README.md

Dimensionality reduction distills complex evolutionary relationships in seasonal influenza and SARS-CoV-2

Sravani Nanduri¹, Allison Black², Trevor Bedford^2,3, John Huddleston^2,4

1. Paul G. Allen School of Computer Science and Engineering, University of Washington, Seattle, WA, USA
2. Vaccine and Infectious Disease Division, Fred Hutchinson Cancer Research Center, Seattle, WA, USA
3. Howard Hughes Medical Institute, Seattle, WA, USA
4. Corresponding author ([email protected])

DOI: https://doi.org/10.1093/ve/veae087

Abstract

Public health researchers and practitioners commonly infer phylogenies from viral genome sequences to understand transmission dynamics and identify clusters of genetically-related samples. However, viruses that reassort or recombine violate phylogenetic assumptions and require more sophisticated methods. Even when phylogenies are appropriate, they can be unnecessary or difficult to interpret without specialty knowledge. For example, pairwise distances between sequences can be enough to identify clusters of related samples or assign new samples to existing phylogenetic clusters. In this work, we tested whether dimensionality reduction methods could capture known genetic groups within two human pathogenic viruses that cause substantial human morbidity and mortality and frequently reassort or recombine, respectively: seasonal influenza A/H3N2 and SARS-CoV-2. We applied principal component analysis (PCA), multidimensional scaling (MDS), t-distributed stochastic neighbor embedding (t-SNE), and uniform manifold approximation and projection (UMAP) to sequences with well-defined phylogenetic clades and either reassortment (H3N2) or recombination (SARS-CoV-2). For each low-dimensional embedding of sequences, we calculated the correlation between pairwise genetic and Euclidean distances in the embedding and applied a hierarchical clustering method to identify clusters in the embedding. We measured the accuracy of clusters compared to previously defined phylogenetic clades, reassortment clusters, or recombinant lineages. We found that MDS embeddings accurately represented pairwise genetic distances including the intermediate placement of recombinant SARS-CoV-2 lineages between parental lineages. Clusters from t-SNE embeddings accurately recapitulated known phylogenetic clades, H3N2 reassortment groups, and SARS-CoV-2 recombinant lineages. We show that simple statistical methods without a biological model can accurately represent known genetic relationships for relevant human pathogenic viruses. Our open source implementation of these methods for analysis of viral genome sequences can be easily applied when phylogenetic methods are either unnecessary or inappropriate.

Citation

@article{10.1093/ve/veae087,
    author = {Nanduri, Sravani and Black, Allison and Bedford, Trevor and Huddleston, John},
    title = {Dimensionality reduction distills complex evolutionary relationships in seasonal influenza and {SARS-CoV-2}},
    journal = {Virus Evolution},
    pages = {veae087},
    year = {2024},
    month = {11},
    issn = {2057-1577},
    doi = {10.1093/ve/veae087},
    url = {https://doi.org/10.1093/ve/veae087}
}

Phylogenetic trees and embeddings

Explore the phylogenetic trees and embeddings on Nextstrain.

• Early influenza H3N2 HA (2016-2018)
  • Phylogeny colored by Nextstrain clade
  • PCA embedding
  • MDS embedding (1 and 2)
  • MDS embedding (2 and 3)
  • t-SNE embedding
  • UMAP embedding
• Late influenza H3N2 HA (2018-2020)
  • Phylogeny colored by Nextstrain clade
  • PCA embedding
  • MDS embedding (1 and 2)
  • MDS embedding (2 and 3)
  • t-SNE embedding
  • UMAP embedding
• Influenza H3N2 HA and NA (2016-2018)
  • HA phylogeny colored by maximum compatibility clades (MCCs) representing HA/NA reassortment groups
  • NA phylogeny colored by MCCs
  • HA/NA tangletree colored by MCCs
  • PCA embedding
  • MDS embedding (1 and 2)
  • MDS embedding (2 and 3)
  • t-SNE embedding
  • UMAP embedding
• Early SARS-CoV-2 (2020-2022)
  • Phylogeny colored by Nextstrain clade
  • Phylogeny colored by collapsed Nextclade pango lineage
  • PCA embedding
  • MDS embedding (1 and 2)
  • MDS embedding (2 and 3)
  • t-SNE embedding
  • UMAP embedding
• Late SARS-CoV-2 (2022-2023)
  • Phylogeny colored by Nextstrain clade
  • Phylogeny colored by collapsed Nextclade pango lineage
  • PCA embedding
  • MDS embedding (1 and 2)
  • MDS embedding (2 and 3)
  • t-SNE embedding
  • UMAP embedding

Interactive figures

Main figures

• Fig 2. Phylogeny of early (2016--2018) influenza H3N2 HA sequences plotted by nucleotide substitutions per site on the x-axis (top) and low-dimensional embeddings of the same sequences by PCA (middle left), MDS (middle right), t-SNE (bottom left), and UMAP (bottom right). Tips in the tree and embeddings are colored by their Nextstrain clade assignment. Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line colors represent the clade membership of the most ancestral node in the pair of nodes connected by the segment. Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny. Clade labels appear in the tree at the earliest ancestral node of the tree for each clade. Clade labels appear in each embedding at the average position on the x and y axis for sequences in a given clade.
• Fig 4. Phylogenetic trees (left) and embeddings (right) of early (2016--2018) influenza H3N2 HA sequences colored by HDBSCAN cluster. Normalized VI values per embedding reflect the distance between clusters and known genetic groups (Nextstrain clades). Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny.
• Fig 5. Phylogenetic trees (left) and embeddings (right) of late (2018--2020) H3N2 HA sequences colored by HDBSCAN cluster. Normalized VI values per embedding reflect the distance between clusters and known genetic groups (Nextstrain clades). Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny.
• Fig 6. Phylogeny of early (2016--2018) influenza H3N2 HA sequences plotted by nucleotide substitutions per site on the x-axis (top) and low-dimensional embeddings of the same HA sequences concatenated with matching NA sequences by PCA (middle left), MDS (middle right), t-SNE (bottom left), and UMAP (bottom right). Tips in the tree and embeddings are colored by their TreeKnit Maximally Compatible Clades (MCCs) label which represents putative HA/NA reassortment groups. Tips from MCCs with fewer than 10 sequences are colored as ``unassigned''. The first normalized VI values per embedding reflect the distance between HA/NA clusters and known genetic groups (MCCs). VI values in parentheses reflect the distance between HA-only clusters and known genetic groups. MCC labels appear in the tree and each embedding for larger pairs of reassortment events. MCC 9 represents two Nextstrain clades, so its labels appear twice in the tree. MCCs 14 and 11 represent a previously published reassortment event within Nextstrain clade A2 (Potter et al. 2019). Labels for MCC 14 represent the subset of its sequences from clade A2.
• Fig 7. Phylogeny of early (2020--2022) SARS-CoV-2 sequences plotted by number of nucleotide substitutions from the most recent common ancestor on the x-axis (top) and low-dimensional embeddings of the same sequences by PCA (middle left), MDS (middle right), t-SNE (bottom left), and UMAP (bottom right). Tips in the tree and embeddings are colored by their Nextstrain clade assignment. Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny. Clade labels in the tree and embeddings highlight larger clades.
• Fig 9. Phylogenetic trees (left) and embeddings (right) of early (2020--2022) SARS-CoV-2 sequences colored by HDBSCAN cluster. Normalized VI values per embedding reflect the distance between clusters and known genetic groups (Nextstrain clades). Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny.
• Fig 10. Phylogenetic trees (left) and embeddings (right) of late (2022--2023) SARS-CoV-2 sequences colored by HDBSCAN cluster. Normalized VI values per embedding reflect the distance between clusters and known genetic groups (Nextstrain clades).

Supplemental figures

• S4 Fig. MDS embeddings for early (2016--2018) influenza H3N2 HA sequences showing all three components. Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line colors represent the clade membership of the most ancestral node in the pair of nodes connected by the segment. Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny. Clade labels appear in the tree at the earliest ancestral node of the tree for each clade. Clade labels appear in each embedding at the average position on the x and y axis for sequences in a given clade.
• S7 Fig. Phylogeny of late (2018--2020) influenza H3N2 HA sequences plotted by nucleotide substitutions per site on the x-axis (top) and low-dimensional embeddings of the same sequences by PCA (middle left), MDS (middle right), t-SNE (bottom left), and UMAP (bottom right). Tips in the tree and embeddings are colored by their Nextstrain clade assignment. Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line colors represent the clade membership of the most ancestral node in the pair of nodes connected by the segment. Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny. Clade labels appear in the tree at the earliest ancestral node of the tree for each clade. Clade labels appear in each embedding at the average position on the x and y axis for sequences in a given clade.
• S8 Fig. MDS embeddings for late (2018--2020) influenza H3N2 HA sequences showing all three components. Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line colors represent the clade membership of the most ancestral node in the pair of nodes connected by the segment. Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny. Clade labels appear in the tree at the earliest ancestral node of the tree for each clade. Clade labels appear in each embedding at the average position on the x and y axis for sequences in a given clade.
• S11 Fig. Embeddings influenza H3N2 HA-only (left) and combined HA/NA (right) showing the effects of additional NA genetic information on the placement of reassortment events detected by TreeKnit (MCCs). Sequences from MCCs with fewer than 10 sequences are colored as "unassigned". Normalized VI values quantify the degree to which the combination of HA and NA sequences in an embedding reduces the distance of embedding clusters to TreeKnit reassortment groups represented by MCCs. MCC labels for larger pairs of reassortment events appear in each embedding at the average position on the x and y axis for sequences in a given MCC. MCCs 14 and 11 represent a previously published reassortment event within Nextstrain clade A2 (Potter et al. 2019). Labels for MCC 14 represents the sequences from clade A2.
• S12 Fig. PCA embeddings for influenza H3N2 HA sequences only (top row) and HA/NA sequences combined (bottom row) showing the HA trees colored by clusters identified in each embedding (left) and the corresponding embeddings colored by cluster (right). Normalized VI values quantify the degree to which the combination of HA and NA sequences in an embedding reduces the distance of embedding clusters to TreeKnit reassortment groups represented by MCCs.
• S13 Fig. MDS embeddings for influenza H3N2 HA sequences only (top row) and HA/NA sequences combined (bottom row) showing the HA trees colored by clusters identified in each embedding (left) and the corresponding embeddings colored by cluster (right). Normalized VI values quantify the degree to which the combination of HA and NA sequences in an embedding reduces the distance of embedding clusters to TreeKnit reassortment groups represented by MCCs.
• S14 Fig. t-SNE embeddings for influenza H3N2 HA sequences only (top row) and HA/NA sequences combined (bottom row) showing the HA trees colored by clusters identified in each embedding (left) and the corresponding embeddings colored by cluster (right). Normalized VI values quantify the degree to which the combination of HA and NA sequences in an embedding reduces the distance of embedding clusters to TreeKnit reassortment groups represented by MCCs.
• S16 Fig. UMAP embeddings for influenza H3N2 HA sequences only (top row) and HA/NA sequences combined (bottom row) showing the HA trees colored by clusters identified in each embedding (left) and the corresponding embeddings colored by cluster (right). Normalized VI values quantify the degree to which the combination of HA and NA sequences in an embedding reduces the distance of embedding clusters to TreeKnit reassortment groups represented by MCCs.
• S17 Fig. MDS embeddings for early SARS-CoV-2 sequences showing all three components. Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny. Clade labels in the tree and embeddings highlight larger clades.
• S18 Fig. Phylogeny of early (2020--2022) SARS-CoV-2 sequences plotted by number of nucleotide substitutions from the most recent common ancestor on the x-axis (top) and low-dimensional embeddings of the same sequences by PCA (middle left), MDS (middle right), t-SNE (bottom left), and UMAP (bottom right). Tips in the tree and embeddings are colored by their Pango lineage assignment. Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny. Clade labels in the tree and embeddings highlight larger Pango lineages.
• S20 Fig. Phylogenetic trees (left) and embeddings (right) of early (2020--2022) SARS-CoV-2 sequences colored by HDBSCAN cluster. Normalized VI values per embedding reflect the distance between clusters and known genetic groups (Pango lineages). Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods).
• S21 Fig. Phylogeny of late (2022--2023) SARS-CoV-2 sequences plotted by number of nucleotide substitutions from the most recent common ancestor on the x-axis (top) and low-dimensional embeddings of the same sequences by PCA (middle left), MDS (middle right), t-SNE (bottom left), and UMAP (bottom right). Tips in the tree and embeddings are colored by their Nextstrain clade assignment. Tips that could not be assigned to a predefined Nextstrain clade due to recombination were colored as "recombinant". Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny. Clade labels in the tree and embeddings highlight larger clades. Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny.
• S22 Fig. Phylogeny of late (2022--2023) SARS-CoV-2 sequences plotted by number of nucleotide substitutions from the most recent common ancestor on the x-axis (top) and low-dimensional embeddings of the same sequences by PCA (middle left), MDS (middle right), t-SNE (bottom left), and UMAP (bottom right). Tips in the tree and embeddings are colored by their Pango lineage assignment. Line segments in each embedding reflect phylogenetic relationships with internal node positions calculated from the mean positions of their immediate descendants in each dimension (see Methods). Line thickness in the embeddings scales by the square root of the number of leaves descending from a given node in the phylogeny. Clade labels in the tree and embeddings highlight larger Pango lineages.
• S24 Fig. Phylogenetic trees (left) and embeddings (right) of late (2022--2023) SARS-CoV-2 sequences colored by HDBSCAN cluster. Normalized VI values per embedding reflect the distance between clusters and known genetic groups (Pango lineages).

Supplemental tables

• S1 Table. Distances between clusters identified per pathogen dataset and method compared to known genetic groups. Distances are measured by normalized variation of information (VI). Smaller VI values indicate fewer differences between HDBSCAN clusters and known genetic groups. VI of 0 indicates identical clusters and 1 indicates maximally different clusters. Known genetic groups include Nextstrain clades, reassortment groups identified by TreeKnit as Maximally Compatible Clades (MCC), and Pango lineages. Total clusters refers to the number of clusters identified for a given dataset and method not including the "-1" label that HDBSCAN assigns to records that could not be assigned to a cluster. Threshold refers to the minimum distance between initial clusters for HDBSCAN to consider them as distinct clusters. For embedding methods, the threshold represents the Euclidean distance between sequences in each embedding. For the genetic distance method, the threshold represents the number of nucleotide differences between sequences. We identified optimal thresholds per pathogen, genetic group, and method from early influenza and SARS-CoV-2 data and applied these optimal thresholds to the corresponding late datasets for each pathogen. Datasets and methods without a threshold value in this table used the threshold from their corresponding early datasets. Rows appear in ascending order of VI values per method within each combination of pathogen and genetic group type.
• S2 Table. Number of clusters (n_clusters) identified by HDBSCAN, transitions between clusters in the phylogeny, and excess transitions indicating non-monophyletic groups per pathogen dataset and embedding method. This table reports the list of specific clusters per dataset and method (clusters), the number of transitions between different cluster labels on the phylogeny (n_cluster_transitions), and the list of observed transitions for pairs of clusters (transitions). The number of excess transitions between clusters (n_extra_transitions) reflects the number of times that we observed a transition from one source cluster to another beyond the one expected transition. Embeddings without any excess transitions reflect monophyletic groups in the corresponding pathogen phylogeny. Data available at https://zenodo.org/records/13381647/files/S2_Table.csv.
• S3 Table. Mutations observed per embedding cluster or Nextstrain clade relative to a reference genome sequence for each pathogen. Each row reflects the alternate allele identified at a specific position of a given pathogen genome or gene sequence, the pathogen dataset, the genetic group (an embedding cluster or Nextstrain clade), the number of clusters in the genetic group with the observed mutation, and the list of distinct genetic group labels with the mutation. Mutations must have occurred in at least 10 samples of the given dataset with an allele frequency of at least 50% to be reported in the table. Cluster- or clade-specific mutations appear in rows with a cluster_count value of 1. Data available at https://zenodo.org/records/13381647/files/S3_Table.csv.
• S4 Table. Average Euclidean distances between each known recombinant, X, and its parental lineages A and B per embedding method. Distances include average pairwise comparisons between A and B (distance_A_B), A and X (distance_A_X), and B and X (distance_B_X). Additional columns indicate whether each recombinant lineage maps closer to both parental lineages (or at least one) than those parents map to each other (X_maps_closer_to_both_parentals and X_maps_closer_to_any_parental, respectively). Records with values of True in the column X_maps_closer_to_both_parentals represent the expected placement of the recombinant lineage between its two parental lineages. Data available at https://zenodo.org/records/13381647/files/S4_Table.csv.
• S5 Table. Accessions and authors from originating and submitting laboratories of seasonal influenza and SARS-CoV-2 sequences from INSDC databases. Data available at https://zenodo.org/records/13381647/files/S5_Table.tsv.

Full analysis

Installation

First, install Conda with the Miniconda distribution. Until Bioconda supports modern Mac CPUs, Mac users with M1/M2 CPUs (the ARM64 architecture) need to install the Mac Intel x86 Miniconda distribution and install Rosetta, so the workflow can run under Mac's emulation mode.

After installing Conda, create the environment for this project.

conda env create -f cartography.yml

Activate the environment prior to running the workflow below.

conda activate cartography

Next, you need to install Julia and then install TreeKnit following the instructions to install the "CLI" version. The TreeKnit binary installs in your home directory, by default, in the path ~/.julia/bin/treeknit. This path is what the project's workflow calls to run TreeKnit.

Notes for Windows users

If you are a Windows user, you will need to install WSL to run this project's workflow. You cannot put this github repository in the Users file. Snakemake sees /U as a unicodeescape error and will not run, so please make a folder outside of the Users folder (ex. directly in the C drive) where you install this github repository, anaconda, and all other dependencies.

Run the full analysis

Run the full analysis for the project which includes simulations, analysis of natural populations, and generation of the manuscript and its figures and tables. Use the following command to run the analysis on a single compute node (e.g., a local laptop, single cluster node through an interactive shell, etc.).

snakemake --profile profiles/local

Use the following command to run the analysis on a SLURM cluster, submitting no more than 20 jobs at a time.

snakemake -j 20 --profile profiles/slurm

This is a complex workflow, so it will take several hours to run.