Biological causes and impacts of rugged tree landscapes in phylodynamic inference

Jiansi Gao, Marius Brusselmans, Luiz Max Carvalho, Marc A. Suchard, Guy Baele, Frederick A. MatsenProceedings of the National Academy of Sciences 2026 · DOI

TL;DR

This paper reveals that phylodynamic analyses, crucial for understanding pathogen evolution and spread, often struggle with "rugged" tree landscapes. This ruggedness, caused by a few problematic sequences, leads to widespread sampling issues that can significantly distort key biological conclusions, especially for local estimates. The authors propose new diagnostics and strategies to identify and mitigate these problems, paving the way for more reliable phylodynamic inference.

What Problem Does This Paper Address?

Phylodynamic inference uses genetic data and epidemiological information to reconstruct the evolutionary history and spread of pathogens. A core challenge is accounting for phylogenetic uncertainty, especially with highly similar sequences from modern datasets. Bayesian phylodynamics tackles this by integrating over possible tree structures using Markov Chain Monte Carlo (MCMC) methods. However, these MCMC methods can struggle to efficiently explore the vast "tree space" (the space of all possible evolutionary trees). This paper investigates why this exploration is difficult, characterizing the "tree landscape" (the landscape of tree likelihoods) and its impact on inference. Specifically, they ask: How rugged is this landscape in phylodynamic analyses? What causes this ruggedness? How do sampling problems affect biological estimates? And how can we better diagnose and address these issues?

Methods

The researchers performed extensive Bayesian phylodynamic analyses on 15 large, publicly available datasets. They used MCMC simulations, running chains for billions of iterations to thoroughly explore the posterior distribution of trees. To understand the tree landscape, they employed multidimensional scaling (MDS) to visualize the sampled trees in a lower-dimensional space, coloring them by MCMC chain to identify convergence issues. They also developed and utilized clade-specific MCMC diagnostics to pinpoint which parts of the tree (clades, sites, parameters) were poorly sampled. A key innovation was their method of "pruning" problematic sequences from sampled trees post hoc to assess their impact and then re-running analyses to validate findings. They also explored traditional data-quality tests (root-to-tip regression, recombination detection) to see if they could identify problematic sequences.

Datasets

The study utilized 15 large phylodynamic datasets from various viruses, including:

  • Ebola virus (EBOV)
  • Avian influenza (AIV)
  • H1N1 influenza
  • H3N2 influenza
  • B/Victoria influenza
  • HIV
  • Lassa virus
  • Mumps virus
  • West Nile virus (WNV)
  • Zika virus
  • SARS-CoV-2 (multiple datasets)
  • Rabies virus

These datasets represent a range of genetic diversity and evolutionary complexities, allowing for a broad assessment of phylodynamic inference challenges.

Key Insights and Ideas

  • Rugged Landscapes are the Norm: Contrary to intuition, limited genetic diversity in phylodynamic datasets doesn't necessarily simplify the tree landscape. Instead, it often leads to a diffuse yet highly "rugged" landscape with multiple peaks of similar height separated by deep valleys.
  • A Few Sequences Cause Big Problems: These sampling problems are frequently driven by a small number of sequences (sometimes just one or two) that act like "rogue" elements, causing significant topological uncertainty and making it hard for MCMC to explore the full tree space. These problematic sequences can be due to recombination, convergent evolution, or even sequencing/data errors.
  • Standard Data-Quality Tests Fall Short: Traditional methods like root-to-tip regression and basic recombination detection are often insufficient to identify the specific sequences causing MCMC sampling issues.
  • Local Estimates Suffer Most: While global parameters like demographic trajectories might be relatively robust to tree sampling problems, "local" estimates—such as the precise timing of introductions, specific transmission routes, or the age and rate of particular clades—are highly vulnerable to distortion.
  • Clade-Specific Diagnostics are Powerful: The authors developed and highlighted the utility of clade-specific MCMC diagnostics, which can pinpoint problematic clades and, by extension, the sequences driving those problems.

Key Results

  • Widespread Sampling Problems: Seven out of the 15 datasets showed clear nonconvergence, with MCMC chains getting stuck in different parts of the tree landscape, even after billions of iterations. Another five datasets exhibited slow mixing within chains.
  • Pruning Solves Many Problems: Removing just a few identified problematic sequences from the Lassa and Mumps datasets dramatically improved MCMC sampling performance, resolving nonconvergence and making the tree landscape unimodal.
  • Impact on Biological Conclusions: For the HIV dataset, different MCMC chains produced significantly different estimates for the timing of subtype B origin and geographic dispersal patterns, demonstrating how sampling problems can lead to conflicting biological conclusions.
  • New Diagnostics Identified: Parsimony-score based diagnostics were found to be effective and model-independent for identifying tree sampling problems.

Limitations and Future Work

  • Computational Cost: While the study provides valuable insights, the extensive analyses required billions of MCMC iterations, highlighting the computational expense of thorough phylodynamic inference.
  • Identifying Problematic Sequences: While clade-specific diagnostics are effective post hoc, proactively identifying problematic sequences before running lengthy analyses remains a challenge. The authors suggest that quick recombination tests might offer a preliminary screening tool.
  • Developing Better MCMC: The paper strongly implies the need for more efficient MCMC algorithms and proposal schemes that can better navigate these rugged tree landscapes and traverse the "valleys" between peaks.
  • Automated Solutions: Future work could focus on developing automated pipelines that integrate these new diagnostics and potentially incorporate machine learning approaches (like VAEs mentioned in the paper) to better characterize and traverse tree space.
  • Data Quality: The increasing availability of large, diverse datasets means that issues like recombination, contamination, and sequencing errors are becoming more critical to address systematically.

Summary updated

Abstract

Phylodynamic analysis has been instrumental in elucidating epidemiological and evolutionary dynamics of pathogens. Bayesian phylodynamics integrates out phylogenetic uncertainty, which is typically substantial in phylodynamic datasets due to limited genetic diversity. Phylodynamic inference does not, however, scale with modern datasets, partly due to difficulties in traversing tree space. Here, we characterize tree space and landscape in phylodynamic inference and assess its impacts on analysis difficulty and key biological estimates. By running extensive Bayesian analyses of 15 classic large phylodynamic datasets and carefully analyzing the posterior samples, we find that the posterior tree landscape is diffuse yet rugged, leading to widespread tree sampling problems that usually stem from sequences in a small part of the tree. We develop clade-specific diagnostics to show that a few sequences-including putative recombinants and recurrent mutants-frequently drive the ruggedness and sampling problems, although existing data-quality tests show limited power to detect them. The sampling problems can significantly impact phylodynamic inferences or distort major biological conclusions; the impact is usually stronger on "local" estimates (e.g., introduction history) associated with particular clades than on "global" parameters (e.g., demographic trajectory) governed by general tree shape. We evaluate existing Markov chain Monte Carlo diagnostics and diagnostics developed here, and offer strategies for optimizing phylodynamic analysis settings and mitigating sampling problem impacts. Our findings highlight the need and directions to develop efficient traversal over rugged tree landscapes, ultimately advancing scalable and reliable phylodynamics.
BibTeX
@inproceedings{gao2026biological,
  author = {Jiansi Gao and Marius Brusselmans and Luiz Max Carvalho and Marc A. Suchard and Guy Baele and Frederick A. Matsen},
  title = {{Biological causes and impacts of rugged tree landscapes in phylodynamic inference}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2026},
  doi = {10.1073/pnas.2510938123},
}
Figure