A model-based clustering method to detect infectious disease transmission outbreaks from sequence variation

Abstract

Clustering infections by genetic similarity is a popular technique for identifying potential outbreaks of infectious disease, in part because sequences are now routinely collected for clinical management of many infections. A diverse number of nonparametric clustering methods have been developed for this purpose. These methods are generally intuitive, rapid to compute, and readily scale with large data sets. However, we have found that nonparametric clustering methods can be biased towards identifying clusters of diagnosis—where individuals are sampled sooner post-infection—rather than the clusters of rapid transmission that are meant to be potential foci for public health efforts. We develop a fundamentally new approach to genetic clustering based on fitting a Markov-modulated Poisson process (MMPP), which represents the evolution of transmission rates along the tree relating different infections. We evaluated this model-based method alongside five nonparametric clustering methods using both simulated and actual HIV sequence data sets. For simulated clusters of rapid transmission, the MMPP clustering method obtained higher mean sensitivity (85%) and specificity (91%) than the nonparametric methods. When we applied these clustering methods to published sequences from a study of HIV-1 genetic clusters in Seattle, USA, we found that the MMPP method categorized about half (46%) as many individuals to clusters compared to the other methods. Furthermore, the mean internal branch lengths that approximate transmission rates were significantly shorter in clusters extracted using MMPP, but not by other methods. We determined that the computing time for the MMPP method scaled linearly with the size of trees, requiring about 30 seconds for a tree of 1,000 tips and about 20 minutes for 50,000 tips on a single computer. This new approach to genetic clustering has significant implications for the application of pathogen sequence analysis to public health, where it is critical to robustly and accurately identify clusters for the most cost-effective deployment of outbreak management and prevention resources.

Identification of hidden population structure in time-scaled phylogenies

“…The birth rate in the outbreak deme is 5-fold the birth rate in the reservoir, but in one set of simulations, both the birth rate and sampling rate in the outbreak was also increased 5-fold. In these simulations, the performance of treestructure (median ARI 56%) is slightly lower than the CLMP method (McCloskey and Poon 2017) (median ARI 83%) when only the birth-rate differs in the outbreak deme. However treestructure maintains good performance when death and sampling rates also differ.…”

“…Across 100 simulations, treestructure has mean ARI of 41% (IQR: 20-57%). The FastBAPS method (Tonkin-Hill et al 2019) has mean ARI of 2.3% (IQR:1.2-3.3%) and the CLMP method (McCloskey and Poon 2017) has mean ARI 5.2% (IQR:-1-7.5%).…”

“…Figure 5 shows performance of treestructure on previously published tree simulations (McCloskey and Poon 2017). These simulations differ from the structured coalescent simulations because both the reservoir and outbreak demes are growing exponentially at different rates.…”

“…To evaluate the potential for treestructure to detect outbreaks we applied the new method to phylogenies estimated from newly simulated data using a structured coalescent model, as well as previously published simulation data based on a discrete-event branching process (McCloskey and Poon 2017).…”

“…Previously, McCloskey et al simulated 100 genealogies from a discrete-event birth-death process (McCloskey and Poon 2017; Vaughan and Drummond 2013). These simulations were based on a process with heterogeneous classes of individuals with different birth rates.…”