METHODS
Dengue virus serotypes (DENV-1, DENV-2, DENV-3, and DENV-4) WGS with available country and year of sampling were downloaded from GenBank and nextstrain 19 (Supplementary material). All dengue virus WGS were aligned by the MAFFT v7 and visually inspected with AliView v1.26. The best-fitting nucleotide substitution (GTR) model was selected using a hierarchical likelihood ratio, Akaike information criterion, and Bayesian information criterion tests with Model Finder in IQ-TRE web server (http://iqtree.cibiv.univie.ac.at/). Dengue virus maximum likelihood phylogenetic tree was inferred according to the best-fitting model using IQ-TRE web server (http://iqtree.cibiv.univie.ac.at/). We used this tree to obtain root-to-tip regressions in TempEst v1.5.
Time-scaled phylogenetic trees, evolutionary rates, and demographic histories of dengue virus WGS were evaluated using the Bayesian coalescent framework implemented in BEAST v2.6.2, which uses a Markov Chain Monte Carlo (MCMC) sampling method to obtain posterior distributions of tree topologies and parameter estimates20. Bifurcating nodes with posterior probability greater than 0.95 were considered statistically well supported. For each run of 900 million of MCMC, the marginal likelihood was estimated via path sampling (PS) and stepping stone (SS) methods and the resulting Bayes Factors (BF) (ratio of marginal likelihoods) was used to select the best-fitting clock/demographic model. The models can be compared to evaluate the strength of evidence against the null hypothesis (H0) defined in the following way: 2lnBF < 2 indicates no evidence against H0; 2-6, weak evidence; 6-10: strong evidence, and >10 very strong evidence. Both SS and PS estimators indicated the uncorrelated lognormal (UCLN) relaxed molecular clock (Bayes Factor = 30.3) as the best-fitted model to the dataset under analysis. Besides, we have used the GTR substitution model.
MCMC were run for 900 million generations to ensure stationary and adequate effective sample size (ESS) for all statistical parameters. Tracer v.1.6 software was used to diagnose MCMC, adjust initial burn-in. Bayesian coalescent analyses were performed to estimate the viral dynamics and the time to the Most Recent Common Ancestor (tMRCA). The time-scale calibration was based on the isolation date of samples. Uncertainty in parameter estimates was evaluated in the 95% highest posterior density (HPD 95%) interval. TreeAnnotator v1.8.2 was used to summarize the maximum clade credibility (MCC) tree from the posterior distribution of trees.
Dengue virus phylogeographic analysis, incorporating both spatial and temporal information, was performed with BEAST v2.6.2 using a discrete trait, symmetric substitution model with Bayesian stochastic search variable selection (BSSVS). The reversible discrete Bayesian phylogeographic model with a continuous-time Markov chain rate reference prior was performed. The number of viral migrations between locations was estimated using ‘Markov Jump’ counts of location-state transitions along with the posterior tree distribution. Migratory events across time were summarized using the SPREAD v.1.0.7. BFs>3 were considered as well supported diffusion rates constituting the migration graph.