Genetic clustering approaches are increasingly adopted in Public Health practice to identify groups of HIV-infected individuals potentially arising from rapidly growing transmission clusters. Members of these clusters are candidates for targeted HIV and STD prevention activities such as early treatment initiation and linkage to, and reengagement in, care, and pre-exposure prophylaxis and partner services among at-risk partners. The goal of this study is to identify the most effective computational approaches to detect these rapidly growing clusters.
The study utilized the earliest HIV pol sequences among 22,398 persons reported in Los Angeles County Molecular HIV Surveillance database from 2000 to 2016. We evaluated five approaches to characterize cluster growth: (i) number of newly identified cluster members in relation to the cluster size [relative growth], (ii) sigmoidal curve fitting, (iii) phylodynamic estimation of the change in effective population size, (iv) phylodynamic estimation of epidemic reproductive number, (v) randomly selected clusters. Clusters for each year in 2008-2015 were identified using HIV-TRACE (pairwise genetic distance threshold of 0.015 substitutions/site). The number of individuals added to the clusters, selected by each approach, over the subsequent 12 months was evaluated to determine the best method for predicting cluster growth.
Of 22,398 persons, 8,133 (36.3%) were linked in 1,722 clusters ranging from 2 to 116 individuals. All approaches predicted cluster growth better than the random method. On average, these four approaches identified a growth rate of 0.3 newly linked persons within 12 months, compared with 0.15 persons in the random method. Although both phylodynamic reconstruction methods could be used to impute non-tested/non-reported cases within a cluster, they did not perform better than relative cluster growth. Notably, phylodynamic analyses of clusters with fewer than 10 cases often failed to converge, (likely due to lack of signal). When we re-analyzed the network focusing only on large clusters (≥10 individuals), none of the methods performed significantly better than targeting large clusters at random.
Past cluster growth is a reliable predictor of future growth. Cluster growth relative to cluster size was as predictive of future cluster growth as phylodynamic reconstruction and was much faster and more reliable to calculate.