Emerging Concepts of Data Integration in Pathogen Phylodynamics

Abstract

Phylodynamics has become an increasingly popular statistical framework to extract evolutionary and epidemiological information from pathogen genomes. By harnessing such information, epidemiologists aim to shed light on the spatio-temporal patterns of spread and to test hypotheses about the underlying interaction of evolutionary and ecological dynamics in pathogen populations. Although the field has witnessed a rich development of statistical inference tools with increasing levels of sophistication, these tools initially focused on sequences as their sole primary data source. Integrating various sources of information, however, promises to deliver more precise insights in infectious diseases and to increase opportunities for statistical hypothesis testing. Here, we review how the emerging concept of data integration is stimulating new advances in Bayesian evolutionary inference methodology which formalize a marriage of statistical thinking and evolutionary biology. These approaches include connecting sequence to trait evolution, such as for host, phenotypic and geographic sampling information, but also the incorporation of covariates of evolutionary and epidemic processes in the reconstruction procedures. We highlight how a full Bayesian approach to covariate modeling and testing can generate further insights into sequence evolution, trait evolution, and population dynamics in pathogen populations. Specific examples demonstrate how such approaches can be used to test the impact of host on rabies and HIV evolutionary rates, to identify the drivers of influenza dispersal as well as the determinants of rabies cross-species transmissions, and to quantify the evolutionary dynamics of influenza antigenicity. Finally, we briefly discuss how data integration is now also permeating through the inference of transmission dynamics, leading to novel insights into tree-generative processes and detailed reconstructions of transmission trees. [Bayesian inference; birth–death models; coalescent models; continuous trait evolution; covariates; data integration; discrete trait evolution; pathogen phylodynamics.

Figure 1.

Temporal signal in pathogen genomes sampled through time. A) Plot of the probability of observing no changes between two genomes of length () separated by the given number of days () for a given rate of evolution per site per year (). We used and for Influenza A/H1N1 (Stadler et al. 2013), and for MERS-CoV (Cotten et al. 2014), and for EBOV (Park et al. 2015), and for CHIK, and for Salmonella Typhimurium DT104 (Mather et al. 2013), and and for Smallpox. An online application to plot such probabilities is available at http://epidemic.bio.ed.ac.uk/node/79. B. Root-to-tip divergences as a function of sampling time based on publicly available Influenza A/H1N1, MERS-CoV and EBOV genomes as well as for an unpublished CHIK data set. The regression plots were rescaled for each virus such that the oldest genome in each dataset was set at time = 0 and the regression line has zero divergence at time = 0.

Figure 2.

The multihost transmission dynamics of rabies in American bat species. Streicker et al. (2010) identified 18 phylogenetic lineages of rabies virus that were statistically compartmentalized to particular bat taxa. These lineages are represented by differently colored clades in the phylogeny, along with a selection of bat species involved. CSTs are defined as jumps to bat species different from the dominant species within each lineage or clade. These dynamics have been quantified though structured coalescent approaches (Streicker et al. 2010). Host switches on the other hand are defined as the jumps along internal branches to new hosts followed by successful transmission in the new host species. Inferring the history of host jumping along the branches (mostly represented by white branches) has been the subject of discrete trait reconstruction (Streicker et al. 2010). We thank Daniel Streicker for providing the tree and MerlinTuttle.org for granting permission to use the bat portraits.

Figure 3.

Antigenic cartography meets Brownian phylogenetic diffusion. A) Conceptual representation of the integration of genetic and antigenic evolution through a Bayesian MDS approach. In the HI assay (on the right), the antigenic phenotype is investigated by measuring the cross-reactivity of a virus (A, B, C, or D) strain to serum () raised against another strain. Based on these HI measurements, MDS approaches allow to position viruses in lower-dimensional space such that the distances in this space best fit the HI assay titres. A probabilistic interpretation of MDS assumes that the observed differences are centered around their cartographic expectation, in which case the virus locations are estimable parameters. We refer to Bedford et al. (2014) for more information on how these locations are estimated in an integrated Bayesian phylogenetic framework. B) Visualization of antigenic drift dynamics reconstructed using Bayesian MDS in a two-dimensional map. These patterns are inferred from the 2002 to 2011 subset of the influenza A/H3N2 dataset analyzed by Bedford et al. (2014). X and Y represent the first and second antigenic dimensions. The contours represent the 80% HPD region for the node locations (both internal and external nodes). The colors range from green to blue for the lines, points and contours reflects the age between 2002 and 2011. This figure was made using SpreadD3 (Bielejec et al. 2016).

Figure 4.

Contrasting models with completely linked and unlinked parameters to hierarchical modeling without and with fixed effects. Traditionally, two competing approaches were used when performing Bayesian inference to estimate parameters from a potentially large number of partitions, strata, or individuals: total evidence, where all the data across strata are pooled or shared to estimate a single parameter of interest, and unconditionally independent partitioning, where each stratum requires the estimation of a completely independent set of parameters. The latter is associated with independent prior specification on every parameter. Hierarchical phylogenetic models (HPMs) offer a middle ground between these two extremes by sharing a hierarchical prior distribution over all parameters with estimable mean and variance, which are drawn from hyperpriors. Edo-Matas et al. (2011) propose an HPM that employs a Bayesian mixed effects model that pools information across patients, affording more precise individual-patient parameter estimates when the data are sparse for a patient, but also allowing estimation of the effect of patient groups or continuous covariates.

Figure 5.

GLM extension of discrete phylogenetic diffusion and examples of covariates or predictors () in a spatial context. The GLM parameterizes each rate of among-location movement in the phylogeographic model as a log linear function of various potential predictors. For each predictor (; ), the GLM parameterization includes a coefficient , which quantifies the contribution or effect size of the predictor (in log space), and a binary indicator variable , that allows the predictor to be included or excluded from the model. Since predictors are essentially matrices of pairwise measurements, location-specific measurements such as population size or density are separated into origin or destination size.