Phylodynamics of the HIV-1 CRF02_AG clade in Cameroon

Abstract

Evolutionary analyses have revealed an origin of pandemic HIV-1 group M in the Congo River basin in the first part of the XX^th century, but the patterns of historical viral spread in or around its epicentre remain largely unexplored. Here, we combine epidemiologic and molecular sequence data to investigate the spatiotemporal patterns of the CRF02_AG clade. By explicitly integrating prevalence counts and genetic population size estimates we date the epidemic emergence of CRF02_AG at 1973.1 (1972.1, 1975.3 95% CI). To infer their phylogeographic signature at a regional scale, we analyze pol and env time-stamped sequence data from 8 countries using a Bayesian phylogeographic approach based on a discrete asymmetric model. Our data confirms a spatial origin of this clade in the Democratic Republic of Congo (DRC) and suggests that viral dissemination to Cameroon occurred at an early stage of the evolutionary history of CRF02_AG. We find considerable support for epidemiological linkage between neighbour countries. Compilation of ethnographic data suggests that well-supported viral migration was related with chance exportation events rather than by sustained human migratory flows. Finally, using sequence data from 15 locations in Cameroon, we use relaxed random walk models to explore the spatiotemporal dynamics of CRF02_AG at a finer geographical detail. Phylogeographic dispersal in continuous space reveals that at least two distinct CRF02_AG lineages are circulating in overlapping regions that are evolving at different evolutionary and diffusion rates. Altogether, by combining molecular and epidemiological data, our results provide a time scale for CRF02_AG, place its spatial root within the putative root of group-M diversity and propose a scenario for the spatiotemporal patterns of a successful HIV-1 lineage both at a regional and country-scale.

Introduction

Since the identification of HIV-1 in 1983, there has been a growing understanding of the emergence of this virus. The closest simian immunodeficiency virus (SIV) to HIV-1 group M (Major) were identified in Pan troglodytes troglodytes that live in the equatorial forests in the southeast corner of Cameroon [1]. However, the highest sequence diversity of group M has been found in the Democratic Republic of Congo (DRC) [2,3,4]. The explosive urbanization of its capital Kinshasa and the high prevalence of genital ulcer diseases have likely created a unique ground for the generation of a plethora of HIV-1 clades within the first part of the XX^th century [5,6], giving rise to what we know today as subtypes and recombinant forms (circulating and unique recombinant forms, CRF/URF). CRFs are mosaic genomes deriving from two different strains that have been found in at least 3 epidemiologically unrelated individuals, whereas URF have only been identified in one patient (http://www.hiv.lanl.gov/). Globally, subtype C causes the highest number of infections (50%), followed by subtype A (12%), subtype B (10%), CRF02_AG (5%) and CRF01-AE (4.8%)[7].

CRF02_AG is currently the predominant clade in the Congo River basin (DRC, Angola, Republic of Congo, Central African Republic, Gabon, Equatorial Guinea and Cameroon) [8], but its epidemiology there is quite diverse. While this clade is nearly absent in the DRC [9], Republic of Congo [10] [11] and Central African Republic [12] [13], the clade accounts for almost 20% of the infections in Angola [14] and Gabon [15]. Importantly, while half of the infections are caused by CRF02_AG in Equatorial Guinea [16], the highest proportion of infections caused by this clade occurs in Cameroon [17] [18,19]. HIV surveillance in Cameroon started in 1985, when prevalence was low, around 0.5% [20]. Since then, extensive HIV/AIDS surveillance studies have revealed an overall increase in the prevalence to around 6% [21] with approximately 60% of the infections caused by CRF02_AG. This proportion has been stable over time and is identical in urban and rural settings [17] [18] [22]. In addition, in Cameroon, above 10% of the viruses are unique recombinant forms [17] [18] and all HIV-1 groups have been found to date (group N [23], group O [24] and more recently, group P [25]).

Although the molecular epidemiology of HIV-1 strains has been extensively studied in the Congo River basin, the spatial dynamics of successful viral lineages circulating within this region remains to be elucidated. Given geo-referenced sequence data, Bayesian phylogeographic models provide a powerful framework for investigating simultaneously the spatial and temporal dispersion of viral populations [26,27]. While the current nomenclature proposes that the CRF02_AG clade is a subtype A/ G recombinant, phylogenetic analyses of non-recombinant fragments showed a distinct history in which the putative subtype G is actually a recombinant form with CRF02_AG/J being its putative parental lineages [28]. Although the recombinant origin of CRF02_AG is still a matter of debate [29], this study focuses on understanding the evolutionary history of the CRF02_AG lineage using solely sequence data that shares an identical phylogenetic behaviour within the CRF02_AG clade.

To reconstruct the spatiotemporal dynamics of CRF02_AG at a regional and country-scale, we use Bayesian phylogeographic diffusion models that take into account uncertainty both at the phylogenetic and viral migration level. First, we inferred the epidemiological dynamics of CRF02_AG integrating molecular sequence data with prevalence counts to reconstruct the temporal origins of this HIV-1 clade. Second, using sequence data with known country of sampling, we show that the CRF02_AG originated in the DRC and highlight the most significant routes of viral dispersal throughout Middle Africa. Finally, given a fine-scale geographic sampling coverage, we introduce relaxed random walks to model the spatial diffusion of HIV viral populations in Cameroon.

Results

A previously published multilocus dataset of 336 gag-pol-env CRF02_AG nucleotide sequences from the same blood-donor population in the two most populated urban centers in Cameroon, Yaoundé and Douala [17] was used to infer the time of origin and effective population size estimates (Ne) through time for this clade. We assume that the majority of the infections were acquired in Cameroon. We focus on the exponentially growth period of the HIV-1 epidemic in Cameroon (1990–2000) to estimate the lag that maximized the maximumlikelihood (ML) fit between the best-fitting mean estimates for Ne (see also supplementary materials, Table S1) and the HIV-1 prevalence counts from 1990 to 2007 in this country [21] to date the time of the most recent common ancestor (tMRCA) of the CRF02_AG clade (Figure 1). The ML estimate of the lag between Ne and prevalence data was 5.3 (4.3–7.5 years, 95% CI), providing evidence for an origin of the HIV-1 CRF02_AG clade in or around 1973.1 (1972.1, 1975.3 95% CI).

Figure 1. Fitting viral effective population size estimates to HIV prevalence data.

Effective population size estimates (Ne*τ, where τ represents the generation time), that were estimated from a 336 multilocus data set from blood-donor patients in Cameroon [17], were fit to the HIV prevalence in the same country [21]. The optimally lagged tMRCA is bounded by its ML confidence interval. The inferred ML estimate of the lag between Ne and prevalence counts is indicated by an arrow.

To analyze the spatial spread of CRF02_AG at a regional scale, we compiled CRF02_AG pol (n=83, L=692 nt) and env (n=37, L=488 nt) molecular sequences from Angola, Democratic Republic of Congo, Central African Republic, Gabon, Equatorial Guinea and Cameroon. We also included nucleotide data from Chad and previously unpublished sequence data from the island nation Sao Tome and Principe. We use two gene regions that according to the current classification of CRF02_AG derive from subtype G and A respectively (http://www.hiv.lanl.gov/). Most probable geographic locations throughout the phylogenetic histories were estimated by applying a discrete asymmetric Bayesian phylogeographic approach, such that dispersal rates from one to another location are allowed to be different from the reversal, thus delivering more realistic scenarios of viral realizations [30]. Additionally, we used a Bayesian stochastic search variable selection (BSSVS) approach to identify most relevant non-zero rates and identify epidemiological linkage between locations [26]. To maximize spatial information embedded in both data sets, we conducted a joint analysis of both pol and env data sets that allowed independent phylogenies to share the same location-exchange matrix. This analysis placed the majority of the posterior root state probability mass in the DRC with posterior root state probabilities of 0.54 for pol and 0.53 for env phylogenies, compared to a prior probability of 0.XX (Figure 2). The second most probable root location for both phylogenies was Cameroon (with posterior root state probabilities of 0.14 and 0.23 respectively). These estimates are robust to the sampling scheme used here since the majority of sequences for both loci were from Cameroon (n=23, 12) and Gabon (n=17, 8). Moreover, the DRC is the location attaining most of the posterior mass when performing the analyses separately for each genomic region albeit with lower support for env (posterior root state probability of 0.70 and 0.31 respectively, not shown).

Figure 2. Phylogeographic origins and spread of CRF02_AG in Middle Africa.

Bayesian maximum clade credibility phylogeographic trees for pol and env datasets. Each branch is coloured according to the most probable location and the legend for the colours is shown on the right. The ancestral root state probability for pol and env data sets is shown on the right of the respective colour codes in grey. Country-codes: CD: Democratic Republic of Congo, CM: Cameroon, GA: Gabon, GQ: Equatorial Guinea, AO: Angola, TD: Chad, CF: Central African Republic, ST: São Tomé and Príncipe.

The reconstructed phylogenies suggest that the earliest migratory events were directed from DRC to Cameroon and Gabon (Figure 2). However, we only find support for epidemiological linkage between the DRC and Cameroon [Bayes factor (BF) comparing a model with a nonzero rate to one with a zero rate of 5.4] (Suchard et al. 2001) (Figure 3). In an attempt to put the viral migration from the DRC to Cameroon into a historical demographic context, the number of migrants living in Kinshasa was investigated for the period 1967 and 1977 (Supplementary File 1). We find that migrants from Cameroon were nearly absent in the capital of DRC and vice-versa (existing statistics for the periods of 1967 and 1976 suggest that these numbers varied between 100–200 people, see Supplementary File 1). This suggests the viral dissemination from the DRC to Cameroon may have been due to chance exportation of the virus rather than sustained human migratory flows.

Figure 3. Most significant epidemiological links of CRF02_AG dispersal in Middle Africa.

Most significant links plotted in Middle Africa. Sequence data from both data sets were used in a combined analysis. The putative root of CRF02_AG emergence, Kinshasa, is highlighted with a dashed circle. Only epidemiological links supported by Bayes factor rates above 5 are indicated. Legend for the strength of the Bayes factor rates is shown on the bottom-left.

Moreover, sequences from the island nation São Tomé and Príncipe, a former Portuguese colony, were found interspersed with sequences from Angola (also a former Portuguese colony), Gabon and Equatorial Guinea (Figure 2). Although this suggests at least three supported independent sources of this clade in São Tomé and Príncipe, the only supported links were from Equatorial Guinea (BF=12.8) and Angola (BF=5.3) (Figure 3). The majority of the sequences from Gabon are descendent from the neighbouring countries Cameroon and Equatorial Guinea, which is confirmed by a high Bayes factor support (BF=16.91 and BF=9.7). Although Cameroon is also bordered by Chad and Nigeria at the north and Central African Republic in the east (and also by the Republic of Congo in the south; yet the presence of CRF02_AG has not been confirmed there [10,11]), we only detected support for viral migration from Cameroon to Gabon and Equatorial Guinea (Figure 3). By 1976, the majority of the foreign population in Cameroon was from Nigeria (n=56.046 from a total of 2.005.223 people), followed by Chad (n=12.176) and Central African Republic (n=7.946) (Supplementary File 1). Human mobility data would suggest viral intermixing between Cameroon and Nigeria. Instead, when performing a similar analysis including sequences from Nigeria, we obtained supported links from Cameroon to Gabon and subsequently to Nigeria (not shown). Altogether, these observations suggest that human mobility alone cannot address the complexity of viral diffusion. Likely, factors such as population growth and accessibility between locations play also an important role in viral spread [31].

Discrete diffusion models offer insights into the origins and epidemiological links within the set of locations from which viruses were sampled. However, given a more fine-scale geographical coverage (n=15 locations sampled from seven out of ten regions in Cameroon), we are able to estimate the unobserved locations of sequence ancestors in continuous space using recently developed relaxed random walk (RWW) models [27]. To first examine whether the Cameroonian sequences show evidence for a single successful viral population, we performed ML analyses including all available Cameroonian sequences with known sampling locations overlapping with the regional pol data set (supplementary materials, Figure S2). The majority of the sequences from Cameroon fell within two supported clusters (n=48 and n=28, named as clusters 1 and 2, supplementary materials, Figure S2). While modelling the diffusion process of the Cameroonian epidemics, Brownian diffusion (BD) models, that assume a constant variance random walk along each branch in the phylogeny, were compared with RRW models, in which dispersion rates are allowed to vary according to distinct prior distributions. In all cases, the RRW models provided a better fit to the data, with the Cauchy distribution attaining the best fit (Table S4). This is consistent with coefficients of variation for both clusters that indicate mean dispersal rates varying among the branches within about 150% of the mean rate (Table 1).

Table 1.

Comparison of the evolutionary parameters between clusters 1 and 2 of CRF02_AG virus in Cameroon. We report posterior mean and 95% HPD intervals except where indicated.

	Cluster 1	Cluster 2
tMRCA	34.1 (30.2, 38.1)	33.8 (30.0, 37.7)
Substitution rate (s/s/y)	1.41 (1.14, 1.71)	0.94 (0.75, 1.16)
Dispersion rate (km/y)	7.51 (5.97, 9.06)	5.41 (3.84, 7.00)
Coefficient of variation	1.58 (1.26, 1.92)	1.38 (1.06, 1.76)
Root area 80% HPD (degrees2)	14.11	10.97
Total 80% HPD area (degrees2)	115.00	12.01

To compare the dynamics of the two CRF02_AG lineages circulating in Cameroon, we consider their evolutionary and geographic diffusion rates (Table 1). Interestingly, the evolutionary rate for cluster 1 are estimated at 1.41 (1.14, 1.71, 95% highest posterior density (HPD) interval) substitutions per site per year (s/s/y) while cluster 2 yields 0.94 (0.75, 1.16, 95% BCI) s/s/y. The values for the diffusion rates were 7.45 (6.00, 9.02, 95% HPD) km per year (km/y) and 5.61 (4.28, 7.13, 95% HPD) km/y respectively. Finally, the dispersal patterns for both lineages circulating in Cameroon were summarized as 95% HPD regions for the roots of clusters 1 and 2 that represent time-slices of the posterior phylogeny distribution (Figure 4). The root location for cluster 1 was inferred in the Centre region, not far from the root location for cluster 2. Importantly, the continuous diffusion inference shows that the CRF02_AG epidemics ignited in the Centre and spread rapidly to the Littoral and West regions in Cameroon, thereafter spreading to the Northeast and Southeast regions; only after that did this clade diffuse to more remote regions of the South and East. These data show that the diffusion of two distinct CRF02_AG lineages ignited in the most populated regions thereafter spreading to more remote regions in Cameroon.

Figure 4. Spatiotemporal dynamics of the CRF02_AG epidemics in Cameroon.

The dispersal patterns of two distinct lineages of CRF02_AG are indicated for 1975, 1990 and 2005. For each cluster, the 95% HPD are of the root location is highlighted by a dashed line. Red-blue lines represent older-recent branches of the MCC trees projected in the surface. Transparent polygons represent the 80% HPD uncertainty on the location of the viral lineage. White-red gradients indicate older-recent age of dispersal. The figure is based on images made available by Google Earth (http://earth.google.com). A dynamic visualization of the spatiotemporal process can be examined at http://www.phylogeography.org/.

Discussion

We have investigated the spatio-temporal dynamics of HIV-1 CRF02_AG in the Congo River basin with particular focus on Cameroon, to provide a better understanding on the origins and spread of this clade within the roots of group-M diversity. The 336 CRF02_AG gag-pol-env sequences from same blood donors in the two most populated cities in Cameroon provided a suitable data set to estimate the dynamics in effective population size (Ne) through time for CRF02_AG in the general population. We subsequently fitted these estimates to the UNAIDS epidemiologic surveillance data to estimate informatively the divergence time of CRF02_AG at 1973.1 (1972.1, 1975.3 95% CI).

By applying Bayesian phylogeographic inference using discrete non-reversible models to pol and env geo-referenced sequences, we investigated the spatial patterns of this clade at a regional level. Our findings suggest that the CRF02_AG clade originated in the DRC, although infections caused there by this clade are rare [9] [32]. We identified the most significant epidemiological links of CRF02_AG within the Congo River basin; these suggest that the high predominance of this clade in Cameroon [17] [18] might be related with at least two chance exportations of the virus from the DRC to Cameroon in the very early history of this clade. Finally, we explore the CRF02_AG spread in greater detail in Cameroon and show that there are two distinct epidemic lineages of CRF02_AG that seem to have ignited in the most urbanized (Centre) region of Cameroon. These lineages have been spreading at distinct evolutionary and diffusion rates, albeit in somewhat overlapping geographic regions.

Recent studies have shown that the effective population sizes estimated from phylogenetic inference should be interpreted in light of the number of new transmissions rather than the number of infected individuals (or prevalence) [33,34]. However, the authors point out that during the exponential period of an epidemic these entities are linearly correlated. Therefore, we restricted the estimation of the ML fit to the exponential growth period of the HIV-1 epidemic in Cameroon (1990–2000) to achieve higher temporal resolution in our tMRCA estimates. The lag between the Ne estimates and the prevalence counts was calculated to be 5.3 years. That Ne estimations precede serological counts has also been noted previously for another fast evolving virus [35]. It is possible that the lag obtained by our estimates reflects the difference in years between the number of new infections and the total number of infected individuals [33,34]. Nevertheless, ML phylogenetic analyses with the published 336 gag, pol and env sequences [17] and the reference set described herein indicated that 5 (1.48%), 4 (1.19%) and 2 (0.59%) sequences from different patients clustered paraphyletically with respect to the CRF02_AG cluster (not shown). Therefore, the inclusion of such sequence data cannot be ruled-out as an explanation for the lag observed between the estimated Ne and prevalence data. Despite this, we obtained the patterns of Ne over time using a multigene analysis that benefits from higher phylogenetic resolution to estimate the phylodynamic patterns derived from blood donors, the population that most resembles the general population as envisaged by the surveillance counts. In addition, the results were qualitatively similar when analyzing each locus individually with gag yielding the closest estimates to the ones obtained using the multilocus data set (not shown). By making use of prevalence counts to infer the tMRCA of the CRF02_AG lineage, the uncertainty on this estimate achieved through our analysis narrows by over 29% compared to (and are included in) the credible intervals delivered by previous estimates [36]. Although CRF02_AG was only identified in 1994 [37], it has been estimated that by this time over 500 thousand people living in the Congo River basin were infected; after, the proportion of infections stabilized [8]. More generally, our results are in line with the timing of this levelling-off, giving support for the divergence time for CRF02_AG estimated here.

Bayesian phylogeographic estimates of the dispersal patterns were applied to two distinct loci. The inference of a spatial root in the DRC was robust to both an analysis sharing a non-reversible diffusion model across unlinked loci or a single gene analysis and also to the sampling scheme used here, since the majority of the sequences were from Cameroon and Gabon. A combined analysis has the potential to more efficiently use the genetic and geographic information in the two loci. In addition, according to the current classification of CRF02_AG the pol and env data sets used here derive from subtype G and A respectively (http://www.hiv.lanl.gov/). Thus, to ensure that parental sequences were not being used we performed a conservative data selection to restrict the phylogeographic inference to sequence data that shared an identical phylogenetic behaviour within the CRF02_AG clade. Nevertheless, and given that the geo-reference sequence data available for analysis is limited, the phylogeographic inference presented here would benefit from a more comprehensive sampling scheme.

For obvious reasons, viral migration has frequently been explained based on human mobility [38]. For example, a study conducted in Yaoundé showed that the risk of HIV infection in men increased up to five times with more prolonged time intervals away from the town [39]. Our findings suggest that viral migration from the DRC to Cameroon occurred in an early stage of the epidemic. Although the human migration patterns within the Congo River basin during the 1970s are difficult to trace, within 1967–77 the number of people living in Kinshasa grew from 901.520 to 2.440.000 and the great majority of migrants living in this city were from Angola (between 14.98 to 10.73% of the total population; see Supplementary File 1). Demographic surveys show that Cameroonian migrants were nearly absent in the DRC and that the reverse was also true (Supplementary File 1), suggesting that this linkage was due to chance exportation of the virus. Within the inferred intervals estimated for viral flow from the DRC to Cameroon, transnational movement could be accomplished either by waterways (mainly through Congo and Sangha rivers), roads (from Ouesso to Bangui in Central African Republic and from there to Bertoua and Yaoundé) or also by air. In addition, it is likely that temporary labour recruitment from Cameroon might have contributed to shape HIV-1 epidemiology in the mineral-rich Gabon and Equatorial Guinea. For instances, the proportion of CRF02_AG infections in miners working in south-eastern Gabon [40] is similar to the observed in the general population of Cameroon and Equatorial Guinea.

The low dispersal rates within Cameroon are inherent to applying these models to a sample that encompasses a limited geographic range for clades that span over 30 years. If viruses from these clades that were potentially exported to other countries had been included, this would have yielded higher dispersal rates. In fact, despite the large overlap, the somewhat wider sampling range within Cameroon for cluster 1 might explain its higher dispersal rates compared to cluster 2. In general, this demonstrates that such estimates are strongly associated with the sampling range and comparisons of dispersal rates across different sampling ranges are likely to be misleading. In addition, the applicability of continuous diffusion models to human viruses may be limited to confined geographic areas because even relaxed random walk models might be poor approximations for viral diffusion across large geographic ranges. The concomitant differences in evolutionary rates between the clusters remain more difficult to explain. In the absence of information concerning risk groups for the data used in this study we can only speculate that different transmission dynamics in distinct risk groups might be responsible for the rate differences [41]. Alternatively, it may be that the two clusters are spreading in different ethnic groups that have inherently different genetic backgrounds and also distinct mobility patterns. Cameroon is home to an extensive diversity in languages and ethnicity. Analysis of sequence data referenced with patient ethnic information intersected with the complex mobility patterns may provide some insight into viral spread in Cameroon.

Our study sheds light in the emergence and dynamics of an important HIV-1 clade in the Congo River basin, the source location for the HIV-1 group M diversity. The evolutionary history of human viruses can only be fully understood when the intrinsic spatial and temporal components are taken into account. Importantly, understanding the origins and dispersal patterns of successful HIV-1 clades both at a regional and intra-country level not only unites the fragmentary pieces delivered by serological counts but should ultimately become central to improve the characterization and control of HIV spread.

Materials and Methods

Nucleotide datasets compilation

A multilocus alignment of 336 gag (HXB2: 1255 to 1682), pol (HXB2: 4228–5093) and env (HXB2: 7890–8266) comprising CRF02_AG published gene sequences sampled between 1996 and 2004 from blood donors from Yaoundé and Douala [17] was used to investigate in detail the demographic dynamics of the CRF02_AG lineage. Pol (HXB2: 2253 to 2944) and env (HXB2: 7038 to 7463) HIV-1 CRF02_AG gene sequences sampled in Middle Africa as defined by United Nations (Chad, Cameroon, Central African Republic, Equatorial Guinea, Gabon, Democratic Republic of Congo, Angola and São Tomé and Príncipe) were selected from the 2010 LANLdb (http://www.hiv.lanl.gov/) to investigate viral migration patterns at a regional level. To investigate viral migration within Cameroonian locations, we compiled pol HIV-1 CRF02_AG gene sequences (HXB2: 2253 to 3275) with known date of sampling and detailed geographical location (city or village) from the 2010 LANLdb database (http://www.hiv.lanl.gov/) (for details see Table S1). Sequence data belonged to 7 out of the 10 regions of Cameroon. Although there is no available data from the Adamawa, North and Far North, these regions have the lowest HIV-1 prevalence in the country [21].

Subtype assignment and sequence alignment

A recent study has demonstrated that 4.9% of the original subtype assignments in the LANL database need a revision [42]. Therefore, and because the effect of recombination in phylogeographic inference might affect the results [43], we conducted a stringent procedure to ensure that the data used for the phylogeographic analyses were closely related to the CRF02_AG clade. First, we excluded all sequences that did not cluster monophyletically within CRF02_AG cluster using a reference set with all available full genome sequences from the closest phylogenetic clades to CRF02_AG [28] sampled worldwide from subtypes A (n=29), sub-subtypes A1 (n=87) and A2 (n=3), subtype G (n=26), CRF02_AG (n=27) and subtype J (n=4). Sequence alignments including the reference set and i) the Middle African, or the ii) Cameroonian dataset were created using the multiple alignment with the fast Fourier transform algorithm (MAFFT) [44] and manually edited with Se-Al (http://tree.bio.ed.ac.uk). The overlapping regions were then used to perform phylogenetic analyses under a general time reversible model with discrete gamma and invariant among-site rate variation (GTR+4Γ+I) using a ML criterion implemented in PhyML with an approximate likelihood ratio test statistics to search for the best of NNI and SPR tree [46]. Only sequences that grouped monophyletically within the CRF02_AG clade with significant statistical support (bootstrap values above 75%) were used for subsequent analyses. In addition, the subtype assignment was confirmed using the NCBI HIV subtyping tool (http://www.ncbi.nlm.nih.gov/projects/genotyping/formpage.cgi) and the Rega subtyping tool v.2 [47] [48]. After the subtype assignment steps, Middle African datasets comprised 77 pol and 59 env sequences with 691 and 488 nucleotides, spanning the genomic regions 2252 to 2943 and 7037 to 7525 (HXB2). These were sampled in Cameroon (23, 12 for pol and env respectively), Chad (3, 9), Gabon (17, 9), Democratic Republic of Congo (5, 10), Central African Republic (3 pol), Guinea- Equatorial (14 pol) and Angola (6 pol). The Cameroonian dataset consisted in 80 pol sequences (HBX2: 2252 to 3319) from a total of 15 locations. In particular, the sequences originated from the Centre (22), Northwest (7), East (18), West (7), Littoral (10), Southwest (6), and South (9) regions. All data sets for phylogeographic analyses tested negative for inter-subtype recombination using the Phi-test implemented in SplitsTree4 version 4.10 [49]. All sequence alignments are available from the authors upon request.

Timing the introduction of CRF02_AG in Cameroon

To estimate the changes in the effective population size Ne through time for CRF02_AG in Cameroon, we used a Bayesian coalescent approach as implemented in BEAST [50] with BEAGLE to enhance computational speed [Suchard and Rambaut, 2009]. For this purpose, we analyzed 336 gag, pol and env HIV-1 CRF02_AG gene sequences [17] using either a multilocus approach or separate analyses under a GTR+4Γ+I substitution process. We used the uncorrelated lognormal molecular clock model that allows rates to vary among the branches of the inferred phylogenies [51]. For the multilocus analysis we tested constant, exponential and constant-logistic demographic tree prior models. Nucleotide substitution models, molecular clock models and the demographic model were shared among the different partitions, whereas each partition was allowed to have different phylogenetic histories and different coefficients of variations for the molecular clock model [52]. Demographic model fit was assessed by comparison of marginal likelihoods [Suchard et al., 2001 and 53] (supplementary materials, Table S1). For the single loci analyses we used the the semi-parametric Skyride demographic tree prior with time aware smoothing [54]. Markov chain Monte Carlo simulations were run for 10 to 25 × 10⁷ chain steps, sub-sampling parameters every 5000 steps. Convergence of the chains was inspected using Tracer.v.1.5. To estimate accurately the introduction of CRF02_AG in Cameroon, we assess the fit of the posterior mean estimates of Ne from BEAST to the HIV-1 seroprevalence data in a similar fashion to previously described for dengue virus [35]. Particularly, the lag between the prevalence counts and the estimates using virus sequence data was selected in order to maximize the likelihood of a linear regression related the counts to Ne translated by a unknown amount of time in the R package. The optimally lagged estimation of the divergence time of CRF02_AG was used as a calibration prior in subsequent analyses.

Bayesian phylogeographic models

Bayesian phylogeographic analyses [Lemey et al., 2009, Lemey et al., 2010] were carried in under an MCMC framework as implemented in BEAST [50] with BEAGLE [Suchard and Rambaut, 2008]. A Bayesian skyride tree prior was used as a coalescent demographic model with time-aware smoothing [54]. The identification of significant migration pathways was performed using discrete non-reversible diffusion models and a BSSVS approach [26] [30]. For the discrete diffusion models, geographic locations were recorded at the tips of pol and env phylogenies respectively. The unobserved locations of the ancestral nodes until the root were inferred for a posterior distribution of trees. Non-reversible models provided a better-fit then reversible diffusion models as confirmed by a higher BF support (Suchard et al, 2001). To reconstruct the evolutionary history of CRF02_AG in Cameroon explicitly in continuous space, we propose models of continuous diffusion [27]. Exact latitude and longitude for each viral isolate were recorded at the tips of an unknown phylogeny, and the unobserved two-dimensional locations along each node of the posterior distribution of the phylogeny were estimated. Gamma, Cauchy, Lognormal and homogeneous prior distributions to rescale the variance of the random walk were tested and posterior estimates for the parameters were compared (supplementary materials, Table S5).

Significant non-zero rates obtained by the BSSVS approach were spatially projected and converted into a keyhole markup language (KML) file, which can be viewed with Google Earth (http://earth.google.com), are available upon request. The animated continuous phylogeographic descriptions of the epidemics in Cameroon are available at http://www.phylogeography.org/.