A worldwide genomic sampling of P. vivax

In this study, we sequenced the whole genome of 214 new P. vivax strains from Central and South America, West, Central and East Africa, Middle East, South Asia, and Oceania (for more details see Materials and methods; S1 Table). Then, these new genomes were combined with 919 modern genomes from the literature [27–29] (see Materials and methods for more details; S1 Table and S1 Fig) and one ancient sample (Ebro) from Spain sequenced by van Dorp et al. [15]. After filtering out samples with >50% missing data (n = 68), those with multi-strain infections (i.e., multiple P. vivax genotypes, n = 320; S2 Fig), and those excessively related (n = 124), the final dataset included whole genome sequencing data of 619 modern P. vivax samples from 34 countries, one ancient DNA sample (Ebro) from Spain (Fig 1 and S1 Table) and two genomes of P. vivax-like from Cameroon, isolated from chimpanzees [28] and used here as outgroup. The mean sequencing depth ranged from 2.88 to 1,484.71 X for the modern genomes and the mean sequencing depth was 1.13 X for Ebro (S2 Table). The details about the bioinformatic pipelines, data processing, and the number of single nucleotide polymorphisms (SNPs) used are described in Materials and Methods and in S1 and S3 Figs. Some of our analyses considered the genotype likelihood using the ANGSD software ecosystem [30], rather than the classic SNP calling followed by hard quality filtering, to account for the genotyping uncertainty related to the sequencing depth heterogeneity. This strategy using ANGSD allowed considering 20,844,131 sites across the 620 genomes analyzed. Nevertheless, as some analyses required actual SNP calling followed by appropriate quality filtering, a Variant Call Format (VCF) dataset composed of 950,034 SNPs was also generated (see the Materials and methods and S1 and S3 Figs for more details).

Genetic relationships of the American P. vivax with worldwide populations

To explore the genetic relationships of Latin American P. vivax isolates with worldwide P. vivax strains, the population structure patterns of 620 P. vivax isolates collected in 36 countries (Fig 1) were investigated using complementary population genomic approaches (see the Materials and methods and S1 Fig): (i) a principal component analysis (PCA) using PCAngsd [31], (ii) a model-based individual ancestry analysis using PCAngsd [31], and (iii) a maximum likelihood (ML) phylogenetic tree using IQ-TREE [32]. All three approaches showed consistent results splitting first the genetic variation geographically into four distinct genetic clusters (Fig 2): (1) Oceania, East and Southeast Asia, (2) Africa, (3) Middle East and South Asia, and (4) Latin America.

The first three principal components (PC) of the PCA (based on the genotype likelihood of 105,527 filtered and unlinked SNPs) revealed these four major genetic clusters (Figs 2a and S4). PC1 separated Oceania, East and Southeast Asia from Africa, the Middle East, South Asia, and Latin America. PC2 splitted American populations from all the others. African populations split from all the other populations on PC3. The same genetic structure with four genetic clusters was also recovered from the ML tree (Fig 2b). P. vivax populations from East and Southeast Asia, Bangladesh, and Oceania split first and cluster closer to the outgroup and apart from those from the rest of the world. Populations from Latin America, Middle East-South Asia and Africa clustered in three distinct lineages; with the last two being more closely related to each other than with the American lineages. Patterns of individual genetic ancestry (Figs 2c and S5) were also in line with results of the PCA and ML tree, displaying again four distinct genetic clusters. As the number of genetic clusters tested (K) increased, the four different genetic clusters also became distinct in terms of genetic ancestry. All the populations from East and Southeast Asia, Bangladesh and Oceania split first from the others at K = 2. Then, populations from Latin America split from Middle East-South Asia and from Africa, which became differentiated only at K = 4 (Fig 2c).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Worldwide genetic structure of P. vivax.

(a) Principal component analysis of 619 modern P. vivax strains and the Ebro ancient DNA sample from Spain, showing the first three PCs based on the genotype likelihood of 105,527 unlinked SNPs (see S2 Fig for more details). (b) Maximum likelihood (ML) tree of the 620 P. vivax individual genomes obtained with IQ-TREE [32] based on a general time reversible model of nucleotide evolution [33], as determined by ModelFinder [34]. The ML tree includes two P. vivax-like strains from African great apes indicated by an asterisk (*) used to locate the root. Note that the length of the outgroup branch was truncated. Black dots at nodes correspond to highly supported nodes (SH-aLRT ≥ 80% and UFboot ≥ 95%). The reference mark (※) highlight a Brazilian P. vivax isolate, discussed in the text. (c) Individual genetic ancestry assuming K = 4 and K = 9 distinct genetic clusters estimated using PCAngsd (see S5 Fig for the other K values). (d) Geographic distribution of the population average genetic ancestry proportions at K = 9 estimated using PCAngsd. PNG: Papua New Guinea. The base layer of the map was made with Natural Earth (naturalearthdata.com).

https://doi.org/10.1371/journal.ppat.1012811.g002

Data analysis at a finer sub-structuration level at K = 6 (the best K value based on the broken-stick eigenvalues plus one, following the recommendations of Meisner and Albrechsten [31]) (Figs 2c and S5) showed that Latin America remained as a single genetic cluster well separated from the rest of the world. Asia displayed three clusters (South Asia-Middle East, East Asia-Southeast Asia, and Oceania-Malaysia) and Africa was split into East and West Africa. Noteworthy, several American populations (Honduras, Ecuador, Brazil, Venezuela, French Guiana, and Guyana) shared genetic ancestry with the cluster that contains West African isolates (in green, Fig 2c).

An additional sub-structuration among the American P. vivax populations was observed at K = 9 (Fig 2c). Specially, two distinct genetic clusters were identified, as previously observed [13,28,35]: (1) isolates from Mexico, Honduras, and Colombia gathered together as a Central American group, and (2) isolates from the French Guiana, Guyana, and Venezuela formed another distinct cluster as an Amazonian group (Fig 2d). The other populations from Brazil, Peru and Ecuador displayed patterns of admixed genetic ancestry between these genetic clusters.

In Africa, P. vivax populations were distributed in two distinct genetic clusters, one composed of the Mauritanian and Cameroonian samples, and the other including samples from Ethiopia and Eritrea. Sudan and Central African samples displayed admixed genetic ancestry between these African clusters (Fig 2d).

In Asia, a subset of the Malaysian population (n = 16) exhibited no evidence of admixture with other populations. These Malaysian samples were closely clustered in the ML tree (Fig 2b) and they were separated from the other Asian populations on PC 4 (S4 Fig). This pattern has already been reported before in this region [36] and it could be due to a strong bottleneck following a sharp decline in P. vivax prevalence in Malaysia over the past decade [37,38], which may have increased the population genetic drift.

In the PCA, the ancient DNA sample from Spain (Ebro) was genetically closer to the American populations than to any other population (Fig 2a), with which it shared a similar ancestry pattern (Fig 2c and 2d). In the ML tree, it clustered with one Brazilian sample (indicated with this symbol ※ in Fig 2b), the oldest sample from the Latin America dataset, collected in 1980 [29]. Similarly, in a previous study based on mitochondrial DNA, it clustered with South American samples [14]. Nevertheless, the nodes clustering these individuals were poorly supported in the ML tree (Fig 2b), preventing any strong interpretation about the origin of Ebro.

Overall, the population structure analyses (PCA, ancestry plots, and ML phylogenetic tree) supported the hypothesis that the Latin American P. vivax populations are genetically closer to the ancient DNA Ebro isolate from Spain, followed by the present-day populations from Africa and from Middle East & South Asia. Noteworthy, both the American and Spanish Ebro genomes displayed some admixed genetic ancestry with other populations from the rest of the world. For example, at K = 6 and 9 in the Admixture analysis (Fig 2c and 2d), Ebro shared approximately 50% of genetic ancestry with the American populations (light blue), but the rest originated from populations located in West Africa and Middle East (respectively light green and brown). Some American populations (Honduras, Ecuador, Brazil, Venezuela, French Guiana, and Guyana) also displayed shared genetic ancestry, particularly with isolates from West Africa.

Population branching and admixture between American P. vivax and the rest of the world

To further investigate potential admixture events and identify candidate source populations that could have contributed to P. vivax colonization of Latin America, we first estimated population admixture graphs using two complementary approaches implemented in TreeMix [39] and in AdmixtureBayes [40].

The TreeMix analysis determined that the optimal number of migration edges was two or five (S6 Fig), while AdmixtureBayes analysis converged toward an optimal solution with two admixture events (S7 Fig). Both methods recovered the four genetic clusters (Fig 3) identified by the genetic structure analyses (Fig 2): (1) Oceania, East and Southeast Asia, (2) Africa, (3) Middle East and South Asia, and (4) Latin America.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Population graphs describing the genetic relationships and admixture proportions between P. vivax populations.

(a) Population tree estimated using TreeMix for a subset of 18 P. vivax populations with two migration edges (orange arrows) and rooted using the two P. vivax-like genomes indicated with the asterisk (*). The scale bar shows ten times the mean standard error (s.e.) of the covariance matrix. The migration weight is indicated as a percentage in orange on the arrows. (b) Network topology of a subset of 18 P. vivax populations with the highest posterior probability obtained with AdmixtureBayes, rooted with the two P. vivax-like genomes (*). The branch length indicates the genetic divergence between populations (measured by drift), multiplied by 100. The percentages in the nodes are the posterior probability that the true graph has a node with the same descendants. For each admixture event (indicated by the dotted arrows) the percentages illustrate the admixture proportion. PNG: Papua New Guinea.

https://doi.org/10.1371/journal.ppat.1012811.g003

The American populations grouped into a single highly supported monophyletic cluster (≥75% bootstrap support in TreeMix, Fig 3a, and ≥98% posterior probability in AdmixtureBayes, Fig 3b). This suggests that American populations most likely originated from a single introduction or from multiple introductions from genetically similar, but possibly admixed source populations from the rest of world. This second hypothesis is supported by the poorly resolved branching pattern close to the root in the TreeMix and AdmixtureBayes analyses (Fig 3). The branching patterns obtained in both population graphs showed a finer substructuration of the American populations (Fig 3): one cluster grouping Mexico, Colombia, and Honduras populations, another one with the French Guiana, Guyana, Venezuela and Ecuador populations, and in between, Brazil and Peru populations branched at an intermediate position reflecting their admixed genetic ancestry observed in the genetic structure analyses (Fig 2).

The ancient Spanish sample (Ebro) stood out as the isolate most closely related, but genetically distinct, from the American cluster, branching at its base. This ancient isolate had a notable long branch that reflected important genetic drift compared with other present-day populations. Such a long branch may be an artifact because it is a single haploid sample from an ancient population [41]. As rare and common alleles in that ancient population can not be properly estimated, this usually leads to an inflated estimate of the genetic drift as previously reported [41]. Additionally, the TreeMix and AdmixtureBayes analyses detected an admixture event between Ebro and Oceanian populations, corroborating the shared genetic ancestry identified in the population genetic structure analyses (Fig 2d and 2c). The TreeMix analysis also identified a migration edge from the common ancestor of India and Afghanistan into Thailand.

To assess further potential admixture events between the American populations and those from the rest of the world, the admixture f₃-statistics were computed. This involved testing American populations against potential source populations from both within Latin America and from all other geographical regions worldwide. None of the tested f₃ combinations yielded a significantly negative result, as expected if admixtures occurred between the American populations and one of the two tested sources (S8 Fig). As underlined by the authors of the f₃ admixture test [42], these non-significant results do not preclude that admixture could have occurred in a distant past, with a residual signal remaining only in a minor fraction of the genome. This would be also expected if the source population(s) that founded the Latin American populations were already admixed.

Effective population size changes and split time of American P. vivax populations

The next step was to determine whether colonization of Latin America by P. vivax was associated with founder effects and population bottlenecks, as can be expected when colonizing a new environment. To this aim, the nucleotide diversity (π) as well as the Tajima’s D values [43] were computed in the core genome, considering only populations with at least 10 samples. The π values in the American populations (S9 Fig) were marginally lower compared with those observed in other populations, with exceptions in Malaysia and Ethiopia. In Latin America, the π values displayed distinct patterns: populations along the Pacific coast (Colombia and Peru) exhibited significantly lower median π values than populations in the Amazon (Brazil, French Guiana, and Venezuela) (~8×10⁻⁴ versus ~1×10⁻³, p-value <0.001, Wilcoxon signed-rank test and Bonferroni correction, S9 Fig). The Tajima’s D distributions in the same P. vivax populations were largely negative, indicating excess of rare variants over shared variants, consistent with demographic expansions. However, the Tajima’s D distribution of populations from Colombia, Peru, and Ethiopia were closer to zero, with a notable presence of positive values (S9 Fig). Overall, these descriptive statistics are consistent with the demographic expansion of the American populations. They also suggest that no strong bottleneck occurred during P. vivax invasion history of Latin America or that the founding isolates were somewhat admixed and harbored genetic diversity comparable to that of the source populations.

The changes in effective population size (N_e) with time were inferred using the coalescence rates (CR) estimated with the software Relate [44] for modern samples. Consistently with the π and Tajima’s D values, these analyses showed that all modern populations across the world displayed similar trends, characterized by a monotonic decline in Ne values that started roughly ~100,000 generations ago (~20,000 years ago, assuming a generation time of 5.5 generations/year [28] and a mutation rate of 6.43 x 10⁻⁹ mutations/site/generation [28]) followed by a recent and moderate expansion (Fig 4a). The recent growth observed in P. vivax populations around the world might be due to historical bottlenecks, possibly caused by climate changes, human migrations, or development of host immunity. Earlier studies have shown that P. vivax populations expanded after these bottleneck events [45,46].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Coalescent-based inference of the demographic history of P. vivax populations.

(a) The variation in effective population (N_e) size was estimated using Relate (axes are log₁₀ transformed). The period highlighted by the blue rectangle corresponds to the diversification of the Latin American populations. (b) Comparison between the inverse coalescence rates (ICR) and inverse cross-coalescence rates (ICCR) from Relate between Latin America (Colombia on top, French Guiana below) and Eurasia/Africa (left to right: Thailand, Papua-New-Guinea, Mauritania) provides a formal estimate of their split time [46]. The axes are log₁₀ transformed. The periods highlighted by the colored rectangles correspond to the divergence time (Tdiv) between populations and are also specified in the panels. (c) Comparison of the ICR and ICCR between Latin America (Colombia on top, French Guiana below) and Ebro. The 95% confidence interval resulting from 100 bootstraps inferred using Colate is indicated by the envelope surrounding the estimate line. The black dot indicates Ebro sampling time. Axes are in log₁₀ scale units. The period highlighted by the light-yellow rectangles corresponds to the divergence time (Tdiv) between the American populations and Ebro and is specified. PNG: Papua New Guinea.

https://doi.org/10.1371/journal.ppat.1012811.g004

Noteworthy, all American P. vivax populations displayed slightly lower Ne values than the other populations, and only three populations from Africa and Oceania showed comparable trends (Mauritania, Malaysia, and Papua New Guinea). The Ne values of all American populations became noticeably distinct from the other populations since the last ~500 to 2,000 generations (~100–300 years ago). This time range suggests a putative divergence time of the American populations from the other populations, but also with each other. Besides the trees (Figs 2b and 3) where the American populations formed a single well-supported cluster, the synchronous departures of the Ne values of the American populations, compared with those from other populations worldwide, also suggests a single introduction from one source or multiple introductions from genetically similar and possibly admixed sources in Latin America.

The divergence time estimates were refined using the inverse cross-coalescent rate (ICCR) between representative populations from the American clusters (Colombia and French Guiana) and other populations in the world that are representative of distinct genetic clusters (Figs 2 and 3), as well as the ancient isolate from the Ebro region in Spain. This analysis used the program Relate [44] for modern samples (Fig 4b) and the program Colate [47] for the ancient Ebro sample (Fig 4c). The split time between populations can be estimated by comparing the changes in ICCR and ICR values between populations and within population respectively. The split time corresponds to the time when the ICR values within each population depart from each other, and the ICCR values between populations increase. This means that the CR between populations decreases, and therefore they are diverging [48]. According to the ICCR variation (Fig 4b), the divergence times estimated between the American populations and Asian (i.e. Thailand) or Oceanian (i.e. PNG) populations occurred more than 1,000 ybp (or more than 55,000 generations ago). The divergence time between populations from West Africa and Latin America dated back 500–1,000 ybp (~3,000–5,500 generations ago). Lastly, the split between the American populations and the ancient Ebro isolate from Spain would have been more recent (100–300 ybp or 550–1,650 generations ago).

The role of West African and European P. vivax populations in the invasion of Latin America

Our analyses of the population structure and relationships between populations indicated that American P. vivax populations form a highly supported monophyletic cluster (Figs 2b and 3). These results and the concomitant divergence of all American populations (Fig 4a) suggested that they descend from a single introduction from one source or from multiple introductions from genetically similar and possibly admixed sources. The ancient Ebro sample from Spain was the isolate most closely related to the American populations. It displayed an admixed genetic ancestry, half of which was shared with the American P. vivax populations, and the other half originated from multiple sources (Africa, the Middle East, and even Asia) (Fig 2c). Nevertheless, West African P. vivax populations may have also contributed to the genetic make-up of the Latin American populations. This is suggested by the recent human history that involved massive population movements between Europe, West Africa and Latin America [17]. This is also supported by the genetic ancestry analyses showing that to some extent, West African populations have contributed to the genetic ancestry of P. vivax populations in Latin America (Figs 2c and S5). In the literature, analyses using mitochondrial DNA [12] or microsatellites [13] similarly indicate a likely West African influence, introduced during the transatlantic slave trade. However, no evidence of significant admixture was detected in the population graph analyses between American and African populations (Figs 3, S6 and S8). Furthermore, American populations did not share ancestry with Asian or with Oceanian populations (Fig 2c), and the divergence times with Asian and Oceanian populations were much older than with West African or European populations (Fig 4). The European ancient Ebro isolate and the West African populations were the only groups from Eurasia and Africa that showed credible evidence of shared genetic ancestry with the American populations. Therefore, it was important to test formally their putative roles as source populations in the P. vivax invasion history of Latin America.

We used the simulation-based approximate Bayesian computation (ABC) statistical framework that relies on the supervised machine learning Random Forest tree classifier (ABC-RF), as implemented in DIYABC-RF [49]. This approach allowed formal statistical comparisons of distinct P. vivax invasion history scenarios of Latin America that differed in terms of population branching order, admixture events, time of split and admixture, and that involved or not unsampled (or ghost) populations (see Fig 5a and Material and method for details). In an ABC-RF analysis, different demographic scenarios (see parameters in S3 Table) are tested by simulating genetic data under each scenario. Then, summary statistics from the simulated data under each scenario are computed and used to train the ABC-RF tree classifier to assess which scenario best fit with the observed data [49,50]. This analysis considered twelve colonization scenarios that involved three representative populations: one from West Africa (Mauritania), one from Europe (the Ebro sample), and one from America (either Colombia or French Guiana). Colombia and French Guiana were selected as representative populations because they had the largest non-admixed sample sizes within each Latin American subgroup (n = 30 for Colombia and n = 14 for French Guiana). To determine whether these subgroups shared the same invasion history in Latin America, we conducted the analysis twice, each time considering one of the two American populations as representative. An “ghost” population was also introduced in scenarios 5 to 9, 11, and 12 to assess the potential contribution of a unsampled population, which can be also extinct (Fig 5a). Scenario 2 considered only a European origin; scenarios 3, 4 and 12 considered only an African origin; and scenarios 1, 6, 7, 8, 9, 10 and 11 considered various combinations of admixed histories.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Most likely P. vivax colonization history scenarios of Latin America inferred using the Approximate Bayesian Computation Random Forest approach.

(a) Twelve colonization scenarios were tested. Distinct solid colored lines correspond to populations with distinct effective size parameters (labels starting with “N”) and horizontal dashed lines represent admixture events with a contribution ra and 1-ra from each of the two contributors. Time parameters of the different events are displayed along the vertical line on the right of each scenario (labels starting with “t”, not to scale). The parameters of each scenario are characterized by probabilistic distributions detailed in S3 Table. Under each scenario is the mean percentage of RF classification votes ± standard deviation (SD) and the mean type II error ± SD. The three most supported scenarios (scenarios 6, 4 and 8, with 46.0% of the total votes) are highlighted by dotted-line rectangles. (b) The best-supported scenario (scenario 6) in the DIYABC-RF analysis, scaled to relative time-parameter estimates (converted to years assuming a generation time of 5.5 generations/year). On the left, each time parameter estimate is indicated by the mean (circle), median (diamond), and 90% confidence intervals (colored bars). Here, the results are only shown considering the Colombian population as a Latin America representative population. Results considering the French Guiana population were very similar (see S4 Table).

https://doi.org/10.1371/journal.ppat.1012811.g005

In both analyses, whether considering the Colombian or French Guianese American representative population, the most likely scenario identified was scenario 6 (Fig 5). This scenario suggests that the American P. vivax populations originated from an admixture event between the historical European population and an unsampled population derived from an ancestor to the European gene pool that branched with West African ancestral populations. This scenario received the highest proportions of ABC-RF classifier votes and this result was consistent over 10 independent replicate analyses (mean RF vote proportions ± standard deviation = 18.4 ± 1.0% with Colombia and 21.90 ± 1.06% with French Guiana) (Fig 5a and S4 Table). The mean posterior probability for this best supported scenario was 46.2 ± 1.3%. After scenario 6, the two other best scenarios supported by the ABC-RF analysis were Scenarios 4 and 8 with Columbia (mean RF vote proportions = 14.4 ± 0.5% and 13.2 ± 0.9%, respectively) and scenario 8 and 11 with the French Guianese American representative population (mean RF vote proportions = 15.03 ± 1.02% and 12.21 ± 0.94%). Scenario 4 assumes that Ebro could be a migrant from Latin America, while American populations would descend from an ancestral population of West African origin. In scenario 8, the American populations would have derived from an admixture event between the European populations and an unsampled population of direct West African origin. Thisscenario is actually close to scenario 6. Combined, the three best scenarios (6, 4 and 8) identified with Columbia collected 46.0% of all the votes, while each of the other nine scenarios received less than 8.9% of all votes (Fig 5a). Scenario 11 identified as the third best scenario when considering French Guiana was actually a more complex version of scenario 6 and 8 combined together. It assumed that the American populations could result from two successive admixture events, the first and oldest one between the European populations and an unsampled ghost population, followed by a more recent wave of introduction from West Africa (Fig 5a). The combined RF-votes for these three best scenarios (6, 8 and 11) identified with the French Guinanese population was 49.14%(S4 Table). In both analyses, ten replicates of the ABC-RF analysis supported strongly and unanimously this result. Furthermore, the confusion matrix (S10 Fig) and the prior error rate of 42.1 ± 0.0% indicated the overall performance of the RF classification analysis and highlighted high power to distinguish among the twelve competing scenarios.

Model parameter estimates for the optimal scenario 6 indicated that the average divergence time between the ancestral European (Ebro) population and the West African population (in green in Fig 5b) was ~350 years ago (or ~2,000 generations ago). Similarly, the average estimated divergence time between the unsampled (ghost) population and Ebro (in yellow in Fig 5b) was ~400 years ago (or ~2,100 generations ago). Therefore, the ABC-RF analysis suggested that the two split times between West Africa, Europe and the unsampled population were nearly synchronous. These split time estimates from the ABC-RF are also very close to those from the CCR estimates obtained using Relate and Colate between the American populations and those from Europe (Ebro) and West Africa (Fig 4). Furthermore, the admixture event between the unsampled population and the European population (in yellow in Fig 5b) most likely occurred ~200 years ago (or ~1,200 generations ago), with a predominant European contribution (ra = 60%).

Altogether, these ABC-RF results support the hypothesis that American P. vivax populations descended from now-extinct European populations, but also from unsampled population(s), possibly originating from West Africa, an event that happened during the post-colonization migration of the American continent (second half of the 19th century).

Plasmodium vivax genetic structure and diversity in Latin America

The analysis based on PCA, ancestry plots, ML phylogenetic tree, and population admixture graphs revealed a consistent genetic structure among the Latin American P. vivax populations within the Americas and as well as to P. vivax populations worldwide, as previously described [29,36,49]. As previously reported [13,28,35], in Latin America, two separate genetic clusters were identified, representing the Central American region (Mexico, Honduras, and Colombia) and the Amazonian region (French Guiana, Guyana, and Venezuela). Genetic structure analyses also indicated the presence of an admixed cluster, thus suggesting potential gene flows between these clusters in Latin America. Several studies confirm cross-border spread between South American countries [13,35,51]. These results highlight the possibility of gene flows between Peru, Brazil, Guyana, Venezuela, and French Guiana that could be explained by the large asymptomatic human reservoirs and the dormant stage characteristic of P. vivax [52–54].

The geographic distribution of genetic diversity across the Latin American region revealed that P. vivax populations along the Pacific coast (Colombia and Peru) exhibited lower genetic diversity (π) compared with Amazonian populations (Brazil, French Guiana, and Venezuela) (S9 Fig). This is in line with previous studies [35,55], and could stem from transmission pattern variations across regions [35]. Indeed, coastal areas experience more periodic malaria transmission events compared with the Amazon region, where transmission is more stable throughout the year [56,57]. However, some studies found that P. vivax populations can exhibit high genetic diversity during periods of low transmission rates [55,58]. This means that data on spatial variation in genetic diversity must be interpreted with caution.

Shared genetic ancestry between P. vivax from Latin America, Europe, and West Africa

The population genetic structure analyses (PCA, ancestry plots, and ML phylogenetic tree) showed that P. vivax populations from Latin America are genetically closer to the ancient DNA Ebro isolate from Spain, followed by present-day populations from Africa, Middle East and South Asia (Fig 2). Noteworthy, both the American isolates and the Ebro isolate displayed some admixed genetic ancestry with other populations from the rest of the world. At first glance, this may suggest that the European P. vivax may have been the only genetic source involved in the colonization of the American continent, as suggested in some studies [15,16]. Nevertheless, this does not rule out the possibility that other populations may have also contributed, especially since various demographic histories may result in similar admixture patterns [59]. When considering the admixture analysis, >50% of the Ebro genetic ancestry was shared with the American populations (light and dark blue), but the rest originated from West African and Middle Eastern populations (light green and brown, Fig 2c and 2d). In addition, some American populations (Honduras, Ecuador, Brazil, Venezuela, French Guiana, and Guyana) also displayed shared genetic ancestry with those from West Africa. These results support European participation in the genetic make-up of Latin American P. vivax populations, but also indicate that the history of the invasion of Latin America could have been more complex with the potential involvement of West African genetic sources.

No genetic evidence of an Asian contribution to the Latin American P. vivax populations

Previous studies suggested that Asian P. vivax populations might have contributed to the colonization history of Latin America. For instance, a study based on mitochondrial genomes found evidence of genetic ancestry from Oceanian P. vivax strains in American populations [12]. Our genomic analysis also suggested a potential introgression from Indonesia to Mexico, as shown by the TreeMix results when considering up to m = 5 migration edges (the second optimum, see S6 Fig). However, this introgression appears to be minimal, close to 0% (S6 Fig), and was not corroborated by the f₃-statistics (S8 Fig). This introgression could have been the consequence of some historical introductions of Asian P. vivax during human migrations, but it could also mirror part of the history of the European P. vivax populations, that were the main sources of American P. vivax. Indeed, the population graph analysis of TreeMix and AdmixtureBayes (Fig 3) detected admixture events that involved Ebro and the Oceanian populations (Papua New Guinea and Indonesia). These admixture events may be due to genetic exchanges between European and local parasites from the Philippines, an archipelago of islands close to Indonesia that was a Spanish colony for more than three centuries (from 1565 to 1899). The reported contribution of Melanesian P. vivax isolates to American P. vivax genomes may reflect historical introgression, admixture events, or incomplete lineage sorting within the Spanish genomes. Therefore, the likelihood of Asian P. vivax serving as genetic sources in the colonization of Latin America appears relatively low, and possibly also integrated within the history of European populations.

Latin American P. vivax populations shared genetic ancestry with the now-extinct European population, but not only

As the Ebro sample showed closer genetic ties with American P. vivax populations than with Middle East populations (Figs 2 and 3), several scenarios where Ebro came from Latin America (scenarios 4, 5, 10 and 12 in Fig 5a) and scenarios where Ebro contributed to the genetic make-up of Latin America (scenarios 1, 6, 8 and 11 in Fig 5a) were tested. The best scenario (scenario 6: 18.4% of votes) suggested that European P. vivax populations are one of the main sources of the genetic make-up of American P. vivax. Conversely, the second best scenario (scenario 4: 14.4% of votes) suggested that Ebro could be a migrant from Latin America (Fig 5a). This hypothesis could be rejected because Michel et al. [16] recently showed that Ebro is genetically part of the European cluster, composed of samples that predate Christopher Columbus’s arrival in the Americas.

Our ABC-RF scenario testing analysis supported the major contribution of European populations in shaping the genetic ancestry of Latin American P. vivax populations. Yet, these results also suggest that the colonization history of the Latin American continent might have been more complicated. Specifically, our analysis indicates a potential contribution from West African populations and/or from an unsampled population to the origins of P. vivax in Latin America. According to the best supported scenario (scenario 6 in Fig 5), the West African, Ebro and unsampled populations would have all split almost at the same time. Then, the admixture between the European population and the unsampled population would have composed the Latin American populations. To ascertain the role of West African P. vivax as a genetic source in the colonization of Latin America, it is imperative to include additional European P. vivax samples from ancient specimens, as the ones sequenced by Michel et al. [16] in future studies, and also to increase the African sample size and its distribution. Indeed, samples from central African countries such as Angola, where one of the main ports of the transatlantic trade was located [60], could provide valuable insights into the evolutionary history of P. vivax. However, at the time of this study, such samples were not available. Furthermore, more elaborated simulation frameworks using demographic-genetic scenarios including gene flow (and not just admixture events) and considering dormancy may bring new insights into the complicated P. vivax colonization history of Latin America.

Distinct invasion waves of P. vivax in Latin America

The results obtained in this study showed that European, West African, and unsampled populations all diverged at the same time, about 400 years ago (Fig 5b). However, P. vivax was probably already present in Latin America at this time, as demonstrated by Michel et al. [16]. This study reported a sample of P. vivax that clusters with modern Peruvian P. vivax strains. This ancient strain was sampled at Laguna de los Cóndores (Peru) and dated between 1437 and 1617, whose host is thought to be a person of Amerindian descent. Thus, the presence of this ancient American sample suggests that the unsampled population came from a first wave of invasion in Latin America, at least before the 17^th century, and probably during European colonization in the 15^th century.

The second significant event in the P. vivax colonization scenario identified by the DIYABC-RF is the admixture between the European population and the unsampled population in which the European population contributed approximately 60% of the current American genetic pool. Our analyses suggest that this event occurred towards the end of the transatlantic slave trade, in the latter half of the 19^th century. This time frame coincides with the massive post-colonial migrations [61], when more than 10 million Europeans settled in Latin America [62], potentially introducing parasites in this continent.

The presence of P. vivax before the 17^th century in Latin America [16] and the results presented in this study, therefore suggest that there were at least two waves of P. vivax introduction into Latin America: a first wave prior to the 17^th century, then a second from Europe at the end of the 19^th century. There is no way to determine from where, following divergence, this first wave originated. It could be from West Africa. This scenario is partially supported by the observed shared ancestry with some American populations (Fig 2c) and our DIYABC-RF analyses. Indeed, in scenarios 4, 8 and 11 (the second and third most supported scenarios in the DIYABC-RF analysis with Colombia and French Guiana as representative American population) (Fig 5a), American populations came from West Africa or from a population descended from West Africa and a European population. Currently, the only way to identify the exact genetic origin of this unsampled population is to generate more genetic data from these regions and/or using ancient samples (e.g. historical slides, historical mosquito specimens).

Noteworthy, our divergence time estimates between Europe and Latin America are approximately 400 years younger than those estimated by van Dorp et al. [15]. Such a difference may be explained by the different methodologies used in the two studies. We used a generation time of 5.5 generations per year [28], but other generation times have been suggested [63] and can also vary depending on the region [64], thus affecting the estimates of divergence times in years. van Dorp et al. [15] did not use any generation time estimate and inferred the divergence times by relying on the correlation between collection times and root-to-tip distances of the different samples [65]. This correlation was calibrated using only two historical points: the Ebro composite sample from the 1940s and a North Korean sample from 1953. Furthermore, the fact that Ebro is a composite sample may bias the divergence time estimate. Indeed, the samples in Ebro were collected between 1942 and 1944 and this was not taken into account in the correlation. Moreover, the mutations observed correspond to several isolates, and not to a single one. This may lead to a higher divergence time due to the artificial accumulation of mutations. Unlike van Dorp et al. [15], the two analyses used in our study converged toward very similar estimates of the divergence time between Ebro and American populations. Indeed, Relate gave a divergence time of 100–300 ybp (Fig 4c) and DYABC-RF suggested a divergence time of 212 ybp (Fig 5b).

Estimates on the effective population size and divergence time in Relate rely on a choice of mutation rate, which is alway difficult to make. Here we chose a value commonly used in the literature [28], but we acknowledge that mutation rate can vary both spatially and temporally in P. vivax. We expect that it will bring uncertainty on the absolute value, but not relative to each other. Furthermore, DIYABC-RF does not require making assumptions about the mutation rates, since simulations are done considering parameters values that are scaled by the mutation rate, following the Hudson simulation framework for SNPs [66]. Therefore, DIYABC-RF provides marginal parameter estimates that are scaled by the very high uncertainty of the mutation rate [49,50]. Usually, this leads to very high uncertainty as observed in Fig 5b, but the fact that both Relate and DIYABC-RF methods converged toward very similar estimates of divergence times suggest the mutation rate we use may be a reasonable approximation for P. vivax.

The approaches used by van Dorp et al. [15] and in the present study both rely on the Wright-Fisher model for coalescent simulation (a.k.a. the Kingman’s coalescent model [48]), which assumes a panmictic population with non-overlapping generations and neutral evolution [67,68]. However, in the case of P. vivax, the assumption of non-overlapping generation is not met, particularly because of dormancy. Especially since P. vivax emerges from dormancy when the host is infected with new strains [3,69]. As a result, diversity is maintained through generation, which may reduce the effect of genetic drift and lead to somewhat overestimated effective population size [58,70]. This may also explain why the Tajima’s D values (S9 Fig) and population size inference of the American populations (Fig 4a) did not suggest any bottleneck resulting from colonization, as previously observed [55]. Furthermore, the Kingman’s coalescent model allows only two lineages to merge at each generation. This also departs from P. vivax’s life cycle. Its alternation between asexual reproduction in humans and sexual reproduction in mosquitoes can lead to multiple simultaneous coalescent events in a generation [71]. Simulations by Korfmann et al. [72] show that using Kingman’s model underestimates effective population sizes when such events are frequent. However, Kingman-based tools remain accurate if these events are rare [72]. To mitigate this bias, we removed clonal and related individuals (see Materials and methods), but we cannot rule out that some bias may remain.

Additionally, our analysis overlooked migrations between populations, that can influence the estimates of genetic parameters, such as divergence times and population sizes. Omitting migration and overlapping generation can reduce the divergence times estimates and increase the inferred population sizes. Linked selection can also lead to underestimated the effective population size and divergence times [73], although we reduce this bias by using unlinked SNPs for these analyses (see Materials and methods). These factors could also explain why the trifurcation of Europe, West Africa, and the unsampled population is more recent than the earliest P. vivax presence detected in Europe and Latin America [16]. However, this trifurcation may also indicate that the transatlantic trade (from the 15^th to the 19^th centuries) facilitated contact between the European and West African P. vivax populations. Incorporating P. vivax-specific life-history and epidemiologic traits and selection in the population genetic inferences could improve their accuracy, but this currently represents major challenges. This would require theoretical and simulation-based population genetic studies, possibly using heavy forward genetic simulations such as those implemented in SLiM [74] and complex and heavy parametrization on each stage of the P. vivax life cycle, as well as on its temporal and spatial variations.

Ethics statement

In this study, 215 new P. vivax isolates from 24 countries were sequenced (Fig 1). P. vivax infection was detected by microscopy analysis, polymerase chain reaction (PCR) amplification of the Cytochrome b gene, and/or rapid diagnostic testing. Samples were obtained from patients infected with P. vivax following their informed consent and approval from the local institutional review board in each country. The informed consent process for the study involved presenting the study objectives to the community and inviting adults to participate. Before sample collection, each individual was informed about the study purpose and design and provided with a study information sheet. Oral informed consent was obtained from each participant. As the study posed no harm, it did not require written consent for its procedures. This oral consent approach aligned with the ethical standards of each country at the time of enrollment and was endorsed by the local ethics committees. Additionally, measures were taken to ensure the privacy and confidentiality of the collected data through sample anonymization before the study started.

For samples from Honduras, ethical and scientific approvals were obtained from the Ethics Review Committee of the Infectious and Zoonotic Diseases Master’s Program at Universidad Nacional Autónoma de Honduras (CEI-MEIZ 02–2014; 5/19/2014). Data were analyzed anonymously because the study made secondary use of biological specimens originally collected for malaria diagnosis as per the standard of care in Honduras. For samples collected in Venezuela, each patient gave a written informed consent, and ethical clearance was obtained from the Comité Ético Científico del Instituto de Medicina Tropical de la Universidad Central de Venezuela. Ecuadorian samples were collected through the Ministry of Health malaria surveillance program. The protocol was approved by the Ethical Review Committee of Pontificia Universidad Católica del Ecuador (Approval Number: CBE-016-2013). Informed written consent was obtained from all participants and/or their legal guardians in the case of minors before sample collection. For four samples from Mauritania, the study was approved by the pediatric services of the National Hospital, the Cheikh Zayed Hospital, and the Direction régionale à l’Action Sanitaire de Nouakchott (DRAS)/Ministry of Health in Mauritania. No ethics approval number was obtained at this time. For samples from Azerbaijan and Turkey, P. vivax was isolated from patients as part of the routine primary diagnosis and post-treatment follow-up, without unnecessary invasive procedures. The informed consent of each patient or an adult guardian of children enrolled in this study was formally obtained at the moment of blood collection. Written informed consent was collected, which included clear information that the samples would be used to investigate the genetic diversity of Plasmodium parasites as part of the VIVANIS project, supported by the COPERNICUS-2 RTD project contract ICA2-CT-2000-10046 of the European Commission. For child participants, formal written consent was obtained from a parent or legal guardian before sample collection. For the remaining newly sequenced samples, no specific consent was required because, in coordination with the Santé Publique France organization for malaria care and surveillance, the human clinical, epidemiological, and biological data were collected in the French Reference National Center for Malaria (CNRP) database and analyzed in accordance with the public health mission of all French National Reference Centers. The study of the biological samples obtained in the context of medical care was considered as non-interventional research (article L1221-1.1 of the French public health code) and only required the patient’s non-opposition during sampling (article L1211-2 of the French public health code).

DNA extraction and sequencing

Genomic DNA was extracted from each sample using the DNeasy Blood and Tissue Kit (Qiagen, France) according to the manufacturer’s recommendations. P. vivax-like samples were identified by amplifying with nested PCR the Plasmodium cytochrome b, as described in Prugnolle et al. [76]. Then, genomic DNA was analyzed on agarose gel (1.8%) in the TAE buffer. Gel-positive amplicons were then sent for sequencing.

For samples identified as P. vivax, selective whole-genome amplification (sWGA) was used to enrich submicroscopic DNA levels, as described in Cowell et al. [77]. This technique preferentially amplifies P. vivax genomes from a set of target DNAs and avoids host DNA contamination. For each sample, DNA amplification was carried out using the strand-displacing phi29 DNA polymerase and P. vivax-specific primers that target short (6 to 12 nucleotides) motifs commonly found in the P. vivax genome (PvSet1 [77,78]), 30 ng of input DNA was added to a 50-μL reaction mixture containing 3.5 μM of each sWGA primer, 30 U of phi29 DNA polymerase enzyme (New England Biolabs), 1× phi29 buffer (New England Biolabs), 4 mM deoxynucleoside triphosphates (Invitrogen), 1% bovine serum albumin, and sterile water. DNA amplifications were carried out in a thermal cycler with the following program: a ramp down from 35° to 30°C (10 min per degree), 16 hours at 30°C, 10 min at 65°C, and hold at 4°C. For each sample, the products of the two amplifications (one per primer set) were purified with AMPure XP beads (Beckman Coulter) at a 1:1 ratio according to the manufacturer’s recommendations and pooled at equimolar concentrations. Samples with the highest concentration of parasite genome were selected after qPCR analysis using a Light Cycler 96 with the following program: 95°C for 10 minutes; 40 cycles of 95°C for 15 seconds, 60°C for 20 seconds and 72°C for 20 seconds; 95°C for 10 seconds; and 55°C for 1 minute. Last, each sWGA library was prepared using the two pooled amplification products and a Nextera XT DNA kit (Illumina) according to the manufacturer’s protocol. Each sWGA library was prepared using the two pooled amplification products and a Nextera XT DNA kit (Illumina) according to the manufacturer’s protocol. Samples were then pooled and clustered on a HiSeq 4000 or Novaseq 6000 S4 1 lane PE150 with 2 × 150-bp paired-end reads.

Combining newly sequenced genomic data with those from the literature

The newly sequenced dataset for this study included 74 American P. vivax isolates (3 from Brazil, 19 from Ecuador, 22 from French Guiana, 4 from Guyana, 9 from Honduras, and 17 from Venezuela), 63 African isolates (1 from Central Africa, 21 from Comoros, 9 from Djibouti, 1 from Ethiopia, 1 from Gabon, 1 from Mali, 16 from Mauritania, 1 from Mayotte, 8 from Sudan, 2 from Chad, 1 from Uganda, and 1 from an unclear location between Chad and Mali), 40 Middle Eastern isolates (23 from Afghanistan, 4 from Armenia, 11 from Azerbaijan, and 2 from Turkey), 37 South Asian isolates (19 from India, 18 from Pakistan), and 1 Oceanian isolate (PNG).

In addition, genomic data from various sources were added, including those from Daron et al. [28], Benavente et al. [27], and the MalariaGEN project P. vivax Genome Variation [29]. From the large P. vivax Genome Variation dataset of the MalariaGEN project (n = 1,895), only samples accessible for which fastq file data were available, passing the quality check defined by Adam et al. [29], and that were not related were selected (S1 Fig). The relatedness between haploid genotype pairs within each country was measured by estimating the pairwise fraction of the IBD between strains within populations using the hmmIBD program [79], with default parameters for recombination and genotyping error rates, and using the allele frequencies estimated by the program. Isolate pairs that shared >50% of IBD were considered highly related. In each family of related samples, only the strain with the lowest number of missing data was retained. This left 500 samples from the MalariaGEN P. vivax Genome Variation project [29]. As the main focus of this study was on Latin America, samples from Malaysia, Myanmar, Thailand, and Cambodia were not retained because we already had more than 20 samples in each country. In total, 295 samples from the MalariaGen P. vivax Genome Variation project [31], 499 samples from Daron et al. [28], and 125 samples from Benavente et al. [27] were selected. The 499 samples from Daron et al. [28] included 26 African great apes P. vivax-like samples from three African countries (Cameroon, Gabon, and Ivory Coast).

As van Dorp et al. [15] inferred a European origin of the American populations, the ancient Spanish sample (Ebro) from this publication was added. Ebro is a composite sample of four samples, from four different slides dating from the 1940s: CA, POS, CM, and lane-8. Samples were downloaded from ENA (study accession PRJEB30878) and the fastq were single ends. The whole dataset from the literature comprised 920 samples that were added to the 214 newly sequenced genomes. Noteworthy, the European samples from Michel et al. [16] were not available at the time of the preparation of this dataset.

P. vivax and P. vivax-like read mapping and SNP calling

The P. vivax and P. vivax-like read mapping and SNP calling steps are detailed in S3 Fig. All modern sample sequencing reads were trimmed to remove adapters and preprocessed to eliminate low-quality reads (—quality-cutoff = 30) using cutadapt v1.18 [80]. Reads shorter than 50 bp containing “N” were discarded (—minimum-length = 50—max-n = 0). Sequenced reads were aligned to the P. vivax reference genome PVP01 [81] with bwa-mem v0.7.17 [82].

For Ebro, a different pipeline was used to consider the specific features of ancient DNA. The 3’ adapters, reads with low-quality score (—quality-cutoff = 30), and reads shorter than 30 bp were removed with cuadapt [80]. Then the cleaned fastq were merged, to create the composite sample as described in van Dorp et al. [15]. As the European samples were coinfected by P. falciparum, all reads were mapped to the P. vivax reference genome PVP01 [81] and to the P. falciparum reference genome Pf3D7 v3 [83] using bwa-mem [82]. For reads that mapped to both reference genomes, the edit distance was extracted and reads that aligned equally or better with P. falciparum than with P. vivax were removed. Then reads with mapping quality <30 and duplicates were removed with samtools v1.9 [84] and with Picard tools v2.5.0 (broadinstitute.github.io/picard/), respectively. Then, the base quality was rescaled with MapDamage 2 [85,86] to account for postmortem damage at the read ends.

For all samples, the Genome Analysis Toolkit (GATK v3.8.0 [87]) was used to call SNPs in each isolate following the GATK best practices. Duplicate reads were marked using MarkDuplicates from the Picard tools with default options. Local realignment around indels was performed using the IndelRealigner tool from GATK. Variants were called using the HaplotypeCaller module in GATK and reads mapped with a “reads minimum mapping quality” of 30 and minimum base quality of >20. During SNP calling, the genotype information was kept for all sites (variants and invariant sites, option -ERC) to retain the information carried by the SNPs fixed for the reference allele. Then only the nuclear genome was kept, filtered out with BCFtools v1.10.2 [84,88], and all individual VCF files were merged with GATK. Organelle genomes (mitochondria and apicoplast) were not included because they are haploid markers, without recombination and with only maternal heritability [89–91].

Data filtering

Combining data from the literature (n = 919) with the 214 newly sequenced samples resulted in a total of 1,133 samples for the unfiltered dataset: 1,106 modern human P. vivax samples, 26 modern African great apes P. vivax-like samples, and one ancient human P. vivax sample. All filtration steps are detailed in S1 Fig.

For all modern samples, samples with >50% missing data were removed. As several strains can infect the same host, the within-host infection complexity was assessed with the F_WS metric [92], calculated with vcfdo (github.com/IDEELResearch/vcfdo; last accessed July 2022). Samples with mono-clonal infections, i.e. F_WS >0.95 were kept. Highly related samples and clones could have generated spurious signals of population structure, biased estimators of population genetic variation, and violated the assumptions of the model-based population genetic approaches used in this study [93]. The relatedness between haploid genotype pairs was measured by estimating the pairwise fraction of the genome IBD as explained above. Within each country, isolate pairs that shared >50% of IBD were considered highly related. In each family of related samples, only the strain with the lowest number of missing data was retained. The final dataset included 619 modern P. vivax individuals from 34 countries, two P. vivax-like from Cameroon, infecting Nigeria-Cameroon chimpanzees (Pan troglodytes ellioti), and Ebro, an ancient DNA sample from Spain.

Two dataset formats were used in function of the analysis: (i) a VCF containing all the samples, and (ii) bam files of the samples for the ANGSD v0.940 [30] or ANGSD suite software (S1 Fig). For both formats, the analysis focused only on the core genome regions as defined by Daron et al. [8].

For the VCF, only bi-allelic SNPs with ≤20% missing data and a quality score ≥30 were kept. The minimum and maximum coverage were 10 and 106. Moreover, singletons were removed to minimize sequencing errors. Finally, 950,034 SNPs with a mean density of 41.14 SNPs per kilobase were retained.

For the ANGSD suite software, all sites with a base quality ≥20, a maximum total depth of 106 X, a minimum total depth of 5 X, and present in at least five individuals were kept. For analyses that did not require invariant sites, only bi-allelic SNPs with a p-value = 1e-6 were retained. In total 20,844,131 sites from the core genome were considered, including 1,437,273 variants (69.59 SNPs/kb).

Population structure

The PCA and the ancestry plots were computed with ANGSD and PCAngsd v0.98, after selecting only biallelic SNPs present in the core region of the P. vivax genomes and excluding SNPs with a minor allele frequency (MAF) ≤5%. The variants were LD-pruned to obtain a set of unlinked variants using ngsld v1.1.1 [94], with a threshold of r² = 0.5 with windows of 0.5kb. In total, 105,527 SNPs were used for 620 individuals (P. vivax-like samples were removed). For the ancestry plots, obtained from PCAngsd, the k value (number of clusters) was from 2 to 16. Then, pong v1.5 [95] was used to analyze the PCAngsd outputs and compute ancestry proportions.

The ML tree was obtained with IQ-TREE v2.0.3 [34] using the best-fitted model determined by ModelFinder [34]. As the dataset comprised 950,034 SNPs of the core genome and no constant site, the ascertainment bias correction was added to the tested models. The best inferred model was a general time reversible (GTR) model of nucleotide evolution that integrated unequal rates and unequal base frequency. The node reliability was assessed with Ultrafast Bootstrap Approximation [96] and the SH-aLRT test [97].

Relationships between populations

Population networks were estimated using two different approaches: TreeMix v1.13 [39] and AdmixtureBayes (github.com/avaughn271/AdmixtureBayes; last accessed June 2023) [40]. Unlike with the “classic” ML phylogenetic tree at the level of individual strain genomes (see above), these two methods use allele frequency (co-)variation within and among populations to derive the fraction of shared versus private genetic ancestry (or genetic drift) and to infer the population branching order, while taking into account population genetic processes, such as genetic drift, historical migration, and admixture events between populations [98,99].

To be able to compare results, the same dataset was used for both approaches: the VCF dataset was restricted to SNPs without missing data in Ebro (17,696 variants) and to subsampled modern populations as follows. Only populations with the largest sample sizes for each genetic cluster (according to PCA and ancestry plots) were kept, as well as all the American populations. The SNP data set was LD-pruned with With PLINK v2 [100], a threshold of r² = 0.5, with sliding windows of 50 SNPs and a step of 10 SNPs, and MAF filtering with a threshold of 5% (9,022 SNPs remaining). In some populations, there were still positions with missing data, and those positions were removed because TreeMix do not handle missing data. This filtering resulted in 7,109 SNPs for 18 ingroup populations (from n = 1 to n = 80) and one outgroup population (P. vivax-like).

For the TreeMix analyses, the number of migration events (m) that best fitted the data were estimated by running TreeMix 15 times for each m value, with m ranging from 1 to 10. The optimal m value was estimated using the OptM R package [101]. Then, a consensus tree with bootstrap node support was obtained by running TreeMix 100 times using the optimal m value. Results were post-processed using the BITE R package [102].

For the AdmixtureBayes analyses, three independent runs were performed, each including 40 Markov Chain Models and 250,000 steps, to identify convergence. Convergence was checked with a Gelman-Rubin Plot, as recommended by Nielsen et al. [40]. This analysis found that two chains had converged (S7 Fig). The results of these two chains were analyzed after a burn-in of 50% and a thinning step of 10, using the default parameters. The common network among the three best networks in each chain that converged was retained, based on posterior probabilities.

From the dataset generated with PLINK for TreeMix and AdmixtureBayes, the f₃-statistics were calculated using ADMIXTOOLS 2 v2.0.0 [103].

Demographic and invasion history of P. vivax in the Americas

Tajima’s D [43] and nucleotide diversity (π) were measured for each population with ≥10 individuals to avoid biases [104]. The sample size was standardized at 10 (i.e., 10 randomly chosen isolates for each population) to obtain values that could be compared. Tajima’s D and nucleotide diversity (π) values were estimated using ANGSD [30], with a window of 500 bp and a step of 500 bp in the core genome, and only windows with ≥50 sites were kept.

The coalescent-based approach of Relate [44] and the core region of the genome of all chromosomes were used to infer historical changes in N_e and to estimate the divergence time between modern populations. Collate was used to calculate CCR between Ebro and modern populations, because this program was designed to handle ancient DNA samples with low coverage [47]. For both analyses, a mutation rate of 6.43 x 10⁻⁹ mutations/site/generation [28] and a generation time of 5.5 generations/years [28] were used. Alleles were polarized using the two P. vivax-like samples as an outgroup.

A genetic map was required to carry out these analyses. This was generated using LDhat [105] and considering the Thailand population, one of our most diverse populations. First, individuals with ≥5% missing data were removed, leaving 58 samples. The function pairwise was used to estimate a first approximation of Watterson’s Theta and then, the likelihood table was estimated with these parameters: -n 58 -rhomax 100 -n_pts 101 -theta 0.00011 -split 8. The recombination map was then obtained by using the function interval (-its 1100000 -samp 100 -bpen 5) followed by stat with a burn-in of 10,000 samples. The recombination map (4N_er/pb) was converted into a genetic map (cM/Mb) with a N_e estimate of 6,000 (inferred with Stairwayplot 2 on chromosome 14).

Demographic scenarios testing of P. vivax invasion of the Americas using DIYABC-RF

Different possible scenarios of population divergence and admixture were evaluated using DIYABC Random Forest (RF) [49] via the diyabcGUI v1.2.1 R package.

As Ebro is a composite sample, it was made haploid by creating two haploid samples, splitting the “heterozygous” sites between the two samples. To minimize the effect of missing data, the sites genotyped in at least one individual in each population and at least in 40% of individuals in all populations combined were kept. Moreover, individuals with ≥ 50% missing data were removed. Thus, 45 haploid individuals for three populations and 3,139 SNPs were retained.

Twelve scenarios were tested: (1) American populations as a result of West African and European admixture; (2) uniquely European origin; (3) uniquely West African origin; (4) Ebro as a migrant from Latin America with American populations descending from West African ancestors; (5) Ebro resulting from the admixture between American populations and an unsampled West African "ghost" population; the American populations would have split from an ancestral population; (6) American populations stemming from the admixture of European and unsampled populations; (7) American populations from Europe with a secondary West African contribution; (8) American populations from West Africa with a secondary European contribution; (9) American populations from Africa with two introduction waves; (10) American populations originating from Europe, with Ebro resulting from an American-European ghost population admixture; (11) American populations resulting from two successive admixture events, the first and oldest one between the European populations and an unsampled ghost population, followed by a more recent wave of introduction from West Africa; and (12) Ebro resulting from the American-ghost population admixture, with American populations splitting from West Africa.

The scenario parameters were considered as random variables drawn from prior uniform distributions (S3 Table). DIYABC-RF was used to simulate 20,000 genetic datasets per scenario with the same properties as the observed data set (number of loci and proportion of missing data). Simulated and observed datasets were summarized using the whole set of summary statistics proposed by DIYABC-RF for SNP markers to describe the genetic variation of each population (e.g., proportion of monomorphic loci, heterozygosity), pair of populations (e.g., F_ST and Nei’s distances), or trio of populations (e.g., f₃-statistics, coefficient of admixture). The total number of summary statistics was 50.

Then with these 20,000 simulated data sets per scenario, the RF classification procedure was used to compare the likelihood of the competing scenarios. RF is a supervised machine-learning algorithm that uses hundreds of bootstrapped decision trees to perform classifications, using the summary statistics as a set of predictor variables. Then a classification forest of 1500 trees was grown.

Some simulations were excluded from the decision tree building during each bootstrap (i.e., out-of-bag simulations). These were used to calculate the prior error rate and the type II error rate. Additionally, the confusion matrix was computed to provide a more global assessment of the RF procedure performance [106] (S10 Fig). The result convergence was evaluated over ten independent RF runs, as recommended by Fraimout et al. [107]. To determine the compatibility of the formulated scenarios and associated priors with the observed dataset, we plotted the summary statistics for both on the first two principal axes of a linear discriminant analysis and of a PCA (S10 Fig), following the recommendation by Pudlo et al. [50]. The location of the summary statistics from the observed data within the clouds formed by those from the simulated data was visually checked to ensure that the simulations under the tested scenarios are realistic enough to reproduce the observed dataset (S10 Fig).

The RF computation provides a classification vote for each scenario (i.e., the number of times a model is selected from the decision trees). The scenario with the highest classification vote was selected as the most likely scenario.

Then, the posterior distribution values for all parameters for the best model identified were estimated using the derivation of a new RF for each component of interest of the parameter vector [108], with classification forests of 1500 decision trees, and based on a training set of 100,000 simulations. Estimated parameters included the effective size (N_e) for each population, split times among populations, and admixture event timing and rates. Estimates for timing parameters were converted from generations to years assuming a generation time of 5.5 generations per year [28].