Background

Gene regulation is a principal mechanism by which genetic variation contributes to phenotypic variation, making its study essential for understanding human evolution and disease. The genomics era has underscored the importance of noncoding variation in human evolution and disease: ~ 90% of the genotype–phenotype associations identified by genome-wide association studies (GWAS) cannot be explained by coding variation [1, 2], and similarly, genomic regions harboring evidence of selection in humans are significantly more enriched for variants altering expression than protein coding [3].

While GWAS and scans of selection can identify genomic regions of interest, they often lack the resolution to identify the specific genes underlying traits or targeted by selection. To bridge this gap, studies have aimed to identify genetic variation associated with fine-scale, molecular phenotypes, through quantitative trait locus (QTL) mapping [4]. Combining these molecular QTL maps with GWAS through colocalization, transcriptome-wide association studies, or Mendelian randomization continues to prove a fruitful approach for identifying genes causally linked to traits and potential drug targets. Unfortunately, there is a persistent ancestry bias in human genomics research, with nearly 80% of GWAS participants being of recent European ancestry [5, 6], as well as the majority of participants of molecular trait studies [7], greatly limiting our ability to translate findings from GWAS to diverse populations, as well as discover population-specific variation of interest [8].

While there is an established and active field of study identifying novel GWAS associations and genetic variation contributing to gene expression differences across populations [7, 9,10,11,12,13,14], most global populations remain understudied, particularly in sub-Saharan Africa. Africa is the birthplace of anatomically modern humans and harbors the greatest levels of human genetic diversity across continents. The majority of genomic studies of sub-Saharan African individuals have focused on populations of primarily West African descent, which fails to capture much of the genetic and phenotypic diversity within sub-Saharan Africa [15]. Moreover, Africa is home to a large array of biomes and terrains, and indigenous Africans continue to practice diverse cultural and subsistence strategies. Together, these environmental pressures have driven genetic adaptations to infectious disease [16], diet [17], and climate [10, 18], sometimes in a population-specific manner. These adaptive variants can have important implications for human health in Africa, and elsewhere [19], and inclusion of African populations is therefore vital for our understanding of human evolutionary history and health.

In this study, we probe the genetic architecture of gene regulation in whole blood from indigenous East Africans by performing expression QTL (eQTL) and splicing QTL (sQTL) mapping in a cohort of 162 individuals, representing nine ethnic groups, from Ethiopia and Tanzania. We measure the degree to which African architecture is shared with that of non-Africans, test whether Africans harbor functional variation absent from existing cohorts, and investigate the demographic and genetic forces that may contribute to variation in gene regulatory architecture. We test whether fine-mapping of QTL signals is improved in Africans relative to an equivalently sized cohort of European Americans, and highlight individual genes with improved fine-mapping in Africans. Finally, we measure the effect of selective forces on shaping gene regulatory architecture and identify candidate genes under selection.

Results

Population structure

The cohort for this study consists of 162 Ethiopian and Tanzanian individuals belonging to nine ethnically and culturally diverse sub-Saharan groups previously unsampled in gene expression studies, including the Cushitic speaking Agaw and Weyto, the Semitic speaking Argoba and Amhara, the Omotic speaking Dizi, the Nilo-Saharan speaking Mursi, the Chabu who speak an unclassified language similar to Nilo-Saharan, and the Khoesan speaking Hadza and Sandawe (Fig. 1A). These populations practice a variety of subsistence strategies, including foraging (Hadza and Chabu currently, Sandawe and Weyto formerly), pastoralism (Mursi), agriculturalism (Agaw, Amhara, and Argoba), and agropastoralism (Dizi), and live in diverse environments with differing pathogen exposures.

Fig. 1
figure 1

Global and genetic structure of study populations. A Locations of East African populations sampled in this study across Ethiopia and Tanzania. B Principal component analysis of genetic data across 162 East Africans, combined with 20 West African Yoruba (YRI), 20 European Americans (CEU), and 20 Han Chinese (CHB) from the 1000 Genomes Project. The percent variance explained by each principal component is indicated in parentheses. C ADMIXTURE analysis of East African, YRI, CEU, and CHB populations

Full size image

To investigate the genetic diversity and structure of these populations, a subset of 162 individuals were genotyped at approximately 4.5 million SNPs on the Illumina Omni5 Beadchip array. These data were further imputed using a reference panel composed of the 1000 Genomes Project (1kGP) dataset [20] and a dataset of whole genome sequences (WGS) from 180 sub-Saharan African individuals [21]. To place their genetic variation in a global context, genotype data from the nine study populations were merged with 1kGP WGS data from 20 individuals each of Yoruban (YRI), Northern and Western European (CEU), and Han Chinese (CHB) ancestry (methods). Principal component analysis (PCA) of this merged dataset recapitulates a primary separation between African and non-African individuals along the first PC, explaining 3.8% of the variance. The second PC, explaining 1.8% of the variance, further separates CEU and CHB individuals, as well as East Africans and the YRI (Fig. 1B). Higher PCs further separate variation in Africa; PC3 captures variation between the Hadza and YRI, and PC4 between the Hadza and Chabu. Several groups cluster relatively nearer to CEU Europeans along PC1, most notably the Ethiopian Agaw, Amhara, Argoba, and Weyto, which are known to have moderate levels of Eurasian admixture [15, 22, 23]. Inferred ancestry components from ADMIXTURE [24] also indicates non-African admixture among these Ethiopian groups, as well as admixture with Bantu-speaking populations of Western African origin [20], represented by the YRI, in the Sandawe, Mursi, and Hadza (Fig. 1C).

Transcriptomic traits in Africans

To assess the contribution of genetic variation to transcriptomic trait variation, we performed genome-wide QTL mapping for expression (eQTL) and splicing (sQTL) transcriptomic traits in cis for expressed protein-coding and long-noncoding RNA genes; collectively, we refer to eQTLs and sQTLs as transcriptomic QTLs (tQTLs). We first correct our phenotypes (expression and splicing) for a number of covariates, including age, sex, delivery date, hidden covariates inferred by PEER [25], and cell-type fractions inferred by CIBERSORT [26]. Cell-type composition of whole blood is known to vary between individuals, and to be a source of confounding in QTL studies [27]. To account for ancestry and relatedness, we generate a genetic relatedness matrix (GRM) and perform tQTL mapping using the linear mixed model tool GEMMA [28]. Testing all autosomal SNPs with minor allele frequency (MAF) greater than 0.05 and within 100 kb of the target gene transcription start site (TSS) for eQTLs or within 100 kb of the target intron for sQTLs, we identify 99,685 SNPs associated with the expression of 1330 genes (eGenes) and 74,445 SNPs associated with splicing of 1118 introns (sIntrons) in 776 genes (sGenes) at FDR < 0.05 (Methods).

SNPs associated with expression (eSNPs) and splicing (sSNPs) show a characteristic enrichment near the transcription start site or intron boundary of their target gene, respectively [29] (Additional file 1: Fig. S3A and B), and are enriched in a variety of functional categories, including transcription start sites, enhancers, and splice sites, and are depleted in repressed chromatin regions. We also find a significant overlap with chromatin QTLs (caQTLs) identified in lymphoblastoid cell lines (LCLs, Additional file 1: Fig. S3C). Further, alleles associated with increased chromatin accessibility are significantly more likely to be associated with increased gene expression (OR = 2.9, p = 8.2 × 10−37 Fisher’s exact test) and slightly less likely to be associated with increased junction inclusion (OR = 0.82, p = 0.03 Fisher’s exact test), suggesting that regulatory mechanisms altering chromatin accessibility play a greater role in regulation of gene expression than splicing. When we restrict to variants with a greater than 10% probability of being causal (Methods), we find a further enrichment in functional categories, particularly for caQTLs among eQTLs and splice regions among sQTLs, indicating we are capturing true causal variation (Additional file 1: Fig. S3C).

Of the genes tested, 198 have both an eQTL and sQTL in our cohort, suggesting possible shared genetic architecture between these transcriptomic traits. To evaluate whether eQTLs are enriched for sQTLs overall, we first compute the π1 statistic, which measures the estimated fraction of sQTLs that are true positives in the eQTL scan. A π1 value of 0.61 suggests that the majority of sQTLs affect expression or are in LD with variants affecting expression (Additional file 1: Fig. S4), though many of these fail to reach genome-wide significance. To account for the possibility that our findings are related to technical artifacts of RNA-seq mapping across different transcript lengths, we measure π1 across gene-length deciles. We find that smallest transcripts have the strongest replication overall, but all deciles show appreciable π1 (min 0.25, max 0.79, Additional file 1: Fig. S5), suggesting our findings are robust to these artifacts. To further evaluate whether the genome-wide significant eQTL and sQTL signals are driven by shared causal variants, we estimated 90% credible sets for each set of QTLs, defined as the minimal set of variants which have at least a 90% probability of containing the causal variant, using the probabilities estimated above (Methods). Overall, we find overlapping credible sets for 114 of the genes with both a significant eQTL and sQTL, which makes up about 9% (114/1,330) of all eGenes in our cohort, comparable to the 12% overlap observed in GTEx [30]. Taken together, this observation suggests that splicing variants likely cause subtle but detectable changes in gene read counts, but that the genetic variants driving genome-wide significant eQTLs and sQTLs are largely independent.

Replication of tQTLs in non-Africans

To validate our tQTLs, and to assess sharing of molecular trait architecture between cohorts of predominantly African vs. predominantly European ancestry, we compared our tQTL results to whole blood eQTL and sQTL summary statistics from the Genotype-Tissue Expression project (GTEx) v8, which is comprised of 85% European Americans [30]. An advantage of using this dataset for replication is availability of both eQTL and sQTL summary statistics for the same RNA-seq samples, though the post-mortem nature of the samples is known to affect gene expression in whole blood [31]. For those QTLs tested in both cohorts, we find that both eQTLs and sQTLs identified in the African cohort show overall high reproducibility in GTEx, with π1 values for eQTLs and sQTLs of 0.88 and 0.91, respectively (Additional file 1: Fig. S6, Methods). For eQTLs, we also found a high π1 replication of 0.97 with results from the eQTLGen consortium, a meta-analysis of 37 blood expression datasets [32]. In addition to π1, effect sizes between our cohort and GTEx also show overall strong concordance (Pearson’s \(\rho\) = 0.73 for eQTLs and 0.82 for sQTLs, Fig. 2B). To assess whether the observed replication is significantly affected by the different genome versions used between our study and GTEx v8, we also measured π1 of eQTLs in GTEx v7, finding a π1 of 0.83 (Additional file 1: Fig. S6).

Fig. 2
figure 2

Replication of tQTLs between East Africans and GTEx v8. A Minor allele frequency distribution in GTEx v8 of FDR-significant tQTLs identified in East Africans, colored by whether they have a p-value less than 0.01 in GTEx v8. B Comparison effect sizes of tQTLs identified in East Africans. Lines show the best fit regression line between East Africans and GTEx v8 effect sizes, colored by whether the tQTL is shared (i.e., is no longer significant after conditioning) or is independent (remains significant after conditioning)

Full size image

While tQTLs as a whole show strong replication using π1, we also investigate the degree to which individual loci show evidence of shared causal variation. Estimating credible sets for all eGenes and sIntrons in GTEx v8 as described above, we find that 715/1262 (57%) of eGene credible sets and 619/852 (73%) of sIntron credible sets in Africans overlap with credible sets in GTEx v8. While the majority of tQTL credible sets overlap, the many non-overlapping sets suggest many tQTL signals identified in Africans may be driven by independent causal variants. To further evaluate this independence, we remapped tQTLs in Africans, conditioning on sets of independent tQTLs identified in GTEx by forward regression [30]. In cases where there are no genome-wide significant eQTLs or sQTLs in GTEx (169 genes and 541 introns, respectively), we instead condition on the lead eSNP or sSNP in GTEx. Using the original FDR significance thresholds for calling eQTLs and sQTLs, we find that 362 (27%) of eGenes and 224 (20%) of sIntrons remain significant after conditioning on GTEx SNPs, including the top variants for 328 eGenes and 199 sIntrons, suggesting widespread independent causal variation in Africa.

The source of replication differences between our cohort and GTEx may be due to several factors, including MAF differences, false positives, differences in LD between cohorts, differences in expression, and/or differences in effect size. Investigating those African tSNPs that fail to replicate in GTEx (p > 0.01), we find that non-replicating tSNPs have consistently lower MAF in GTEx when compared with replicating tSNPs (Fig. 2A). Of the 308 lead eSNPs that fail to replicate in GTEx, 60 have a p-value > 0.01 in the larger eQTLGen dataset, close to our specified false discovery rate of 5% (60/1330 = 4.5%), suggesting we are not detecting an excess of false positives. Further restricting to those independent tQTLs identified above, we investigate whether MAF and LD differences can account for our findings. For eight genes, INPP5K, TMEM140, ACSM3, CNTNAP3, PPP1R14C, PDZK1TP1, GPR56, and TRAM2, the top eSNP in Africans is untested in GTEx and has a MAF < 0.01 (the threshold used by GTEx) in 1kGP EUR populations. The top eSNP for these eight genes are also non-significant or absent from the FIVEx browser [33]. Similarly, the top sSNPs for introns in four genes, ADAM8, ICAM2, LINC00694, and MAPK1, are absent in GTEx, absent or non-significant in the FIVEx browser, and have a EUR MAF ≤ 0.01. Overall, however, we find that frequency differences between Africans and EUR do not differ significantly between shared and independent eQTLs (p = 0.49, one-sided Kolmogorov–Smirnov (KS) test, Methods), while we do find a significant, though slight, enrichment for larger frequency differences among independent sQTLs (p = 5.74 × 10−3, one-sided KS test). To investigate the impact of LD variation on tQTL replication, we estimate r2 between tQTL lead SNPs and SNPs within 100 kb of lead SNPs in 1kGP CEU and YRI populations. We find that correlations between CEU and YRI r2 values do not differ significantly between shared and independent tQTLs (Additional file 1: Fig. S8, p = 0.25 for eQTLs, p = 0.43, one-sided KS test). Finally, comparing effect size estimates between the African cohort and GTEx at top tSNPs, we find a significantly lower correlation of independent tQTLs when compared with shared signals (Fig. 2B, p < 2.2 × 10−16), which may reflect true effect size variation, GxE effects [34,35,36], or possibly more subtle differences in MAF and local LD between these cohorts [37].

Finally, we investigate whether expression differences may affect replication between cohorts. Of the 1330 eGenes identified in Africans, the expression of 98 in GTEx v8 whole blood is too low to be tested for eQTLs. These 98 genes are significantly enriched in two KEGG pathways, “Hypertrophic cardiomyopathy” (FDR = 0.032) and “Dilated cardiomyopathy” (FDR = 0.038). Investigating what may be driving broader expression differences for testable genes, we identify those genes measured in Africans that fail to reach expression thresholds for testing in GTEx whole blood and vice versa. Altogether, 951 out of 12,377 genes measured in both cohorts and tested for eQTLs in Africans were not tested in GTEx. These genes are enriched for a number of biological processes related to sensory perception, including perception of smell (FDR = 2.85 × 10−6), sound (FDR = 1.60 × 10−5), mechanical stimulus (FDR = 5.60 × 10−5), and chemical stimulus (FDR = 5.22 × 10−4). Similarly, 6728 out of 18,168 tested for eQTLs in GTEx were not tested in Africans and are enriched for several biological processes related to immunity, including “complement activation, classical pathway” (FDR = 1.78 × 10−22), “humoral immune response mediated by circulating immunoglobulin” (FDR = 7.32 × 10−18), and “B cell mediated immunity” (FDR = 2.02 × 10−2). This observation suggests that disease status, sample collection, and response to environmental factors, in addition to genetics, may account in part for incongruent findings between eQTL cohorts.

Fine-mapping

In addition to assessing the replication of transcriptional QTLs in the larger GTEx v8 dataset, we are interested in the relative power to detect and fine-map tQTLs between cohorts of predominantly African versus European ancestry. To account for sample size differences between our cohort and GTEx, we performed eQTL mapping in a size-matched sample of 162 European American (EA) individuals from GTEx v8 using FastQTL [38], with sex, sequencing platform, PCR batch, the top 15 PEER factors, and top 5 genotype PCs as covariates. The number of PEER factors and genotype PCs was chosen based on prior GTEx analyses [30]. Testing all SNPs with MAF > 0.05 within 100 kb of the target TSS, we identify 1029 eGenes in the 162 EA individuals at FDR < 0.05, compared with 1330 identified in Africans, of which 326 eGenes are FDR-significant in both cohorts. Despite only 326 eGenes being shared, we find consistently high replication in an independent whole blood meta-analysis [32]; eQTLs that are FDR-significant in both cohorts reach a π1 of 0.999, while eQTLs discovered only in Africans reach a π1 of 0.958 and eQTLs discovered only in EAs reach a π1 of 0.989. This observation suggests that the greater number of eGenes discovered in Africans is not driven by an increase in false positives and that, at similar sample sizes, there is greater power to detect eQTLs in samples from African individuals when compared with samples from individuals of European ancestry.

We next investigate the relative ability to fine-map eQTLs between our African cohort and the 162 EA individuals from GTEx v8. Considering eGenes that are FDR-significant in either cohort (Methods), we perform fine-mapping in both our African cohort and the 162 EAs using the approach described above. Overall, most genes do not fine-map well at this modest sample size, with 57% of genes having a credible set larger than 50 in both cohorts (Fig. 3A). Excluding these genes, we find that Africans have a smaller credible set in 63% of cases (437/697, p = 2.06 × 10−11 binomial test), with a median credible set size of 25 in Africans vs 58 in EAs, and 23 genes fine-mapped to a single variant in Africans vs. 13 in EAs, demonstrating that using ethnically diverse populations facilitates fine-mapping, as has been shown previously [39]. One possible explanation of the smaller credible sets in Africans is that Africans simply have fewer SNPs tested per gene; however, we find the opposite, with 94% of genes have fewer tested SNPs in EAs.

Fig. 3
figure 3

Fine-mapping in East Africans vs. GTEx v8. A Credible set (CS) sizes for eGenes identified in 162 East Africans (Afr) or a subset of 162 European Americans from GTEx v8 (EA162). B Locus plot of NR1D1 eQTLs identified in 162 East Africans (Afr) or the full GTEx v8 cohort (v8). P-values are overlaid with African (YRI) and European American (CEU) recombination rates, GENCODE v19 [40] gene models from the UCSC genome browser [41] (http://genome.ucsc.edu) and inferred ChromHMM [42] states for GM12878 [43]. The top SNP in Africans, rs883871, disrupts a nucleotide for the core motif of ETS family transcription factors (motif of ETS1 shown)

Full size image

We further compare our credible sets in African eQTLs to credible sets estimated in the full GTEx dataset. As expected, the majority of eGenes have smaller credible sets in GTEx due to the considerably larger sample size (670 vs 162), though we do identify several examples of greatly reduced credible sets in the African cohort. For 18 eGenes and 32 sGenes, we are able to fine-map the QTL signals to a single variant in Africans and find that these variants overlap a lead GWAS association for 10 eGenes and 3 sGenes (supplement). We highlight rs883871 (Fig. 3B), an eQTL for both THRA and NR1B1, which is FDR-significant in GTEx whole blood but is not the lead eSNP. SNP rs883871 is a strong chromatin QTL in lymphoblastoid cell lines (LCLs) [44], overlaps the binding sites of numerous transcription factors (TFs) in the LCL GM12787 [43], is predicted to disrupt a consensus motif for the ETS family of TFs, which share a core “CCGGAA” motif, and is the lead SNP for a Multiple Sclerosis GWAS association [45]; variants in ETS1 itself have been previously associated with multiple sclerosis [46].

Signatures of selection

Gene regulation is known or suspected to underlie many adaptive traits in humans, including diet [17, 47], immunity [48], and skin pigmentation [10], and transcriptomic traits show evidence of both purifying and positive selection [35, 36, 49]. Consistent with previous tQTL studies, we find decreasing effect size with increasing MAF among eQTLs and sQTLs, indicative of negative selection against variants of large effects (Additional file 1: Fig. S9). To identify QTLs with evidence of positive selection, we measure genome-wide FST between our broader African dataset and the 1kGP European (EUR) individuals, with the expectation that selection for expression-altering alleles will lead to increased differentiation at these loci. To assess whether tQTLs are enriched for evidence of positive selection, we identify the highest FST value for all SNPs in high LD (r2 > 0.8) with the top eQTL or sQTL and compare these values with null SNPs matched on MAF and the number of SNPs in LD (Methods). Overall, we do not find an enrichment of high FST among eQTLs or sQTLs, either when combining all populations or testing populations individually, suggesting that selection has not driven significant frequency differentiation at the majority of tQTLs (Additional file 1: Fig. S10 and S11).

We next investigate evidence of selection at individual loci. To account for the fact that the top eSNP may not be the true causal SNP, we score an individual gene’s evidence of selection by taking a weighted sum of each SNP’s FST value multiplied by the probability of that SNP being causal. We also perform repeated permutations between FST and causal probabilities at a locus to generate a locus-specific background expectation (Methods). Considering as candidates loci with a score within the 99th percentile threshold of all SNP FST values, and greater than 99% of background values, we identify 23 eGenes and 20 sGenes with evidence of selection (supplement). The most differentiated eGene is TTC26 (weighted FST = 0.59); a mutation in this gene has been associated with abnormal cilia in model organisms and biliary ciliopathy in human liver [50]. We also identified a strong signature of selection at ARPC1B (weighted FST = 0.59), deficiency of which can result in severe immunodeficiency [51]. Other highly differentiated loci include Platelet Factor 4 Variant 1 (PF4V1, FST = 0.50), IL8 (FST = 0.49), a major inductor of immune cell chemotaxis and activation [52], and CCR1 (FST = 0.43), a chemokine receptor. Among the most differentiated sGenes, we find several related to immunity and metabolism, including NADSYN1 (weighted FST = 0.50), a gene associated with vitamin D concentration [53], BTN3A3 (weighted FST = 0.50), a butyrophilin gene implicated in activation of T cells [54], and GANC (weighted FST = 0.43), a member of the glycosyl hydrolase family 31, which play a key role in glycogen metabolism [55].

Given our genetically and culturally diverse cohort, we are also interested in tQTLs with evidence of population-specific differentiation and selection. For each of the nine populations in the African dataset, we calculate a modified version of the d-statistic [56], a summation of normalized, pairwise FST, which tests for variants that are highly differentiated in a focal population versus other populations (Methods). As above, we weight these d-statistics by the probability of a SNP being causal to derive a “d-score” for each gene or intron. Genes with high d-scores in populations with evidence of non-African admixture (i.e., Agaw, Amhara, Argoba, and Weyto) are more genetically similar to EUR samples from the 1kGP, based on FST. Conversely, populations with evidence of West African admixture (i.e., the Hadza, Mursi, and Sandawe) are more genetically similar to YRI samples at high d-score genes, suggesting that in many cases the genetic differentiation at these loci is driven by population-specific patterns of admixture. We therefore calculate the population branch statistic (PBS) [57] between individual populations in our study and 1kGP CEU and YRI populations. Considering genes with a weighted d and PBS score in the top 99.5th percentile as significant, we identify 22 eGenes and 22 sGenes with significant evidence of population-specific selection (Fig. 4A, B).

Fig. 4
figure 4

Population-specific selection in East Africa. Gene scores for the d-statistics plotted against the population branch statistics (PBS) for each population. PBS is calculated for each focal population versus the CEU and YRI populations from the 1000 Genomes Project. Genes with a score above the 99.5th percentile of genome-wide statistics for d and PBS are highlighted in red for eGenes (A) and sGenes (B)

Full size image

Among the top eGenes with evidence of population-specific selection is TMEM216 among the Nilo-Saharan speaking Mursi pastoralists (Fig. 4A). This gene is located near a skin pigmentation GWAS locus discovered in a cohort with the same sub-Saharan African populations [10]. This association signal overlaps the UV-repair gene DDB1, as well as several other genes expressed in melanocytes. Colocalization analyses show strong overlap between the African TMEM216 eQTL and pigmentation GWAS signals (PP4 = 0.95, Fig. 5, Methods), suggesting possible shared causal variation between TMEM216 expression and pigmentation variation. LD patterns around TMEM216 shows evidence of three independent eQTLs segregating for this gene, tagged by rs7948623, rs11230664, and rs3741265. Two of these SNPs, rs7948623, rs11230664, are also genome-wide significant GWAS SNPs for pigmentation variation in Africans, while the third, rs3741265, is marginally significant (p < 10−5, Fig. 5). All three SNPs show strong population-specific differentiation in Ethiopian Nilo-Saharan groups, who have amongst the highest levels of skin melanin of any global population (Additional file 1: Fig. S12). Previous analyses of these populations have shown evidence of a selective sweep near this pigmentation GWAS locus, including high PBS and d values among GWAS variants (Additional file 1: Fig. S14) and extreme negative Tajima’s D values overlapping the TMEM138/TMEM216 locus [10].

Fig. 5
figure 5

Colocalization of whole blood eQTLs and pigmentation GWAS. Comparison of pigmentation GWAS p-values from Crawford et al. [10] against eQTL p-values from our study (East Africa), GTEx v8 Whole Blood, or GTEx v8 Sun-exposed skin (lower leg), in the style of LocusCompare [58]. Variants are colored by their degree of LD with three top pigmentation GWAS variants, rs7948623, rs11230664, and rs2512809. Colocalization probabilities from coloc [59] (PP4) are indicated for each eQTL group

Full size image

The top GWAS variant, rs7948623, overlaps an active enhancer in keratinocytes and melanocytes and has been demonstrated to alter enhancer activity in melanocytes via luciferase reporter assays [10]. SNPS rs7948623 is a significant eQTL for TMEM216 in our study, as well as an eQTL mapping study of stimulated monocytes from European and African ancestry individuals [35], but is not significant in GTEx whole blood, though it has been identified in ovary, nerve, and exposed skin. In addition, rs7948623 is a significant sQTL for TMEM216 in multiple GTEx tissues, including exposed skin (Fig. 5). A second group of TMEM216 eQTL and pigmentation GWAS variants are tagged by rs11230664 and include the indel rs148172827, which overlaps an active melanocyte enhancer and shows significant correlation with TMEM216 expression in GTEx exposed skin (Fig. 5). We do not identify significant sQTLs in Africans for TMEM216; however, the top sSNP for TMEM216 in GTEx exposed skin, rs3741265 (p = 1.43 × 10−322), is in high LD with the top TMEM16 eQTL in Africans, rs7934229 (r2 = 0.99). Both of these SNPs are moderately associated with skin pigmentation in Africans (p < 5 × 10−6) but do not reach genome-wide significance (Additional file 1: Fig. S14).

Discussion

This study extends our understanding of the genetic basis of human gene regulation, with the inclusion of whole blood samples for 162 ethnically diverse sub-Saharan Africans from Ethiopia and Tanzania. We find that variation underlying expression and splicing is broadly shared between African and European cohorts, though there is considerable independent variation at individual loci in Africans, often driven by variation in frequency and effect sizes of tQTLs. When matched for sample size, Africans show improved fine-mapping of molecular traits, facilitating the identification of causal variants and candidate genes underlying GWAS traits. This diverse cohort also allows for inference of tQTLs with evidence of local adaptation, identifying TMEM216 as a target of selection in Nilo-Saharan speakers and a candidate gene that may play a role in skin pigmentation.

We find that the majority of tQTLs replicate between Africans and GTEx v8, with π1 values near 0.9 among both eQTLs and sQTLs, on par with the 0.919 value estimated between African Americans in the GENOA cohort [60] and EUR populations from the Geuvadis project [13]. We also observe strong effect size correlation between tQTLs in our study and GTEx v8. Investigating individual loci, however, we find that many genome-wide signals are driven by distinct causal variation; 43% of eQTL and 27% of sQTL credible sets in Africans do not overlap those in GTEx v8, and 27% of eGenes and 20% of sIntrons have QTL signals that remain significant after conditioning on all tQTLs in GTEx.

Investigating what may account for QTL differences between Africans and non-Africans, we find that genes relating to sensory perception and immunity show differential expression between our African cohort and the GTEx cohorts, pathways known to vary across populations and environments [11, 61], though the explained variability is generally low. Additionally, the post-mortem nature of GTEx samples may contribute to expression differences. An analysis of the effects of death on gene expression in GTEx found that immune genes in whole blood are significantly dysregulated following death; however, this change was characterized by an overall deactivation of immune genes, along with an overall increase in NK cells and CD8 T cells and a reduction in neutrophils [31]. In addition to expression differences, we find an enrichment for low-frequency variants in GTEx among non-replicating tQTLs. However, the majority of tQTLs that are conditionally independent show similar frequency differences with shared tQTLs, suggesting that frequency variation alone cannot account for independent tQTLs. This issue of trans-ethnic GWAS replication is an ongoing area of research [62, 63], and non-replication may occur for many reasons including frequency variation, differences in power, LD, or true differences in effect size, including G × E effects. While we do not find a significant difference in local LD structure between shared and independent QTL signals, we do find significant differences in estimated effect sizes. Using a Bayesian approach to account for frequency and LD variation, Brown et al. also found eQTL effect size differences between EUR and YRI individuals from Geuvadis [13], which become more pronounced as genetic effects become weaker [64]. However, for strong, genome-wide significant effects, Zanetti and Weale demonstrated using simulations that most trans-ethnic differences in GWAS effect sizes can largely be accounted for by a combination of frequency and LD variation, though they could not rule out effect size differences [37]. More recently, Patel et al. leveraged local ancestry information to infer differences in causal effect sizes between variants on European and African ancestry-derived haplotypes, finding a significant effect of haplotype background on variant effect size [65].

Beyond replication, we demonstrate that at comparable sample sizes, African cohorts have improved sensitivity to detect tQTLs and improved ability to fine-map causal variants, compared with cohorts of European ancestry. It is well established that non-African populations have more extensive LD relative to Africans [66, 67], resulting from the out-of-Africa bottleneck [68, 69], and that multi-population analyses can improve causal variant detection[39], which likely account for the observed improvement in fine-mapping in African populations. As to the increased sensitivity to detect tQTLs in Africans, one hypothesis is a higher false-positive rate in the African cohort. However, we find comparable replication of African-specific tQTLs in a large, independent meta-analysis [32], suggesting that false positives do not account for the observed improvement. Moreover, Quach et al. found a similar pattern of improved sensitivity to detect eQTLs in individuals of self-reported African ancestry in an analysis of stimulated and unstimulated monocytes from 200 Belgians, 100 of European and 100 of African ancestry [35]. Among African Belgians, they found 13% more eQTLs in unstimulated monocytes, and 10% more eQTLs across all conditions. While several other studies have mapped eQTLs across multiple ancestry groups [12, 13, 36, 70], variation in sample size precludes direct comparison of sensitivities across ethnicities.

In addition to the inclusion in our study of ancestral groups not represented in existing reference cohorts (e.g., the 1kGP), which enables the detection of novel regulatory variation, these populations live in diverse climates and have distinct cultural and subsistence practices, which may have driven unique local adaptations. Using an outlier approach based on the FST based d and PBS statistics [56, 57], we identify population-specific differentiation of tQTLs among East African populations. One notable example is the eQTL TMEM216 among the Mursi, which is near a recently identified pigmentation locus specific to sub-Saharan Africans [10]. TMEM216, and the nearby TMEM138 gene, form an evolutionarily conserved cis-regulatory module vital for ciliogenesis and have been identified as causal genes underlying Joubert and Merkel syndromes [71, 72]. TMEM216 has not been previously associated with pigmentation variation, though activation and suppression of primary cilia have been shown to inhibit and activate melanogenesis, respectively, in a human skin model [73]. Consistent with this, we find that the expression decreasing allele is associated with increased melanin levels for rs7948623, rs11230664, and rs3741265 and is most common in the Mursi, a population with darkly pigmented skin (Additional file 1: Fig. S12) [10]. In addition, recurrent somatic mutations driving alternative splicing of TMEM216 are significantly associated with melanoma in The Cancer Genome Atlas (TCGA), suggesting possible tumor suppressor function for this gene [74]. While the strong colocalization between the TMEM216 eQTL and pigmentation GWAS signals suggests TMEM216 as a possible pigmentation gene, there are several haplotypes segregating in this region, some of which carry tQTLs for other genes in GTEx (Additional file 1: Fig. S16 and S17). In addition, several nearby genes show melanocyte-specific expression or have been previously associated with pigmentation in other organisms, complicating identification of the gene or genes that are causally associated with pigmentation variation [10, 75].

There are several limitations to our study, foremost being our modest sample size of 162 individuals, with current eQTL datasets reaching sample sizes an order of magnitude larger [60]. Many of the populations participating in this study live at considerable distances from medical or scientific facilities, and all necessary tools and supplies must be transported to field sites, greatly limiting the capacity for sample collection. Additionally, we are limited to studying blood tissues among these populations. Generation of induced pluripotent stem cells (iPSC) may allow for the study of gene regulation across developing tissues or differentiated cells within diverse populations [76, 77], but such approaches remain technically difficult. This study is also restricted to steady-state gene expression, which may miss cell-type- or dynamic, environment-specific genetic effects, which cannot be captured in bulk and/or steady-state tissues [34,35,36, 78,79,80]. Despite these limitations, this study makes important contributions to our understanding of gene expression variation and the molecular basis of human adaptation in sub-Saharan Africa.

Conclusion

We have presented a comprehensive analysis of transcriptomic variation in a cohort of previously unstudied indigenous sub-Saharan Africans. We identify extensive novel regulatory variation in Africans and show that the study of African populations improves the detection of transcriptomic QTLs and fine-mapping of causal variation. Studying diverse populations within Africa also allows for the detection of genes targeted by population-specific selection, including evidence of selection on TMEM216 expression in the Mursi and strong colocalization between TMEM216 eQTLs and a pigmentation GWAS locus.

Methods

Sample collection

Phenotypic, genealogical, and biological data were collected from individuals belonging to nine populations in Ethiopia and Tanzania. Prior to sample collection, IRB approval for this project was obtained from the University of Pennsylvania. Written informed consent was obtained from all participants and research/ethics approval and permits were obtained from the following institutions prior to sample collection: the University of Addis Ababa and the Federal Democratic Republic of Ethiopia Ministry of Science and Technology National Health Research Ethics Review Committee; COSTECH, NIMR, and Muhimbili University of Health and Allied Sciences in Dar es Salaam, Tanzania. To obtain DNA and RNA data, whole blood was collected using vacutainers and RNA was stabilized in the field using LeukoLOCK Total RNA Isolation System (Ambion life Technologies). The Poly(A)Purist Kit (Ambion Life Technologies, CA) was used for mRNA selection, and Ampure XP magnetic beads (Beckman Coulter, CA) were used for size selection after amplification.

Genotyping and imputation

A subset 162 individuals were genotyped as part of the 5 M dataset using the whole genome Illumina Omni5 Beadchip array, which includes approximately 4.5 million SNPs. The full 5 M dataset was phased using Beagle 4.0 [81] and the 1kGP reference panel [20]. These data were further imputed using minimac3 [82] and a reference panel consisting of the 1kGP and 180 WGS from the Tishkoff lab [21]. The 180 WGS data include 15 individuals from each of the following populations used in our study: Amhara, Dizi, Hadza, Mursi, Chabu, and Sandawe.

PCA and ADMIXTURE

To identify related individuals, relatedness was inferred in the imputed 5 M dataset using the KING extension of plink 2.0 [83]. To place the genetic variation in this study within a global context, the 5 M imputed dataset was merged with the 1KGP. Individuals from the 162 in this study with inferred relatedness more distant than third degree were then extracted from the merged dataset (145 total), along with 20 individuals each from the YRI, CEU, and CHB populations, restricting to unambiguous SNPs (i.e., excluding A/T and C/G) with MAF > 0.01 and with imputation accuracy (r2) greater than 0.99 reported from minimac3. SNPs were LD-pruned using plink v1.90 [84] and parameters “–indep-pairwise 50 10 0.1.” PCA was performed on this dataset using smartpca from EIGENSOFT v6.1.4 [85], with “numoutlieriter” set to 0. ADMIXTURE [86] was run on the same dataset using parameters “–cv -j8 -B100 -s7.”

mRNA sequencing and molecular trait quantification

Samples were sequenced on an Illumina HiSeq to a median depth of 56,122,076 reads (11,727,716 min., 228,660,534 max.). Prior to mapping, all reads aligned to rRNA genes with BLAST [87] were removed. Remaining reads were mapped to the hg19 genome with STAR v2.5.3a [88] and the GTEx GENCODE v19 gene annotations [40] using two-pass mapping. Expression was quantified at the gene level using featureCounts v1.5.3 [89] as fragments per gene, as well as using RSEM v1.2.31 [90] as transcripts per million (TPM). Splicing was quantified using leafcutter [91] as fraction of intron exclusion reads per cluster (JPC).

Cell-type inference

Cell-type fractions for each individual were inferred using CIBERSORT [26]. The LM22 signature gene file from Abbas et al. [92] was used to infer frequencies of 22 immune cell types for a mixture file of TPM values for all 171 individuals with RNA-seq data. Quantile normalization was disabled, and 1000 permutations were used.

Quantile normalization and hidden factor inference

Prior to hidden factor inference and QTL mapping, molecular phenotype matrices were first filtered and quantile-normalized. For eQTL mapping, only lncRNA and protein-coding genes with more than 5 reads in at least 20 individuals and with mean TPM > 0.1 across all populations were considered. For sQTL mapping, introns from lncRNA and protein-coding genes with no more than 5 individuals with 0 reads were included. Furthermore, clusters were required to have at least 20 reads in at least 100 individuals and have 0 reads in fewer than 10 individuals. These filtered phenotype matrices (TPM for eQTL mapping and JPC for sQTL) were then quantile-normalized using the two-stage procedure implemented by GTEx [30]. Briefly, the distribution of the phenotypes per individual were first quantile-normalized to the mean of the phenotypes across individuals. Next, the distribution of each phenotype was quantile-normalized to the standard normal. Hidden covariates were inferred using PEER [25] for these quantile-normalized phenotype matrices.

eQTL and sQTL mapping

Expression and splicing quantitative trait loci were mapped using a linear mixed modelling approach, using the quantile-normalized gene or intron fractions as phenotypes, while correcting for sex, age, cell-type composition, delivery date, latent PEER factors, and genetic relatedness. Mapping was performed for SNPs with MAF > 0.05, imputation r2 > 0.3, and within 100 kb of the target phenotype (gene TSS for eQTLs and intron for sQTLs) using GEMMA [28] and a genetic relatedness matrix (GRM) generated from all biallelic SNPs across the imputed, 162 individual genotype dataset. tQTL mapping was repeated across a range of PEER factors: 0–5, 10, 15, 20, 25, and 30 factors for eQTL mapping, and 0–10 factors for sQTL mapping, and the number of factors maximizing the number of eQTLs or sQTLs discovered were chosen for downstream analysis.

To identify significant QTLs, tested SNPs for each phenotype were first FDR-corrected using Benjamini-Hochberg (BH), yielding single-corrected p-values (P’) for each tested SNP-phenotype pair. The minimum P’ per phenotype were again FDR-corrected using BH, yielding double-corrected p-values (P’’) per phenotype, and phenotypes with P’’ < 0.05 were considered significant. To identify significant SNPs, a threshold was set equal to the lowest P’ for the phenotype with highest significant P’’, and all SNPs with P’ lower than this threshold were deemed significant.

Credible sets

For each gene or intron of interest, Approximate Bayes Factors were calculated for each tested SNP using the function “approx.bf.estimates” from the coloc package [59], or the function “approx.bf.p” in cases where effect size or standard error information was not available. The posterior probability of each SNP n being causal (PPn) was then taken as:

$$P{P}_{n}= \frac{AB{F}_{n}}{\sum_{p}AB{F}_{p}}$$

Similar to The Wellcome Trust Case Control Consortium et al. [93], where \(AB{F}_{n}\) is the Approximate Bayes Factor of SNP \(n\), and \(p\) indexes all tested SNPs for a given feature of interest. A 90% credible set was then defined as the minimal number of SNPs whose sum of posterior probabilities was > 0.9.

Functional enrichment

All SNPs in the imputed genotype dataset of 162 individuals were annotated for functional consequences using the Variant Effect Predictor (VEP) [94] with parameters “–per_gene –most_severe.” In addition, SNPs were overlapped with 15 state ChromHMM tracks for PBMCs (E062) from the Roadmap Epigenomics Consortium [75], transcription factor binding sites for lymphoblastoid cell lines (LCLs, GM12878) from ENCODE [43], and chromatin QTLs from Tehranchi et al. [44]. To test for enrichment, each FDR-significant eQTL or sQTL was matched on MAF and distance to nearest TSS or intron boundary, respectively, and the log ratio of tQTL SNPs to matched background SNPs overlapping each functional category was taken as an enrichment score. This was repeated 10,000 times, producing an empirical distribution of enrichment scores for each functional category.

Replication with GTEx v8

All SNPs and intron boundaries were converted to hg38 coordinates using liftOver [95]. For eQTLs, those hg19 SNPs that successfully mapped to locations in hg38 (81,928/82,144) and genes with Ensembl IDs shared between GENCODE v19 and GENCODE v26 (1291/1330) were considered (96,903/99,685 of possible eQTLs). Of these, 77,238 eQTLs were tested in GTEx v8 and could be compared. For sQTLs, SNPs and Ensembl IDs were required to successfully map between versions (49,706/49,794 and 772/776, respectively), and intron boundaries were required to map between GENCODE versions (738/1118). Of these, 55,046 sQTLs were tested in GTEx. The fraction of true positives for successfully mapped tQTLs in GTEx, π1, was estimated using the R package qvalue [96].

Conditional tQTL mapping

To identify tQTLs in the African cohort that are independent of GTEx v8 tQTLs, we performed eQTL and sQTL scans conditioning on independent GTEx eQTLs and sQTLs identified via step-wise regression [30]. In cases where there are no significant tQTLs in GTEx, we instead use the top variant per feature. To account for these variants, we residualize the quantile-normalized feature matrices used in the original QTL mapping against the genotypes of independent GTEx QTLs. We then perform identical eQTL and sQTL scans and consider genes and introns with variants that pass the original FDR threshold as independent.

LD variation across populations

To compare LD structure between Africans and Europeans at tQTL loci, LD was estimated (using r2) between lead SNPs for eQTLs and sQTLs and all tested SNPs in the YRI and CEU 1kGP samples, restricting to those variants polymorphic in both, resulting in an r2 vector per group (YRI and CEU) per locus (eGenes and sIntrons). For each tQTL locus, we estimated the Pearson correlation \(\rho\) between the YRI and CEU r2 vectors, and the distribution of these \(\rho\) values was compared for tQTLs shared between East Africans and GTEx and independent tQTLs.

Testing differences in allele frequency and LD between shared and independent tQTLs

To test whether independent tQTLs show greater allele frequency differences between Africans and EUR samples compared with shared tQTLs, we perform a one-sided Kolmogorov–Smirnov (KS) test, with the alternative hypothesis being that the absolute frequency difference for independent tQTLs is right skewed (i.e. has an enrichment of large frequency differences) compared with the distribution of shared tQTLs. Similarly, to test whether independent tQTLs show weaker LD-structure correlation between African and Europeans compared with shared tQTLs, we perform a one-sided KS test using the \(\rho\) values calculated above, with the alternative hypothesis being that independent tQTLs are left skewed (i.e., has an enrichment of low \(\rho\) values) compared with shared tQTLs.

eQTL mapping in 162 European Americans from GTEx v8

eQTL mapping was performed on 162 individuals of European ancestry from GTEx v8 using FastQTL [38] with 10,000 permutations for all SNPs with MAF > 0.05 and within 100 kb of the target TSS. Covariates included the top 15 PEER factors, top 5 genotype PCs, sex, platform, and PCR batch. Significance was evaluated using the hierarchical Benjamini–Hochberg procedure used for African samples.

Scans of natural selection

To test for genetic differentiation between our African dataset and Europeans, all individuals belonging to the 9 populations in our study were extracted from the full 5 M dataset (664 total) and allele frequencies were combined with frequency information for EUR populations from the 1KGP, restricting to SNPs polymorphic in both datasets. FST was estimated using the Hudson estimator [97], and SNPs within the top 99th percentile (FST > 0.36) were considered outliers. To test for overall enrichment of FST outliers among tQTLs, we use an approach similar to that of Quach et al. [35]. The maximum FST value of SNPs in LD with lead tQTL SNPs (r2 > 0.8) was found, and the fraction of outliers among these maximum FST values was calculated. To generate a null expectation, each lead tSNP was matched with a random SNP, matching on MAF (bins of 0.05) and number of SNPs in LD (bins of [0], [1, 2], (2,5], (5,10], (10,20], (20,50], and > 50). The maximum FST of SNPs in LD with these matched SNPs was found, and the fraction of outliers among these matched maximum FST SNPs calculated. This procedure was repeated 10,000 times, generating a null distribution of expected number of outlier SNPs.

To identify individual eGenes and sGenes with evidence of selection, weighted FST scores were generated for each eGene and sIntron. For each feature of interest (gene or intron), the posterior probability of each tested SNP was calculated using the approach used to define credible sets, and for each feature, a weighted FST score was calculated as:

$$\overline{{F }_{ST}}=\sum_{p}{PP}_{p} {F}_{ST}^{p}$$

where \(P{P}_{p}\) is the posterior probability of SNP \(p\) being causal and \({F}_{ST}^{p}\) is the FST of SNP \(p\). Scores higher than the 99th percentile of genome-wide FST values were considered as candidate genes under selection. To further account for background selection, or other factors that may alter the FST of the surrounding region, we permuted FST across variants for each candidate and re-calculated the FST score. This was repeated 10,000 times to generate a null expectation of weighted FST scores for each candidate gene, which we compared observed scores against to generate empirical p-values. Those candidate genes with an empirical p-value less than 0.01 were considered significant.

To detect population-specific selection, we use an adapted, polarized version of the d-statistic for each SNP:

$${d}_{i}=\left|{\sum }_{j\ne i}{{\varvec{I}}}_{{p}_{i}\ge {p}_{j}}\frac{{F}_{ST}^{ij}-E\left[{F}_{ST}^{ij}\right]}{sd\left[{F}_{ST}^{ij}\right]}\right|$$

where \({p}_{i}\) and \({p}_{j}\) are the allele frequencies in populations \(i\) and \(j\), respectively, \({{\varvec{I}}}_{{p}_{i}\ge {p}_{j}}\) is an indicator function that returns 1 if \({p}_{i}\ge {p}_{j}\) and − 1 otherwise, \({F}_{ST}^{ij}\) is the FST between focal population \(i\) and population \(j\), and \(E\left[{F}_{ST}^{ij}\right]\) and \(sd\left[{F}_{ST}^{ij}\right]\) are the expected value and standard deviation of FST between populations \(i\) and \(j\) across all SNPs. We implement this polarization procedure because SNP frequencies that are at an intermediate frequency in the focal population, but strongly differentiated in others, can show up as strong \({d}_{i}\) outliers in the focal population due to the symmetry of FST. To identify individual eGenes and sGenes with evidence of population-specific selection, we generate weighted \({d}_{i}\) scores as described above for FST.

Due to differential levels of admixture across populations, some \({d}_{i}\) outlier loci show genetic similarity with non-African and West African populations, suggesting that these loci are uniquely differentiated in the focal population due to admixture. To eliminate candidates that may be driven by admixture, we also calculate the population branch statistic (PBSi) [57] between each focal population \(i\) and the CEU (a proxy for non-Africans) and the YRI (a proxy for sub-Saharan Africans):

$${PBS}_{i}= \frac{{T}^{i, YRI}+{T}^{i,CEU}-{T}^{YRI,CEU}}{2}$$

where \({T}^{A,B}=-\mathrm{log}\left(1-{F}_{ST}^{A,B}\right)\) and \({F}_{ST}^{A,B}\) is FST calculated between populations \(A\) and \(B\). We then go on to create a weighted \(PB{S}_{i}\) statistic per gene or intron as above. Candidates of selection are then defined as those features with a weighted \({d}_{i}\) and \(PB{S}_{i}\) score above the 99.5th percentile of genome-wide \({d}_{i}\) and \(PB{S}_{i}\) SNP-wise statistics.