115 UK celiac disease individuals were recruited from Barts and the London NHS Trust and the Oxford Radcliffe Hospitals NHS Trust after informed consent and with ethical approval. Individuals all had a small bowel endoscopic biopsy diagnosis of celiac disease, median age of 51 (23–88), a median age at diagnosis of 42 (1–75), a male to female sex ratio of 1:3, and a median length of treatment on a gluten free diet of 9.4 years (1–47). Celiac individuals responding to a gluten free diet typically show no detectable inflammation (which we confirmed at the mRNA level in peripheral blood). We also enrolled 22 healthy unrelated UK controls, with a male to female sex ratio of 1:1.2, but of unknown age.

2.5 ml of peripheral blood was collected into a PAXgene tube (Becton Dickinson, UK, 762165). PAXgene vials were chosen to prevent density gradient centrifugation, immortalization or in vitro cell culture artifacts changing mRNA profiles.

PAXgene tubes were mixed gently and incubated at room temperature for two hours. After collection, tubes were frozen at -20°C for at least 24 hours followed by storage at -80°C. RNA was isolated using the PAXgene Blood RNA isolation kit (Qiagen, UK, 762174). RNA was quantified using the Nanodrop (Nanodrop Technologies, USA). Total RNA integrity was analyzed using an Agilent Bioanalyzer (Agilent Technologies, USA).

Anti-sense RNA was synthesized, amplified and purified using the Ambion Illumina TotalPrep Amplification Kit (Ambion, USA) following the manufacturers' protocol. Complementary RNA was hybridized to Illumina HumanRef-8 v2 Whole Genome BeadChips and scanned on the Illumina BeadArray Reader. Data was handled through the Illumina BeadStudio Gene Expression module v3.2. All gene expression (MIAME compliant) and genotype data has been made freely available by submission to GEO under GSE11501.

Five celiac disease samples were excluded from subsequent analysis due to poor median probe intensity correlation with all other samples or incorrect sex assignment, based on an analysis of all the probes that mapped to the non-pseudoautosomal region of chromosome Y. A dataset of 110 celiac disease patients and 22 controls was then used for analysis.

Expression probes were mapped to the cDNA sequence from Ensembl v45_36g 23 and the NCBI build 36 genome assembly if necessary. Probes that had less than 96% sequence homology or that mapped to multiple loci were removed. Subsequent analyses were confined to autosomal probes, in order to prevent sex specific effects on gene expression. After removal of probes that map to sex chromosomes, data was quantile-quantile normalized 24.

All celiac disease patients were genotyped as previously described 20 using Illumina Infinium HumanHap300v1.0 BeadChips.

257,013 autosomal SNPs were tested for association with expression levels in the 110 celiac disease samples that met analysis criteria of minor allele frequency (MAF) > 0.1, exact Hardy-Weinberg equilibrium P-Value > 0.0001 and call-rate > 0.95. Analyses were confined to those probe-SNP pairs for which the distance from probe genomic midpoint to SNP genomic location was less than 250 kb or 500 kb, depending on the analysis performed. To prevent spurious associations due to outliers, a non-parametric Spearman's rank correlation analysis was performed. In order to correct for multiple testing we controlled the false discovery rate (FDR) 25. The distribution of all the observed p-values was used to calculate the FDR, by comparing this distribution to a null-distribution, obtained from an identical analysis where the expression phenotypes, relative to the genotypes had been permuted. Through 1,000 permutations the Spearman's rank correlation P-value threshold could be determined that corresponded to an FDR of 0.01 or 0.05.

We compared the identified cis-eQTLs in the celiac peripheral blood dataset to a published human genetical genomics dataset 3. We reanalyzed expression data from EBV-transformed B cell lines (further described as HapMap B cell line dataset) for 90 CEU HapMap samples 3. Analyses were performed as described for the celiac disease samples. To enable a comparison between the celiac peripheral blood dataset and the HapMap B cell line dataset, only SNPs were tested that had been successfully called within HapMap and that were present on the Illumina HumanHap300 platform (257,013 SNPs). Although this is only a subset of all the SNPs that have been called for these HapMap samples, this subset of SNPs is known to capture most genetic Caucasian variation well 26.

We investigated over- or underrepresentation of certain biological processes or functions through an analysis of all significant cis-eQTL genes using the Panther Classification System 27 (Binomial P-Value, Bonferroni corrected).

Many cis-eQTLs have been detected in human datasets, but only a few trans-eQTLs have been found 13810. These trans-eQTLs imply a biological relationship between the trans-locus and the trans-gene and as such can provide valuable biological insight. If several of these were to be identified, they would allow for the reconstruction of gene networks.

A different, regularly used approach for reconstructing gene networks is by systematically assessing co-expression between pairs of genes. We performed such a co-expression analysis for each probe that comprised a cis-eQTL and contrasted this to an analysis where the genetic component of the cis-eQTL on the probe intensity level had been removed.

We first investigated whether the identified probe that comprised a cis-eQTLs (probe distance of 250 kb, FDR of 0.05) showed co-expression with all other probes that mapped to different chromosomes for both the celiac peripheral blood dataset and the HapMap B cell line dataset. We correlated the measured intensity levels with all other probes that mapped on different chromosomes, through a Spearman's rank correlation analysis. By applying Bonferroni correction to account for multiple testing, sets of significantly correlated transcript-pairs could be identified.

Subsequently we used partial-correlations, enabling us to redo this co-expression analysis while effectively removing the genotypic effect on the probe intensity level for the probes that constitute the significant cis-eQTLs. Again, by applying Bonferroni correction, sets of significantly correlated transcript-pairs could be identified.

To obtain the most accurate reflection of mRNA levels in peripheral blood leukocytes, whole blood RNA was immediately fixed during venepuncture in PAXgene vials, giving a reflection of in vivo RNA expression from whole blood. One hundred and fifteen treated celiac patients, all of whom were successfully treated and compliant with a gluten free diet for at least six months, were enrolled. Seventy percent of celiac disease patients were female, (66.6% of adult cases diagnosed with celiac disease in the population are female 29). In addition, 22 random healthy control samples were enrolled to obtain a background RNA expression profile, enabling us to determine whether the patients indeed did not show ongoing inflammation. In a sex matched comparison no known inflammatory disease associated cytokines, including IFNG 30 and IL2 31 showed significantly increased expression (sex matched Wilcoxon Signed-Ranks Test P value < 0.05) in celiac versus control samples, as was expected since these patients had been treated with a gluten free diet.

Gene expression levels for 19,867 transcripts were analyzed for significant genotypic effects at 257,013 autosomal SNPs, mapping within 500 kb of the centre of the transcript probes, resulting in 1,850,599 tests. By using a Spearman's rank correlation coefficient statistic and an FDR of 0.01 or 0.05, 1,360 and 2,178 different SNP-probe effects were detected, respectively. These reflect 394 and 658 unique probes, and 1,273 and 2,035 unique SNPs, respectively (see Table 1).

An identical analysis was performed on publicly available EBV-transformed B cell line expression and genotype data for 90 CEU HapMap individuals 3. Using this dataset and a SNP-probe distance of 500 kb and FDR of 0.01 or 0.05, 1,799 and 2,839 different SNP-probe effects were detected, respectively. These reflect 536 and 821 unique probes, and 1,601 and 2,464 unique SNPs, respectively (see Table 1). Details of all detected individual cis-eQTLs are available in Additional file 1.

Our analysis was initially confined to probes that had a midpoint distance to the tested SNPs less than 500 kb. Analysis of the significant cis-eQTLs SNP-probe distances (Figure 1) suggests few cis-eQTLs have been missed by imposing this threshold, as in both datasets for 95% of the cis-eQTLs, the SNPs map within 250 kb of the probes. As such it is expected that an increase in statistical power can be achieved by reducing the distance to 250 kb as less tests will be performed. Indeed in the celiac peripheral blood dataset more cis-eQTLs were identified (in total 2,487 different SNP-probe effects were observed, compared to 2,178 in the 500 kb analysis, based on an FDR of 0.05), reflecting 765 unique probes and 2,315 unique SNPs. For the HapMap B cell line dataset 3,226 different SNP-probe effects were identified (compared to 2,839 in the 500 kb analysis, based on an FDR of 0.05), reflecting 994 unique probes and 2,826 unique SNPs (see Table 1). Subsequent analyses were confined to cis-eQTLs that had been detected using this 250 kb window analysis.

4,681 Illumina expression probes had oligonucleotide sequences that are shared between the two different oligonucleotide arrays used (Human Ref8 v2 & Human WG-1 v1). By limiting the analysis to a window size of 250 kb and only to SNPs that had been successfully genotyped in both studies, 576 different SNP-probe effects in the celiac peripheral blood data at a FDR = 0.05 (338 at an FDR = 0.01) were detected. These reflect 573 different SNPs and 197 different probes. In the HapMap B cell line data, 573 different SNP-probe effects were identified at an FDR = 0.05 (290 at an FDR = 0.01), reflecting 573 different SNPs and 189 different probes. A combined meta-analysis of both cohorts (weighted-Z method) identified 1,133 different SNP-probe associations at an FDR = 0.05 (428 at an FDR = 0.01) (see Figure 2). These reflect 1,120 SNPs and 328 unique probes. 440 SNP-probe pairs, reflecting 217 different probes and 439 different SNPs (FDR = 0.05) were not detected when either dataset was analyzed separately.

135 identical SNP-Probe effects (reflecting 135 different SNPs and 51 different probes) had been identified in both the 110 celiac disease samples and in the 90 HapMap B cell line samples (FDR = 0.05). In all cases, the combined meta-P-Value for each of these shared cis-eQTLs was more significant due to the larger sample size, indicating allelic effects in the same direction. A comparison between cis-eQTLs that had been detected in a large scale linkage based genetical genomics study using RNA obtained from peripheral blood mononuclear cells 2, and the cis-eQTLs we detected, indicates that 57.3% (distance 250 kb, 0.05FDR) of probes, displaying a genotypic effect in the celiac peripheral blood dataset, also show a linkage cis-eQTL signal. Of note is that CCR3 and IL18RAP are both identified as cis-eQTLs in both these datasets.

While it can be assumed that for most of the detected cis-eQTLs indeed the probe expression is affected by genetic variation, it can also be that SNPs, mapping to regions to which the probe hybridizes, may affect hybridization efficacies and result in cis-eQTLs 32 that are not due to expression differences. For the celiac peripheral blood dataset, 10.0% of all Illumina HumanRef-8 v2 probes map to regions that contain known dbSNP polymorphisms. For the HapMap B cell line dataset, 20.5% of all Illumina HumanRef-6 v1 probes map to known SNPs. For the probes that make up the identified cis-eQTLs (distance 250 kb, FDR = 0.05) this percentage is significantly higher for both the celiac peripheral blood analysis (12,5%, Fisher's Exact test P = 0.02) and even more pronounced for the HapMap B cell line dataset (29.1%, P = 7.76 × 10^-11). Numbers within brackets in table 1 denote the number of cis-eQTLs, probes or genes that are potentially due to these primer polymorphisms. If these primer polymorphisms are responsible for different hybridization efficacies, these SNPs should be in LD with the SNPs that make up the cis-eQTLs. We could assess this for those cis-eQTLs where genotype data was available within HapMap Phase II for both the SNP that makes up the eQTL and for the SNPs that map within the probe that constitutes this particular eQTL. We compared the resulting distribution of r² values to a null distribution, generated by calculating the LD between the SNP that makes up an eQTL and the SNP that mapped immediately adjacent to the primer polymorphism SNP. While the mean LD (mean r² = 0.27) in the celiac peripheral blood was not significantly higher than the mean LD in the null distribution (mean r² = 0.25, Wilcoxon Mann-Whitney P-Value = 0.86), this was observed for the HapMap B cell line dataset (mean r² = 0.46, mean r² of null distribution = 0.30, Wilcoxon Mann-Whitney P-Value = 6.3 × 10^-8).

To enable a comparison of gene expression levels between the HapMap B cell line samples and the celiac peripheral blood samples, we limited our analysis to 12,401 transcripts that map to identical exons. The expression levels for these transcripts were quantile normalized. cis-eQTLs that have only been significantly detected in the celiac peripheral blood dataset, reflect probes that on average show higher expression in the celiac peripheral blood samples compared to HapMap B cell line samples (Wilcoxon Signed-Ranks test, P = 7.6 × 10^-7) (see Figure 3). Conversely, probes that comprised HapMap B cell line specific cis-eQTLs were higher expressed than within the celiac peripheral blood dataset (P = 2.0 × 10^-4). As expected, probes that comprised cis-eQTLs that were common to both data sets did not show differences in expression (P = 0.30). Noteworthy is that IL18RAP and CCR3, both recently identified as being genetically associated with celiac disease 21, exhibited significant cis-regulation in the celiac peripheral blood dataset but not in the HapMap B cell line dataset. Cis-effects for these two genes could be detected as these two genes showed markedly upregulated expression in the celiac peripheral blood dataset (expression rank order 12,072 vs. 6,431 and 11,104 vs. 3,602, respectively). These results indicate that the analysis of different cell types is valuable to gain insight in the potential functional consequences of disease associated genetic variants.

The genes, comprising the significant cis-eQTLs (distance 250 kb, FDR = 0.05), showed an overrepresentation of hydrolase and transferase functions for both datasets, but 'immunity and defense' cis-eQTLs genes were more predominantly detected in the celiac disease peripheral blood dataset than in the HapMap B cell line dataset (see Additional file 2), as might be expected from the differential RNA profiles of the cells under investigation in each dataset.

The observed concordance in allelic direction for the 135 significant SNP-probe effects, detected in both the 110 celiac peripheral blood and the 90 HapMap B cell line samples was perfect when probe sequences were identical. We also assessed whether the allelic directions for cis-eQTLs probes was the same when multiple significant cis-eQTL probes mapped within the same genes. We identified 14 genes (see Table 2) where this was not the case: significant cis-eQTL probes with different oligonucleotide sequences mapped to the same genes but showed opposite allelic directions. For three of the genes, MRPL43, OAS1 and TIPRL, we re-sequenced the probe regions in four, four and three individuals with different genotypes, respectively, and did not discover any polymorphisms, confirming that these cis-eQTLs do not result from artificial hybridization differences.

We used the celiac disease dataset to assess whether significant trans-eQTLs could be identified (Spearman's rank correlation test P-Value < 1 × 10^-12, FDR = 0.05). Two trans-eQTLs were detected: rs2318331 (mapping within COL22A1) with GI_48093066 (NBPF3) and rs12634559 (mapping downstream of IL1RAP) with GI_16306505 (CASP8AP2). However, these trans-eQTLs were not detected in the HapMap B cell line dataset, even after relaxing the nominal Spearman rank correlation P-value to 0.05.

As trans-eQTLs implicate (direct or indirect) biological relationships, it is somewhat disappointing that none were clearly identified. An alternative way to identify biological relationships is by using co-expression analysis (although such an analysis does not have the aim to identify or explain trans-eQTLs). We conducted such a co-expression analysis and compared it to a co-expression analysis that also takes the genetic variation into account. First, co-expression was assessed between each of the probes that constitute cis-eQTLs (SNP-probe midpoint distance < 250 kb, FDR = 0.05) and all other probes (Figure 4a) that do not map to the same chromosome, by calculating all pair-wise Spearman correlation coefficients. This resulted in 13,003,505 and 42,270,807 pair-wise tests for the celiac peripheral blood dataset (using 765 different cis-probes) and the HapMap B cell line dataset (using 994 different cis-probes), respectively. Through Bonferroni correction we accounted for multiple testing, resulting in the identification of 50,821 and 168,292 significantly correlated transcript-pairs respectively (Spearman's correlation test P-Value < 0.05 after multiple testing correction).

Subsequently, we redid this co-expression analysis after having removed the genotypic effect on the probe intensity level for the probes that constitute the significant cis-eQTLs (Figure 4b). We observed a considerably increased number of significantly correlated transcript-pairs: 54,773 (an increase of 7.8%) and 258,874 (an increase of 53.8%), respectively.

To validate whether this increase in co-expressed transcripts reflects known biology, we assessed a collection of 80,350 known biological interactions (derived on 17 April 2007 from KEGG 33, BioGrid 34, Reactome 35, BIND 36, HPRD 37 and IntAct 38). For the celiac peripheral blood dataset 91 of the 50,821 (0.179%) significantly correlated transcript-pairs that had been identified in the unconditioned co-expression analysis are known to interact. In the conditioned co-expression analysis 106 of the 54,773 (0.194%) transcript pairs were known to interact, indicating an 8.0% increase in the proportion of known biological interactions that was identified. Comparable results were obtained for the HapMap B cell line dataset: For the unconditioned co-expression analysis 334 of the 168,292 (0.195%) significant transcript-pairs are known to interact. In the conditioned co-expression analysis 634 of the 258.874 (0.245%) significant transcript-pairs are known to biologically interact, representing an increase of 23.4% in the proportion of known interactions among the most significantly co-expressed transcript-pairs.

Figure 4c provides an example for Ubiquitin (UBA52), for which a strong cis-eQTL was observed for probe GI_15451941 with rs2314664 (Spearman's correlation coefficient P-Value = 1.11 × 10^-16) in the HapMap B cell line dataset. Within this probe three other polymorphism map (rs6554, rs34040670 and rs3209501), of which rs6554 is in near perfect LD with rs2314664 (r² = 0.98, HapMap CEU population), suggesting this cis-eQTL does not reflect a real expression difference, but rather a difference in hybridization efficacy. Co-expression with other genes is already present in an unconditioned co-expression analysis, because the distribution of co-expression for UBA52 with all other genes differs significantly from a theoretical null-distribution (Wilcoxon-Mann Whitney P-Value < 1 × 10^-12). However, in the conditioned co-expression analysis, this difference is more pronounced (see figure 4c). When confining this analysis to a set of 156 known interactions for UBA52, overall co-expression was significantly stronger for these pairs of genes in the conditioned (mean absolute Spearman's correlation coefficient = 0.39) than in the unconditioned co-expression analysis (mean absolute Spearman's correlation coefficient = 0.20, Wilcoxon-Mann Whitney P-Value < 10^-50).