Variants in CXCR4 associate with juvenile idiopathic arthritis susceptibility

The JIA cases in our study were recruited from five sites in USA, Australia, and Norway: Texas Scottish Rite Hospital for Children (TSRHC; Dallas, Texas), Children’s Mercy Hospitals and Clinics (CMHC; Kansas City, Missouri), the Children's Hospital of Philadelphia (CHOP; Philadelphia, Pennsylvania), the Murdoch Childrens Research Institute (MCRI; Royal Children’s Hospital, Melbourne, Australia), and Oslo University Hospital (OUH; Oslo, Norway). (Table 1, Additional file 1: Table S1). A subset of subjects from these sites has been described previously [15‐19]. JIA diagnosis was made according to the International League of Associations for Rheumatology (ILAR) revised criteria [4] and confirmed using the JIA Calculator^TM software (URLs) [20], an algorithm-based tool adapted from the ILAR criteria. All JIA cases were of age of onset <16 years old.

The clinical data of JIA case in the CHOP cohort were collected from the JIA Registry maintained within the CHOP Division of Rheumatology; clinical data of case samples from TSRHC, CMHC, MCRI, and OUH were drawn from medical records provided by the respective sites and stored in a de-identified database at the Center for Applied Genomics of the CHOP Research Institute.

The control subjects used are unrelated and disease-free children recruited within the CHOP Healthcare Network. Control subjects had no history of JIA or other chronic illnesses and were screened as negative for a diagnosis of autoimmune diseases, based on data from CHOP’s electronic health record and by intake questionnaires obtained by the recruiting staff from the Center for Applied Genomics. A total of 6500 pediatric controls passed stringent quality control (QC) filtering, as detailed below; post-QC, cases and controls were matched based on the multidimensional scaling (MDS) analysis [21, 22]. For OUH cohort, the 3000 well-characterized subjects from the Wellcome Trust Case–control Consortium (WTCCC) [21] were used as controls.

We combined TSRHC and CMHC samples to form the discovery cohort, and kept CHOP, MCRI and OUH cohorts as three independent replication cohorts.

The study was approved by the institutional review boards of TSRHC, CMHC, CHOP, MCRI, OUH, and CCHMC, and was compliant with HIPAA regulations. Parental written informed consent was obtained from all participants prior to inclusion in this study for the purpose of DNA collection and genotyping.

All samples except those in the OUH replication cohort were genotyped using Illumina HumanHap550 BeadChip or the Human610-Quad arrays. The 530,000 SNPs that are overlapped by these two platforms were included in the study. Samples in the OUH replication cohort were genotyped using the Affymetrix GeneChip 500 K Mapping Array Set.

After the above sample and SNP QC, we conducted principal component analysis, trying to resolve relationship within-European samples. Using PLINK, we performed LD pruning and SNP exclusion so that only independent SNPs (r² < 0.2) on the autosomes were kept for PCA. We then performed PCA using EIGENSTRAT [23]. MDS analysis was also performed with the independent SNPs using PLINK.

For genotyped SNPs, association was tested by basic allelic test (chi-square test) and the odds ratio was calculated with respect to the minor allele using PLINK [22]. Logistic regression analyses were additionally performed, including the first 10 coordinates from the MDS analysis as covariates. Similar analyses were also performed including the first 10 principal components from the PCA analysis as covariates. Conditional association analysis was performed by including the genotype of the most associated SNP rs953387 as a covariate.

Imputation was carried out using software MACH 1.0 [24], with the reference panel of the HapMap CEU samples (HapMap release 22, NCBI build 36). The default two-step procedure was adopted for imputation. Imputed SNPs with MAF <0.01 in either cases or controls and SNPs with poor imputation quality (r² < 0.3) were excluded from further analysis. We also zeroed out imputed genotypes with a posterior probability of <0.9.

We performed meta-analysis using the inverse variance based method implemented in software METAL [25], which accounts for the direction of association relative to a consistent reference allele and adopts a fixed effect model. In this method, the effect size estimate of each cohort is weighted by its corresponding standard error. All meta-analyses comply with MOOSE guidelines (URLs).

We selected 480 patients with JIA and 480 controls without history of any autoimmune or inflammatory diseases, all of European ancestry based on the above MDS analysis. Samples were pooled in batches of 8 cases or 8 controls. One control pool was excluded for final analysis because it failed QC. Library preparation for targeted resequencing was performed according to the TruSeq (Illumina) sample-preparation protocol. DNA libraries were then hybridized to customized probes for capturing CXCR4 with NimbleGen SeqCap EZ Choice Library (Roche NimbleGen). The captured region is chr2:136871907–136895725, including introns. CXCR4-enriched libraries were sequenced on the HiSeq 2000 (Illumina).

We performed sequence alignment using BWA against the reference human genome (UCSC hg19). We achieved ~320X coverage per pool or ~40X per individual. We performed variant calling using SNVer [26], a statistical tool designed for pooled sequencing data. We used ANNOVAR software [27] to annotate variants. Each pool has 2 × 8 = 16 haplotypes, so we estimated allele frequency by rounding X/K*16, where X is the number of reads carrying the alternative allele, and K is the total coverage.

For the sequencing data, we employed the SUM test [28] for testing the association of the identified multiple non-synonymous variants. We computed the SUM association p value using the R package AssotesteR. Specifically, we used a permutation version of SUM, in order to prevent an inflated Type I error.

We performed Sanger sequencing to validate the rare non-synonymous and stop-gain sequence variants identified by targeted resequencing. Primers for all the five variants were designed using software Primer3 [29, 30] (URLs). Purification of PCR products was conducted using ExoSAP-it (USB, Affymetrix), and Sanger sequencing on both strands was performed using Big Dye Terminator Cycle Sequencing Kit v3.1 Kit (Applied Biosystems) with ABI 3130xl Genetic Analyzer (Applied Biosystems).

To test association between SNP genotypes and gene expression quantified in immortalized B-lymphocytes and T-cells, we performed in silico analysis using publicly available data from genome-wide expression analysis of quantitative trait loci (eQTL) of the 270 individuals genotyped in the HapMap Project (including 30 Caucasian trios of Northern and Western European origin [CEU]) [31‐34] and the 85 individuals of the GenCord project (a collection of cell lines from umbilical cords of individuals of Western European origin) [35]. Linear regression was used to test the association between gene expression and SNP genotypes under additive model [32]. SNP genotype was coded as 0, 1, and 2, corresponding to the counts of the minor allele in each genotype.

We performed a genome-wide association study in the discovery cohort (TSRHC + CMHC) including 388 children with JIA and 2500 genetically matched controls of European ancestry with high-quality SNP array data (Table 1, Additional file 1: Table S1). Then three independent JIA cohorts (CHOP, MCRI, OUH) were analyzed for replication and followed by meta-analysis of all the cohorts in a total of 1166 cases and 9500 controls (Table 1, Additional file 1: Table S1). The age at onset, gender, and subtype distributions of all case cohorts are similar to those reported previously for JIA [1].

We focused on the phenotypic commonality amongst our patients, that is, chronic inflammation of the joints. We observed a strong genome-wide significant association at the HLA locus on 6p21 (Additional file 1: Table S2, Additional file 1: Figure S1), A total of 24 SNPs surpassed genome-wide significance threshold at this region in the discovery cohort. Furthermore, three loci which previously have been implicated in JIA and several other autoimmune diseases [10, 36, 37] – the PTPN22 locus on 1p13, the IL2RA locus on 10p15, and the ANTXR2 locus on 4q21.21 – were nominally associated with JIA (Additional file 1: Table S3). Replication of these known JIA susceptibility loci demonstrated the validity of our study.

In addition to these known loci, we found a novel association signal at 2q22.1, with the most significant marker being rs953387 (p = 2.07 × 10⁻¹⁰; OR 0.59, 95 % CI 0.50–0.69). Six genotyped SNPs and ten additional imputed SNPs located on 2q22.1 of p value < 5 × 10⁻⁸ (Additional file 1: Table S4–S5, Additional file 1: Figure S1–S2) in the discovery cohort. Conditional on the top SNP rs953387, the significant association of the other SNPs were ablated, suggesting a single association signal at this locus. These SNPs were all nominally significant in each of our three independent replication cohorts. (Additional file 1: Table S4). Combined meta-analysis of all four cohorts indicated that four SNPs at the 2q22.1 locus were of p values < 5 × 10⁻⁸ (Additional file 1: Table S4).

However, as CXCR4 is near a known stratified locus (nearby the lactase gene) [38], logistic regression analysis including the first ten coordinates from the MDS analysis as covariates brought the locus p value below genome-wide significance (top associated SNP rs1016269, P = 4.70 × 10⁻⁵). Similarly, using the first 10 principal components as the covariates in the logistic region association, we obtained similar results (top associated SNP rs1016269, P = 2.80 × 10⁻⁵). Thus, our original association results were inflated due to subpopulation stratification within the subjects of European ancestry. The association of CXCR4 was nominally significant with the best p-value <0.0001. When we evaluated the association by another method, permutation tests with the inclusion of the first ten coordinates from the MDS analysis as covariates in each association analysis, we obtained similar results of nominally significant results based on 10,000 permutations. CXCR4 is an interesting gene because of its known role in immune regulation in a variety of immune cell types as a chemokine receptor. Gene CXCR4, is expressed on the surface of T-cells, B-cells, monocytes, neutrophils and dendritic cells (Additional file 1: Figure S3). The important function of CXCR4 during B cell development and activation [39, 40], and the key role of B cells in pathogenesis of JIA [41] make CXCR4 an intriguing target worth of further examination. Because of sub-population stratification, additional replication is still necessary for this locus to be considered a true risk locus for JIA. Its association with JIA need to be further investigated by stratified analysis in more refined homogeneous sub-populations of European ancestry, with a larger sample size. Further functional studies would also be required to demonstrate its potential contribution to JIA etiology.

Additionally, rs953387 was significantly associated with the mRNA levels of CXCR4 (p = 0.014; Additional file 1: Figure S4) among HapMap dataset of 30 CEU children in the Sanger Institute GENe Expression VARiation (Genevar) public database (URLs), which profiles gene expression in immortalized B-lymphocyte samples [31]. Analysis of an independent dataset of T-cell lines from umbilical cords of 75 individuals of Western European origin [35] also demonstrated association between genotypes of another significant SNP rs1016269 at CXCR4 locus (Additional file 1: Table S5) and levels of CXCR4 expression (p = 0.0054; Additional file 1: Figure S3). These data suggest that common variants at CXCR4 locus may regulate mRNA expression of this chemokine receptor.

In addition to common variants, rare variants also contribute to disease susceptibity and usually confer a larger effect than common variants. We further examined rare variants at this locus by conducting deep resequencing in 480 JIA cases and 480 controls. Table 2 shows the variants identified in the coding region of CXCR4. Of interest are 4 non-synonymous and 1 stop-gain SNVs. These are all rare variants that are not catalogued in current variant databases (1000G and ESP6500) or with an extreme low frequency (0.0003). Four of these five rare variants are present only in JIA patients, not in controls. The SUM test produced a significance p value of 0.015 when testing the association of these 5 variants, suggesting that rare variants in the CXCR4 gene may be contributing to the pathogenesis of JIA.

Then we performed Sanger sequencing to validate their genotype status. The primer sequences for each rare variant are shown in Additional file 1: Table S6. Among 40 samples of the 5 pools from which the rare variants were detected, 33 samples were left with enough DNA for Sanger sequencing. No genomic DNA was left for 4 samples in pool from which variant chr2:136872970 was identified and 3 samples in pool from which variant chr2: 136873341 was found. Among the 33 samples with enough genomic DNA, 4 samples were found carrying the rare variant identified in deep sequencing (Fig. 1).