Introduction

Mosquitoes of the Anopheles gambiae complex are the principal vectors of malaria in sub-Saharan Africa. The complex comprises six sibling species that differ in their ecology and epidemiological importance. Two species, An. gambiae s.s. and An. arabiensis, are regarded as major vectors; another three species, An. merus, An. melas and An. bwambae, are of localized importance, whereas the sixth, An. quadriannulatus, is zoophilic and hence not a vector. A knowledge of population structure in the major vector species is fundamental to an understanding of malaria epidemiology and the spread of insecticide resistance. To date most published work on population structure within the complex has focused upon An. gambiae s.s. (Coluzzi et al., 1985; Lehmann et al., 1997; and refs. therein). Although the two species exist sympatrically over much of their species range, An. arabiensis extends into more arid environments, is more likely to rest in outdoor structures than in the interior of houses and will feed preferentially on animals rather than humans (Charlwood et al., 1995). Such differences in ecology and behaviour between An. arabiensis and An. gambiae may be reflected in differences in their population structure.

The aim of this study was to determine how the population structure of An. arabiensis differs from that previously reported for An. gambiae, where possible relating this to known differences in the biology of the two species. Our approach has been to use primers for six polymorphic microsatellite loci, developed for use in An. gambiae s.s., to investigate genetic differences and possible substructure in populations of An. arabiensis from Tanzania and Mozambique in East Africa.

We have investigated mosquito population substructure at three levels: (i) in samples collected from different houses within a village in Tanzania; (ii) in samples collected from two villages, 20 km apart, in Mozambique; (iii) between samples from Mozambique and Tanzania separated by 2000 km. From the derived allele and genotype frequencies we have compared three estimates of population differentiation and we discuss their appropriateness and their significance for our understanding of the biology of this species.

Materials and methods

Sample sites

Mozambique Samples were collected from two villages close to Maputo (25°58′S, 32°36′E), Mozambique. Matola is a coastal peri-urban area on the outskirts of Maputo, whereas Boane (25°2′S, 32°19′E) is approximately 26 km from Maputo and 20 km from Matola. The climate of this region is dry tropical with monthly rainfall fluctuating between 0 and 326 mm and an annual average of 650 mm. The average monthly temperature varies from a low of 19°C in July to a high of 26°C in February. The wet season lasts from November to April with a dry season from May to October and malaria is classified as holoendemic. Mosquitoes were collected inside houses by human landing catches and resting collections between December 1995 and March 1996.

Tanzania Resting collections of mosquitoes were carried out from 18 to 21 July 1995. Samples came from three houses in the village of Kivukoni (8°09′S, 36°24′E), 7 km south of the town of Ifakara on the banks of the Kilombero River, in an area of intense malaria transmission. There are approximately 12 houses in the settlement of Kivukoni all within a radius of 250 m. The area has two rainy seasons. The main rains last from March to May and are followed by a dry season that extends from June to November. A shorter rainy season occurs in December and January and is followed by a brief dry season before the onset of the main rains. The majority of the population are subsistence farmers, predominantly cropping maize and rice (Charlwood et al., 1995).

DNA extraction, species identification and microsatellites

All mosquitoes were stored dry over silica gel prior to DNA extraction. DNA from all specimens was extracted using a modification of the phenol:chloroform technique of Ballinger-Crabtree et al. (1992) and resuspended in 100 (L of TE buffer. Each specimen was then identified to species level using species-specific PCR primers, following the technique of Scott et al. (1993). Of 13 locus-specific primers (Primers 1D1, 2A1, 29C1, 33C1 from Lehmann et al., 1996a; Ag2H26, Ag2H46, Ag2H102, Ag2H175, Ag2H1010, Ag3H83, Ag3H88, Ag3H93, Ag3H249, from Zheng et al., 1996) designed to amplify regions of the An. gambiae s.s. genome, 10 were found to amplify consistently in An. arabiensis (Primers 1D1, 29C1, 33C1 Ag2H26, Ag2H46, Ag2H102, Ag2H175, Ag2H1010, Ag3H88, Ag3H249) and eight in An. merus (Primers 1D1, 29C1, 33C1, Ag2H26, Ag2H102, Ag2H175, Ag2H1010, Ag3H249). Six primers were selected for use in a more detailed study of An. arabiensis population substructure and gene flow (Table 1). All of these loci, except locus 33C1, lie outside the major known chromosomal inversions (Zheng et al., 1996).

Table 1 Microsatellite loci studied in more detailed survey of population differentiation and gene-flow in Anopheles arabiensis
Full size table

PCR using fluorescent markers

PCR products were labelled fluorescently using one of two methods. Either one of the primer pair was labelled with a 5′ fluorescent dye (6-FAM, HEX or TET; GIBCO, Palo Alto, CA) or fluorescent UTPs were added to reactions with unlabelled oligonucleotide primers. PCR reactions were carried out in 25 μL in 0.5 mL Eppendorf tubes with a mineral oil overlay. Each reaction contained approximately 1/100th of the genomic DNA of an individual mosquito, 1 × reaction buffer, 200 μM of each dNTP, 1.5 mM MgCl2, 0.5 Units of DNA Polymerase (DyNAzymeII DNA Polymerase, Finnzymes Oy) and 12.5 pmols of each primer. In reactions where fluorescently labelled primers were not used, fluorescently labelled UTP was included in the reaction mix at a concentration of 6.4 nM. All amplifications were performed in a Hybaid Omnigene thermal cycler using the following amplification programme: a 5-min 94°C denaturation step followed by 30 cycles of 30 s at 94°C, 30 s at 55–60°C (primer dependent) and 30 s at 72°C; there was a final extension step of 5 min at 72°C. Products were run on a Perkin-Elmer ABI PRISM 377 Automatic Sequencer using the default settings. Data were automatically collected and analysed using GENESCAN and GENOTYPER software (Applied Biosystems). The alleles were sized using a ‘local southern’ option. This technique estimates allele sizes by using the three molecular weight markers closest in size to the product to determine a best fit estimate.

Data analysis

Linkage disequilibrium Linkage equilibrium was tested using a contingency table test for genotypic linkage disequilibrium between pairs of loci in a population, based upon the null hypothesis that genotypes at one locus are independent of genotypes at the other locus. Calculations were performed using the GENEPOP V3–1 program (Raymond & Rousset, 1995) which performs a significance test using Markov chain procedures.

Hardy–Weinberg equilibrium (HWE) Tests of deviation from Hardy–Weinberg proportions were performed using the GENEPOP V3–1 program (Raymond & Rousset, 1995) and based on the score test (U-test) with a null hypothesis of random union of gametes and an alternative hypothesis of heterozygote deficiency.

Nei’s unbiased estimates of heterozygosity The within-population heterozygosity at each locus was estimated using Nei’s (1978) unbiased estimator he=2n(1 − Σpi2)/(2n − 1), where pi is the frequency of the ith allele and n is the number of individuals in the sample. Standard errors were calculated based upon a variance of Σ(he − He)2/n(n − 1), where He=Σ(he)/n.

Estimates of population differentiation and gene flow The differences between populations were investigated by three methods: (i) exact tests of genotype homogeneity among populations (Goudet et al., 1996); (ii) Weir & Cockerham’s (1984) estimates of F statistics; (iii) Slatkin’s (1995) RST statistic.

Genotype-based exact tests of population differentiation were calculated using the GENEPOP V3–1 program (Raymond & Rousset, 1995). Tests of between-population genotype frequency homogeneity were calculated using an unbiased estimate of the P-value of a log-likelihood based exact test. A Markov chain method was used to generate an unbiased estimate of the exact probability. The parameters of the Markov chain were the default values of the program. If more than two sample populations were present, all population pairs were tested. A simple comparison of genotypic frequencies can be used to compare populations or loci even when they are not in HWE. Although exact tests of genotype and allele frequencies may be the most sensitive detectors of population differentiation, they provide no estimate of the magnitude of the differences: to assess population structure in a quantitative manner it is necessary to use either FST or RST estimators.

F-statistics (FIS, FIT and FST) were calculated using the FSTAT program (Goudet, 1995). This program calculates unbiased estimates of F-statistics and performs numerical re-sampling by bootstrap and jackknife procedures in order to estimate confidence intervals and the significance of values. The average FST value was calculated following Weir & Cockerham (1984).

Slatkin’s (1995) RST statistic is similar to FST but developed specifically for microsatellite loci to account for the assumed differences in mutational processes at these loci. FST-based measures of differentiation are based on an infinite allele model (IAM) or k-allele model (KAM) and may not be applicable to microsatellite loci which are thought to follow a stepwise mutation model (SMM) more closely. Calculation of RST, based upon the repeat numbers of alleles, assumes that there is no restriction on allele size and that the mutation rate is relatively constant across all alleles. In practice there is emerging evidence to suggest that both these assumptions may sometimes be violated (Amos et al., 1996; Lehmann et al., 1996b).

RST statistics were calculated by hand. Average, across loci, values for RST were calculated following Slatkin’s procedure of averaging the denominators and numerators for all loci prior to calculation. The significance of RST values was determined by a simple ANOVA test. In our discussion we consider the most appropriate measure of population differentiation with reference to both the possible underlying mutation processes at individual loci and the methods of calculating each statistic.

As an estimate of gene flow, the number of migrants per population per generation (Nm), was calculated for FST according to the equation Nm=(1 − FST)/4FST. Calculations of Nm for RST followed Slatkin (1995), Nm=(d −1) (1 − RST)/4d RST, where d is the number of populations in the study. These calculations assume an island model of population structure and mutation–drift equilibrium. As discussed below, this may not be the most appropriate model for these populations and, indeed, there may be no simple relationship between estimators of genetic differentiation and gene flow.

Mutation model The underlying mutation model influences which method of calculating gene flow is most appropriate for the data set. The appropriateness of the mutation model for these loci was examined following the example of Garcia de Leon et al. (1997), based on the simulations of Shriver et al. (1993) who demonstrated that when values of heterozygosity are greater than 0.5, the number of alleles for a given heterozygosity can be significantly larger for a locus following the Infinite Alleles Model (IAM) than for one following the Stepwise Mutation Model (SMM). This permits a simple empirical test of which model is more likely to fit the data, although no estimates of statistical confidence are possible by this test.

In all cases where multiple tests were performed, significance levels were adjusted following Bonferroni procedures (Holm, 1979).

Results and discussion

All An. gambiae s.l. from Mozambique were identified as either An. arabiensis or An. merus, the relative proportions of An. arabiensis being 103/122 in Matola and 44/85 in Boane. In Kivukoni, Tanzania all specimens were identified as either An. gambiae (54) or An. arabiensis (89). Subsets of the An. arabiensis samples were taken and used in the following analysis of population substructure and gene flow.

All loci save one (29C1) were highly polymorphic with between four and 14 alleles per locus, excluding null alleles (see later discussion). For the pooled samples, excluding locus 29C1, the average number of alleles per locus was 9.6 and the average unbiased estimate of heterozygosity was 0.739 (Table 2). There were no significant associations, indicative of linkage disequilibrium, between any pair-wise combination of alleles across loci at any population level (P > 0.05; 120 pair-wise comparisons comprising 15 pair-wise comparisons between loci for eight samples; significance level adjusted for multiple tests).

Table 2 Observed (upper value) and expected (lower value) estimates of heterozygosity of Anopheles arabiensis for each sample by loci, the number of alleles found at each locus for pooled populations and details of which mutation model the loci most closely fit: the Infinite Allele Model (IAM) or the Stepwise Mutation Model (SMM)
Full size table

Deviations from HWE

Five, of the six loci (except 29C1) showed significant deviation (P<0.05) from HWE as a result of heterozygote deficit. In Table 3 values of FIS for each locus at each population level are given together with the significance of the test of deviation from HWE based on an alternative hypothesis of heterozygote deficit. It can be seen that deviation from HWE as a result of heterozygote deficit is detected at all population levels although not necessarily within all sample/loci combinations at a population level.

Table 3 Estimates of FIS and significance level for goodness of fit tests to HWE, based upon a null hypothesis of heterozygote deficit, for all Anopheles arabiensis populations. Also FIT value for pooled population
Full size table

Deviation from HWE is a common finding in studies that utilize microsatellite loci. This deficit is usually attributed to null alleles (Callen et al., 1993; Lehmann et al., 1996a; Garcia de Leon et al., 1997), selection (Garcia de Leon et al., 1997), or grouping of gene pools (Wahlund effect) (Gibbs et al., 1997). In addition, inbreeding or nonrandom mating may also result in heterozygote deficits. Individuals in the Tanzanian sample were collected from three houses and it was therefore possible to investigate whether there was indication of grouping at the village level. Only one locus (Ag3H88) gave evidence of between-house differences but this result was no longer significant when significance levels were adjusted to account for multiple testing. Thus any heterozygote deficit is unlikely to result from pooling samples from different houses into village collections.

Null alleles, that is alleles that are not amplified because of mutations at the primer binding sites, can result in underestimation of heterozygosity because they are detected only in the homozygous state. An estimate of the expected frequency of the null allele can be made by assuming that the observed heterozygote deficit is attributable to null alleles. Following Chakraborty et al. (1992) the estimated frequency of the null allele, r=(He − Ho)/(He + Ho), where He= expected frequency of heterozygotes and Ho = the observed frequency of heterozygotes. From this figure it is possible to calculate the expected number of homozygous null individuals likely to be observed in a sample of a given size. Table 4 details both the observed and expected number of null allele homozygotes. A specimen which was shown to have amplifiable DNA but would not produce a product for a given locus on three occasions was regarded as a null allele homozygote. It can be seen that in general, the frequency of the putative null allele homozygotes agrees with the expected value indicating that in these cases null alleles are the likely cause of heterozygote deficit. However, for two loci (Ag3H88 and Ag3H249) there are higher than expected numbers of putative null allele homozygotes. Examination of the raw data showed that in 37% of cases, an individual that was a putative null allele homozygote at locus Ag3H88 was also an apparent null allele homozygote at locus Ag3H249. This suggests that a number of these supposed null homozygotes are caused by a failure of amplification because of variations in template quality. Exclusion of null allele homozygotes from the analysis will affect the allele and genotype frequency distributions and this could lead to inaccuracies in measures of population structure. Previous cytological studies have found no evidence for deviation from HWE in east African populations of An. arabiensis (Petrarca & Beier, 1992), which suggests that there is unlikely to be nonrandom mating in our samples. This is also consistent with the observation that populations of An. gambiae s.l. are normally in HWE (Petrarca & Beier, 1992; Lehmann et al., 1996a). The latter authors found null alleles in locus Ag2H46 in An. gambiae s.s. (Lehmann et al., 1996a) and concluded that they were likely to be present also at locus 33C1 (Lehmann et al., 1997). It seems likely that the presence of null alleles at these loci is the cause of the heterozygote deficit we see in our An. arabiensis populations.

Table 4 The numbers of observed (upper value) and expected (lower value) null allele homozygotes for each locus at each Anopheles arabiensis population level
Full size table

Population substructure between villages in Mozambique

Allele frequencies at the six loci were similar for both populations from Mozambique. There was no evidence for population differences between the samples from either FST or RST (mean FST=0.010; mean RST=0.008) or from analysis of genotypic frequencies in each population (P > 0.05), although it should be noted that the sample from Boane is relatively small (n=23). Thus the indications are that populations of An. arabiensis are effectively panmictic in this region within an area of radius 20 km Lehmann et al. (1997) also found no evidence for between-house differentiation in An. gambiae from western Kenya and concluded that the area occupied by a deme of An. gambiae was in excess of a radius of 25 km.

Estimation of genetic differentiation between Mozambique and Tanzania

The sample from Matola in Mozambique was compared with the population from Kivukoni in Tanzania. At only three of the six loci were the most common alleles the same in both populations (Fig. 1) and population-specific alleles were found in four of the loci examined. For only three of the 14 population-specific alleles was the allele frequency per locus per population greater than 5%. For all but one locus the sample from Tanzania produced higher estimates of heterozygosity than did the sample from Matola. This suggests that the Mozambique sample may be drawn from a parent population that is genetically depauperate, perhaps as a result of lower effective size. The Mozambique samples are from the extreme southern end of the geographical range of An. arabiensis where habitat fragmentation may reduce effective population size. Apparent decreases in genetic diversity towards the geographical edges of a species’ range have been observed in Drosophila (Dobzhansky, 1970 and refs. therein).

Figure 1
figure 1

Allele frequency arrays in Anopheles arabiensis for (a) locus 29C1, (b) locus 33C1, (c) locus Ag2H46, (d) locus Ag3H88, (e) locus Ag2H175, (f) locus Ag3H249.

Full size image

Exact tests showed that the complement of genotypes at five out of six loci was highly significantly different between Tanzania and Mozambique (P<0.05 for genotypes; Table 5); the locus that showed no difference (locus 29C1) was close to fixation for the same allele in both populations. Because it contributes nothing to the study of population structure, this locus has been excluded from further analysis and discussion. The FST values of four of the five remaining loci, were significantly different from zero (P<0.05–P<0.001) (Table 5). However, there was considerable heterogeneity between loci, with estimates of FST ranging from 0.015 to 0.239. RST values also showed a large variation across loci (0.002–0.098, mean 0.027) and were only significant at three out of five loci (Table 5). There was no detectable correlation between FST and RST values at each locus or across all loci averaged.

Table 5 Estimates of genetic differentiation and geneflow (Nm) between samples of Anopheles arabiensis from Kivukoni, Tanzania and Matola, Mozambique
Full size table

Because the genotype frequencies at five of the six loci were shown to be significantly different, the nonsignificant values of FST and RST statistics at some loci raises the question whether FST and RST and the tests of their significance are sufficiently sensitive. The large between-locus variation in values of both FST (0.015–0.239) and RST (0.002–0.098) may simply reflect sampling effects but may be caused by selection upon some of the loci; possibly a result of linkage effects with nearby coding regions or chromosomal inversions. Lehmann et al. (1996b) inferred that locus Ag2H46 was under mutational constraint in An. gambiae but it is not known whether this locus, or others, are similarly constrained in An. arabiensis.

This study highlights the variations in sensitivity of FST and RST for detecting evidence of differentiation. For example, the allele frequency distribution for locus 33C1 (Fig. 1b) shows a large difference between populations, with the modal and the next most frequent allele being different in each population. This results in high estimates of population differentiation and low estimates of gene flow (θ=0.239, Nm=0.79) when the FST statistic is used. However, the RST value for this locus is low, implying high levels of gene flow (RST=0.022, Nm=5.55). The converse is observed at locus Ag2H175 (θ=0.026, Nm=9.39; RST=0.098, Nm=1.15) with allele arrays that are apparently very similar. FST, which measures variation across the allele array, is more sensitive to variation in allele frequencies across the array than is RST which may give undue weight to alleles furthest from the median class. In addition, only one of the loci studied here conformed primarily to a SMM, consistent with the growing evidence for deviation from a strict SMM for microsatellite loci (Di Rienzo et al., 1994; Garcia de Leon et al., 1997) (Table 2). Furthermore, we have found a number of polymorphisms involving insertions or deletions that result in the formation of electrophoretically identical but nonhomologous alleles (Donnelly & Townson, in prep.). This homoplasy may affect both FST and RST estimators (Viard et al., 1998). Recent data suggest that mutation rates at microsatellite loci are lower than has previously been assumed. Thus Schug et al. (1997) have observed much lower rates of mutation at microsatellite loci in Drosophila melanogaster (6.3 × 10–6) than in mammalian species (10–3–10–5). If this lower mutation rate holds for An. arabiensis, populations that have recently become isolated may not be detected using RST estimators. Constraints upon mutation at each locus would also exert a similar effect in reducing the accumulation of mutations in each population. A lower rate of mutation at these loci would increase the influence of drift on allele frequencies and thereby make FST a more appropriate estimator.

There is a growing body of evidence that RST may not be an appropriate estimator of population substructure for all microsatellite data. Forbes et al. (1995) found that FST was more sensitive to differences between allopatric populations and Perez-Lezaun et al. (1997) observed that genetic distance measures such as FST, which do not consider mutational relationships among alleles and which have known relationships to differentiation by drift, better reflect currently understood patterns of human evolution than do mutation-based distance measures, such as RST.

Estimates of gene flow

Large estimates of geneflow (Nm) were derived from both FST and RST. The calculation of gene flow from FST statistics assumes that migration occurs at random among all populations (an island model), that the populations are in migration–drift equilibrium and that the loci are neutral. Our estimates of gene flow based on FST statistics may be too large as they account solely for drift and do not account for mutation rates at these loci. However, notwithstanding our caveat about using values of gene flow averaged across loci, the values of Nm in excess of 3, as determined both by FST and RST (Nm=3.37 and 4.9, respectively), appear to suggest that there are apparently few barriers to gene flow over 2000 km. However, we would argue that these high values of Nm are more likely to reflect large effective population sizes and recent separation than large-scale migrations. This result is in broad agreement with studies based upon multi locus allozyme electophoresis and chromosome polymorphisms which showed low levels of population differentiation in An. arabiensis from south of the Rift Valley (Petrarca & Beier, 1992). In comparison, Lehmann et al. (1996a) obtained similar values for gene flow between populations of An. gambiae from east and west Africa separated by a distance of approximately 6000 km (Nm 7.7 and 3.4 for FST and RST, respectively). The lower levels of gene flow observed in this study are heavily influenced by locus 33C1. Removal of this locus from the analysis (Table 5) results in higher estimates of gene flow (mean Nm=11.11). Locus 33C1 is within the 3Ra chromosomal inversion strongly suggesting that loci within inversions may behave differently.

Conclusion

Appetitive flights in anopheline species are usually 1–2 km and this is not readily reconciled with the apparently high levels of gene flow observed in this and other studies. It has been proposed that the low values of FST and RST for An. gambiae are attributable to either constraints on microsatellite loci or a large amount of human-mediated transport (Lehmann et al., 1996a). Although there are numerous examples of long-range transport of members of the An. gambiae complex by land, sea and air (for review see Gillies & De Meillon, 1968), the large estimates of the numbers of migrants per year implicit in our calculated values of FST and RST seem too great to be attributable to such a cause. Anopheles arabiensis commonly feeds on cattle and larvae are often found in rice-fields. It seems likely that this species has extended its range relatively recently, aided by the expansion of human settlement and intensive agriculture. We would argue that the population structure we observe in An. arabiensis reflects the relatively recent expansion of its population range and the large effective size of its populations, rather than a large amount of contemporary gene flow.