Gene-Expression Variation Within and Among Human Populations

Abstract

Understanding patterns of gene-expression variation within and among human populations will provide important insights into the molecular basis of phenotypic diversity and the interpretation of patterns of expression variation in disease. However, little is known about how gene-expression variation is apportioned within and among human populations. Here, we characterize patterns of natural gene-expression variation in 16 individuals of European and African ancestry. We find extensive variation in gene-expression levels and estimate that ∼83% of genes are differentially expressed among individuals and that ∼17% of genes are differentially expressed among populations. By decomposing total gene-expression variation into within- versus among-population components, we find that most expression variation is due to variation among individuals rather than among populations, which parallels observations of extant patterns of human genetic variation. Finally, we performed allele-specific quantitative polymerase chain reaction to demonstrate that cis-regulatory variation in the lymphocyte adaptor protein (SH2B adapter protein 3) contributes to differential expression between European and African samples. These results provide the first insight into how human population structure manifests itself in gene-expression levels and will help guide the search for regulatory quantitative trait loci.

Gene expression is the primary mechanism by which information encoded in the genome is converted into developmental, morphological, and physiological phenotypes.¹ Gene expression is also an important source of evolutionary change within and among species,² and aberrant gene expression has been implicated in the pathogenesis of numerous diseases.^3,4 Thus, understanding the amount, structure, and patterns of gene-expression variation is of fundamental importance to both biomedical research and evolutionary biology.⁵

Although it is well known that 85%–95% of human genetic variation is due to variation among individuals within a population, whereas 5%–15% is attributable to variation among populations,^6–9 it remains unclear whether similar levels of within- versus among-population components of variation will extend to higher-level phenotypes such as gene-expression levels. Whereas some prior work on gene-expression differences among human populations has been done in the context of disease studies,^10,11 to our knowledge, there have been no systematic and quantitative attempts to apportion natural variation in gene-expression levels into within- and among-population components akin to several excellent studies in model organisms.^5,12–14

To begin to address these issues, we used Affymetrix Human Focus Arrays to study gene-expression levels in B lymphoblastoid cells derived from eight unrelated individuals of northern and western European ancestry (CEU) and eight unrelated individuals from the Yoruba of Ibadan, Nigeria (YRI). These samples are a subset of the CEU and YRI individuals used in the International HapMap project,¹⁵ and cell lines were obtained from the Coriell Cell Repositories (samples GM06995, GM07029, GM07349, GM10845, GM10851, GM10856, GM10857, GM10860, GM19138, GM18516, GM18859, GM18871, GM18501, GM18504, GM18507, and GM18522). All study individuals were males, to eliminate the potential confounding effects of sex on gene-expression levels.

We performed tissue culture and RNA extraction as described elsewhere^16,17 and assessed RNA integrity by measuring the optical density 260/280 ratio and subjecting the sample to analysis with the Agilent Bioanalyzer 2100. Extracted RNA was labeled and hybridized according to the manufacturer’s protocol (Affymetrix). We performed quantile normalization and used the RMA algorithm to combine probe-set intensities into a single measure of expression for each gene.¹⁸ Low-intensity probe sets that were deemed absent in ⩾50% of the arrays with use of the algorithms implemented in MAS5¹⁹ were discarded in subsequent statistical analyses, resulting in 5,194 analyzable probe sets. All reported results were robust to different normalization methods and definitions of low-intensity genes (results not shown). Technical replicates were obtained for each individual, resulting in a total of 32 microarrays.

Of the ∼8,500 genes on the array, 5,194 were expressed in lymphoblastoid cells, which is comparable with previous observations.¹⁷ We used a fully nested, mixed-model analysis to identify genes differentially expressed among individuals within populations and genes differentially expressed among populations (see appendix A). This model allows tests of differential expression among individuals to be performed while properly accounting for population effects and technical variation. Similarly, it allows tests of differential expression among populations while properly accounting for individual effects and technical variation.

We used methodology described elsewhere^22,25 to analyze the complete distribution of P values resulting from tests of differential expression, to estimate the proportion of all genes that are differentially expressed either within or between the CEU and YRI samples. Under the null hypothesis of no differential expression, we expect the P values to be uniformly distributed between 0 and 1. Conversely, if the data set contains differentially expressed genes, the distribution of P values will be skewed toward 0.²² We estimated that ∼83% of genes are differentially expressed among individuals and ∼17% of genes are differentially expressed between the CEU and YRI samples (fig. 1). That these two percentages add up to 100% is a coincidence; we found that the significance of tests for differential expression within and among populations was uncorrelated (see appendix A). The estimated proportion of genes that possess interindividual variation is consistent with previous studies describing pervasive cis-regulatory variation in humans.^26–28 To our knowledge, there have been no systematic studies of gene-expression differences among human populations. Thus, these results demonstrate substantial natural variation in gene-expression levels both within and among populations and show that population structure exists in levels of transcript abundance.

Figure 1. .

Estimates of the proportion of genes differentially expressed within and among populations. Histograms of all P values calculated for tests of differential expression among individuals within populations (a) and between the CEU and YRI samples (b) are shown. The Y-axis is drawn to reflect a histogram density, such that the total area of all rectangles is 1. Under the null hypothesis of no differential expression, we expect the P values to be uniformly distributed between 0 and 1, forming a histogram with frequencies following the dashed black line. However, the observed P values in each graph are skewed toward 0, suggesting that these data sets contain differentially expressed genes. Using methodology described elsewhere,^22,25 we estimated that 82.6% of genes are differentially expressed among individuals and that 17.4% of genes are differentially expressed between the CEU and YRI samples. The dashed red lines indicate these estimates by showing that the P values close to 1 flatten out at a height of ∼17% (a) and 83% (b).

We next investigated the magnitude of expression differences observed within and among populations, which, in general, was relatively small (fig. 2). For example, of the ∼83% of genes estimated to be differentially expressed among individuals, 1,210 were significant at a false-discovery rate (FDR) ⩽1%. These genes varied by an average factor of 1.25 across individuals (see fig. 2), although 20 loci varied by a factor of 2. Of the ∼17% of genes estimated to be differentially expressed between the CEU and YRI samples, 50 were significant at an FDR ⩽20%. The average absolute log₂ difference in mean expression levels between samples for these 50 genes was 0.73 (corresponding to a 1.65-fold change). Although the majority of observed gene-expression differences within and between populations are modest, even small perturbations in expression can have significant functional and phenotypic consequences.^29,30 The results for all tests of differential expression are presented in a tab-delimited txt file.

Figure 2. .

Magnitude of expression differences within and between populations. a, Magnitude of gene-expression differences among individuals, shown as the factor of variation (X-axis) versus Q value (Y-axis). The Q value is a measure of statistical significance in terms of the FDR.^22,25 For each gene, the factor of variation is calculated as the ratio of the maximum:minimum log₂ expression level across all individuals.⁵ b, Magnitude of gene-expression differences between populations, shown by a volcano plot of the average log₂-fold change between the CEU and YRI samples (X-axis) versus Q value (Y-axis). In panels a and b, the horizontal dashed lines at 0.17 and 0.83 indicate the estimated proportion of truly null hypotheses in tests of differential expression among individuals and between the CEU and YRI samples, respectively.

To get a broad overview of the types of pathways that differentially expressed genes participate in, we tested whether they were overrepresented among PANTHER biological pathways.³¹ In this analysis, we considered the top 10% of genes differentially expressed either between individuals or between populations. Only two pathways were nominally significant (P=.05) for genes differentially expressed among individuals, and no pathways remained significant after correction for multiple hypothesis tests (table 1). Thus, gene-expression differences among individuals are found in a wide variety of pathways, which is consistent with our estimate that ∼83% of genes are differentially expressed among individuals. Examples of genes with large interindividual variation in expression include RAGE (MIM *605762) and LRAP (MIM *609497), the expression levels of which correlate with diabetic complications³² and improper antigen processing,³³ respectively. Genes differentially expressed between the CEU and YRI samples were strongly enriched in inflammatory pathways, even after a strict Bonferroni correction for multiple-hypothesis tests (table 1). Included in this set of genes are several cytokines and chemokine receptors (CCL22 [MIM *602957], CCL5 [MIM *187011], CCR2 [MIM *601267], CCR7 [MIM *600242], and CXCR3 [MIM *300574]) that have been implicated in numerous cardiovascular, infectious, and immune-related diseases.^34,35

Table 1. .

Enrichment of PANTHER Biological Pathways among Differentially Expressed Genes^[Note]

Sample Comparison and PANTHER Biological Pathway	P
Individuals:
Inflammation mediated by chemokine and cytokine	1.91×10−2
T-cell activation	3.01×10−2
Populations:
Inflammation mediated by chemokine and cytokine	2.91×10−4
Histamine H1 receptor–mediated signaling pathway	3.90×10−3
Toll-receptor signaling pathway	1.02×10−2
Fibroblast growth factor–signaling pathway	1.11×10−2
Vascular endothelial growth factor–signaling pathway	1.14×10−2
T-cell activation	1.32×10−2
EGF receptor–signaling pathway	1.53×10−2
B-cell activation	2.70×10−2
Notch-signaling pathway	2.99×10−2
Enkephalin release	2.99×10−2
5HT2 type receptor–mediated signaling pathway	4.24×10−2

Note.— All pathways nominally enriched at P=.05 are shown; bold type indicates significance after a Bonferroni correction for multiple hypothesis tests.

Simply identifying genes differentially expressed within or among populations may provide an incomplete view of the quantitative details of gene-expression variation. For instance, we found examples where expression variation was observed primarily between populations but not individuals, both among individuals and between populations, or among individuals but not between populations (fig. 3). Classifying genes by differential versus no differential expression fails to accurately reflect the quantitative patterns of how expression variation is apportioned into within- and among-population components; therefore, it is important to investigate how much of total gene-expression variation is explained by individual and population effects. To this end, for each gene, we estimated the proportion of total gene-expression variation due to either differences among individuals or differences between populations, while properly taking into account technical variation (see appendix A). The median proportion of variation due to interindividual variation is 0.85 (fig. 4), which is nearly identical to levels of population structure observed in extant patterns of human genetic variation.^6–9 In addition, similar to estimates of genetic structure at individual loci,^7,36 the distribution of population structure in expression levels across genes varies considerably (fig. 4).

Figure 3. .

Patterns of gene-expression variation within and between the CEU and YRI samples. The log₂ expression levels (averaged across replicates) of each individual are shown for 20 genes that are differentially expressed between the CEU and YRI samples but exhibit little within-population variation (top), for 20 genes that are differentially expressed both within and between populations (middle), and for 20 genes that are differentially expressed among individuals but not between populations (bottom). For each gene, the expression of each individual is expressed as the deviation from the overall mean across all 16 individuals. Values range in color from magenta to green, indicating expression levels relatively smaller or larger, respectively, compared with the overall mean. Black values indicate expression levels close to the overall mean across individuals. The dendrograms on the X- and Y-axes correspond to individuals and genes, respectively. The topology of the dendrogram is based on the 50 genes differentially expressed between the CEU and YRI samples.

Figure 4. .

Distribution of the proportion of total gene-expression variation explained by variation among individuals. The percentile of the proportion of all genes (X-axis) versus the proportion of total expression variation explained by interindividual differences (Y-axis) is shown for the observed (solid black line) and randomized (solid gray line) data. For example, the median (50th percentile) (dashed red line) proportion of variation explained by differences among individuals is 85%, leaving 15% explained by differences between populations. The dashed blue lines indicate the interquartile range. Note that the magnitude of total gene-expression variation attributable to interindividual differences in the observed data is considerably greater compared with the randomized data.

To fully understand the genetic architecture of gene-expression levels and how population structure influences patterns of gene-expression variation, it will be necessary both to confirm predictions from microarray data and to delineate the molecular mechanisms governing regulatory variation. To begin to address these issues, we performed quantitative allele-specific PCR (qPCR) on SH2B3 (MIM *605093), which encodes for an adaptor protein that regulates growth factor and cytokine receptor-mediated pathways³⁷ and was in the top 1% of genes differentially expressed between the CEU and YRI samples. We generated double-stranded cDNA and used TaqMan (Applied Biosystems) allelic discrimination assays to interrogate the expression level of each allele for a SNP (rs1107853) in the coding region of SH2B3, as described elsewhere.³⁸ We constructed a dilution series of heterozygous genomic DNA to estimate dye effects and differences in allele-specific hybridization efficiency, as described elsewhere.³⁸ To obtain the overall expression level for each individual, we summed the expression levels of the two alleles after adjusting for dye effects and hybridization effects. All qPCR experiments were performed in triplicate.

In our microarray experiments, SH2B3 showed an average log₂-fold change between the CEU and YRI samples of 0.52 (P=6.5×10^-4; FDR=0.134). Consistent with the microarray data, the qPCR results also demonstrate that SH2B3 is differentially expressed between the CEU and YRI samples (P=.0157) (fig. 5a). To better understand the molecular basis for the observed difference in expression, we asked whether the expression level of one allele was different from the other in heterozygous individuals. If so, this provides evidence of cis-regulatory effects.²⁶ There was a significant difference (P=1.18×10^-3) in expression between alleles in heterozygous cDNA versus genomic DNA, strongly suggesting cis-regulatory effects (fig. 5b).

Figure 5. .

Allele-specific qPCR analysis of SH2B3. a, Log₂-fold change of SH2B3 expression for all CEU and YRI individuals, relative to the average expression level in the YRI sample obtained from allele-specific qPCR. The distribution of SH2B3 expression is significantly different between samples (t-test, P=.0157), which confirms the microarray results. b, Allele-specific qPCR of a coding polymorphism (rs1107853), which demonstrates that the log₂-fold change of the G allele relative to the A allele is significantly different between heterozygous DNA (Het DNA) and heterozygous cDNA (Het cDNA) samples (t-test, P=.00118). c, The gene structure of SH2B3, shown in blue with rectangles denoting exons. The arrow indicates transcriptional orientation. The graph below shows the distribution of conservation scores (dark gray) (University of California–Santa Cruz Genome Browser) across the SH2B3 gene and pairwise F_ST values (yellow) between the CEU and YRI samples for each SNP in this region (derived from HapMap phase II release 21 data).

Interestingly, these observations coincide with patterns of genetic variation at SH2B3, since there are 13 SNPs with large allele-frequency differences (F_ST⩾0.45) between the CEU and YRI samples (fig. 5c). Five of these highly differentiated SNPs occur in conserved regions, as determined by alignment of 17 vertebrate genomes, making them strong candidates for future functional studies. We calculated the empirical probability of observing a SNP with a pairwise F_ST⩾0.45 between the CEU and YRI samples, on the basis of all autosomal markers contained in HapMap release 21, to be ∼0.05, and this magnitude of allele-frequency difference is consistent with a signature of local adaptation.^7,39 SH2B3 also possesses unusually large levels of linkage disequilibrium compared with the rest of the genome,⁴⁰ which provides additional support for the hypothesis that this locus has been subject to adaptive evolution, although additional studies will be necessary to make more-definitive inferences about its evolutionary history.

In summary, consistent with previous studies of model organisms,^5,12–14 our results demonstrate that considerable natural variation in gene-expression levels exists within and among human populations. Genes differentially expressed among populations may be particularly relevant to explore as candidate susceptibility loci for diseases whose prevalence varies as a function of ethnicity and may be amenable to genetic dissection by admixture linkage-disequilibrium mapping.⁴¹ Importantly, we also show that simply focusing on differentially expressed genes can lead to an incomplete understanding of how gene-expression variation is apportioned within and among human populations. By decomposing expression variation into its component sources, we find that, similar to that observed for genetic variation, the majority of gene-expression variation is due to differences among individuals rather than among populations.

These observations are subject to several caveats, including the fact that a relatively small number of individuals and populations were studied and that expression levels were measured only in a single (transformed) cell type. For example, of the ∼17% of genes expressed in B lymphoblasts that we estimate to be differentially expressed between the CEU and YRI samples, 50 could be identified at an FDR <20%. Thus, to fully catalog the specific genes that are differentially expressed, it will be necessary to increase the sample size. In addition, it is plausible that probes that overlap SNPs could lead to biased estimates of gene-expression levels⁴² and confound our interpretation of gene-expression variation within and among populations. Although the algorithms we used to normalize the raw expression data and to combine individual probe sets into an overall measure of gene expression should be relatively robust to low levels of sequence divergence, it remains a formal possibility that probes interrogating sequences with SNPs contribute to the observed patterns of gene-expression variation. However, recent work suggests that this is unlikely to have a large influence on estimates of gene-expression levels among closely related populations.^43–45 Despite these limitations, our results and methodology provide the foundation for building a more principled understanding of natural variation in gene-expression levels that will be useful for testing hypotheses of regulatory evolution and interpreting patterns of expression variation in disease.

Appendix A

Tests of Differential Expression Within and Among Populations

We used the recently developed optimal discovery procedure²⁰ (ODP), available in the EDGE (J.D.S.'s Web site) software package,²¹ to test for differential expression between the CEU and YRI samples. We averaged technical replicates for each individual and tested for a difference in mean expression between samples, as detailed in the ODP method and EDGE manual (J.D.S.'s Web site). This produced an FDR Q value (J.D.S.'s Web site)²² for differential expression between samples for each gene, as well as a conservative estimate of the total proportion of population–differentially expressed genes (∼17%). The ODP is derived from the same principles yielding more-traditional methods, such as a t test and its popular microarray extensions.²³ However, the ODP is aimed at optimizing a more relevant balance between true positive and false positive results, to yield substantially greater power to identify genes, as has been shown elsewhere.²⁰ Although the global estimate of the proportion of differentially expressed genes, on the basis of a t test, was not significantly different from the ∼17% estimate produced by the ODP, the power of the ODP to identify genes as significant was substantially better here as well.

We formulated a new model of gene-expression variation from individuals in structured populations, to estimate the proportion of expression variation due to differences among individuals and among populations and to identify genes showing differential expression within populations. Specifically, we employed a mixed model in which population effects were treated as fixed and individual effects were treated as random. The model for each gene can be written as: expression=baseline+population+individual+error, where “baseline” is the fixed baseline expression level, “population” is the fixed effect due to potential differences between populations, “individual” is the random effect representing each individual’s potentially different average level of expression, and “error” represents the remaining random fluctuations in expression due to technical and measurement variation. Treating individual effects as fixed⁵ attributes too much variation to interindividual differences, which leads to substantial underbias in estimating the proportion of variation due to population effects. The mixed model allows for unbiased estimates of both population and interindividual differential expression effects and allows us to separate the technical and measurement errors from the biological signals of interest.

We fit the above model to each gene, by maximum likelihood under the assumption that the individual random effects and error terms are normally distributed, using the statistical software package R.²⁴ From this, we obtained point estimates of the fixed population effect (Δ), the variance of the individual random effect (τ²), and the variance of the error term (σ²) for each gene. We can show that the total variance for a gene’s expression is equal to Δ²+τ²+σ². We calculated the proportion of variance explained by population differences as the ratio of the variance due to population differences to the sum of the variances due to population differences and interindividual differences: Δ²/(Δ²+τ²). The proportion of variance explained by interindividual differences is equal to one minus this quantity: τ²/(Δ²+τ²). Note that the mixed model allowed us to remove the nonbiological variance component when partitioning the variance into within- and among-population components.

To test for differential expression among individuals, we performed a hypothesis test for each gene, to determine whether the variance corresponding to the individual random effect (τ²) is zero, where a nonzero variance indicates the presence of interindividual differences in expression. For each gene, the full model was fit by maximum likelihood, as described above, as well as the analogous model with no individual random-effect term. The two models were compared by a generalized likelihood-ratio statistic. We simulated >500,000 statistics from the null distribution by permuting the individual labels within each population and recomputing the generalized likelihood-ratio statistics on these permuted data. The observed and null statistics were then used to estimate an FDR Q value (J.D.S.'s Web site) for each gene as described elsewhere.²² This also provides a conservative estimate of the total proportion (83%) of interindividual differentially expressed genes.

Assessing Data Quality

We performed several diagnostic procedures to make sure that the results derived from tests of differential expression within and among populations were genuine and not confounded by technical artifacts. First, we observed that the significance of these two types of differential expression appeared to be independent. Genes showed both types of differential expression, only one type, or neither type in proportions expected by chance, given the fact that the two types of differential expression occur independently. Second, the correlation of log[p/(1-p)] of the two sets of P values was only 5%, which is well within the range observed under random permutations of the P values. Third, the error-variance estimates did not show any functional relationships with the estimated proportion of variation due to interindividual or population differences, indicating that our model successfully separated the biological signal of interest from the technical and measurement errors.