Abstract Body

Molecular epidemiology is increasingly used to understand HIV transmission and monitor outbreaks and can be a critical tool in their prevention. Different methods are available to identify transmission networks, and their choice can be arbitrary. The impact of method choice on cluster identification has not been comprehensively assessed and may affect public health interventions that rely upon their results.

We studied 8 commonly used methods (7 model-based, and distance-based HIV-TRACE) and used them to identify clusters of HIV-1 subtype B pol sequences from 1,656 persons, ~80% of a densely-sampled Rhode Island epidemic during 2004-2018. For each method, we compared proportion of clustered sequences within and between methods using various distance and bootstrap thresholds; and clustering concordance between methods (including percentages of identical sequence pairs that cluster together; percentages of cluster similarity at 100% and at 80%; and percentages of identical non-clustered sequences). We conducted comparisons under (i) strict (bootstrap ≥0.95; TN93 distance 0.015 substitutions/site) and (ii) relaxed (bootstrap 0.8-0.9; distance 0.03-0.045 substitutions/site) thresholds, intended to address different public health objectives (e.g. strict for outbreak; relaxed for epidemic characterization). 

Of the 1,656 sequences, 18-53% formed 114-217 clusters, depending on thresholds used. Clustering proportion within methods depended on bootstrap and distance, with distance having stronger effects. Variation in clustering proportion across methods was more pronounced with stringent bootstraps and relaxed distances. For strict thresholds, HIV-TRACE identified 5-15% higher proportion of clustered individuals than model-based methods (p<0.005 for all pairwise comparisons). In contrast, for relaxed thresholds, HIV-TRACE identified 3-19% lower proportion of clustered individuals than model-based methods (p<0.05 for all pairwise comparisons). Distributions of percent concordance between methods, stratified by threshold type (strict, relaxed), are presented in the Table.

Clustering similarity between common molecular epidemiology methods varied, with some substantial discordance. In the context of integration of molecular epidemiology into public health, this implies that the choice of clustering method and threshold may impact precision of public health interventions.