Kaitlyn E. Wade

Phylogenetics has played a pivotal role in the genomic epidemiology of severe acute respiratory syndrome coronavirus 2, such as tracking the emergence and global spread of variants and scientific communication. However, the rapid accumulation of genomic data from around the world—with over two million genomes currently available in the Global Initiative on Sharing All Influenza Data database—is testing the limits of standard phylogenetic methods. Here, we describe a new approach to rapidly analyze and visualize large numbers of SARS-CoV-2 genomes. Using Python, genomes are filtered for problematic sites, incomplete coverage, and excessive divergence from a strict molecular clock. All differences from the reference genome, including indels, are extracted using minimap2 and compactly stored as a set of features for each genome. For each Pango lineage (https://cov-lineages.org), we collapse genomes with identical features into ‘variants’, generate 100 bootstrap samples of the feature set union to generate weights, and compute the symmetric differences between the weighted feature sets for every pair of variants. The resulting distance matrices are used to generate neighbor-joining trees in RapidNJ that are converted into a majority-rule consensus tree for each lineage. Branches with support values below 50 per cent or mean lengths below 0.5 differences are collapsed, and tip labels on affected branches are mapped to internal nodes as directly sampled ancestral variants. Currently, we process about 2 million genomes in approximately 9 h on 52 cores. The resulting trees are visualized using the JavaScript framework D3.js as ‘beadplots’, in which variants are represented by horizontal line segments, annotated with beads representing samples by collection date. Variants are linked by vertical edges to represent branches in the consensus tree. These visualizations are published at https://filogeneti.ca/CoVizu. All source code was released under an MIT license at https://github.com/PoonLab/covizu.

The prevailing abundance of full-length HIV type 1 (HIV-1) genome sequences providesan opportunity to revisit the standard model of HIV-1 group M (HIV-1/M) diversitythat clusters genomes into largely nonrecombinant subtypes, which is not consistentwith recent evidence of deep recombinant histories for simian immunodeficiency virus(SIV) and other HIV-1 groups. Here we develop an unsupervised nonparametricclustering approach, which does not rely on predefined nonrecombinant genomes, byadapting a community detection method developed for dynamic social network analysis.We show that this method (dynamic stochastic block model [DSBM]) attains a signif-icantly lower mean error rate in detecting recombinant breakpoints in simulated data(quasibinomial generalized linear model (GLM), P < 8 × 10−8), compared to otherreference-free recombination detection programs (genetic algorithm for recombinationdetection [GARD], recombination detection program 4 [RDP4], and RDP5). Whenthis method was applied to a representative sample of n = 525 actual HIV-1 genomes,we determined k = 29 as the optimal number of DSBM clusters and used change-pointdetection to estimate that at least 95% of these genomes are recombinant. Further,we identified both known and undocumented recombination hotspots in the HIV-1 genome and evidence of intersubtype recombination in HIV-1 subtype referencegenomes. We propose that clusters generated by DSBM can provide an informativeframework for HIV-1 classification.

Many viruses are organized into taxonomies of subtypes based on their genetic similarities. For human immunodeficiency virus 1 (HIV-1), subtype classification plays a crucial role in infection management. Sequence alignment-based methods for subtype classification are impractical for large datasets because they are costly and time-consuming. Alignment-free methods involve creating numerical representations or genetic sequences and applying statistical or machine learning methods. Despite their high overall accuracy, existing models perform poorly on less common subtypes. Furthermore, there is limited work investigating the impact of sequence vectorization methods, in particular natural language-inspired embedding methods, on HIV-1 subtype classification. We present a comprehensive analysis of sequence vectorization methods across machine learning methods. We report a k-mer-based XGBoost model with a balanced accuracy of 0.84, indicating that it has good overall performance for both common and uncommon HIV-1 sub-types. We also report a Word2Vec-based support vector machine that achieves promising results on precision and balanced accuracy. Our study sheds light on the effect of sequence vectorization methods on HIV-1 subtype classification and suggests that natural language-inspired encod-ing methods show promise. Our results could help to develop improved HIV-1 subtype classification methods, leading to improved individual pa-tient outcomes, and the development of subtype-specific treatments.