TreeFix-TP: Phylogenetic Error-Correction for Infectious Disease Transmission Network Inference

Samuel Sledzieski, Chengchen Zhang, Ion I. Mandoiu, Mukul S. Bansal

TL;DR

This paper introduces TreeFix-TP, a novel method for reconstructing more accurate phylogenetic trees of infectious disease strains by correcting errors in existing maximum likelihood trees. By leveraging within-host genetic diversity and host information, TreeFix-TP significantly improves the accuracy of downstream transmission network inference, crucial for understanding and controlling disease outbreaks.

What Problem Does This Paper Address?

Understanding how infectious diseases spread within populations is critical for public health interventions. A key tool for this is inferring transmission networks, which often rely on phylogenetic trees that represent the evolutionary history of pathogen strains. However, reconstructing these phylogenies accurately is challenging, especially for rapidly evolving viruses that exist as "clouds" of closely related strains within a single host (quasispecies). Existing methods for phylogenetic reconstruction can be error-prone and struggle to scale to large datasets, negatively impacting the accuracy of subsequent transmission network inference. Furthermore, most methods fail to utilize the rich information available from sequencing multiple distinct strains from the same infected host, a capability now enabled by next-generation sequencing. This paper addresses the need for a scalable and accurate method that can leverage this within-host diversity to improve phylogenetic reconstruction and, consequently, transmission network inference.

Methods

The core innovation of this paper is TreeFix-TP, an error-correction framework for phylogenetic reconstruction. It starts with a maximum likelihood (ML) phylogeny (generated by tools like RAxML) as a baseline. TreeFix-TP then iteratively refines this tree by performing local searches (using NNI and SPR operations) and evaluating candidate trees based on two criteria:

  1. Statistical Likelihood: The Shimodaira-Hasegawa (SH) test is used to determine if a candidate tree is statistically equivalent to the ML tree, ensuring it's well-supported by the sequence data.
  2. Transmission Cost: This is a novel criterion that minimizes the number of implied inter-host transmissions. It's calculated using Fitch's algorithm for the small parsimony problem, where the "states" at the tree's leaves are the hosts from which the sequences were sampled.

By accepting trees that are statistically equivalent to the ML tree but have a lower transmission cost, TreeFix-TP aims to find a more biologically realistic phylogeny that better reflects the actual transmission events. The method is designed to be scalable and can leverage the information from multiple sequences sampled from each host.

Datasets

The authors evaluated TreeFix-TP using both simulated and real-world datasets:

  • Simulated Data: They generated a large number of simulated datasets using the FAVITES framework. These simulations varied parameters such as transmission models (SEIR, SIR), sequence lengths, the number of viral sequences sampled per host, and evolutionary rates (scale factors). This allowed for a systematic evaluation of TreeFix-TP's robustness and performance under different conditions.
  • Real-World Data: They used 10 real-world Hepatitis C Virus (HCV) outbreak datasets from the CDC. These datasets were used to assess TreeFix-TP's ability to improve source detection in actual outbreaks.

Key Insights and Ideas

  • Within-host diversity is a goldmine for phylogeny: The paper strongly demonstrates that leveraging multiple sequences from a single host significantly improves phylogenetic accuracy, a resource largely untapped by previous methods.
  • Minimizing transmissions as a phylogenetic criterion: Introducing "transmission cost" as a key metric for evaluating phylogenies is a novel approach. It shifts the focus from purely sequence-based likelihood to a more biologically grounded measure of evolutionary history.
  • Error correction has a downstream impact: Improved phylogenetic accuracy directly translates to better transmission network inference, highlighting the importance of robust phylogenetic reconstruction as a foundational step.
  • TreeFix-TP is robust: The method performs well across a wide range of simulated parameters, indicating its general applicability.
  • The importance of sampling: The results underscore that comprehensive sampling of infected hosts is crucial for accurate transmission network inference, and the benefits of TreeFix-TP are most pronounced with good sampling.

Key Results

  • Phylogenetic Accuracy: TreeFix-TP consistently reconstructed more accurate phylogenies than RAxML, as measured by normalized Robinson-Foulds (RF) distance. For example, under the SEIR model, the average RF distance decreased by 6.78% with TreeFix-TP. This improvement was particularly notable with shorter sequence lengths.
  • Transmission Network Accuracy: Using TreeFix-TP phylogenies as input for the Phyloscanner tool led to improved transmission network inference. The average F1 score for network reconstruction increased from 0.645 (with RAxML) to 0.652 (with TreeFix-TP) on baseline SEIR datasets. This improvement was more pronounced when sequence information was limited.
  • Source Detection: In real-world HCV outbreaks, TreeFix-TP correctly identified the source host in 8 out of 10 outbreaks, compared to 6 or 7 with RAxML, depending on the rooting strategy.
  • Scalability: TreeFix-TP is scalable, with running times that scale linearly with tree size, number of hosts, and sequence length.

Limitations and Future Work

  • Absolute Error Rate: While TreeFix-TP significantly reduces error rates, the absolute error rate of the reconstructed phylogenies can still be quite high. This is partly because the effectiveness of error correction depends on the number of distinct hosts represented in the phylogeny, not just the tree size.
  • Need for Advanced Network Inference: The paper suggests that while improved phylogenies lead to better network inference, more sophisticated network inference methods might be needed to fully exploit the gains in phylogenetic accuracy.
  • Future Directions: The authors propose developing even more advanced, scalable methods for transmission phylogeny construction. They also suggest exploring the use of additional information about within-host strain evolution to further enhance transmission phylogeny inference.

Summary updated

BibTeX
@inproceedings{sledzieski2021treefix,
  author = {Samuel Sledzieski and Chengchen Zhang and Ion I. Mandoiu and Mukul S. Bansal},
  title = {{TreeFix-TP: Phylogenetic Error-Correction for Infectious Disease Transmission Network Inference}},
  booktitle = {Psb},
  year = {2021},
  doi = {10.1142/9789811232701\_0012},
}
Figure