Background
Despite increased global efforts to control malaria,
Plasmodium falciparum malaria remains a major public health issue [
1], and a priority of the Millennium Development Goals. Attempts to eliminate malaria in Africa have been hampered by resistance of
P. falciparum, first to chloroquine and later to sulphadoxine-pyrimethamine (SP). The recent emergence in Southeast Asia of resistance to artemisinin, a key component of current first-line therapy, is a major worldwide public health concern [
2,
3], and a significant impediment to future malaria eradication [
4].
Chloroquine was the most common first-line therapy until the late 1990s. Resistance to chloroquine was first detected in Asia in the 1950s [
5,
6] and spread to Africa in the late 1970s [
7]. As chloroquine failure became widespread throughout sub-Saharan Africa in the following decades, it was replaced with SP in some countries as the first-line therapy. Resistance to SP rapidly followed its introduction [
8] and has been extensively documented in many regions of Africa [
9].
The emergence and spread of parasite resistance is a two-staged process:
de novo appearance of a parasite genotype that confers better survival in the presence of the drug, followed by the preferential transmission of organisms with the acquired resistance [
10]. Molecular markers alone cannot be used to predict treatment outcomes in individual patients, because other factors such as immunity, nutritional status, haemoglobinopathies and variation in drug absorption and metabolism can also affect clinical outcomes [
11‐
14]. However, validated molecular markers are useful tools for mapping and monitoring anti-malarial resistance at a population level and as a surveillance tool, to indicate an aggregated measure of increased risk for clinical failure [
15].
Point mutations occur
de novo, independently within individual parasites, but resistance to SP is a complex trait, requiring a specific constellation of changes in two unlinked genes. Molecular studies have shown that the combination of the three mutations in
dhfr (S108N, C59R, N51I) defines a key highly pyrimethamine-resistant combination or haplotype. A parasite that also carries a ‘double’ mutant allele of
dhps (A437G, K540E) is strongly associated with increased risk of SP treatment failure in Africa [
16‐
19].
These molecular markers have been productively used in population analysis of the molecular changes underlying the development of SP resistance in Africa. The correct set of point mutations in the
dhps and
dhfr genes occur together relatively infrequently but once assembled, parasites that carry these combinations then spread over large-scale geographic regions. In fact, the emergence of parasites with mutations in the
dhps gene, is almost always observed in populations that already carry the triple mutant allele of
dhfr [
8]. Most commonly in Africa, mutant alleles of
dhps are selected in a stepwise fashion; an intermediate allele that carries A437G alone, followed by selection of the A437G + K540E associated with clinical SP resistance. Maps of the distributions of observed
dhps mutant haplotypes show that the 437G single mutant haplotype is commonly found in West Africa while the 437G + 540E double mutant haplotype is prevalent throughout East Africa [
20]. The study further analysed the number of independent origins of these alleles and showed that the dispersal of five major mutant lineages (three different 437G alleles and two 437G + 540E double mutants) accounted for the majority of the observed haplotypes in African populations. It is worth noting again that mutations in the
dhps gene only arise in populations that carry a high prevalence of the triple
dhfr mutation [
8]. The addition of A581G and/or A613T/S confers even higher levels of resistance
in vitro[
21]. Although comparatively rare, the 581G in combination with 437G and 540E in the form of a triple mutant allele is responsible for a measureable deterioration in SP efficacy [
22,
23]. Currently the WHO recommends SP for intermittent prophylactic treatment of pregnant women and infants, but in populations where 50% or more of the parasites carry a
dhps 540E allele, this is no longer recommended [
24,
25].
Lessons learned from the spread of these markers of SP resistance in Africa can be used as a model of anti-malarial spread in the continent. This work was motivated, not solely to investigate the spatiotemporal spread of molecular markers that confer resistance to SP, but also to predict the likely spread of resistance to artemisinin combination therapy (ACT) when and where it arises.
Methods
For each of the A437G, K540E and A581G mutations, the literature was searched and data extracted on the time and location of the survey, the number of people who were tested and the number of people who were positive for that mutation: that information was recorded in a database [
9,
25]. For visualizations of the molecular marker data, see [
26] and [
27]. Additional file
1 contains a summary of the timing and location of the surveys captured. In this paper the aim was to use the A437G, K540E and A581G data to infer a continuous surface for the prevalence of K540E, in order to inform an understanding of the emergence and dispersal of drug resistance in African populations of of
P. falciparum. A continuous surface was inferred based upon the observations of prevalence, and a statistical approach employed to estimate the prevalence of
dhps 540E at locations between the observations.
Throughout this paper, the prevalence of dhps 540E refers to the proportion of infected individuals in a population that are infected with one or more resistant clones. The distinction between the prevalence and frequency of a molecular marker is an important one. Briefly, the frequency of a molecular marker is the proportion of parasites in the population that carry the marker in question (taking into consideration that a single person can be infected with multiple clones) while the prevalence is the proportion of all individuals that are infected with one or more resistant clones. Even if the frequency of the molecular marker is the same across space and/or time, individuals will tend to be infected with more clones in regions of high malaria transmission. For this reason, frequency is the measure upon which the genetics of allele spread should ideally be modelled. However, genetic studies typically measure or report marker prevalence and while it is possible to use a statistical model to infer marker frequency from prevalence data, for the purposes of this paper the primary, individual patient level data were not available. Only a single aggregate prevalence from each study site and location was available.
The purpose of the modelling approach used in this paper was to generate a continuous surface, in both space and time, which approximates the prevalence of the
dhps 540E marker. That is, the
dhps data, available only at discrete study locations and times, were used to predict the changing prevalence of
dhps 540E across the entire African continent from 1990-2010, thus providing insight on the spread of resistance, in both space and time, in a way that observations from the discrete data points alone can not provide. The model included estimates of the
P. falciparum transmission intensity in 2010, as estimated by the spatiotemporal models developed by the Malaria Atlas Project [
1]. Since multiplicity of infection (MOI) is an indicator of, and positively correlated with, transmission intensity, the inclusion of transmission compensates, to some extent, for the omission of MOI. Full details of the model are provided in Additional file
2.
There were two main stages to the statistical methodology for the spatiotemporal prediction of the
dhps 540E molecular marker prevalence, which are outlined briefly here (see Additional file
2 for details). Firstly, based on the observed data, the model parameters were estimated. Secondly, given the model parameters in the first stage,
dhps 540E prevalence was predicted on a 25 × 25 km grid of sub-Saharan Africa in each year from 1990 to 2010. For each location, a distribution of prevalences were drawn from the model and summarized using the median statistic to create a single continuous surface. The standard deviation surface is also presented alongside the median maps as a summary of the associated uncertainty in the predictions at each location.
The model validity was assessed to ensure that the interpretation of the model output was valid. The
dhps 540E dataset was divided into five groups, at random; each subset was treated as a validation set to test the model’s predictive ability. For each of the five subsets of data, the model was run with one dataset withheld and the ability of the model to predict that subset was tested against the actual withheld data. The predictive results for each of the five subsets of data were pooled, so that each
dhps 540E observation had an associated predictive validation value. Further details about the validation procedures are given in Additional file
3.
Discussion
In this paper a spatiotemporal model was developed for the prevalence of the dhps 540E marker in sub-Saharan Africa. Continuous maps in space and time were presented for the median predicted dhps 540E prevalence, along with the associated uncertainty in these predictions. Based on the data available, the emergence and spread of the dhps 540E mutation has been visualized at all space-time points within a 25 x 25 km gridded sub-Saharan Africa domain for each year between 1990 and 2010. In modelling the spatiotemporal emergence and spread of drug resistance in Africa, the results of this work have significance for public health and the future management of artemisinin resistance in Africa.
The maps presented here (see Figures
1,
2 and
3) visualize the changing prevalence of the
dhps 540E mutation in sub-Saharan Africa and are a useful tool to inform public health policy on the continuing use of SP. It could be used to advise on regions where SP should not be used as a partner drug, as a seasonal malarial chemoprevention or as an intermittent preventative treatment. For example, currently the WHO recommends SP for intermittent prophylactic treatment of pregnant women and infants, but in populations where 50% or more of the parasites carry a
dhps 540E allele, this approach is no longer recommended [
24,
25]. These regions could easily be predicted using the model developed here.
Previous attempts to spatially map the molecular markers associated with SP have been made [
28], however the model presented here extends this work since it is spatiotemporal and also because it is embedded in a Bayesian framework, allowing the uncertainty to be quantified. The predicted surface generated as part of this previous work differs considerably from the maps of
dhps 540E prevalence presented here because the previous study considers all the prevalence data to influence a single continuous map equally, regardless of the year in which the data were collected. Whereas here a space-time model framework was considered that ensured that studies conducted in a particular year influence the map for that year more than a study conducted many years before.
The limited number of spatial and/or temporal data points available in certain regions of Africa, greatly affect the predictive value of the model, and as a result, the level of uncertainty of the median
dhps 540E prevalence estimates can be relatively high (see Figure
3). This observation in itself is important. While SP has been the most studied anti-malarial, in terms of molecular markers over the last three decades, major gaps of information remain. These gaps reflect the absence of research activities in particular regions, a lack of systematic reporting of available data, and/or a limited access to unpublished data.
There are several ways that this work could be extended: in addition to transmission intensity, which was incorporated into the current model, other, informative covariates, such as human population estimates within Africa and spatial environmental variables could be included within the model framework. The temporal trends of
dhps 540E at certain locations (see Figure
4 for the predictive distributions of
dhps 540E prevalence in the Kilifi region, Kenya and the Bamako region, Mali in 1990, 2000 and 2010) are likely to be dependent both on the national anti-malarial policy and actual drug use. Likewise, the temporal trends shown in Figure
5 (illustrating the proportion of sub-Saharan Africa with a predictive median
dhps 540E prevalence exceeding various thresholds) will be related to both factors. For instance, the period of rapid increase in the temporal trends shown in Figure
5 corresponds to the years when the highest number of African countries were recommending SP as a first-line therapy [
29]. An extension of this work will add SP drug pressure and national anti-malarial drug policy to inform more accurately the spatiotemporal spread of SP resistance.
Many malaria endemic countries are working to eliminate malaria and up-to-date intelligence on the various parameters that are likely to impede such progress is critically important. The gaps in data reflect not only technical, human, political and financial constraints but also difficulties in establishing the optimal sites to survey, in terms of predictive value and representativeness. The work presented here will be extended to investigate the design of future surveillance, informed by the level of uncertainty, the level of
dhps 540E prevalence, malaria transmission [
1] and human population density [
30]. The aim would be to define the minimum spatiotemporal set of data necessary to design comprehensive surveillance matrices, allowing resistance mapping with acceptable level of uncertainties. This concept, called “smart surveillance”, has the potential to inform a guided surveillance plan, less based on current research capacities and more on where informative data are most needed. By adopting such an approach, mathematical modelling can facilitate the information systems needed to optimize the current efforts in malaria elimination and eradication.
The methodology outlined in this paper serves as a proof of concept for the application of geospatial modelling techniques to other anti-malarial drugs, as well as forms of data other than molecular markers, and will allow anti-malarial resistance to ACT or novel drugs to be monitored in space and time.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
SH, CR and PG designed the study. JF implemented the algorithms, with the help of AP. JF wrote the first draft of the manuscript. CR and IN supplied the molecular data. CS and MV helped with the interpretation of the results. All authors read and approved the final manuscript.