Falciparum malaria from coastal Tanzania and Zanzibar remains highly connected despite effective control efforts on the archipelago

WGS was attempted on 106 P. falciparum isolates collected from subjects with uncomplicated malaria or asymptomatic infection from 2015 to 2017. Forty-three of these were leukodepleted blood collected as part of an in vivo efficacy study of artemether–lumefantrine (AL) in paediatric uncomplicated malaria patients collected from 2015–2017 in Yombo, Bagamoyo District. A remaining 63 isolates were from dried blood spots (DBS) collected in Zanzibar in 2017. These came from cross-sectional surveys of asymptomatic individuals (n = 34) and an in vivo efficacy study of artesunate–amodiaquine (ASAQ) with single low dose primaquine (SLDP) in paediatric uncomplicated malaria patients (n = 29). These isolates essentially represent a convenience sample. Isolates were not selected for sequencing on the basis of specific clinical or epidemiologic characteristics; however, sequencing was more likely to be successful on isolates from subjects with high parasitaemia. Study participants from Zanzibar were asked to report any overnight travel away from home in the past 4 months. Responses were coded as yes (overnight travel to mainland Tanzania or Kenya) or no (no overnight travel off islands of Zanzibar). Clinical characteristics of the attempted and sequenced samples from each cohort from Zanzibar is provided in Additional file 1: Table S1.

Leukodepleted blood samples and DBS were extracted using QIAmp 96 DNA blood kits per the manufacturer protocol (Qiagen, Hilden, Germany). DNA from leukodepleted blood was acoustically sheared using a Covaris E220 instrument, prepared for sequencing without enrichment using Kappa Hyper library preps, and individually barcoded per manufacturer’s protocol (Kappa Biosystems, Columbus, OH). DNA extracted from DBS was enriched for P. falciparum DNA before library prep using two separate selective whole genome amplification (sWGA) reactions. The sWGA approach was adapted from previously published methods and employed two distinct sets of primers designed for P. falciparum, including the Probe_10 primer set described previously by Oyola et al. and another set of custom primers (JP9) designed using ‘swga’ [19‐21]. Phosphorothioate bonds were included between the two most 3’ nucleotides for all primers in both sets to prevent primer degradation. Design and evaluation of these custom primers and the sWGA approach are described in Additional file 1: Table S2. The two sWGA reactions were carried out under the same conditions. The products of the two sWGA reactions were pooled in equal volumes and acoustically sheared using a Covaris E220 instrument before library preparation using Kappa Hyper library preps. The indexed libraries were pooled and sequenced on a HiSeq 4000 using 2 × 150 chemistry at the University of North Carolina High Throughput Sequencing Facility. Sequencing reads were deposited into the NCBI SRA (Accession numbers: pending).

Illumina short read WGS data for P. falciparum isolates was downloaded from public databases. This included 68 isolates from other regions of Tanzania, collected between 2010 and 2013, as well as 179 isolates from other regions, including Southeast Asia, South Asia, East and West Africa (Additional file 1: Table S3).

Raw paired-end reads were trimmed for adapter sequences with ‘cutadapt’ v1.18 and aligned to the P. falciparum 3D7 reference genome (assembly version 3, PlasmoDB version 38: https://​plasmodb.​org/​common/​downloads/​release-38/​Pfalciparum3D7/​fasta/​data/​PlasmoDB-38_​Pfalciparum3D7_​Genome.​fasta) with ‘bwa mem’ v0.7.17-r1188. Duplicates were marked with ‘samblaster’ v0.1.24. A position was defined as “callable” if it was covered by ≥ 5 high-quality reads (MQ ≥ 25, BQ ≥ 25), and computed the proportion of callable sites in each isolate was calculated with the Genome Analysis Toolkit (GATK) ‘CallableLoci’ tool v3.8-0. Only isolates with ≥ 70% of the genome callable were used for further analysis.

Short sequence variants (including SNVs, indels and complex multi-nucleotide variants) were ascertained in parallel in each isolate using GATK ‘HaplotypeCaller’ v.4.0.3.0, then genotyped jointly across the entire cohort with GATK ‘GenotypeGVCFs’ according to GATK best practices. Variant discovery was limited to the core (non-hypervariable) nuclear genome as defined by Miles et al. [22]. Putative SNVs only were filtered using the GATK Variant Quality Score Recalibration (VQSR) method. For training sets, the following datasets were used: QC-passing sites from the P. falciparum Genetic Crosses Project release 1.0 (ftp://​ngs.​sanger.​ac.​uk/​production/​malaria/​pf-crosses/​1.​0/​; [22]) (true positives, prior score Q30); QC-passing sites from the Pf3K release v5.1 (ftp://​ngs.​sanger.​ac.​uk/​production/​pf3k/​release_​5/​5.​1/​) (true positives + false positives, prior score Q15). Site annotations QD, MQ, MQRankSum, ReadPosRankSum, FS, SOR were used and the model was trained with 4 Gaussian components. A VQSLOD threshold − 0.0350 achieved 90% sensitivity for re-discovering known sites in the training sets. All biallelic SNVs with VQSLOD at or above this threshold were retained.

Isolates may contain multiple strains that are haploid resulting in mixed infections with arbitrary effective ploidy. To account for this complexity of infection (COI), prior literature was followed [23] and the following quantities were calculated at each variant site: for each isolate, the within-sample allele frequency (WSAF), the proportion of mapped reads carrying the non-reference allele; the population-level allele frequency (PLAF), the mean of within-sample allele frequencies; and the population-level minor allele frequency (PLMAF), the minimum of PLAF or 1-PLAF. These calculations were performed with ‘vcfdo wsaf’ (https://​github.​com/​IDEELResearch/​vcfdo).

Ancestral versus derived alleles at sites polymorphic in P. falciparum were assigned by comparison to the outgroup species Plasmodium reichenowi. Briefly, an approximation to the genome of the P. reichenowi–P. falciparum common ancestor (hereafter, “ancestral genome”) was created by aligning the P. falciparum 3D7 assembly to the P. reichenowi CDC strain assembly (version 3, PlasmoDB version 38: https://​plasmodb.​org/​common/​downloads/​release-38/​PreichenowiCDC/​fasta/​data/​PlasmoDB-38_​PreichenowiCDC_​Genome.​fasta) with ‘nucmer’ v3.1 using parameters “-g 500 -c 500 -l 10” as in [24]. Only segments with one-to-one alignments were retained; ancestral state at sites outside these segments was deemed ambiguous. The one-to-one segments were projected back into the 3D7 coordinate system. Under the assumption of no recurrent mutation, any site polymorphic in P. falciparum is not expected to also be mutated on the branch of the phylogeny leading to P. reichenowi. Thus, the allele observed in P. reichenowi is the ancestral state conditional on the site being polymorphic. Transitions-transversion (Ti:Tv) ratios and mutational spectra were tallied with ‘bcftools stats’ v1.19.

VQSR-passing sites were filtered more stringently for PCA to reduce artifacts due to rare alleles and missing data. Genotype calls with GQ < 20 or DP < 5 were masked; sites with < 10% missing data and PLMAF > 5% after sample-level filters were retained for PCA, which was performed with ‘akt pca’ v3905c48 [25]. For calculation of f₃ statistics, genotype calls with GQ < 10 or DP < 5 were masked; sites with < 10% missing data and PLMAF > 1% after sample-level filters were retained. Then f₃ statistics were calculated from WSAFs rather than nominal diploid genotype calls, using ‘vcfdo f3stat’.

Estimates of sequence diversity and differentiation were obtained from the site-frequency spectrum (SFS), which in turn was estimated directly from genotype likelihoods with ‘ANGSD’ 0.921-11-g20b0655 [26] using parameters “-doCounts 1 -doSaf 1 -GL 2 -minDepthInd 3 -maxDepthInd 2000 -minMapQ 20 -baq 1 -c 50.” Unfolded SFS were obtained with the ‘ANGSD’ tool ‘realSFS’ using the previously-described ancestral sequence from P. reichenowi. All isolates were treated as nominally diploid for purposes of estimating the SFS because systematic bias against mixed isolates was noted when using ‘ANGSD’ in haploid mode. Four-fold degenerate and zero-fold degenerate sites were defined for protein-coding genes in the usual fashion using transcript models from PlasmoDB v38. SFS for all sites, fourfold and zerofold degenerate sites were estimated separately in mainland Tanzania and Zanzibar isolates in non-overlapping 100 kb bins across the core genome. Values of sequence diversity (theta_pi) and Tajima’s D were estimated for these bin-wise SFS using ‘sfspy summarize’ (https://​github.​com/​IDEELResearch/​sfspy), and confidence intervals obtained by nonparametric bootstrap. F_st was calculated from the joint SFS between mainland Tanzania and Zanzibar. The distribution of local F_st values was calculated in 5 kb bins for purposes of visualization only.

Complexity of infection (COI) and strain deconvolution (phasing) were performed jointly using ‘dEploid’ v0.6-beta [14]. These analyses were limited to 125 isolates from mainland Tanzania and Zanzibar (57 new in this paper and 68 previously published). On the basis of the analyses shown in Figs. 1 and 2, these isolates appeared to constitute a reasonably homogeneous population, so the set of 125 was used for determination of PLAFs to be used as priors for the phasing algorithm. Phasing was performed using population allele frequencies as priors in the absence of an external reference panel known to be well-matched for ancestry. The analysis was further limited to very high-confidence sites: VQSLOD > 8, 75% of isolates having GQ ≥ 10 and DP ≥ 5, ≥ 10 bp from the nearest indel (in the raw callset), ≥ 10 total reads supporting the non-reference allele, and PLMAF ≥ 1%. The ‘dEploid’ algorithm was run in “-noPanel” mode with isolate-specific dispersion parameters (“-c”) set to the median coverage in the core genome, and default parameters otherwise. Within-isolate IBD segments were extracted from the ‘dEploid’ HMM decodings by identifying runs of sites with probability ≥ 0.90 assigned to hidden states where at least two of the deconvoluted haplotypes were IBD. The total proportion of strain genomes shared IBD (within-isolate F_IBD) for isolates with COI > 1 was obtained directly from ‘dEploid’ log files, and agreed closely with the sum of within-isolate IBD segment lengths.

Between-isolate IBD segments were identified by applying ‘refinedIBD’ v12Jul18 [27] to the phased haplotypes produced by ‘dEploid’. For a genetic map, a constant recombination rate of 6.44 × 10⁻⁵ cM/bp (equal to the total genetic length of the P. falciparum map divided by the physical size of the autosomes in the 3D7 assembly) was assumed. Segments > 2 cM were retained for analysis. The proportion of the genome shared IBD between phased haplotypes (between-isolate F_IBD) was estimated by maximum likelihood described in [28] using ‘vcfdo ibd’.

Curves of recent historical effective population size were estimated from between-isolate IBD segments with ‘IBDNe’ v07May18-6a4 [29] using length threshold > 3 cM, 20 bootstrap replicates and default parameters otherwise. Local age-adjusted parasite prevalence point estimates (PfPR_2–10) and credible intervals were obtained from the Malaria Atlas Project [30] via the R package ‘malariaAtlas’ [31].

More remote population-size histories were estimated with ‘smc++’ v1.15.2 [32]. Phased haplotypes from ‘dEploid’ were randomly combined into diploids and parameters estimated separately for mainland Tanzania and Zanzibar populations using fivefold cross-validation via command ‘smc++ cv’, with mutation rate set to 10⁻⁹ bp⁻¹ gen⁻¹. Marginal histories from each population were then used to estimate split times using ‘smc++ split’.

The distribution of fitness effects (DFE) was estimated within mainland Tanzania and Zanzibar populations with ‘polyDFE’ v2.0 using fourfold degenerate sites as putatively-neutral and zerofold degenerate sites as putatively-selected [33]. “Model C” in ‘polyDFE’ parlance—a mixture of a gamma distribution on selection coefficients of deleterious mutations and an exponential distribution for beneficial mutations—was chosen because it does not require a priori definition of discrete bins for selection coefficients, and the gamma distribution can accommodate a broad range of shapes for the DFE of deleterious mutations (expected to represent the bulk of polymorphic sites). Confidence intervals for model parameters were obtained by non-parametric bootstrap via 20 rounds of resampling over the 100 kb blocks of the input SFS. Because ‘polyDFE’ fits nuisance parameters for each bin in the SFS, computation time increased and numerical stability decreased for SFS with larger sample sizes. Input SFS were, therefore, smoothed and rescaled to pre-specified sample size of 10 chromosomes each using an empirical-Bayes-like method (https://​github.​com/​CartwrightLab/​SoFoS/​) re-implemented in ‘sfspy smooth’. Smoothing of input SFS had very modest qualitative effect on the resulting DFE.

The cross-population extended haplotype homozygosity (XP-EHH) statistic was used to identify candidate loci for local adaptation in mainland Tanzania or Zanzibar. Because the statistic requires phased haplotypes and is potentially sensitive to phase-switch errors, only isolates with COI = 1 were used (n = 18 mainland Tanzania, n = 12 Zanzibar). XP-EHH was calculated from haploid genotypes at a subset of 103,982 biallelic SNVs polymorphic among monoclonal isolates with the ‘xpehhbin’ utility of ‘hapbin’ v1.3.0-12-gdb383ad [34]. Raw values were standardized to have zero mean and unit variance; the resulting z-scores are known to have an approximately normal distribution [35] so nominal p-values were assigned from the standard normal distribution. The Benjamini-Hochberg method was used to adjust nominal p-values for multiple testing.

Pipelines used for WGS read alignment, variant calling, variant filtering, haplotype deconvolution and SFS estimation are available on Github: https://​github.​com/​IDEELResearch/​NGS_​Align_​QC_​Pipelines.

Genomic data for P. falciparum was generated using leukodepleted blood collected from 43 subjects from Yombo, Tanzania (“mainland”) and from DBS collected from 63 subjects from the Zanzibar archipelago (“Zanzibar”; Fig. 1a) using selective whole-genome amplification (sWGA) followed by Illumina sequencing. Thirty-six isolates (84%) from the mainland and 21 isolates (33%) from Zanzibar yielded sufficient data for analysis. These 57 genomes were combined with an additional 68 published genomes from other sites in Tanzania in the MalariaGEN P. falciparum Community Project (PfCP) and 179 genomes from other sites in Africa and Asia, representing a broad geographic sampling of Africa and Asia [36]. Single-nucleotide variants (SNVs) were ascertained jointly in the global cohort. After stringent quality control on 1.3 million putative variant sites, a total of 387,646 biallelic SNVs in the “core genome”—the 20.7 Mb of the 3D7 reference assembly lying outside hypervariable regions and accessible by short-read sequencing [22]—were retained for further analysis. The frequency spectrum was dominated by rare alleles: 151,664 alleles (39.1%) were singletons and 310,951 (80.2%) were present in < 1% of isolates in the dataset. Ancestral and derived states at 361,049 sites (93.1%) were assigned by comparison to the P. reichenowi (CDC strain) genome, treating the reichenowi allele as ancestral. Similar biases were observed in the mutational spectrum as have been estimated directly from mutation-accumulation experiments [37]: transitions are more common transversions (Ti:Tv = 1.12; previous estimate 1.13), with a large excess of G:C>A:T changes even after normalizing for sequence composition (Additional file 1: Fig. S1). Consistency in the mutational spectrum between independent studies, using different methods for sample preparation and different bioinformatics pipelines, supports the accuracy of genotype calls.

In order to place new isolates in the context of global genetic variation in P. falciparum, principal components analysis (PCA) was performed with existing isolates from around the globe (Fig. 1b). A subset of 7122 stringently-filtered sites with PLMAF > 5% (see “Methods”) were retained for PCA to minimize distortion of axes of genetic variation by rare alleles or missing data. Consistent with existing literature, isolates separated into three broad clusters corresponding to southeast Asia, east Africa and west Africa. Mainland Tanzania and Zanzibar isolates fell in the east Africa cluster. This observation was formalized using f₃ statistics [38, 39], which measure shared genetic variation in a pair of focal populations A and B relative to an outgroup population O. By calculating f₃ across different combinations of comparator populations and holding the outgroup fixed, one can build up an idea of the ancestry of the populations of interest: pairs with relatively larger positive values of f₃ are more genetically similar than pairs with relatively smaller f₃. The new isolates from Yombo and Zanzibar and published Tanzanian isolates shared mutually greater genetic affinity for each other than for other populations in the panel (Fig. 1c–e); isolates from neighbouring countries Malawi and Kenya were next-closest. Together these analyses support an east African origin for parasites in mainland Tanzania and in Zanzibar.

In order to better understand the population demography and effects of natural selection in the parasite populations, indices of genetic diversity within populations, and the degree to which that diversity is shared across populations, were examined. The genome was partitioned into four sequence classes—all sites in the core genome; fourfold degenerate (“synonymous”) sites; zerofold degenerate (“nonsynonymous”) sites; and coding sites in genes associated with resistance to antimalarial drugs—and several estimators of sequence diversity were calculated in each class (see "Methods"). Levels of sequence diversity at synonymous (putatively neutral) sites were very similar within mainland Tanzania and Zanzibar isolates (theta_pi = 9.0 × 10⁻⁴ [95% CI 8.6 × 10⁻⁴–9.4 × 10⁻⁴] vs. 8.4 [95% CI 8.0 × 10⁻⁴–8.7 × 10⁻⁴ per site) and 1.3-fold lower than among previously-published Tanzanian isolates (Fig. 2a). As expected, diversity was lower at non-synonymous sites, which are more likely to be under purifying selection. Tajima’s D took negative values in all three populations and across all sites classes (Fig. 2b); demographic explanations for this pattern are investigated later in the manuscript. Minimal evidence was found for differentiation between parasites in mainland Tanzania and Zanzibar. Genome-wide F_st was just 0.0289 (95% bootstrap CI 0.0280–0.0297); the distribution of F_st in 5 kb windows is shown in Fig. 2c. For comparison, genome-wide F_st between southeast Asian and African isolates is on the order of 0.20 [23]. Thus there exists minimal evidence for genetic differentiation between parasites in mainland Tanzania and Zanzibar.

Long segments of the genome shared identical by descent (IBD)—that is, inherited intact from the same recent common ancestor—provide a powerful and fine-grained view of relationships in the recent past. Recent methodological innovations [14] allow estimation of complexity of infection (COI)—the number of distinct parasite strains in a single infection—and simultaneous deconvolution the component haplotypes. The F_ws statistic, an index of within-host diversity that is conceptually similar to traditional inbreeding coefficients, was also calculated for comparison [23]. Approximately half of isolates had COI = 1 (“clonal”) and half had COI > 1 (“polyclonal” or “mixed”) in both populations, and the distribution of COI was similar between the mainland and Zanzibar (Chi squared = 0.27 on 2 df, p = 0.87; Additional file 1: Table S4). Ordinal trends in F_ws were qualitatively consistent with COI but show marked variation for COI > 1 (Fig. 3a). Phased haplotypes were used to identify segments shared IBD between isolates and, in the case of mixed infections, within isolates. This revealed substantial relatedness between infecting lineages within mixed isolates (Fig. 3b): the median fraction of the genome shared IBD (F_IBD) within isolates was 0.22 among mainland and 0.24 among Zanzibar isolates, with no significant difference between populations (Wilcoxon rank-sum test, p = 0.19). The expected sharing is 0.50 for full siblings and 0.25 for half-siblings with unrelated parents [40]. F_IBD was then estimated between all pairs of phased haplotypes. F_IBD between pairs of isolates was then defined as the maximum over the values for all combinations of haplotypes inferred from the isolates (Fig. 3c). As expected, most pairs were effectively unrelated (median F_IBD ≤ 0.001, on the boundary of the parameter space), but a substantial fraction were related at the level of half-siblings or closer (F_IBD > 0.25, 4.0% of all pairs), including 1.3% of mainland-Zanzibar pairs.

Long segments of the genome are shared IBD both within and between isolates. Mean within-isolate segment length was 5.7 cM (95% CI 4.1–7.3 cM, n = 117) on the mainland and 3.7 cM (95% CI 2.8–4.6 cM, n = 80) on Zanzibar in a linear mixed model with individual-level random effects; the full distributions are shown in Fig. 3d. Segments shared between isolates within the mainland population (6.2 cM, 95% CI 5.9–6.6 cM, n = 3279) were longer than segments shared within Zanzibar (4.5 cM, 95% 4.1–4.8 cM, n = 592) or between mainland and Zanzibar populations (4.1 cM, 95% CI 3.9–4.3 cM, n = 6506). After accounting for differences in segment length by population, difference in lengths of IBD segments detected between versus within individuals are not significant (mean difference − 0.038 cM, 95% CI − 0.10 to 0.023 cM). In a random-mating population the length of a segment shared IBD between a pair of individuals with last common ancestor G generations in the past is exponentially-distributed with mean 100/(2*G) cM. The shared haplotypes that observed, with length on the order of 5 cM, are thus consistent with shared ancestry in the past 10 generations—although as many as half of such segments probably date back at least 20 generations [41]. In the presence of inbreeding, IBD sharing persists even longer in time.

Close relationships between isolates from the archipelago and the mainland suggest recent genetic exchange. A threshold of F_IBD > 0.25 (half-siblings) was chosen because it implies that two isolates shared at least one common parent in the last outcrossing generation and, therefore, are related as recently as the last 1–2 transmission cycles, depending on background population dynamics. In principle this could result from importation of either insect vectors or human hosts. To investigate the latter possibility, a travel-history questionnaire completed by subjects from Zanzibar was used. Nine subjects reported travel to the mainland in the month before study enrollment; their destinations are shown in Fig. 4a. Ten pairs with F_IBD > 0.25 (marked by orange triangles in histogram in Fig. 4b) were identified; all involved a single Zanzibar isolate from a patient who travelled to the coastal town of Mtwara (orange arc in Fig. 4a). It is very likely that this individual represents an imported case. Overall, isolates from travellers had slightly higher mean pairwise relatedness to isolates from the mainland (mean F_IBD = 0.0020, 95% CI 0.0018–0.0021) than did isolates from non-travellers (mean F_IBD = 0.0015, 95% CI 0.0014–0.0016; Wilcoxon rank-sum test p = 1.8 × 10⁻¹² for difference). But these relationships—spanning 10 or more outcrossing generations—are far too remote to be attributed to the period covered by the travel questionnaire. The pattern likely represents instead the presence of subtle population structure within Zanzibar.

The distribution of IBD segment lengths carries information about the trajectory of effective population size in the recent past, up to a few hundred generations before the time of sampling. The site frequency spectrum and patterns of fine-scale linkage disequilibrium carry information about the more remote past. Complementary methods were used to infer recent and remote population demography from phased haplotypes. First, a non-parametric method was applied [29] to infer recent effective population size (N_e) from IBD segment lengths separately in mainland Tanzania and Zanzibar populations (Fig. 5a). The method infers a gradual decline of several orders of magnitude in N_e over the past 100 generations to a nadir at N_e ~ = 5000 around 15–20 outcrossing generations before the time of sampling. Although the confidence intervals are wide, similar trajectories are inferred in all three populations (Zanzibar, new mainland Tanzania isolates and published Tanzanian isolates).

Second, more remote population size histories were inferred jointly for mainland Tanzania and Zanzibar and used to estimate the split time between these populations using a sequentially Markovian coalescent method [32]. This family of models has good resolution for relatively remote events, but less precision in the recent past than models based on IBD segments. The result (Fig. 5b) supports a common ancestral population with N_e ~ = 10⁵ individuals that underwent a sharp bottleneck followed by rapid growth around 50,000 generations before the present. The time at which the mainland and Zanzibar populations diverged could not be estimated precisely and may have been as recent as 50 or as ancient as 50,000 generations before the present. Trends in N_e were compared to local trends in parasite prevalence from the Malaria Atlas Project [30] (Fig. 5c). Assuming an interval of approximately 12 months per outcrossing generation [42], the contraction in N_e may correspond in time to the decrease in prevalence brought about by infection-control measures over the past two decades.

Finally, several approaches were taken to characterize the effects of natural selection on sequence variation in mainland and Zanzibar populations. The fate of a new mutation—whether it spreads and ultimately becomes fixed, or is lost—is determined by its selection coefficient (s), scaled by the effective population size (N_e). The distribution of fitness effects (DFE) describes the distribution of s and can be estimated from the frequency spectrum at putatively-neutral (synonymous) and putatively-selected (non-synonymous) sites (Fig. 6a). Building on previous work in other organisms, the DFE was modelled in each population as a mixture of a gamma distribution (for deleterious mutations, N_es < 0) and an exponential distribution (for beneficial mutations, N_es > 0) [33]. The inference was performed using both the raw SFS and a smoothed representation of the SFS that is more numerically stable and found that results to be similar with both methods. Fitted parameter values are provided in Additional file 1: Table S5, but the discretized representation of the DFE is more amenable to qualitative comparisons (Fig. 6b).

Differences in the DFE between mainland Tanzania and Zanzibar populations are not statistically significant. The great majority of new mutations (mainland: 74%; Zanzibar: 76%) are expected to be very weakly deleterious (− 0.01 < 4N_es < 0), and only a small minority are expected to be beneficial (4N_es > 0) (mainland: 4.5% [95% CI 2.7–29%]; Zanzibar: 2.4% [95% CI 0.56–50%]). The DFE also allows us to estimate that 8.8% (mainland) and 5.2% (Zanzibar) of substitutions since the common ancestor with P. reichenowi have been fixed by positive selection; this quantity is known in some contexts as the “rate of adaptive evolution.”

Although the DFE tells us the proportion of polymorphic sites under positive selection, it does not pinpoint which sites those are. To identify signals of recent, population-specific positive selection, the XP-EHH statistic between mainland and Zanzibarian isolates were used [35]. Outliers in the XP-EHH scan, defined as standardized XP-EHH scores above the 99.9th percentile, represent candidates for local adaptation (Additional file 1: Fig. S2). One-hundred four biallelic SNPs in 20 distinct genes passed this threshold (Additional file 1: Table S6). None of these have been associated with resistance to anti-malarial drugs—an important form of local adaptation in this species—but one (PF3D7_0412300) has been identified in a previous selection scan [43]. Prevalences of 54 known drug-resistance alleles are shown in Additional file 1: Table S7 and are similar to previous reports in East Africa [44‐46]. None of these loci had F_st > 0.05 between mainland Tanzania and Zanzibar.