ideally, researchers perform "joint inference," using Bayesian frameworks like BEAST to simultaneously estimate the evolutionary tree (topology) and the parameters of interest. However, as dataset sizes explode—particularly evident during the COVID-19 pandemic—the computational cost of searching through the astronomical number of possible tree shapes becomes prohibitive.

To manage this, many researchers adopt a two-step shortcut: first, they infer a single "best" tree using Maximum Likelihood (ML) methods, and then they fix that topology to estimate parameters like the molecular clock rate or effective population size. This study systematically evaluates whether this common shortcut is a safe approximation or a source of fundamental distortion in our understanding of pathogen evolution.

Methodology

The researchers evaluated 15 empirical viral datasets, including Ebola, Influenza, HIV, SARS-CoV-2, West Nile Virus, and Zika. They compared "unconstrained" Bayesian analyses (the ground truth) against several fixed-topology strategies:

  1. ML-based trees: Trees generated by IQ-TREE and dated using heuristic tools like LSD or TreeTime.
  2. Posterior-informed trees: Using the Maximum Clade Credibility (MCC) or Maximum Sum of Posterior (MSP) trees from a full Bayesian run.
  3. Maximum Rooting Credibility (MRC): A hybrid strategy using an ML tree rooted based on the Bayesian posterior.

Robustness of Substitution Models

A key finding of the study is the striking divide between different types of parameters. Parameters describing the biochemical process of mutation (substitution and site models) appear remarkably robust to the choice of tree topology. Across nearly all datasets, these parameters remained within a ±25% error range of the Bayesian reference, regardless of whether the tree was an ML-based heuristic or a posterior-informed summary tree.

Relative error of site and substitution model parameters across 15 viral datasets, demonstrating their general robustness to fixed topologies.

Sensitivity of Temporal and Coalescent Parameters

In contrast, parameters that depend on the temporal structure of the tree—such as molecular clock rates, node ages, and demographic histories—showed high sensitivity and significant bias. When using ML-based fixed trees, only 62% of these temporal estimates fell within the ±25% error band. The bias was particularly pronounced when using heuristic dating tools like LSD and TreeTime, which often selected root positions with zero support in the full Bayesian analysis.

Relative error of temporal (clock, coalescent, and tree) parameters, showing significant bias and variance when using heuristic dating methods compared to posterior-informed trees.

This "rooting bottleneck" cascades into massive errors in clock rate estimates and population growth histories. For instance, the error (RMSRE) for population size estimates was nearly six times higher on average when using trees rooted with TreeTime or LSD compared to using Bayesian summary trees.

Conclusion

The study highlights a critical trade-off: while fixing the topology saves time, it can severely compromise the accuracy of the phylodynamic parameters researchers care about most. The results suggest that ML-inferred trees often sit far outside the cluster of trees supported by the temporal signal in the data. To bridge this gap, the authors advocate for the development of "time-aware" ML methods—algorithms that can infer a tree while simultaneously accounting for the molecular clock, offering the speed of ML with the inferential accuracy of Bayesian methods.

Summary updated

Abstract

Fixed tree topologies are widely used in phylodynamic analyses to reduce computational burden, yet the consequences of this assumption remain insufficiently understood. Here, we systematically assess the impact of various fixed-topology strategies on phylogenetic and phylodynamic parameter estimates across a diverse set of viral datasets. We compare fully Bayesian joint inference with fixed-topology strategies, including conditioning on maximum likelihood trees subsequently dated with LSD or TreeTime. Our analyses show that global parameters of the substitution and site models are largely robust to the fixed-topology assumption, whereas parameters that depend on the temporal structure of the tree, such as molecular clock rates, node ages, and demographic histories, can exhibit substantial biases. We do treat unconstrained Bayesian analyses as the reference, although we recognize that these too are model-based approximations. Nevertheless, our results highlight serious discordance associated with fixing the topology and underscore the need for faster, time-aware methods that simultaneously integrate topology and parameter estimation. These findings raise important questions about the balance between computational efficiency and inferential accuracy in phylodynamic studies.
BibTeX
@misc{fourment2025assessing,
  author = {Mathieu Fourment and Jiansi Gao and Marc A. Suchard and Frederick A. Matsen},
  title = {{Assessing the Validity of the Fixed Tree Topology Assumption in Phylodynamic Inference}},
  year = {2025},
  eprint = {2511.22005},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.22005},
  howpublished = {arXiv},
}

Loading related content...

Figure