Meta-analysis of GWAS on both Chinese and European populations identifies GPR173 as a novel X chromosome susceptibility gene for SLE

The discovery panel in the present study includes two cohorts with Chinese origin and one cohort with European origin (Table 1). All cases used in this study fulfilled the revised criteria of American College of Rheumatology for SLE. Informed consent was given by all individuals involved. All studies were approved by the corresponding institutional review boards.

The genome-wide genotyping data from Hong Kong [13] and Anhui Province, China [14], were generated using the Human610-Quad BeadChip array (620,901 markers; Illumina, San Diego, CA, USA), whereas the UK GWAS [10] was conducted using the HumanOmni1-Quad BeadChip array (1,140,419 markers; Illumina) and the HumanOmni2.5 BeadChip array (2,443,179 markers; Illumina). Quality control on the X chromosome was conducted in all datasets according to the following procedure. SNPs with a genotyping call rate < 90%, a minor allele frequency < 1%, or violating Hardy-Weinberg equilibrium (P < 1 × 10^− 4 in female controls) were discarded. Individuals with ambiguous gender or with SNP calling rate < 90% were also excluded.

On the basis of genotyping data, we imputed the X chromosome SNPs for all three datasets. First, SHAPEIT [15] was used to prephase each of the datasets. Subsequently, in order to obtain genotypes of additional SNPs, imputation on X chromosome SNPs was performed using IMPUTE v2.3.2 [16] on the three studies separately, using samples from the 1000 Genomes Project (phase 3, released in October 2014, build 37) as the reference. Both of the programs have a specialized algorithm to deal with X chromosome data. In all the studies, SNPs with an imputation score < 0.9 were removed from further analysis.

The X chromosomal SNPs passing quality control were analyzed for association using SNPTEST [16], fitting a logistic regression model in males and females separately. The association for European data, which comprise samples derived from different cohorts [10], was adjusted using four principal components. The two Asian cohorts were adjusted by the top two principal components. Meta-analysis was conducted using METAL [17], which employs a method based on inverse variance and weights the effect size estimates of each SNP by its SE. For each SNP, Cochran’s Q statistic and the I² index were used to test for any evidence of genetic heterogeneity between the Chinese and European data.

After meta-analysis, the top novel SNP with prominent association signal was selected for replication in three different cohorts, including 738 cases and 952 controls from Hong Kong (HK replication panel [HK_REP]), 460 cases and 965 controls from Thailand (Thailand replication panel [TH_REP]), and 1102 cases and 2327 controls from Anhui (Anhui replication panel [AH_REP]) (Table 1). The replication was performed using a TaqMan SNP genotyping assay (Thermo Fisher Scientific, Waltham, MA, USA). The missing genotype rate was less than 10% and was similar between cases and controls.

All of the known SLE risk loci in the X chromosome identified in previous association studies, as well as the SNPs in the surrounding regions (± 200 kb), were closely examined. Pairwise linkage disequilibrium (LD) among the SNPs was calculated to detect any potential independent signals. SNPs with association P_meta < 1 × 10^− 3 and low LD with the reported SNP (r² < 0.5) were selected for further examination, and the SNPs with a conditional P value less than 0.01 were considered as potential independent signals.

Because the association signals in the L1CAM- MECP2 region were much stronger than the others, the analysis was different for this region. Eighty-eight SNPs with genome-wide significance (P < 5 × 10^− 8) were selected, instead of using the P < 1 × 10^− 3 threshold used in the other regions. An LD block was plotted, and the top SNPs in each block were selected for a conditional analysis to test for independent associations. Only the SNPs with a conditional P value less than 0.01 were considered as potentially independent.

In order to measure the disease risk for SLE for each individual and compare it between the sexes and populations, we performed a modified calculation of genetic risk scores described by Hughes et al. [18]. Based on the 63 SLE autosomal susceptibility SNPs characterized by Morris et al. [19], a genetic risk score was calculated using the effect size of the risk alleles and the number of copies carried by each individual. Specifically, as shown in eq. (1), the number of risk alleles for SNP i (n_i) is multiplied by the natural logarithm of the corresponding OR, which is then summed and divided by k (total number of risk SNPs available for this individual) to obtain the average genetic risk score (aGRS):

We used the intragenomic replicates (IGR) method described by Cowper-Sal lari et al. [20] to predict the functional impact of a single-nucleotide variant (SNV) on transcription factor (TF) binding. This method takes a 7-bp short DNA sequence (7-mer) containing the target SNV (the 7-mer containing the reference allele is then referred to as “reference 7-mer,” and the 7-mer containing the alternative allele is then referred to as “alternative 7-mer”) and does genome-wide searches for the reference 7-mers and the alternative 7-mers. Then it compares the average chromatin immunoprecipitation sequencing (ChIP-seq) signal intensity of all reference 7-mer matches and all alternative 7-mer matches. A sliding window was used to find all 7-bp short DNA sequences containing the SNV. The reference 7-mer with the highest average intensity and the alternative 7-mer with the highest average intensity would be used for final comparison. All of the matches genome-wide would be filtered to exclude sites outside open chromatin (marked by DNase I hypersensitivity site [DHS]). Two cell lines—GM12878 (lymphoblastoid cell line) and K562 (human immortalized myelogenous leukemia cells)—were used in this IGR analysis. For the TF data in the GM12878 cell line, the corresponding GM12878 DHS data were used as the filter, and for the TF data in K562 cell line, the K562 DHS data were used as the filter. ChIP-seq files used in the analysis were downloaded from the ENCODE [21] website, and a complete list of files is provided in Additional file 1: Table S1.

Our study consisted of three sets of GWAS data, including two GWASs from China [13, 14] and one from the United Kingdom [10] (Table 1). Considering the sex difference in X chromosome dosage, we analyzed the female and male samples separately. After imputation by IMPUTE2 [16] and association analysis by SNPTEST [15], meta-analysis of the six datasets (female and male being analyzed separately) was conducted by METAL [17] using an inverse-variance-based method. A quantile-quantile plot (Additional file 1: Figure S1) was generated to evaluate association signals on the X chromosome. After removing all SNPs within the known X-linked SLE susceptibility loci and those within the range of ± 200 kb, the data (Additional file 1: Figure S1b) still deviated from the null expectation, which suggests that there are more novel X-linked SLE susceptibility loci to be discovered. A Manhattan plot (Fig. 1) was also generated to gain a systematic view of the association signals. Apart from the reported X-linked SLE susceptibility loci and the genes to which they are mapped, there seemed to be more novel signals to be further investigated.

On the basis of our meta-analysis results, after removing known associated regions as mentioned above, 48 SNPs with P_meta < 1 × 10^− 4 level of significance were further analyzed (Additional file 1: Table S2). Pruning based on LD (r² < 0.3) excluded 39 SNPs from the list, leaving 9 SNPs with potential independent signals (Additional file 1: Table S3). Among them, SNP rs13440883 showed the most prominent signal. LocusZoom [22] was used to plot the regional (± 200 kb) association signal (Additional file 1: Figure S2). The association of rs13440883 was detected in both Asian and European data, with a slightly higher effect size in Asians (OR_HK = 1.18; OR_AH = 1.18; OR_EUR = 1.13).

SNP rs13440883, as the most prominent signal, was selected for further replication. Three independent cohorts were used in the replication, with a total of 2300 cases and 4244 controls. After analysis of the replication results together with the GWAS data, the selected SNP (rs13440883) showed a final P_meta value of 7.53 × 10^− 9, reaching genome-wide significance (Fig. 2).

Conditional analyses were performed on Asian and European GWAS data to test for potential independent signals in this locus. Within the ± 200-kb window, there were 60 SNPs with association signals (P_meta < 1 × 10^− 3). However, after conditional analysis, none of them remained significant. Thus, based on the samples available, rs13440883 is the only independent association signal confirmed in this region.

Data on histone modification, DNase hypersensitivity, and TF binding, accessible using ENCODE [21], were used to identify functionally important SNVs, and rs13440883 was found to be within a region marked by H3K27ac and H3K4me1 in CD19 primary cells. The SNPs in high LD (r² > 0.8) with rs13440883 in both European and Asian populations and different TFs with binding peaks overlapping the corresponding SNPs are highlighted in Additional file 1: Figure S3a. Among them, SNP rs11091720, which showed high LD with rs13440883 (r²_EUR = 0.9384; r²_Asian = 0.9915), was a TF binding hot spot in the coding region of GPR173.

The IntraGenomic Replicates tool [20] was used to predict potential TF binding affinity differences that may lead to differential gene expression for both rs13440883 and rs11091720 (Additional file 1: Figure S3b and c). For rs13440883, the prediction showed a significant increase (P = 3.54 × 10^− 5) in the chromatin-binding intensity of SMARCA4 for the alternative allele. For rs11091720, the prediction showed a significant decrease in the intensity of L3MBTL2 (P = 2.31 × 10^− 13) and CTCFL (P = 3.67 × 10^− 10), as well as an increase in the intensity of CTCF (P = 6.22 × 10^− 11) and JUN (P = 0.0064) for the alternative allele. SMARCA4 is a component of a large ATP-dependent chromatin-remodeling complex (SNF/SWI) that is required for transcriptional activation of genes normally repressed by chromatin. CTCF forms methylation-sensitive insulators that may regulate X chromosome inactivation. The function of L3MBTL2, CTCFL, and JUN was not previously known. However, before experimental validation, we had to treat the results of this in silico analysis carefully, which is discussed later.

We further examined the associations of the reported susceptibility loci in previous studies, including PRPS2, LINCO1420, CXorf21, and IRAK1. The susceptibility gene TLR7 is not examined in this analysis due to poor coverage of this region. All the SNPs within a ± 200-kb window centered on the reported SNPs were filtered by association P values and pairwise LD (r²) with each other (see detailed criteria in the Methods section above) before analysis of independence. For the first three loci, we found no SNP passing the filtering criteria.

Meta-analysis of the GWAS datasets showed 88 SNPs in the L1CAM-MECP2 region (153,284,192 ± 200 kb, hg19) attained genome-wide significance (P < 5 × 10^− 8) (Additional file 1: Table S4), including the previously reported SNPs [9]. Independent contributions of each risk-associated SNP in this region was further examined. The LD pattern of the 88 SNPs was plotted separately for Anhui, Hong Kong, and U.K. data (Additional file 1: Figure S4). In the two Asian datasets, similar patterns were found, and four LD blocks were observed. The top SNPs from each block were then selected for a conditional test (Additional file 1: Table S5). We noted that rs5987175 (P_meta = 1.50 × 10^− 9) in LCA10 exhibited an independent contribution toward SLE susceptibility. After adjusting for the effect of the known independent SNPs (rs1059702, rs17422, rs2071128) reported in our previous study [9], rs5987175 was still significant in Asians (Table 2). However, the independence could not be replicated in the U.K. cohort, probably owing to higher LD between the blocks in Europeans, which is discussed later.