Unifying Viral Genetics and Human Transportation Data to Predict the Global Transmission Dynamics of Human Influenza H3N2

Lemey, Philippe; Rambaut, Andrew; Bedford, Trevor; Faria, Nuno; Bielejec, Filip; Baele, Guy; Russell, Colin A.; Smith, Derek J.; Pybus, Oliver G.; Brockmann, Dirk; Suchard, Marc A.

In order to assess the extent to which evolutionary analyses such as ours benefit from integrating host mobility data, we examine their predictive performance by using them to predict the relative timing of the geographic spread of the pandemic H1N1 influenza variant that emerged in 2009. We conduct simulations of the spread of a novel pathogen out of Mexico using an SIR model whose transmission parameters are informed by epidemiological estimates obtained for pandemic H1N1 [31] and whose spatial spread is determined by one of four different migration rate models, each defined by a different matrix of movement rates among all pairs of locations (Methods). We measure the relative correspondence between the simulated and observed H1N1 peaks for each location except Mexico using a Spearman's rank correlation coefficient (r) and mean absolute error (MAE; in days) (Fig. 4 ). An equal rates model (A), which does not express any migration rate preference, results in a weak match (r~0:11, P = 0.73, MAE = 40.9 days) between the simulations and the observed spatial spread of H1N1 (Fig. 4) , indicating that the population sizes included in the SIR model for each region offer limited predictive performance. As expected, adding information on the number of airline passengers (model B) yields a large improvement in correspondence between simulations and observations (r~0:61, P = 0.03, MAE = 35.8 days). In contrast, a standard parameterrich phylogeographic model that is only informed by sequence data and not air traffic information (model C) yields only part of this improvement in predictive performance (r~0:47, P = 0.10, MAE = 39.4 days). However, if inference under model C is made more efficient by focusing on a small set of parameters (using BSSVS, [21] ; see Methods) then phylogeographic estimates yield a predictive performance (r~0:62, P = 0.02, MAE = 36.4 days, Fig. S5 ) that is close to that of the air travel model (B). Finally, The inclusion probabilities are defined by the indicator expectations E½d because they reflect the frequency at which the predictor is included in the model and therefore represent the support for the predictor. Indicator expectations corresponding to Bayes factor support values of 10 and 100 are represented by a thin and thick vertical line respectively in these bar plots. The contribution of each predictor, when included in the model (bDd~1), where b is the coefficient or effect size, is represented by the mean and credible intervals of the GLM coefficients on a log scale. NA 1 : no conditional effect size available because the predictor was never included in the model. We tested different population size and density measures, different incidence-based measures and different seasonal measures (Text S1), but only list the estimates for a representative predictor for the sake of clarity. The estimates for the full set of predictors are summarized for each sub-sampled data set in Fig. S5 . NA 2 : no indicator expectation or conditional effect size available because the predictor was not available for this discretization of the sequence data. doi:10.1371/journal.ppat.1003932.g002 the GLM model (D) predicts the observed spread of H1N1 more accurately than all other models (r~0:82, P,0.01, MAE = 32.3), suggesting that global influenza transmission is best predicted by combining passenger flux data with the information on viral lineage movement contained in sequence data. The simulations generally correspond better with obse