Mapping of cis-acting expression quantitative trait loci in human scalp hair follicles

About 50 HFs were plucked from the occipital scalp of 100 (discovery sample), and the frontal and occipital scalp of 25 (replication sample, previously described in [5]) unrelated male volunteers. Peripheral venous blood samples were collected from all study participants. All volunteers were German residents of European descent, and showed a collective mean age of 27.9 years.

DNA was extracted from whole blood samples using the Chemagic Magnetic Separation Module I (Perkin Elmer Chemagen Technology Inc., Baesweiler, Germany). Total RNA was extracted from HFs using the RNeasy Micro Kit (Qiagen, Hilden, Germany), and the quality and quantity of the RNA were assessed using a BioAnalyzer 2100 (Agilent Technologies, Waldbronn, Germany), and a NanoDrop ND-1000 spectrophotometer (Peqlab Biotechnologie, Erlangen, Germany), respectively. Only total RNA samples with an RNA integrity number (RIN) ≥ 8 were further analyzed in the study.

DNA extracts were hybridized onto the Human OmniExpress-12v1.0 bead array (Illumina, San Diego, CA, USA) (N = 100) or the Illumina PsychArray v1.0 (N = 25) for genome-wide genotyping, while RNA extracts from HFs were amplified and biotinylated using the TotalPrep™-96 RNA Amplification Kit (Illumina, San Diego, CA, USA) prior to whole transcriptome profiling, performed on the Illumina HT-12v4 bead array.

SNP array raw data was initially analyzed using the Genotyping module within the GenomeStudio software (Illumina). Genotype calls were exported for basic quality control in PLINK v1.9 ([6]; www.​cog-genomics.​org/​plink/​1.​9/​) to eliminate bad quality data (e.g. SNPs and individuals with high degree of missing data, very rare SNPs) prior to genotype imputation. Imputation was performed on the Michigan Imputation Server [7] using the 1000 Genomes Project Phase 3 v5 reference panel and Eagle v2.3 phasing [8]. Post-imputation data processing was performed using VCFtools [9] and quality control was carried out in PLINK 1.9. Briefly, only biallelic single nucleotide variants with high imputation quality score (Rsq) > 0.7, minor allele frequency (MAF) ≥ 1% and under Hardy-Weinberg equilibrium (HWE p > 1 × 10^− 8) were further analyzed. A principal component analysis (PCA) was performed with PLINK to identify potential outlier samples and use the generated principal components (PCs) as covariates for eQTL analysis. Two final autosomal genotype datasets consisted of 5,887,234 SNPs and 98 individuals for the discovery sample, and 1,044,566 SNPs and 24 individuals for the replication sample.

Raw data from the expression microarrays was initially analyzed using the Gene Expression module within the GenomeStudio software to generate calls and detection p-values. The probe-level gene expression data was exported for pre-processing by background correction, quantile normalization, log2 transformation, probe quality filtering and identification of potential outlier samples by PCA using R. Probes were considered expressed when showing a detection p-value < 0.01 in at least 5% of the samples. Probe quality filtering included the retention of only “good” and “perfect” quality probes mapping to only one gene with a valid identifier, according to annotations retrieved from the illuminaHumanv4.db package [10]. Three final gene expression datasets consisted of 13,217 probes and 98 individuals for the discovery sample, and 13,091 probes from frontal scalp and 12,814 probes from occipital scalp and 24 individuals for the replication sample.

Genome-wide associations in cis (1 Mb window) between the expression levels in scalp HFs and SNP genotypes were tested using QTLtools [11]. Three covariate files prepared for the analyses included the first 10 (discovery sample) or 5 (replication sample) PCs for the genotype datasets and the first 10 or 5 PCs for the phenotype datasets. Initially, the full spectrum of eQTLs for each of all three gene expression datasets was identified through nominal pass analysis. After exploration of the nominal significant results (p < 0.05), only those eQTLs with false discovery rate (FDR) < 1 × 10^− 4 in the discovery sample were considered true eQTLs, while all eQTLs with p < 0.01 in the replication sample were retained for further analyses. To identify independent signals within our set of true eQTLs, a permutation pass analysis (1000 permutations), followed by a conditional pass analysis on grouped phenotypes (i.e. a gene-level output from the probe-level analysis) were applied.

We explored reported eQTL effects for our independent HF eQTLs using the Variant Annotation tool from SNiPA (Single Nucleotide Polymorphisms Annotator) ([12]; http://​snipa.​org), noting whether or not the HF eQTL has been previously reported to have cis-eQTL effects in at least one tissue, and whether or not eQTL effects have been observed on the same gene as in our study. Additionally, we searched for trait associations of our true HF eQTLs that have been reported in the GWAS Catalog ([13]; https://​www.​ebi.​ac.​uk/​gwas/​).

To investigate differential regulatory effects between frontal and occipital scalp areas, non-overlapping eQTLs with p-value < 5 × 10^− 5 were investigated in the replication sample. These included eSNPs that showed different effects between both datasets (i.e. association to a different eGene or opposing direction of the effect for the same SNP-gene pair; different-effect eQTLs), and those eQTLs that were unique to the frontal or occipital dataset (i.e. no overlaps in eSNPs or eGenes; region-specific eQTLs).

The lists of eGenes obtained from the discovery sample and the differential eQTL analysis were submitted for enrichment analysis using the GENE2FUNC function of the Functional Mapping and Annotation of Genome-Wide Association Studies (FUMA GWAS) platform ([14]; http://​fuma.​ctglab.​nl/​). Each analysis used the provided option to include all human genes as background, a significance cutoff of FDR < 0.05 and a threshold of minimum overlapping genes with the gene-sets = 3. In addition, a comparative analysis of biological pathways for the differential eQTLs was performed using the lists of the identified region-specific eGenes and the Gene Enrichment Compare function within the FunRich (Functional Enrichment analysis tool) software [15]. From the resulting pathway analysis using the default FunRich database, only frontal- and occipital-specific terms with at least 2 genes from the dataset present in the pathway term, and p < 0.01 (from hypergeometric test) in one scalp area and p > 0.05 in the other, were retained for the purposes of the present study.

To test the informativeness of our HF eQTLs for the interpretation of genetic findings for hair phenotypes, we used three sets of published genome-wide significant variants (p < 5 × 10^− 8) associated with (i) hair shape (discovery meta-analysis, supplementary table 2 from [1]), (ii) hair color (meta-analysis, supplementary table 2 from [2]), and (iii) AGA (reported lead SNPs from [16‐26]). These three variant sets were subjected to analysis by the SNP2GENE function of FUMA GWAS to assign GWAS findings to genomic regions. The analysis for each phenotype was set to include SNPs from the 1000 Genomes Project (Phase 3, European population) [27] that are in linkage disequilibrium (LD) with the reported GWAS variants. LD blocks were set to include variants of minor allele frequency ≥ 0.01, r² ≥ 0.6 and a distance < 500 kb for merging into a locus. Afterwards, we searched for overlaps between our true HF eQTLs and the resulting loci for hair shape, hair color and AGA.

We identified almost 3 million nominally significant eQTL signals in HFs from the occipital scalp (Fig. 1a). However, when we plotted the p-values and distances to the transcription start site (TSS) of the probe for each top SNP-probe pair with FDR < 0.05, we observed that a large number of SNPs were lying over 250 kb from the probe’s TSS, suggesting many of these might actually be noise signals (false positives). For this reason, only eQTL-findings that showed an FDR < 1 × 10^− 4 were considered the likely true eQTLs (14,497 probe-SNP pairs) (Fig. 1b). From this set of true eQTLs, we identified a total of 374 independent HF eQTLs (Fig. 2a, Suppl.Table.1 [Supplementary Tables]). The top 20 independent eQTL findings are presented in Table 1. Functional enrichment analyses using FUMA revealed that associated eGenes were highly enriched (FDR < 0.05) for a variety of Hallmark and Reactome pathways related to metabolism, immune functions, cellular proliferation, apoptosis, adipogenesis and responses to sex hormones (Fig. 1c, Suppl.Table.2 [Supplementary Tables]). Additionally, the annotation of reported eQTL effects with SNiPA showed that 266 of 374 independent eSNPs have been previously reported to have cis-eQTL effects in at least one other tissue, from which 228 were reported to affect the same eGene. Annotation of GWAS Catalog trait associations identified 13 independent HF eSNPs that are associated with 15 traits, including type 2 diabetes mellitus and chronic inflammatory diseases. From the set of true HF eQTLs, 185 eSNPs showed associations with 134 traits, with the highest number of associations found for body mass index (associated with 29 eSNPs) and blood protein levels (associated with 13 eSNPs) (data not shown).

We achieved the replication of 68 independent signals (Fig. 2) in either occipital and/or frontal scalp HFs from a total of 255 overlapping eQTLs (Suppl.Table.3 [Supplementary Tables]). In total, 188 and 157 eSNPs, and 264 and 276 eGenes, overlapped between our true eQTLs and the frontal and occipital HF eQTLs from the replication sample, respectively. Although only 92 eSNPs overlapped between all three eQTL datasets, these accounted for 249 eGenes (Fig. 1d and e); however, not all eQTL effects were consistent between the three datasets.

We included all eQTLs with p < 5 × 10^− 5 in the analysis of the replication sample. However, as the power to detect associations is markedly reduced due to the limited sample size (N = 24), this is reflected in the number and statistical significance of eQTLs that we detected in frontal and occipital scalp, as well as in their distributions with respect to the probe’s TSS (Suppl.Figure.1 [Supplementary Figures]). In general, we found little overlap between these two datasets, and will limit ourselves here to briefly present an overview of regional differences between frontal and occipital scalp.

We identified 71 HF eQTLs with inconsistent effects between frontal and occipital HFs. These were considered “different-effect eQTLs” and affected 11 genes (Suppl.Table.4 [Supplementary Tables]). Furthermore, we identified 289 frontal (Suppl.Table.5 [Supplementary Tables]) and 339 (Suppl.Table.6 [Supplementary Tables]) occipital (region-specific) eQTLs, from which the top 10 for each scalp region are shown in Table 1. To identify differentially-enriched pathways, we performed a separate FUMA pathway analysis for frontal and occipital differential eGenes (i.e. region-specific + different-effect) and compared the results for both scalp regions. This analysis suggested differences in the genetically-determined regulation of the responses to steroid hormones, cell cycle control, cellular metabolism and immune functions between the scalp regions (Suppl. Figure.2 [Supplementary figures]). However, as several of the region-specific pathway terms seemed redundant, we sought to further elucidate the differences between both scalp regions through a comparative pathway analysis using FunRich. This analysis suggested that the more important differences between frontal and occipital scalp HFs are on pathways related to the metabolism of amino acids and signaling by histone deacetylases (HDACs) in the frontal scalp, and to proliferative processes in the occipital area (Table 2).

To illustrate the applicability of our set of true eQTLs to interpret GWAS findings for hair phenotypes, we investigated for a potential overlap of HF eSNPs with genomic risk loci for AGA (98 loci), hair shape (12 loci) and hair color (122 loci) (Table 3). We identified one genomic locus at 1p36.13 that had previously been associated with hair phenotypes. This comprised 31 eSNPs overlapping association signals for hair shape (overlapping lead SNP: rs11203346) and hair color (overlapping lead SNP: rs72646785) (Fig. 3a). For this locus, the nearest genes were mapped to PADI3 and PADI4 in both GWAS, while the eGene for this region corresponded to PADI3. Moreover, another locus at 4q21.21, comprised by 26 eSNPs linked to the signal of rs6533756 (a non-eSNP in LD r² ≥ 0.8 with all eSNPs), was associated with hair shape. The overlapping eSNP most strongly associated with hair shape at this locus was rs7695038 (p = 3.9 × 10^− 10). While eSNPs at this locus locate within a region coding for the FRAS1 gene, the associated eGene, ANXA3, was located 66,182 bp downstream of FRAS1 (Fig. 3b). A third locus at 17q25.3 comprised the eSNP, rs8070929, which was in LD (r² ≥ 0.97) with rs34872037 (a non-eSNP) that is associated with hair color. While the nearest gene for the SNP at this locus was NPLOC4, the HF eGene corresponded to TSPAN10 (Fig. 3c). All eSNPs overlapping association signals for hair shape and/or hair color can be found in Suppl.Table.7. We found no overlaps of our true HF eQTL loci with the reported genome-wide significant findings for AGA. However, we observed that the eGene ATP2B4 was annotated as the nearest gene to an AGA genomic locus at 1q32.1. Moreover, it is perhaps worth to notice that, at the nominal significance level (p < 0.05), we found 93 eGenes to overlap with reported AGA candidate genes, of which 18 were associated with eSNPs at an FDR < 0.05 (ALPL, B3GNT8, BCL2, DIP2B, FAM136A, HDAC9, LCLAT1, LHPP, NSF, PRRX1, RPTN, RSPO2, SUCNR1, TCHH, TMEM50A, TWIST2, MSIG2 and SPAG17).