Assessing microsatellite variations in Plasmodium falciparum following a decade-long antimalaria campaign in Kenya

Abstract

Anti-malaria interventions typically reduce the intensity of Plasmodium transmission, but the effects of reduced transmission on P. falciparum population biology remain unclear. Highly polymorphic microsatellite markers in P. falciparum were used to investigate genetic diversity, polyclonality, and genetic structure among populations in areas of varying malaria transmission intensity across Kenya. We also assessed relationships between metrics derived from genetic data, transmission intensity estimates, and bioclimatic variables. Despite an overall reduction in transmission intensity across Kenya from 2005 to 2014, we found that parasite populations maintained high genetic diversity and that genetic diversity correlated more closely with past transmission intensity estimate in the year 2000 as compared to contemporary estimate in 2014. In contrast, we found genetic structuring to be significant, consistent with our observation of shifting parasite migration patterns in western Kenya. Both genetic diversity and polyclonality increased with higher precipitation in the dry season, revealing the potential impacts of changing climate patterns on parasite population dynamics. Whereas fragmentation of P. falciparum populations increases opportunities for spatially targeted interventions in Kenya, the high genetic diversity of isolates in our study signals enhanced adaptability of parasites.

1.0. Introduction

Malaria caused by Plasmodium falciparum remains a major public health problem, especially in developing African countries like Kenya, where the disease is responsible for about 19% of clinical outpatient visits (WHO 2022; PMI 2019). In Kenya, climatic conditions vary markedly between lowlands and highlands and influence the intensity of malaria risks across geographical landscapes. Populations of malaria parasites generally display higher genetic diversity, infection complexity, and polyclonal infection prevalence in areas of intense malaria transmission than in areas experiencing mild and low risks of infections (Paul et al 1998; Babiker et al 1997; Anderson et al 2000). High genetic diversity indices (for examples nucleotide diversity and expected heterozygosity) in intense transmission areas are due to large effective population sizes of parasites (Anderson et al 2000), as well as increased opportunities for sexual recombination (Nkhoma et al 2020; Mu et al 2005).

During malaria transmission, superinfection events result from successive multiple infective mosquito bites, while co-transmission events result from a single infective mosquito transmitting multiple related parasite strains during feeding. Both superinfection and co-transmission events are common in areas of intense malaria transmission and impact the complexity of infections (the number of genetically distinct parasite clones in an individual). Genetic diversity and infection complexity are often reflective of transmission intensities (Koepfli & Mueller 2017; Gunawardena et al 2014; Branch et al 2011; Morris et al 2015; Roh et al 2019; Schultz et al 2010; Bendixen et al 2001; Lopez et al 2021). Assessing the relationship between transmission intensity, parasite genetic diversity, and infection complexity is important for improving our understanding of disease epidemiology. This information could provide insights into the effects of transmission intensity on human-immunity to blood-stage parasitemia, symptomatic malaria, and dynamics of drug resistance (Anderson et al 2000; Nassir et al 2005; Awadalla et al 2001).

Human travels facilitate parasite gene flow between geographical areas and introduce new gene variants into travel destinations. Travel is considered an important risk factor for malaria in some highland areas, where malaria transmission is low and less stable (Zhong et al 2007; Shanks et al 2005). At the same time, the increase in global temperature between the 1980s and 2000s has been linked to increased infections in highlands (Omumbo et al 2011; Alonso et al 2011). In western Kenya highlands, a high genetic diversity of P. falciparum has been observed, and the extent to which local transmission and human travel drive the genetic diversity in the area remains unclear (Zhong et al 2007; Vardo-Zalik et al 2013). Malaria interventions may also impact parasite genetic diversity (PMI 2019). Kenya increased the distribution and use of insecticide-treated bed nets (ITNs) and artemisinin-based combination therapy (ACT) for malaria control interventions since 2005, but the effects of ITNs and ACT on parasite diversity in the country remain understudied.

Measurements of genetic diversity and relatedness among populations are useful for inferring both historic and recent parasite dispersal patterns, and assessing how wild populations respond under changing epidemiological and ecological contexts (Lo et al 2017). In this study, we investigated the genetic diversity and polyclonality among P. falciparum malaria parasite populations in areas with varying transmission intensities across Kenya using microsatellite markers, which are useful for population assessment due to being neutral and highly polymorphic. We also investigated the parasite population structure with the objective of inferring historical and contemporary gene flow or migration among geographically distinct populations. Finally, we evaluated the relationships between parasite diversity, polyclonality, malaria transmission intensity, and bioclimatic variables.

2.0. Methods

2.1. Sample collection

Plasmodium falciparum-infected blood samples were collected between May 2014 and January 2015 from twelve sites within four geographical areas in Kenya: western lowlands, western highlands, rift valley, and eastern coastal (Figure 1; Supplementary Table 1). Peripheral blood samples were collected from symptomatic patients seen at local clinics (one clinic per study site) who tested positive for P. falciparum via rapid diagnostic testing. Individuals were qualified for the study at any age; the median and mean age of the study population was 7 and 12 years of age, respectively, and ranged from 5 months to 83 years old. Study population age varied across study sites (Supplementary Table 1; p < 0.001; Kruskal-Wallis). The overall male to female sex ratio of the study population was 1.2 and did not significantly vary among populations (p = 0.646; Kruskal-Wallis). Sample sizes in this study are specified in supplementary table 1. Ethical clearance was given by the Institutional Scientific and Ethical Review Boards of the Kenya Medical Research Institute, and the University of California, Irvine, CA. Written informed consent/assent for study participation was obtained from each individual who was willing to participate in the study and the parent or guardian for study participants under 18 years of age. Blood dots were collected on Whatman filter paper. Filter paper samples were stored at −20°C until they were analyzed.

Figure 1. Study site localities, Kenya.

Colors of study site points indicate region.

2.2. DNA extraction and parasite detection

The Saponin/Chelex method was used to extract DNA (Bereczky et al 2005). Quantitative polymerase chain reaction (qPCR) that amplifies P. falciparum-specific 18s rRNA was used to detect P. falciparum DNA using previously published primers (Rougemont et al 2004). Amplification of Pf18s rRNA was conducted in a 20μL mixture consisting of 2μL of DNA, 10μL of 2xSYBR Green qPCR Master Mix (Thermo Scientific, USA), and 0.3μL of 10μM primer. Cycling conditions were as follows: initial denaturation at 95°C for 3 min, followed by 45 cycles at 94°C for 30 sec, 55°C for 30 sec, 68°C for 1 min, and 95°C for 10, followed by a melting curve step from 65°C to 95°C with 0.5°C increments. Reactions were performed in the CFX96 Touch Real-Time PCR Detection System (Bio-Rad).

2.3. Microsatellite genotyping

Eight microsatellite loci namely, Polyα, TA42, TA81, TA87, TA109, 2490, ARA2, and Pfg377 were selected for genotyping based on evidence of polymorphism in previous studies, reliable amplification, having an even distribution across chromosomes, and comparability to previous studies in the region (Supplementary Table 2) (Anderson et al 1999; Zhong et al 1999). The M13 tailed primer method was used to fluorescently label our primers (Zhong et al 2007). Amplification for microsatellite loci was conducted in a total volume of 10μL with 5μL of 2x DreamTaq Green PCR Master Mix (Thermo Fisher, USA), 0.5μL of 10μM primer (forward primer with M13 tail), and 1μL of DNA template. Thermocycling conditions were as follows: initial denature of 95°C for 5 min, followed by 40 amplification cycles of 95°C for 30 sec, 50°C for 30 sec, and 65°C for 45 sec, and then a final extension of 65°C for 7 min. PCR products were analyzed on an automated 4300 DNA analyzer (Li-Cor, Lincoln, NE), and alleles were quantified with the use of Gene ImagIR 4.33 software (Li-Cor). Minor peaks which were at least one-third as high as the predominant peak were counted for assessing polyclonal infections. For other analyses, only the primary peak was used in polyclonal infections.

2.4. Genetic variation and polyclonality

To describe population genetic metrics and polyclonality across populations and regions of Kenya, the following metrics were measured: allelic richness (mean number of alleles calculated with rarefaction, measure of genetic diversity) (El Mousadik & Petit 1996), expected heterozygosity (a common measure of genetic diversity due to its independence from sample size), linkage disequilibrium (non-random association of alleles at multiple loci), and percent of infections which were polyclonal (infections with > 1 unique parasite strain). The genetic variables were estimated from the average of all 8 microsatellite loci selected for amplification. Expected heterozygosity was calculated in Arlequin 3.5.2.2 (Excoffier & Lischer 2010). Allelic richness at each study site was calculated in the R package PopGenReport (El Mousadik & Petit 1996; Adamack & Gruber 2014). Linkage disequilibrium was estimated in Genepop 4.2 (Rousset 2008; Raymond & Rousset 1995) using a test of the composite linkage disequilibrium described by Weir et al (1996). Polyclonality and expected heterozygosity in the western highland and lowland populations were compared to a previous study in which samples were collected in 2005 (Zhong et al 2007), prior to the scale-up of ITNs, from individuals experiencing clinical malaria (much the same as this study) residing in the nearby geographic vicinity. We acknowledge that microsatellite data is particularly difficult to compare across studies and any such comparisons should be cautiously interpreted. We have taken several measures to promote comparability across the two studies, including that the microsatellite data from 2005 was collected using the same instrument and protocols in the same research group. In addition, the lead author on the previous study trained the person conducting the genotyping and analysis for this study.

2.5. Population structure

Genetic clustering was assessed using two methods: principal component analysis (PCA) and rmaverick, a Bayesian clustering technique. PCA is an ordination technique that involves applying a transformation to a rectangular matrix to rotate it into a space where the columns are uncorrelated. By analyzing these new columns, or components, and in particular the columns that explain the most variance, the information content of the data can be understood more easily. For this analysis, the microsatellite data was reformatted as a binary matrix, with one column per allele and one row per sample. Missing data was imputed as the mean allele frequency for the column in question. No more than two missing loci were allowed per sample. The PCA was performed with the R programming language (R Core Team 2013) and visualized with the aid of the GGally R package (Schloerke et al 2021).

rmaverick is similar to the popular program STRUCTURE, in that it is a Bayesian clustering method that seeks to find the set of clusters that best explains observed deviations in the data from Hardy-Weinberg and linkage disequilibrium (Verity & Nichols 2016). A number of clusters, K, is specified by the user, and the program estimates K admixture coefficients for each individual that represent the fraction of membership that individual has in each cluster. STRUCTURE and rmaverick both use a Markov chain Monte Carlo method to find these coefficients, but rmaverick differs from STRUCTURE in that it uses multiple, “Metropolis-coupled” Markov chains simultaneously (Verity & Nichols 2016). Information is passed between chains during the model fitting process, improving mixing and model convergence (Verity & Nichols 2016). The admixture coefficients produced by rmaverick were visualized with the pophelper R package (Francis 2017).

To assess the significance of population structuring, sources of variation among regions, among populations, and within populations were evaluated by calculating F-statistics via an analysis of molecular variance (AMOVA) in Arlequin 3.5.2.2 (Excoffier & Lischer 2010). Specifically, the following F-statistics were assessed in the AMOVA for our study: F_SC, the degree to which alleles within populations in the same region are more similar to each other than would be expected if the population were composed of random alleles from the region; F_CT the degree to which alleles within a region are more similar to each other than would be expected if the population was composed randomly from alleles from all the regions together; and F_ST, the degree to which alleles within populations are more similar to each other than would be expected if the population were composed of random alleles from the entire study population. Furthermore, pairwise F_ST and significance of F_ST was estimated by bootstrapping over loci (nboot=1000) in the R package hierfstat (Goudet 2005).

2.6. Relatedness

To understand patterns of individual- and population-level relatedness, the Dcifer method was employed to model relatedness between infections (Gerlovina et al 2022). This approach estimates relatedness based on allele frequencies using population-level allele frequencies to differentiate chance matching of common genotypes from genotype sharing due to genetic relatedness. For inference, Dcifer uses a likelihood-ratio approach to assess the null hypothesis of unrelated infections for each pair of samples in the dataset.

To understand relatedness at the population level, the sample-level pairwise relatedness values were aggregated to two coarser levels of organization: site and region (Figure 1). Two aggregation methods were used in each case: mean relatedness and the fraction of sample pairs that are significantly related. The former is likely more intuitive but tends to be skewed by relatedness values of zero. The latter metric does not have this issue and should show recent gene flow more clearly. Note that because the p-values estimated by Dcifer were only used for this purpose of aggregating from the sample to the geography level, they were not corrected for multiple testing.

To investigate whether the relatedness shows a pattern of isolation-by-distance, a Mantel test was performed with the R package vegan (Dixon 2003). This analysis is a test for correlation between two distance matrices – in this case, genetic and geographic distance. To create an appropriate genetic distance matrix for this analysis, the mean relatedness between each pair of sites was subtracted from one.

2.7. Historical migration

The direction and magnitude of historical parasite migration between populations was estimated in Migrate-N v3.7.2, which uses coalescent theory, with the Brownian motion model and Bayesian inference search strategy (Beerli et al 2019). Four independent runs were conducted with a burn-in of 10⁴ steps, sampling increment of 100 steps, and 5,000 recorded steps in each chain for a total of 2×10⁶ visited parameter values. The number of immigrants per generation (Nm) was estimated by multiplying the estimated mutation-scaled effective population size (Θ) by the mutation-scaled effective immigration rate (M), divided by the inheritance parameter (x, 2 for haploid data). Credible intervals for M were assessed from the posterior distributions. The Wilcoxon rank sum test with a Holm correction for multiple pairwise comparisons was used to assess significant differences in M among populations and regions.

2.8. Relationships between genetically-derived metrics, transmission intensity, climate, and study population age

As a measure of population-level transmission intensity, P. falciparum parasite rate in 2–10 year olds (PfPR_2–10) for the years 2000–2014 were obtained from The Malaria Atlas Project (malariaatlas.org) at a 5km spatial resolution (Weiss et al 2019).. Correlations with annual PfPR_2–10 rates between 2000 and 2014 were evaluated to assess which between genetic diversity and polyclonality correlated more closely with current and past PfPR_2–10. To assess relationships between parasite population metrics derived from genetic data and climate, historical (‘near recent’) bioclimatic data for 1970–2000 were obtained from WorldClim at a 30 second (~1km) spatial resolution (worldclim.org) (Fick & Hijmans 2017). The WorldClim bioclimatic dataset, which was released in 2020, was selected because the dataset is commonly used in species distribution modeling (Booth 2018), and malaria transmission is driven by precipitation and temperature variables which are represented in the dataset. All nineteen of the standard bioclimatic variables from WorldClim were obtained for analysis (see Supplementary Table 3 for description of the nineteen variables).

PfPR_2–10 and bioclimatic values at study sites were extracted in the R package raster (Hijmans et al 2015). For bioclimatic values, values were extracted and averaged within a 2.5km radius buffer area surrounding each study site point to reflect the spatial resolution of the PfPR_2–10 data. Due to potential collinearity issues in multivariate models, collinearity was assessed among candidate predictor variables in each model by the Pearson correlation coefficient. For variable pairs that had a correlation coefficient > 0.7, one variable in the pair was dropped from consideration in future analyses. For PfPR_2–10, we kept the variable that resulted in the largest spread in years while including the concurrent year of sample collection, 2014. For bioclimatic variables, we kept the variable in the correlated pair that we hypothesized to affect malaria transmission more directly in the study area. Next, we included all variables that passed the collinearity assessment into grouped (PfPR_2–10 and bioclimatic) multivariate linear models for genetically-derived metrics and performed variable selection via backward elimination to systematically select variables for final models using a p-to-remove of 0.1. The variables from the best fit models from group variable selection were then combined in linear models for both outcome metrics and backward elimination repeated on the combined models. Linear models were generated using the lm() function in R (Wilkinson & Rogers 1973; Chambers et al 1992).

Since the study population age varied among sites, we assessed relationships between study population age and other variables in our study. First, a Pearson collinearity assessment among mean study population age, genetically-derived metrics, and selected PfPR_2–10 and bioclimatic variables (via backward selection) was performed. Second, a multivariate linear model was built for study population age by backward elimination with selected PfPR_2–10 and bioclimatic variables as dependent variables. Lastly, the impact of including mean study population age as a dependent variable into models for expected heterozygosity and proportion polyclonal was assessed. Since age data was missing for the KT site, backward elimination for the combined model with selected PfPR_2–10 and bioclimatic variables was repeated with this site removed. Then, an initial multivariate model with mean population age in addition to the same variables was built and backward elimination repeated to select variables for final models. The final models with and without age were compared by ANOVA.

The Python programming language (Van Rossum 2007), the R programming language (R Core Team 2013), and the tidyverse R packages (Wickham et al 2019) were used throughout the project to automate analyses. The adegenet R package (Jombart 2008) was used for manipulating genetic data. The workflow was partially automated with Snakemake (Mölder et al 2021).

3.0. Results

3.1. Genetic diversity and polyclonality

Allelic richness (A_R) was generally highest in the eastern coastal sites (3.938–4.224) and lowest in the rift valley sites (3.087–3.483; Table 1). Similarly, expected heterozygosity (H_E) was high in the eastern coastal sites (0.725–0.788) and low in the rift valley sites (0.582–0.693; Table 1). The highest percentage of polyclonal infections occurred in western lowland and highland sites (MA: 88.2% (30/34); KB: 78.9% (15/19); HB: 83.3% (40/48); Table 1). In contrast, the lowest percentage of polyclonal infections occurred in the rift valley (KT: 25.0% (4/16); MT: 36.4% (4/11); NA: 47.1% (8/17). Analysis of linkage disequilibrium for each locus pair across all populations and for each pair by population did not indicate that any pairs of loci were in linkage disequilibrium by Fisher’s exact test with P-values adjusted by the false discovery rate method (α = 0.05). Comparisons of parasite populations between 2005 and 2014 revealed a decrease in the proportions of polyclonal infections in both lowlands and highlands in western Kenya (Figure 2). These differences were significant for both the lowlands and highlands (p=0.0028 and p =0.0023, respectively; Wilcoxon signed rank test). However, the expected heterozygosity of parasites did not significantly differ between2005 and 2014 in the western Kenyan lowlands and highlands (p=0.51 and p=0.33, respectively; Wilcoxon signed rank text).

Table 1.

Genetic indices among Plasmodium falciparum populations, Kenya.

Region	Site	N	NA	AR	HE	Polyclonality
Western lowland	PB	31	47	3.541	0.634	51.6% (16/31)
	HB	48	59	3.885	0.69	83.3% (40/48)
	KB	19	49	4.153	0.786	78.9% (15/19)
Western highland	EL	35	57	3.959	0.722	68.6% (24/35)
	MA	34	53	3.983	0.729	88.2% (30/34)
Rift valley	KT	16	33	3.087	0.582	25% (4/16)
	MT	11	31	3.307	0.687	36.4% (4/11)
	NU	17	39	3.483	0.693	47.1% (8/17)
Eastern Coastal	KK	34	51	3.938	0.754	73.5% (25/34)
	JU	30	56	4.224	0.788	50% (15/30)
	MI	38	57	4.207	0.747	55.3% (21/38)
	JA	30	53	4.005	0.725	73.3% (22/30)

Figure 2. Comparison of genetic diversity (expected heterozygosity) and polyclonality between 2005 and 2014 in western Kenya.

Values were averaged across populations in western Kenya (n=1 for lowland-2005; n=2 for highland-2005; n=3 for lowland-2014, n=2 for highland 2014) for each locus and then compared. Statistical significance of the difference in expected heterozygosity and proportion of polyclonal infections detected at matching loci from 2005 (pre-ITN scale-up) to 2014 (post-ITN scale-up) were assessed by Wilcoxon signed-rank test for the eight matching loci. In western lowland sites, compared to previously published expected heterozygosity values for parasites collected in 2005 (Zhong et al 2007), the mean value was lower in 2014 ($\underset{\_}{x}$ = 0.703) than in 2005 ($\underset{\_}{x}$ = 0.754).

3.2. Population structure

Based on the amount of variance explained by each, the first three components of the PCA results were chosen for visualization. None of these show a clear pattern of clustering (Supplementary figure 2). The log likelihoods corresponding to the rmaverick models for each K value suggest that K=1 is the best supported model (Supplementary Figure 1). However, this method cannot necessarily identify the “true” K value (Verity & Nichols 2016), so for the sake of thoroughness K values from two through four were visualized (Figure 3). These visualizations do not suggest any pattern of clustering by geography.

Figure 3:

Admixture plot from Bayesian clustering by rmaverick.

The analysis of molecular variance (AMOVA) indicated that a significant source of variation occurred among populations within regions (F_SC = 0.0320; P < 0.001), but variation among regions was not significant (F_CT = 0.00245; P = 0.217). Variation among populations among regions was significant (F_ST = 0.0178; P = 0.001). Further evidence for significant population structuring across Kenya is provided by significant pairwise F_ST values among populations (Supplementary Table 4). Despite evidence of genetic structuring among populations, these analyses do not reveal clear patterns of genetic clustering, geographic or otherwise, in our data.

3.3. Relatedness

Sample-level relatedness is fairly low, with most values falling below 0.25 and almost all values below 0.5 (Supplementary Figure 3). It is therefore not surprising that the mean relatedness between sites and regions is also fairly low (Supplementary Figure 4). Within-site and within-region relatedness tend to be higher than relatedness between sites and between regions (Supplementary Figure 4). The same patterns generally hold for the fraction of significant pairs at the site- and region-level, although there is more variability in this metric (Figure 4). Among-populations, the fraction of highly related isolates, which is more indicative of recent migration, was highest between the following population pairs: Junju (JU) & Kakayuni (KK), Junju (JU) & Msambweni (MI), Kabernet (KT) & Marigat (MT), Marigat (MT) & Port Bunyala (PB), Kendu Bay (KB) & Jaribuni (JA), and Kakayuni (KK) & Marigat (MT) (Figure 4A). Among these population pairs, three of the pairs contained two populations within the same region, one pair contained a western lowland population and a rift valley population, one pair contained an eastern coastal and rift valley population, and one pair contained an eastern coastal and western lowland population. Among regions, the fraction of highly related isolate pairs was generally highest between the rift valley and western lowlands and lowest between eastern coastal and other regions as well as between western lowlands and western highlands (Figure 4B).

Figure 4: Heatmap of fraction of pairs with significant relatedness out of all the sample pairs corresponding to each pair of populations.

A) Color swatches along the diagonal correspond to sample pairs from the same population and thus show within-population values. B) As A, except sample pairs are grouped by region, rather than population.

With both metrics, Marigat is an outlier at the site level and the Rift Valley is an outlier at the region level in terms of within-site/region relatedness. Marigat is in the Rift Valley region (Figure 1), and there are relatively few samples from this site and region (Table 1). This means that the higher within-region relatedness of the Rift Valley is likely being driven by Marigat. As for the high relatedness within Marigat itself, it is difficult to conclude from this data whether transmission in this population is more clonal than other populations or whether it is an artifact of the smaller sample size (n=11) at this location. Notably, malaria transmission at this site has remained relatively low throughout the years, compared to other populations in this study, with a PfPR_2–10 well below 0.2 since 2000.

The Mantel test revealed a low correlation between genetic and geographic distance at a moderate level of significance (r = 0.261; P = 0.013). This result suggests a weak but significant pattern of isolation-by-distance in our data. Note that one of the assumptions of this test is that of a linear relationship between the distances. Visual inspection of our data suggests that this assumption may not be met (results not shown), which would serve to reduce the power of the test. It is therefore possible that the true pattern of isolation-by-distance is stronger than we were able to detect here.

3.4. Historical migration

Of the migration routes having the highest 10% mutation-scaled migration rates (M), nine of the thirteen routes contained a western lowland site (Figure 5A), indicating the western lowland populations were relatively well connected to other populations in Western Kenya. Of the top thirteen migration routes, twelve of the routes had a 50% credible interval of M larger than other migration routes based on the posterior distributions; these migration routes were MA to KB (>131 routes), PB to HB (>125 routes), HB to PB (>123 routes), KB to HB (>108 routes), HB to KB (>84 routes), MI to EL (>76 routes), KK to JU (69 routes), JU to EL (> 37 routes), KB to PB (>37 routes), PB to KB (>37 routes), PB to MA (>37 routes), MI to KT (>14 routes) (Supplementary Table 5). Using the 95% credible interval, only MA to KB (>127 routes) had a M higher than other routes. When comparing population-level migration rates, there were few significant differences among migration rates; JA (Jaribuni, eastern coastal) was the only population that significantly differed, which had a lower immigration rate than HB (Homa Bay, western lowland) and a lower emigration rate than six populations in the eastern coastal and rift valley regions (Figure 5B–C). Comparing migration rates between populations grouped by region, western lowland populations had the highest within region migration rates, which were significantly higher than within the rift valley and eastern coastal regions (P = 0.01 and P = 0.0026, respectively; Wilcoxon rank sum test; Figure 5D). Among regions, several significant differences were detected among the comparisons (Figure 5D); migration rates were highest from the western highland populations to the western lowland populations, and migration rates were lowest from the western lowland populations to the rift valley (Figure 5D). At the region-level, the western lowlands had the highest immigration (sink) rate, which was significantly higher than the western highlands, rift valley, and eastern coastal regions (P = 0.024, P = 0.01, and P < 0.001, respectively; Wilcoxon rank sum test; Figure 5E). In contrast, differences among emigration rates by region were found to be smaller (Figure 5F).

Figure 5. Estimation of historical migration directionality and intensity among populations, Kenya.

A) Map showing the migration routes which have the highest 10% migration rates. Thickness of lines is roughly proportional to the estimated migration rate (M). B) Immigration (sink) rates (box and whisker) and estimated number of immigrants (bar plot) by population. C) Emigration (source) rates (box and whisker) and estimated number of emigrants per generation (bar plot) by population. D) Total migration rates (box and whisker) and estimated number of migrants per generation (bar plot) within and among regions. E) Immigration (sink) rate (box and whisker) and estimated number of immigrants per generation per population (bar plot) by region. F) Emigration (source) rate (box and whisker) and estimated number of emigrants per generation per population (bar plot) by region.

3.5. Relationship between transmission intensity and metrics derived from genetic data

An assessment of multicollinearity among yearly PfPR_2–10 revealed that PfPR_2–10 was highly correlated among years at our study sites; 78 variable pairs of a possible 105 variable pairs had a correlation coefficient that exceeded our threshold of 0.7 (Supplementary Table 6). Likewise, correlations among past PfPR_2–10 to the concurrent PfPR_2–10 (2014) were high. The only years that did not exceed our correlation threshold to 2014 were 2000 and 2001; these years also correlated with each other by our threshold. We therefore selected PfPR_2–10 from 2000 and 2014 for multivariate linear models, as this pair contained the largest spread in years and lowest correlation coefficient. Following variable selection, the predictor variable in the PfPR_2–10 model for expected heterozygosity was PfPR_2–10 in 2000 and for proportion polyclonal was PfPR_2–10 in 2014 (Table 2). However, neither of these models were statistically significant at a significance level of 0.05.

Table 2.

Group variable selection results by linear models with backward elimination

Variable group	Metric from genetic data	Independent variable	Estimate	T value	P value	Adjusted R2
PfPR	Expected heterozygosity	PfPR in 2000	0.19	1.9	0.094 ^
		Overall			0.094 ^	0.18
	Proportion polyclonal	PfPR in 2014	0.77	1.8	0.10
		Overall			0.10	0.17
Bioclimatic	Expected heterozygosity	Mean diurnal range (BOI2)	−0.017	−2.8	0.022*
		Precip. of driest quarter (BIO17)	0.00067	1.9	0.089 ^
		Overall			0.055 ^	0.36
	Proportion polyclonal	Precip. of wettest quarter (BIO16)	0.00087	5.0	0.00071***
		Precip. of driest quarter (BIO17)	0.0024	4.8	0.00098***
		Overall			6.4×105***	0.86

3.6. Relationship between bioclimatic variables and metrics derived from genetic data

An assessment of multicollinearity among bioclimatic variables revealed that 48 variable pairs of a possible 171 variable pairs had a correlation coefficient that exceeded our threshold of 0.7 (Supplementary Table 7). Of variables pairs or groups that were highly correlated, we selected the variable which we hypothesized to affect malaria transmission most directly in the study area. For example, predicted minimum and maximum temperatures for malaria transmission are 17°C and 34°C, respectively (Mordecai et al 2013). Thus, when choosing amongst several temperature variables that correlated (BIO1, BIO5, BIO6, BIO8, BIO9, BIO10, and BIO11), we selected ‘min temperature of coldest month’ (BIO6) after inspecting the temperature variable values at our study sites, as BIO6 contained the most site values outside of the malaria transmission temperature range. After eliminating variables from variable pairs or groups that were highly correlated, six of the nineteen variables remained for further analysis: ‘mean diurnal range’ (BIO2), ‘min temperature of coldest month’ (BIO6), precipitation of the wettest quarter’ (BIO16), ‘precipitation of driest quarter’ (BIO17), ‘precipitation of warmest quarter’ (BIO18), and ‘precipitation of coldest quarter’ (BIO19). These predictors were included together in initial multivariate linear models for expected heterozygosity and proportion of polyclonal infections.

Following variable selection, the predictor variables in the final bioclimatic model for expected heterozygosity were ‘mean diurnal range’ (BIO2) and ‘precipitation of the driest quarter’ (BIO17), where ‘mean diurnal range’ negatively correlated with expected heterozygosity and ‘precipitation of driest quarter’ positively correlated with expected heterozygosity (Table 2). However, this model is not statistically significant at a significance level of 0.05. For proportion of polyclonal infections, ‘precipitation of the driest quarter’ (BIO17) and ‘precipitation of the wettest quarter’ (BIO16) were included in the final bioclimatic model where both variables positively correlate with polyclonal infections (Table 2). This model was statistically significant (P = 6.4×10⁵).

3.7. Relationship between genetically-derived metrics, climate, transmission intensity, and study population age

Combining selected PfPR_2–10 variables with bioclimatic variables for models of expected heterozygosity and proportion of polyclonal infections revealed that bioclimatic variables were the only significant predictors for both genetically-derived metrics; PfPR_2–10 variables were eliminated from both models by backward elimination. Thus, the best-fit models considering both PfPR_2–10 and bioclimatic variables for both genetically-derived metrics were the final bioclimatic models (Table 2).

Since study population age varied among study sites, we investigated the potential relationships between study population age and other variables in our study, which were expected heterozygosity, proportion of polyclonal infections, selected bioclimatic variables (BIO2, BIO16, and BIO17), and selected PfPR_2–10 variables (PfPR_2–10 in 2000 and 2014). A collinearity assessment revealed that the mean age of the study population negatively correlated with PfPR_2–10 in 2000 (correlation coefficient = −0.81; Supplementary Table 8). Next, the five bioclimatic and PfPR_2–10 variables were included as independent variables in a multivariate linear model for mean study population age. Following backward elimination, the variables remaining in the final model for mean age of the study population were PfPR_2–10 in 2000 and precipitation of the driest quarter (BIO17); this model was significant (P = 0.0022; Supplementary Table 9). Furthermore, we assessed the impact of including mean study population age as a dependent variable in models for expected heterozygosity and proportion polyclonal without the KT site, for which we did not have age data. Before adding age data, removing KT from the analysis led to a different final combined model for expected heterozygosity, which included only ‘mean diurnal range’ (BIO2) and was not significant (R²=0.12; P = 0.16; Table 3). Removing KT had no effect on the final model for proportion of polyclonal infections, and the model remained significant (P = 3.9×10⁵). Adding age as a variable to the initial multivariate model for expected heterozygosity led to a different final model which included PfPR_2–10 in 2000, mean study population age, and ‘precipitation of the driest quarter’ (BIO17); this model was significant (P = 0.011) and a significantly better fit than the final model without age (P = 0.012; ANOVA). Adding age to the initial proportion polyclonal model did not affect the final combined multivariate model.

Table 3.

Comparison of combined multivariate linear models of genetically derived metrics with and without the inclusion of mean study population age as an independent variable.

Metric from genetic data	Age inclusion	Dependent variable	Estimate	T value	P value	Adjusted R2
Expected heterozygosity	Without age	Mean diurnal range (BOI2)	−0.0078	−1.5	0.16
		Overall			0.16	0.12
	With age	PfPR in 2000	0.52	4.9	0.0017 **
		Mean study population age	0.013	4.3	0.0036 **
		Precip. of driest quarter (BIO17)	0.00059	2.8	0.029 *
		Overall			0.011 *	0.68
	Age model comparison by ANOVA				0.012 *
Proportion polyclonal	With and without age	Precip. of wettest quarter (BIO16)	0.0023	6.3	0.00025***
		Precip. of driest quarter (BIO17)	0.00074	5.7	0.00045***
		Overall			3.9×105***	0.90

4.0. Discussion

In this study, we found that eastern coastal P. falciparum populations had the highest genetic diversity, while western lowland and highland populations had the highest percentage of polyclonal infections. Despite a reduction of transmission intensity since 2005 throughout Kenya, genetic diversity was comparable to values obtained from clinical samples collected in 2005. Furthermore, we did not find significant correlations between contemporary transmission intensity estimates and measures of genetic diversity or infection complexity. In contrast, historical climate data correlated with infection complexity, and historical transmission intensity and climate correlated with genetic diversity. Populations were significantly genetically structured, although the structuring was subtle and no clear patterns in structuring were observed aside from an isolation-by-distance signal. Interestingly, the fraction of isolates from the western highlands and lowlands which were found to be highly related by identity-by-descent was relatively low, indicating low levels of recent migration between these nearby regions. In contrast, historic migration between the western highlands and lowlands was relatively high, suggesting that migration patterns could have shifted.

Little to no population genetic differentiation has been found for P. falciparum in Kenya, which is thought to be influenced by frequent human travel maintaining extensive gene flow between populations (Zhong et al 2007; Bonizzoni et al 2009). In this study, we detected significant genetic structuring by F_ST, which suggests that populations are more genetically isolated than observed in previous studies in Kenya (Zhong et al 2007; Bonizzoni et al 2009; Omedo et al 2017). With that said, this structure appears subtle as F_ST values were still low ($\overline{x}$ = 0.053) and structuring was not detected in other structuring analyses; one explanation for these results could be that populations are not in a stable equilibrium. A trend of increasing fragmentation, should that be occurring, could be the result of increased public health interventions since 2006 (Zhou et al 2011; Hemming-Schroeder et al 2018) decreasing parasite transmission intensity throughout the country (Anderson et a 2000; Pumpaibool et al. 2009). Malaria admissions in Kenya decreased from more than 500 admissions per 100,000 people in 2009 to less than 100 admissions per 100,000 people in 2014 (the year parasite isolates were collected for this study) (WHO 2022). Likewise, significant P. falciparum genetic structuring following a period of decreasing malaria transmission was reported in Malaysia (Anthony et al 2005). This finding of increasing fragmentation of P. falciparum populations in Kenya may have implications for malaria control. For example, control of infections is likely to be more effective at reducing local malaria incidence when interventions are designed to target isolated populations.

We note that our analyses did not indicate a clear geographic pattern in the apparent genetic structuring, other than a weak pattern of isolation-by-distance, which suggests the structure that does exist may be spatially complex and/or in the early stages of development. A recent landscape genetics study in western Kenya revealed that Lake Victoria provided a barrier to gene flow between P. falciparum populations north and south of the Winam Gulf of Lake Victoria (Hubbard et al 2023). This result provides additional evidence that human movement drives the population structure of P. falciparum in Kenya, as human movement, based on mobile phone data, was also found to be lower between populations north and south of the Winam Gulf, as compared to populations on the same side of the gulf (Hubbard. et al 2023; Wesolowski et al 2012). With that said, gene flow between vector populations has been found to be low across Lake Victoria, as well (Kayondo et al 2005; Hemming-Schroeder et al 2020), though land cover, specifically tree cover and cropland, appears to be an important driver of vector connectivity (Hemming-Schroeder et al 2020). Additional population and landscape genetic studies utilizing systematic sampling designs and spatial scales relevant to vector and human movement would greatly improve our understanding of the factors influencing parasite structure and movement (Hemming-Schroeder et al 2018).

Our results suggest that genetic diversity remains high in parasite populations despite dramatic reductions in transmission intensity. Specifically comparing expected heterozygosity and proportion of polyclonal infections detected from western lowland and highland specimens collected in this study to values from an earlier study at nearby sites (within 60 kilometers), which also used clinical malaria samples and the same protocols (Zhong et al 2007), expected heterozygosity values in 2014 were similar to those observed in 2005, prior to the scale-up in use of ITNs (PMI 2019). However, the proportion of polyclonal infections observed in 2014 was noticeably lower than in 2005 (Figure 2). In addition to expected heterozygosity remaining high in the western lowlands and highlands, expected heterozygosity correlated with the PfPR_2–10 of 14 years prior. Notably, PfPR_2–10 data was not available prior to 2000, so it is possible that expected heterozygosity had a higher correlation with PfPR_2–10 more than 14 years prior to 2014. A potential lag time between decreased transmission intensity and a reduction in genetic diversity is consistent with findings from other studies. For example, in the Kingdom of Eswatini, high diversity in P. falciparum persisted despite a 94% reduction in malaria cases (Roh et al 2019). In the Peruvian Amazon, high P. falciparum genetic diversity was maintained after five years of sustained low transmission (< 0.5 infections/person/year) (Branch et al 2011). Furthermore, a modeling study revealed that genetic diversity could be maintained for decades following a decrease in transmission intensity whereas metrics related to polyclonal infections responded more quickly (within months) to a change in transmission intensity (Hendry et al 2021). This difference in response time is due to the nature of the event required to modify the values; infection complexity is directly altered by individual infection events, whereas diversity metrics respond to genetic drift (Hendry et al 2021). This lag time would help explain previous studies (Schultz et al 2010; Bendixen et al 2001) reported a lack of correlation between genetic diversity and contemporaneous transmission intensity across geographical areas.

A more accurate indicator of current transmission levels than genetic diversity may be the proportion of polyclonal infections or the complexity of infections (Hendry et al 2021). Polyclonal infections can result from two processes. Firstly, superinfection, or infections from subsequent mosquito bites in which infective mosquitoes carry unique parasite strains. Secondly, co-transmission, an infection from a single mosquito bite in which an infective mosquito is positive for multiple parasite strains. In high transmission areas where there are more parasites circulating, there are generally more opportunities for superinfection and co-transmission to occur, which leads to a higher incidence of polyclonal infections. In this study, the proportion of polyclonal infections more closely correlated with current PfPR_2–10 in 2014 than expected heterozygosity, a finding which is consistent with previous studies (Morris et al 2015; Nkhoma et al 2013). However, the correlation we found was weak and not statistically significant and may be limited in part by the number of populations in the study and the temporal resolution of the PfPR_2–10 data evaluated (Hendry et al 2021). Indeed, a recent systematic review found polyclonal infection prevalence to be weakly correlated with Plasmodium prevalence (Lopez & Koepfli 2021). Additionally, on a smaller scale of ~10 km area in western Kenya, infection complexity, but not genetic diversity was found to reflect parasite prevalence in areas surrounding an irrigation scheme (Hemming-Schroeder et al 2021). On a temporal scale in Senegal, a dramatic decrease in transmission intensity was followed by a decrease in infection complexity and increased evidence of shared genotype clusters (Bei et al 2018). Notably, the ability to detect superinfection and co-transmission events is affected by the underlying genetic diversity of the population, which is likely still affected by historical transmission levels. With that said, monitoring infection complexity may be a useful supplemental indicator of malaria transmission dynamics, particularly for monitoring temporal changes as models suggest the metric is quite responsive to transmission reductions (Hendry et al 2021) or differences across a small spatial area in which populations share similar historical transmission levels (Hendry et al 2021; Hemming-Schroeder et al 2021; Tusting et al 2014; Neafsey & Volkman 2017).

Genetic variation is influenced by mutation, recombination, drift, natural selection, and gene flow. Gene flow —parasite migration from surrounding areas— can help to counteract the effects of inbreeding and genetic drift resulting from low transmission, thus maintaining genetic diversity in these areas despite dramatic reductions in transmission (Roh et al 2019; Bendixen et al 2001; Zhong et al 2007; Hedrick 2004). Parasite migration may help to partly explain the maintenance of genetic diversity in western Kenya despite the fact that transmission intensity decreased during this time. We assessed recent migration through significance of Identity-By-Descent (IBD) between isolate-pairs as well as historic migration through estimating population genetic parameters by coalescent theory. Interestingly, we found historical migration to be relatively high between the western lowlands and western highlands, but recent migration between the regions to be low. Similarly, historical migration between the western lowlands and rift valley was found to be low, but evidence for recent migration between the two regions was relatively high. These results suggest that parasite migration patterns across Kenya may have shifted, perhaps due to changing patterns in human movement or changing transmission patterns as a result of antimalarial interventions or climate change. Parasite migration between populations in Kenya may help to sustain genetic diversity, and shifting parasite migration patterns may lead to a reduced association between local genetic diversity and transmission intensity. It has been proposed that diversity in the western highlands is driven by parasite migration from the lowland areas (Zhong et al 2007; Shanks et al 2005), but our results do not provide evidence to support that this theory holds true in 2014. This result may be due to the specific highland sites in this study. With that said, local transmission and resulting recombination may be largely responsible for maintaining diversity in some western highland populations. Importantly, sustained high genetic diversity has implications for malaria control, as parasite diversity provides opportunities for adaptation and survival, which facilitates transmission (Gunawardena et al 2014).

Climate variables are known to affect malaria incidence through altering the abundance of suitable mosquito breeding habitats and the duration of mosquito and parasite life cycles. It is therefore not surprising that we found relationships between climate variables and each of expected heterozygosity and proportion of polyclonal infections. We specifically found a positive correlation between ‘precipitation of the driest quarter’ and each of expected heterozygosity and proportion of polyclonal infections. Given that Kenya has a warm, moist climate, low precipitation in the dry season would reflect a higher seasonality in malaria transmission compared to year-round malaria transmission (Mordecai et al 2013). Parasite population genetic diversity and prevalence of complex infections have been found to decrease during seasonally low transmission (Adjah et al 2018), so it would be reasonable that these metrics would be generally lower in areas with marked precipitation seasonality. The majority of models explaining the relationships between temperature and parasite and mosquito development are based on laboratory studies conducted in constant temperature conditions, and monthly temperature averages are often used to describe environmental conditions (Craig et al 1999). Here we note that ‘diurnal temperature range’ had a weak correlation with genetic diversity, in which a high mean diurnal range weakly correlated with reduced parasite diversity in Kenya. These findings are consistent with a life history study, which found that temperature fluctuations around high mean temperatures resulted in slower development time of malaria parasites, which would inhibit malaria transmission (Paaijmans et al 2010). The relationships between temperature and precipitation and genetic indicators of transmission intensity have implications for understanding the potential impacts of climate change on malaria transmission.

Malaria burden shifts toward older ages in areas of lower transmission intensity, as acquisition of immunity is slow compared to high transmission areas (Carneiro et al 2010). Therefore, it is possible that the age of the population in this study was affected by local transmission intensity, as parasite samples were collected passively through a clinic setting. While study population age did differ among study sites, these differences were not associated with contemporary transmission intensity estimates (PfPR_2–10 in 2014). Furthermore, host age may correlate with infection complexity and genetic diversity due to age-dependent acquired immunity (Sondo et al 2020), but several studies have not found such relationships (Mara et al 2013; Sakihama et al 2007; Engelbrecht et al 2000; Aubouy et al 2003; Abamecha et al 2020; Mohammed et al 2019). In our study, we did not find a relationship between host age and proportion of polyclonal infections at the population level, but we did find a positive relationship between population-level age and genetic diversity measured by expected heterozygosity when included with PfPR_2–10 in 2000 and precipitation of the driest quarter. On its own, mean age of the study population very weakly negatively correlated with expected heterozygosity (correlation coefficient = −0.0077) as well as had a negative relationship with precipitation of the driest quarter (correlation coefficient = −0.32) and PfPR_2–10 in 2000 (correlation coefficient = −0.81). These results along with the multivariate modeling for mean age of the study population indicate that study population age was higher in areas with a drier dry season and areas with historically low PfPR_2–10, which could be linked to malaria transmission levels. Thus, site-specific study population age could have been somewhat influenced by local transmission, as well as study population age increased with expected heterozygosity when background historical climate and transmission was accounted for in multivariate models. Notably, incorporating study participant age improved our model of expected heterozygosity, but not for proportion of polyclonal infections. Additional studies parsing out the effects of historical and contemporary transmission, genetic diversity and polyclonality among age groups, and malaria burden among age groups would be helpful to improving our understanding of the impacts of changing malaria transmission on parasite populations.

This study had certain limitations. First, PfPR_2–10 as an indicator of malaria transmission has inherent biases. For example, PfPR_2–10 is affected by acquired immunity and antimalarial drug use and does not capture multiple infections (Hemming-Schroeder et al 2018). In addition, PfPR_2–10 does not reflect prevalence in individuals older than 10 years old, and these older individuals may contribute significantly to sustaining transmission and genetic diversity (Tusting et al 2014). Second, this study was limited to a single time point and few populations for assessing metrics derived from genetic data. Further studies are needed to determine whether the patterns observed in this study are maintained over time with additional populations. Third, parasite populations collected from clinical infections may differ from corresponding parasite populations in subclinical infections. Thus, results from an analysis of subclinical infections may differ from the results presented in this study. Fourth, the comparison of genetic diversity between 2005 and 2014 was done with parasites collected from different populations in the region and from different studies (albeit with the same equipment, protocols, and loci). It is possible that the observed temporal trends are affected by site-specific differences and/or the comparability of microsatellite-based data across studies. Fifth, the bioclimatic variables used do not account for recent climate change since the year 2000. It is possible that genetic diversity or polyclonality correlate more closely with recent weather than historical climate. Sixth, the migration analyses may be limited by the observed subtle population structure and/or population disequilibrium. Finally, detecting changes or differences in genetic diversity and polyclonality may be limited by the choice of genetic marker and/or number of loci, isolates, or populations which were genotyped. It is conceivable that by modifying these variables in the study design, we would be able to detect more patterns.

To conclude, a reduction in transmission intensity is expected to result in reduced diversity, increased linkage disequilibrium, increased geographical differentiation, and decreased polyclonal infection prevalence. In Kenya, malaria transmission decreased from 2000 to 2014, and as expected, we found evidence of increased geographical differentiation since the early 2000s. This result is reinforced by the finding that while historical migration between western lowland and highland populations was relatively high, recent migrations between western lowland and highland populations is relatively low. In contrast, we found no evidence of linkage disequilibrium in 2014, and we did not observe a decrease in parasite diversity between 2000 and 2014. Rather, we found that genetic diversity was maintained and may be more reflective of past transmission rates than current rates. Polyclonal infection prevalence weakly correlated with contemporary transmission rates, but significantly correlated with historical climate variables. Additional population genetic studies at different spatial and temporal scales would improve our understanding of how changes in climate, landscape, and transmission intensity impact malaria populations. Knowledge of these relationships is important to understanding how malaria interventions and climate and land-use change impact the dynamics of malaria transmission and biology.

Data availability

The data that support the findings of this study are openly available in Dryad at http://doi.org/10.5061/dryad.0zpc8677c.