A handful of genetic variants have already been associated with risk of or protection against severe malaria infection.⁵ Our first objective in this study was to test variants of β-globin (HbAS^9,10), intercellular adhesion molecule (ICAM TT¹¹), CD36 (CD36 GT¹²), nitric oxide synthase (NOS2A 1659 AA¹³), tumor necrosis factor (TNF 238 A¹⁴ and TNF 308 A^14–16), Fc γ-receptor IIA (CD32 AA^17,18), interferon-α receptor-1 (IFNARI LI168V CC¹⁹ and IFNARI 17470 CC¹⁹), and Toll-like receptor (TLR4²⁰), which had previously been associated with malaria susceptibility. The particular phenotype we focused on was high levels of parasitemia in young children due to malaria infection.

Second, we compared the frequency differentiation in populations in which malaria is endemic and in closely related populations in which it is not endemic, searching for the differences that would be expected if natural selection had affected those alleles in one population but not in the other, because malaria began to affect only one group. Finally, we formally combined the evidence of association from case-control studies with evidence of natural selection in populations that have been exposed to malaria infection. We note that there has been discussion elsewhere of how one could formally combine case-control association studies with statistical weights obtained on the basis of evidence of natural selection.²¹ Our goal in this study was to empirically demonstrate the power of this approach.

Table 1. Characteristics of the Populations Included in This Study

For the selection study, we assembled population control samples from the Masai, Kikuyu, and Yoruba ethnic groups. We collected samples from the Masai and Kikuyu from secondary schools in Narok and Nyeri, Kenya, respectively (table 1). The Yoruba samples were from the International Haplotype Map project²²; we analyzed data from unrelated men and women, the parents in HapMap mother-father-child trios.

About 2 ml of blood was obtained by venipuncture for all the samples we collected in Kenya. We extracted DNA within 10 h of blood collection, using a Qiagen DNA Blood mini kit, and then stored it at −20°C. All the participants provided informed consent, and, for children, informed consent was obtained from the parents and/or guardians. The study was reviewed and approved by the Harvard Medical School and Kenyatta University ethical review boards and by the Kenyan government.

Table 2. Replication Analysis for 10 Genotypes or Alleles Previously Associated with Malaria Susceptibility

As an assessment of genotyping quality, we observed that, for 85 genotypes obtained in duplicate, there were 2 discrepancies, for a discordance rate of 2.4%. After removing samples with <80% genotyping completeness, we found that the average completeness of genotypes was 97.8%. We also compared genotyping of 10 SNPs in Yoruba samples with data from HapMap.²² Of 30 genotypes obtained in duplicate, there was 1 discrepancy (3.33%). All SNPs were in Hardy-Weinberg equilibrium in all the ethnic groups we studied (P>.05).

To estimate c empirically, we used data from the 1,454 randomly chosen markers. For a given pair of populations, we estimated c as the empirical variance of the difference in population frequencies, after normalizing by p(1−p) and accounting for sampling noise, which has variance p(1-p)(1/N₁+1/N₂), where N₁ and N₂ are total allele counts for the two populations at a given marker. We approximated the normalization term p(1−p) by setting p equal to the average of observed frequencies of the two populations, and we approximated binomial sampling noise as normally distributed. The same approximations were applied both to our estimation of c and to our subsequent analysis of individual markers. SNPs with average minor-allele frequency <5% for the two populations being compared were omitted from all computations, since the normal approximation becomes less reliable (table 3).

Table 3. Tests for Differentiating Selection between Malaria-Endemic and -Nonendemic Populations

To test whether an individual marker was more differentiated than expected between two populations, we compared the observed difference in frequency with the expected distribution N[0,p(1-p)(c+1/N₁+1/N₂)], using the value of c estimated above, and computed a χ² statistic with 1 df. A feature of this test is that the χ² statistic has a mean value of 1 across the set of markers used to infer c. The test appropriately handles different sample sizes for candidate markers versus random markers used to infer c. A detailed statistical treatment will appear elsewhere (A.L.P, N.P., and D.R., unpublished data).

Table 4. Formal Combination of Case-Control Association Analysis and Tests of Natural Selection

We note in passing that NOSA (rs8078340) gives a nominally significant P value (by a two-tailed test), but the association is in the opposite direction to previous reports (P=.99) (table 2). Our null findings at the other variants do not necessarily mean that they are unassociated; the CIs for the ORs are broad (table 2) and are often consistent with substantial association. We also note that our study included only individuals with parasitemia; we had no power to detect associations that were specific to cerebral malaria, a phenotype that was the focus of some previous studies.^13,27,28

Finally, we tested for epistatic interactions between each pair of variants,²⁹ but no pair showed a statistically significant interaction by a Wald test (not shown). We also tested for different strengths of association by sex but found no evidence of this (table A1).

Table A1.

Statistical Tests for Association^[Note]

				P in Two-Tailed (One-Tailed) Tests of Previously Associated
Genotype/Allele, Reference SNP, and Sex	Frequency in Luo Controls (%)	No. of Cases/Controls	OR (95% CI)	Genotype	Allele	P in Tests Based on 3×2 Genotype Table
HbAS AT:
rs334:
All	25	447/454	.57 (.41–.79)	.0008 (.0004)	.006 (.003)	.0031
Female	25	232/175	.55 (.33–.90)	.02 (.01)
Male	25	215/279	.59 (.38–.93)	.02 (.01)
CD36 GT:
rs3211938:
All	12	456/457	1.50 (1.03–2.18)	.03 (.015)	.06 (.03)	.061
Female	12	232/174	1.40 (.85–2.33)	1.19 (.60)
Male	13	224/283	1.57 (.90–2.74)	.11 (0.06)
ICAM TT:
rs5491:
All	7	460/455	.71 (.42–1.21)	.20 (.10)	.89 (.45)	.22
Female	8	235/176	.43 (.18–1.00)	.04 (.02)
Male	7	225/279	1.05 (.53–2.09)	.90 (.45)
NOS2A 1659 AA:
rs8078340:
All	6	450/455	.42 (.21–.83)	.01 (.99)	.69 (.5)	.015
Female	6	229/179	.55 (.22–1.40)	.21 (.11)
Male	6	221/276	.28 (.09–.85)	.02 (.01)
TNF 238 A:
rs361525:
All	9	459/457	1.00 (.73–1.39)	.28 (.14)	.97 (.49)	.47
Female	8	233/173	1.33 (.82–2.17)	.24 (.12)
Male	9	226/284	.74 (.47–1.16)	.20 (.10)
CD32 AA:
rs1801274:
All	25	455/447	.95 (.71–1.29)	.76 (.38)	.81 (.45)	.95
Female	24	229/176	1.06 (.67–1.67)	.37 (.19)
Male	25	226/271	.45 (.30–.70)	.66 (.33)
IFNARI LI168V CC:
rs2257167:
All	3	455/457	1.18 (.54–2.07)	.48 (.76)	.93 (.47)	.86
Female	2	231/180	1.37 (.45–4.17)	.37 (.19)
Male	3	224/277	.37 (.11–1.24)	.66 (.33)
TNF 308 A:
rs1800629:
All	9	450/433	1.13 (.82–1.56)	.21 (.11)	.42 (.21)	.25
Female	6	234/166	1.33 (.75–2.37)	.33 (.17)
Male	11	216/267	1.17 (.78–1.74)	.45 (.23)
IFNARI 17470 CC:
rs1012335:
All	3	455/452	.85 (.53–1.36)	.68 (.34)	.70 (.35)	.78
Female	2	234/180	1.73 (.52–5.70)	.34 (.17)
Male	3	221/272	.46 (.15–1.43)	.66 (.33)
TLR4 AG:
rs4986790:
All	10	407/299	1.36 (.86–2.17)	.20 (.10)	.067 (.02)	.20
Female	9	201/123	1.48 (.70–3.16)	.67 (.34)
Male	11	206/176	1.33 (.74–2.40)	.17 (.09)

Table A2.

Allele Counts in Cases and Controls for the 10 Polymorphisms^[Note]

To test whether these allele-frequency differences are greater than what could be explained in the absence of selection, we compared them with a panel of 1,454 random SNPs³⁰ for which we obtained genotypes in 45 Luo, 47 Masai, 37 Kikuyu, and 59 Yoruba. (We first assessed whether there was evidence of population substructure in the Luo,³¹ which could, in principle, confound our case-control tests of association. No structure was detected, indicating that population stratification is not likely to cause false-positive or false-negative results in the association analysis.) We also used the data to assess the genetic relationships among the populations; understanding this is crucial to the tests for differentiating selection.

The genetic differentiation among populations ranges from F_ST=0.0012 between Masai and Kikuyu (lowest differentiation) to F_ST=0.021 between Yoruba and Masai (highest differentiation). We found that the Luo and Masai do not cluster genetically, despite the fact that they both speak Nilotic languages, whereas the Masai and Kikuyu are closely related (despite the fact that the Kikuyu speak a Bantu language) (fig. 1). These results show that the linguistic patterns in Kenya do not correlate with the genetic patterns, which is at odds with what has been suggested elsewhere.³² Sampling of more populations should elucidate the relationships between genetic and linguistic groups in East Africa.³³

Figure 1. Principal-components analysis of samples from four different populations genotyped for 1,454 SNPs. The first eigenvector clusters the Yoruba, Luo, Kikuyu, and Masai. This is contrary to the expectation based on linguistics (the Luo and Masai both speak (more ...)

To formally test for differentiating selection, we computed a χ² statistic for frequency differentiation at each tested SNP, assuming it was drawn from the empirical distribution defined by 1,454 random SNPs (see the “Material and Methods” section and table 3). Allele-frequency differentiation between malaria-endemic and -nonendemic populations is significant at HbAS T (P=.00036 for the most extreme Luo-Kikuyu comparison) and CD36 G (P=.00096 for Yoruba-Kikuyu), with the results significant even after use of a Bonferroni correction for testing 40 comparisons of malaria-endemic and -nonendemic populations at 10 SNPs (this essentially involves multiplying the nominal P values by a factor of 40). By contrast, the eight SNPs that do not give positive case-control association show no evidence of differentiating selection (P>.25 for each SNP and pair of populations after correction for multiple hypotheses tested; P=.39 for the sum of 32 χ² statistics at these eight SNPs). We further evaluated the robustness of our selection test by computing χ² statistics for each of the 1,454 random SNPs for each of the four pairs of populations. If the test is robust, we would expect to achieve a χ² value >3.84, with probability 0.05. Restricting the analysis to SNPs in which the average allele frequency across the two populations tested was at least 5%, we observed a χ² value >3.84 in 255 (5%) of 5,090 of tests performed. Similarly, only 4 of 5,090 tests produced a P value <.001, and the lowest P value was not statistically significant after correction for 5,090 hypotheses tested (P>.16). These results show that our test for differentiating natural selection is not prone to false-positive results in a large selection of randomly chosen SNPs.

We note that both HbAS T and CD36 G have been identified elsewhere as targets of recent positive natural selection.^22,34–36 However, the long-range haplotype test used to detect selection at these alleles detects evidence of selection from any cause and thus is not specific to a particular type of selection (e.g., for malaria resistance). The tests of allele-frequency differentiation we present here are much more specific to malaria. By comparing malaria-endemic and -nonendemic populations, we increase the probability that the loci detected as being affected by selection are specifically associated with malaria resistance. Of all the SNPs we tested for population differentiation—1,454 random SNPs and 10 candidates for malaria susceptibility—2 of those that achieve a nominal P value <.001 for at least one pair of populations were among the candidate malaria-resistance SNPs.

We performed a case-control association study of malaria resistance in the Luo, an East African population, analyzing 10 previously implicated variants. We replicated associations at HbAS (OR 0.57 [95% CI 0.41–0.79]) and CD36 (OR 1.50 [95% CI 1.05–2.18]). Our OR for CD36 is in agreement with the results published elsewhere.¹² Similarly, the OR for HbAS is in agreement with the previously reported longitudinal study in the same population (OR 0.45 [95% CI 0.24–0.84]; P=.0001).¹⁰ For HbAS, the protective effect that we observed is smaller than in some previous reports, which is potentially due to the fact that the cases we studied were young (average age 2.6 years) and thus lacked an immune basis for HbAS protection. (Williams and colleagues showed that HbAS has a more protective effect for older individuals.^9,10,26) A possible reason why we did not replicate all the previous associations is that, in our study, the phenotype was parasitemia, whereas previous studies sometimes focused on cerebral malaria (table 2). We also show in table 2 that the CIs for the ORs are broad; thus, many of the variants we tested are consistent with an effect on malaria susceptibility, even if we could not reject the hypothesis of no association.

A particularly striking observation is that, at CD36, where we observe significant case-control association and highly significant allele-frequency differentiation, the variant increasing susceptibility actually has higher frequency in malaria-endemic populations. A possible historical explanation is that the selection pressures on this variant may have changed over time because of host-parasite genetic interactions. For example, the variant may have historically reduced susceptibility to malaria, and then, as the parasite evolved to adapt to the human immune system, the allelic association might have reversed. This hypothesis would be consistent with the known temporal and geographical heterogeneity in CD36 binding and pathogenicity.¹² For example, genetic variation at the PfEMP1 gene in the malaria parasite has been shown elsewhere to be associated with the pathogenicity,^37,38 and parasite PfEMP1 and human CD36 are known to interact.^39–41 In future studies, it will be interesting to explore whether human variants at CD36 have different interactions with genetically different malaria parasite strains.⁴²

These results finally provide empirical validation for a long-standing idea.²¹ The idea is that, to increase power in case-control studies, one can combine the evidence of association with that from tests of natural selection. Previous studies have prioritized SNPs by natural selection on the basis of a combination of the alleles being frequent and being surrounded by a long-range haplotype^3,22,43; the present study adds to this in several ways. First, we provide a formal χ² test of statistical significance, which can be combined with a case-control statistic to provide evidence that a SNP is a statistical outlier and, thus, a strong candidate for being associated with malaria. Second, our selection evidence is more specific to our phenotype of interest, since we are comparing frequency variants in populations differentiated by whether malaria has been historically endemic or nonendemic. Tishkoff et al.³³ recently applied a similar strategy to the phenotype of lactase persistence. They compared pastoral and nonpastoral populations in East Africa that have been differently exposed to diets including cow's milk. This analysis demonstrated high allele-frequency differences at variants near the lactase gene LCT and simultaneously showed that these highly differentiated variants also conferred the phenotype of lactase persistence.

We conclude that, in future whole-genome association scans, evidence from case-control comparisons can be combined with allele-frequency differentiation between differently exposed populations—and, potentially, other sources of evidence about recent selection^22,43—to provide increased sensitivity and power in tests to detect disease-related genetic variants. It has been suggested that the identification of targets of selection may soon become a mainstream approach to finding genetic variants affecting human disease; our results provide empirical validation for this idea.⁴⁴ In our study, P values for HbAS and CD36 were enhanced by several orders of magnitude with the use of <60 samples from each population analyzed, suggesting that this strategy may be cost effective relative to the number of additional samples needed to obtain a similar increase in power within the conventional case-control paradigm.

We are grateful to the human subjects and their families who participated in this study. We thank the medical officers and nursing staff in Bondo District Hospital, as well as the staff and students of the secondary schools where controls were collected. We are grateful to Pardis Sabeti, for thoughtful comments, and to Gavin McDonald, Alicja Waliszewska, Julie Neubauer, and Christine Schirmer, for technical advice. This work was supported by funds from Harvard Medical School, National Institutes of Health (NIH) grant R21-AI64519, and a Burroughs Wellcome Career Development Award in the Biomedical Sciences to D.R. A.L.P. is supported by a Ruth Kirschstein K-08 award from the NIH. We declare that we have no competing financial interests.