Biased phylodynamic inferences from analysing clusters of viral sequences

Abstract

Phylogenetic methods are being increasingly used to help understand the transmission dynamics of measurably evolving viruses, including HIV. Clusters of highly similar sequences are often observed, which appear to follow a 'power law' behaviour, with a small number of very large clusters. These clusters may help to identify subpopulations in an epidemic, and inform where intervention strategies should be implemented. However, clustering of samples does not necessarily imply the presence of a subpopulation with high transmission rates, as groups of closely related viruses can also occur due to non-epidemiological effects such as over-sampling. It is important to ensure that observed phylogenetic clustering reflects true heterogeneity in the transmitting population, and is not being driven by non-epidemiological effects. We qualify the effect of using a falsely identified 'transmission cluster' of sequences to estimate phylodynamic parameters including the effective population size and exponential growth rate under several demographic scenarios. Our simulation studies show that taking the maximum size cluster to re-estimate parameters from trees simulated under a randomly mixing, constant population size coalescent process systematically underestimates the overall effective population size. In addition, the transmission cluster wrongly resembles an exponential or logistic growth model 99% of the time. We also illustrate the consequences of false clusters in exponentially growing coalescent and birth-death trees, where again, the growth rate is skewed upwards. This has clear implications for identifying clusters in large viral databases, where a false cluster could result in wasted intervention resources.

Growth of HIV-1 Molecular Transmission Clusters in New York City

“…Therefore, intervention and prevention efforts targeted at the parts of the transmission network with the greatest potential for future growth are expected to have disproportionately large effects on preventing onward transmission [23][24][25]. However, there is reason for concern that the largest clusters in a molecular transmission network represent sampling bias and not growth potential [26,27]. Furthermore, it remains unclear whether molecular epidemiology approaches can provide superior predictions of future cluster growth, compared to predictions from standard transmission risk factor and demographic data.…”

Identification of hidden population structure in time-scaled phylogenies

“…For example, we have applied non-parametric methods to estimate the effective population size through time in HIV outbreaks detected using treestructure which highlighted particular groups that appear to be at higher risk of transmission. Such analyses would be more problematic using other partitioning or clustering algorithms because phylogenetic clusters can appear by chance in homogeneous populations of neutrally evolving pathogens, and this can give the false appearance of recent growth (Dearlove et al 2017). This application of phylodynamics analysis methods is possible because the statistical test used in treestructure provides theoretical justification for treating each partition as a separate unstructured population.…”

“…Their most recent common ancestor (MRCA) is at the root of the tree, but they have a very similar distribution of coalescent times suggesting that they were generated by similar demographic or epidemiological processes. For example, this can happen in infectious disease epidemics, when lineages independently colonise the same host population with greater susceptibility or higher risk behaviour (Dearlove et al 2017). It is therefore also desirable to have an automated method for identifying polyphyletic taxonomic groups defined by shared inferred population histories as opposed to genetic or phenotypic traits.…”

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

“…Genetic clustering can be an important resource for retrospective epidemiological investigations [ 50 ] and may eventually play a central role in the near-real time monitoring and prediction of infectious disease outbreaks [ 3 , 4 ]. However, the growing popularity of applying genetic clustering to detect outbreaks of transmission needs to be tempered with greater skepticism about the underlying methods [ 42 , 51 ]. Our model simulations represent a highly simplified hypothetical scenario where many clustering methods could potentially misdirect public health efforts away from groups suffering from higher rates of transmission, and towards groups where new infections were diagnosed sooner than the population average.…”