Abstract
Autoimmune disease-associated variants are preferentially found in regulatory regions in immune cells, particularly CD4+ T cells. Linking such regulatory regions to gene promoters in disease-relevant cell contexts facilitates identification of candidate disease genes. Here we show that, within four hours, activation of CD4+ T cells invokes changes in histone modifications and enhancer RNA transcription that correspond to altered expression of the interacting genes identified by promoter capture Hi-C (PCHi-C). By integrating PCHi-C data with genetic associations for five autoimmune diseases we prioritised 252 candidate genes, of which 116 were related to activation-sensitive interactions. This included IL2RA, where allele-specific expression analyses were consistent with its interaction-mediated regulation, illustrating the utility of the approach.
Genome-wide association studies (GWAS) in the last decade have associated 324 distinct genomic regions to at least one and often several autoimmune diseases (http://www.immunobase.org). The majority of associated variants lie outside genes1 and presumably tag regulatory variants acting on nearby or more distant genes2,3. Progress from GWAS discovery to biological interpretation has been hampered by lack of systematic methods to define the gene(s) regulated by a given variant. The use of Hi-C4 and capture Hi-C to link GWAS identified variants to their target genes for breast cancer5 and autoimmune diseases6 using cell lines, has highlighted the potential for mapping long range interactions in advancing our understanding of disease association. The observed cell specificity of these interactions indicates a need to study primary disease-relevant human cells, and investigate the extent to which cell state may affect inference.
Integration of GWAS signals with cell-specific chromatin marks has highlighted the role of regulatory variation in immune cells7, and in particular CD4+ T cells, in autoimmune diseases8. CD4+ T cells are at the centre of the adaptive immune system and exquisite control of activation is required to guide CD4+ T cell fate through selection, expansion and differentiation into one of a number of specialised subsets. Additionally, the prominence of variants in physical proximity to genes associated with T cell regulation in autoimmune disease GWAS and the association of human leukocyte antigen haplotypes have suggested that control of T cell activation is a key etiological pathway in development of many autoimmune diseases9.
Here, we explored the effect of activation on CD4+ T cell gene expression, chromatin states and chromosome conformation. PCHi-C was used to map promoter interacting regions (PIRs), and to relate activation-induced changes in gene expression to changes in chromosome conformation and transcription of PCHi-C linked enhancer RNAs (eRNAs). We also fine mapped the most probable causal variants for five autoimmune diseases, autoimmune thyroid disease (ATD), coeliac disease (CEL), rheumatoid arthritis (RA), systemic lupus erythematosus (SLE) and type 1 diabetes (T1D). We integrated these sources of information to derive a systematic prioritisation of candidate autoimmune disease genes.
Results
A time-course expression profile of early CD4+ T cell activation
We profiled gene expression in CD4+ T cells from 20 individuals across a 21 hour activation time-course, and identified eight distinct gene modules by clustering these profiles (Fig. 1, Supplementary Table 1). This time-course focused on much earlier events than previous large time-course studies (eg 6 hours - 8 days10) and highlights the earliest changes that are either not seen or are returning towards baseline by 6 hours (Supplementary Fig. 1). Gene set enrichment analysis using MSigDB Hallmark gene sets11 demonstrated that these modules captured temporally distinct aspects of CD4+ T cell activation. For example, negative regulators of TGF-beta signalling were rapidly upregulated, and returned to baseline by 4 hours. Interferon responses, inflammatory responses and IL-2 and STAT5 signalling pathways showed a more sustained upregulation out beyond 6 hours, while fatty acid metabolism was initially downregulated in favour of oxidative phosphorylation.
PCHi-C captures dynamic enhancer-promoter interactions
We examined activated and non-activated CD4+ T cells in more detail at the four-hour time point, at which the average fold change of genes related to cytokine signalling and inflammatory response was maximal, using total RNA sequencing, histone mark chromatin immunoprecipitation sequencing (ChIP-seq) and PCHi-C. Of 8,856 genes identified as expressed (see Methods) in either condition (non-activated or activated), 25% were up- or down-regulated (1,235 and 952 genes respectively, FDR<0.01, Supplementary Table 2). We used PCHi-C to characterise promoter interactions in activated and non-activated CD4+ T cells. Our capture design baited 20,676 HindIII fragments (median length 4 kb) which contained the promoters of 29,131 annotated genes, 18,202 of which are protein coding (Supplementary Table 3). We detected 283,657 unique PCHi-C interactions with the CHiCAGO pipeline12, with 55% found in both activation states, and 22% and 23% found only in non-activated and only in activated CD4+ T cells, respectively (Supplementary Table 4). 11,817 baited promoter fragments were involved in at least one interaction, with a median distance between interacting fragments of 324 kb. Each interacting promoter fragment connected to a median of eight promoter interacting regions (PIRs), while each interacting PIR was connected to a median of one promoter fragment (Supplementary Fig. 2).
We compared our interaction calls to an earlier ChIA-PET dataset from non-activated CD4+ T cells13 and found we replicated over 50% of the longer range interactions (100 kb or greater), with replication rates over ten-fold greater for interactions found in non-activated CD4+ T cells compared to interactions found only in erythroblasts or megakaryocytes (Supplementary Fig. 3). We also compared histone modification profiles in interacting fragments in CD4+ T cells to interacting fragments found in erythroblasts or megakaryocytes. Both promoter fragments and, to a lesser extent, PIRs were enriched for histone modifications associated with transcriptionally active genes and regulatory elements (H3K27ac, H3K4me1, H3K4me3; Supplementary Fig. 4), and changes in H3K27ac modifications at both promoter fragments and PIRs correlated with changes in gene expression upon activation. PIRs, but not promoter fragments, showed significant overlap with regions previously annotated as enhancers14.
We found that absolute levels of gene expression correlated with the number of PIRs (Supplementary Fig. 5a, rho=0.81), consistent with recent observations13. We defined a subset of PCHi-C interactions that were specifically gained or lost upon activation (2,334 and 1,866 respectively, FDR<0.01) and found that the direction of change (gain or loss) at these differential interactions agreed with the direction of differential expression (up- or down-regulated) at the module level (Fig. 2). We further found that dynamic changes in gene expression upon activation correlated with changing numbers of PIRs. Notably, the effect was asymmetrical, with a gained interaction having approximately twice the effect of a lost interaction (Fig. 3a). Given the >6 hour median half life of mRNAs expressed in T cells15 (Supplementary Fig. 5b), it is possible that the relatively weaker effects of lost interactions are due to the persistence of downregulated transcripts at the early stages of T cell activation.
As we sequenced total RNA without a poly(A) selection step, we were able to detect regulatory region RNAs (regRNAs), which are generally non-polyadenylated and serve as a mark for promoter and enhancer activity16. We defined 6,147 “expressed” regRNAs (see Methods) that mapped within regulatory regions defined by a 15 state ChromHMM17 model (Supplementary Fig. 6) and validated them by comparison to an existing cap analysis of gene expression (CAGE) dataset18 which has been successfully used to catalog active enhancers.19 We found 2,888/3,897 (74%) regRNAs expressed in non-activated cells overlap CAGE defined enhancers. This suggests that the combination of chromatin state annotation and total RNA-seq data presents an alternative approach to capture active enhancers.
Almost half (48%) of expressed regRNAs showed differential expression after activation (2,254/681 up/down regulated; FDR<0.01). To determine whether activity at these regRNAs could be related to that at PCHi-C linked genes, we focused attention on a subset of 640 intergenic regRNAs, which correspond to a definition of eRNAs20. Of these, 404 (63%) overlapped PIRs detected in CD4+ T cells and we found significant agreement in the direction of fold changes at eRNAs and their PCHi-C linked protein coding genes in activated CD4+ T cells (p<0.0001, Fig. 3b). We also observed a synergy between eRNA expression and the effect of a PIR on expression with a gain or loss of a PIR overlapping a differentially regulated eRNA having the strongest effect on gene expression (Fig. 3c), supporting a sequential model of gene activation21. While eRNA function is still unknown20, our results demonstrate the detection, by PCHi-C, of condition-specific connectivity between promoters and enhancers involved coordinating gene regulation.
PCHi-C-facilitated mapping of candidate disease causal genes
We defined an experimental framework to integrate PCHi-C interactions with GWAS data to map candidate disease causal genes for autoimmune diseases (Fig. 4). First, to confirm that PCHi-C interactions inform interpretation of autoimmune disease GWAS, we tested whether PIRs were enriched for autoimmune disease GWAS signals in CD4+ T cells, compared to non-lymphoid PIRs. We used blockshifter, which accounts for correlation between (1) neighbouring variants in the GWAS data and (2) neighbouring HindIII fragments in the interacting data by rotating one dataset with respect to the other. This method appropriately controls type 1 error rates, in contrast to methods based on counting associated SNP/PIRs which ignore correlation, such as a Fisher’s exact test (Supplementary Fig. 7). We found autoimmune GWAS signals were enriched in CD4+ T cell PIRs compared to non-autoimmune GWAS signals (Wilcoxon p = 2.5×10−7) and preferentially so in activated versus non-activated cells (Wilcoxon p = 4.8×10−5; Fig. 4).
Next, we fine-mapped causal variants for five autoimmune diseases using genetic data from a dense targeted genotype array, the ImmunoChip (ATD, CEL, RA, T1D), and summary data from GWAS data imputed to 1000 Genomes Project (RA, SLE; Supplementary Table 5). For the ImmunoChip datasets, with full genotype data, we used a Bayesian fine mapping approach22 which avoids the need for stepwise regression or assumptions of single causal variants and which provides a measure of posterior belief that any given variant is causal by aggregating posterior support over models containing that variant. Variant-level results are given in Supplementary Table 6, and show that of 73 regions with genetic association signals to at least one disease (106 disease associations), ten regions have strong evidence that they contain more than one causal variant (posterior probability > 0.5), among them the well studied region on chromosome 10 containing the candidate gene IL2RA22. For the GWAS summary data, we make the simplifying assumption that there exists a single causal variant in any LD-defined genetic region and again generate posterior probabilities that each variant is causal23. To integrate these variant level data with PCHi-C interactions and prioritize protein coding genes as candidate causal genes for each autoimmune disease, we calculated gene-level posterior support by summing posterior probabilities over all models containing variants in PIRs for a given gene promoter, within the promoter fragment or within its immediate neighbour fragments. Neighbouring fragments are included because of limitations in the ability of PCHi-C to detect very proximal interactions (within a region consisting of the promoter baited fragment and one HindIII fragment either side). The majority of gene scores were close to 0 (99% of genes have a score <0.05) and we chose to use a threshold of 0.5 to call genes prioritised for further investigation. Having both ImmunoChip and summary GWAS data for RA allowed us to compare the two methods. Overlap was encouraging: of 24 genes prioritised for RA from ImmunoChip, 19 had a gene score > 0.5 in the GWAS prioritisation, a further four had GWAS scores > 0.3. The remaining gene, MDN1, corresponded to a region which has a stronger association signal in the RA-ImmunoChip than RA-GWAS dataset, which may reflect the greater power of direct genotyping versus imputation, given that the RA-ImmunoChip signal is mirrored in ATD and T1D (Supplementary Fig. 8). We prioritised a total of 252 unique protein coding genes, 116 of which related to activation sensitive interactions (Supplementary Table 7, Fig. 4). Of 135 prioritised genes which could be related through interactions to a known susceptibility region, 64 (48%) lay outside that disease susceptibility region. The median distance from peak signal to prioritised gene was 152 kb. Note that prioritisation can be one (variant)-to-many (genes) because a single PIR can interact with more than one promoter, and promoter fragments can contain more than one gene promoter. Note also that the score reflects both PCHi-C interactions and the strength and shape of association signals (Supplementary Fig. 9), therefore a subset of prioritised genes relate to an aggregation over sub-genomewide significant GWAS signals. This is therefore a “long” list of prioritised genes which requires further filtering (Table 1). One hundred and eighty six (of 252) prioritised genes were expressed in at least one activation state; we highlight specifically the subset of 120 expressed genes which can be related to a genome-wide significant GWAS signal through proximity of a genome-wide significant SNP (p<5×10−8) to a PIR. Of these, 83 were differentially expressed, 49 related to activation-sensitive interactions and 29 showed overlap of GWAS fine-mapped variants with an expressed eRNA (Supplementary Table 7).
Taken together, our results reflect the complexity underlying gene regulation, and the context-driven effects that common autoimmune disease-associated variants may have on candidate genes. Our findings are consistent with, and extend, previous observations7,8 and we highlight six examples which exemplify both activation-specific and activation-invariant interactions.
PCHi-C may prioritise additional genes lying some distance from peak association signals. For example, CEL has been associated with a region on chromosome 1q31.2, for which RGS1 has been named as a causal candidate due to proximity of associated variants to its promoter24. Sub-genome-wide significant signals for T1D (min. p=1.5×10−6) across the same SNPs which are associated with CEL have been interpreted as a colocalising T1D signal in the region25. RGS1 has recently been shown to have a role in the function of T follicular helper cells in mice26, the frequencies of which and their associated IL-21 production have been shown to be increased in T1D patients27. However, our analysis also prioritises, in activated T cells, the strong functional candidate genes TROVE2 and UCHL5, over half a megabase distant and with three intervening genes not prioritised, for CEL and T1D (Fig. 5). UCHL5 encodes ubiquitin carboxyl-terminal hydrolase-L5 a deubiquitinating enzyme that stabilizes several Smad proteins and TGFBR1, key components of the TGF-beta1 signalling pathway28,29. TROVE2 is significantly upregulated upon activation (FDR=0.005) and encodes Ro60, an RNA binding protein that indirectly regulates type-I IFN-responses by controlling endogenous Alu RNA levels30 A global anti-inflammatory effect for TROVE2 expression would fit with its effects on gut (CEL) and pancreatic islets (T1D).
A similar situation is seen in the chromosome 1q32.1 region associated with T1D in which IL10 has been named as a causal candidate gene31. Together with IL10, prioritised through proximity of credible SNPs to the IL10 promoter, we prioritised other IL10 gene family members IL19, IL20 and IL24 as well as two members of a conserved three-gene immunoglobulin-receptor cluster (FCMR and PIGR, Supplementary Fig. 10). While IL19, IL20 and PIGR were not expressed in CD4+ T cells, FCMR was down- and IL24 and IL10 were up-regulated following CD4+ T cell activation. IL-10 is recognised as an important anti-inflammatory cytokine in health and disease32 and candidate genes FCMR and IL24 are components of a recently identified proinflammatory module in Th17 cells33. At this, and other regions, we found candidate causal variants interacting with multiple genes. Parallel results have demonstrated co-regulation of multiple PCHi-C interacting genes by a single variant34, suggesting that disease related variants may act on multiple genes simultaneously, consistent with the finding that regulatory elements can interact with multiple promoters35–37. This region also shows that clusters of multiple adjacent PIRs can be detected for the same promoter. It remains to be further validated whether all PIRs detected within such clusters correspond to ‘causal’ interactions or whether some such PIRs are ‘bystanders’ of strong interaction signals occurring in their vicinity. The use of PCHi-C nonetheless adds considerable resolution compared to simply considering topologically associating domains (TADs), which have a median length of 415 kb in naive CD4+ T cells34 compared to a median of 37.5 kb total PIR length per gene in non-activated CD4+ T cells (Supplementary Fig. 11).
Three neighbouring genes on chromosome 16q24.1, EMC8, COX4I1 and IRF8, were prioritised, the last only in activated T cells, for two diseases: RA and SLE (Supplementary Fig. 12). Candidate causal variants for SLE and RA fine-mapped to distinct PIRs, yet all these PIRs interact with the same gene promoters, suggesting that interactions, possibly specific to different CD4+ T cell subsets, may allow us to unite discordant GWAS signals for related diseases6,38,39. EMC8 and COX4I1 RNA expression was relatively unchanged by activation, whereas IRF8 expression was upregulated 97-fold, coincident with the induction of 16 intergenic IRF8 PIRs, four of which overlap autoimmune disease fine-mapped variants. Although the dominant effect of IRF8 is to control the maturation and function of the mononuclear phagocytic system40, a T cell-intrinsic function regulating CD4+ Th17 differentiation has been proposed41. Our data further link the control of Th17 responses to susceptibility to autoimmune disease including RA and SLE42.
Other notable examples include CCR7 and RARA, prioritised for T1D through a GWAS signal which maps to chromosome 17q21.2 (Supplementary Fig. 13a) and AHR, which was prioritised in rheumatoid arthritis (RA), specifically in activated T cells rather than non-activated T cells (Supplementary Fig. 13b). Both CCR7 and RARA are strong functional candidates with key roles in trafficking of CD4+ T cells and immune homeostasis43 and modulating T cell differentiation44, respectively. AHR is a high affinity receptor for toxins in cigarette smoke that has been linked to RA previously through differential expression in synovial fluid of patients, though not through GWAS45. Our analysis prioritises AHR for RA due to a sub-genome-wide significant signal (rs71540792, p=2.9×10−7) and invites attempts to validate the genetic association in additional RA patients.
Interaction-mediated regulation of IL2RA expression
We focused on the gene IL2RA and attempted to confirm predicted functional effects of fine-mapped variants on IL2RA expression. IL2RA encodes CD25, a component of the key trimeric cytokine receptor that is essential for high-affinity binding of IL-2, regulatory T cell survival and T effector cell differentiation and function46. Multiple variants in and near IL2RA have been associated with a number of autoimmune diseases31,47–49. We have previously fine-mapped genetic causal variants for T1D and multiple sclerosis (MS) in the IL2RA region22, identifying five groups of SNPs in intron 1 and upstream of IL2RA, each of which is likely to contain a single disease causal variant. Out of the group of eight SNPs previously denoted “A”22, three (rs12722508, rs7909519 and rs61839660) are located in an area of active chromatin in intron 1, within a PIR that interacts with the IL2RA promoter in both activated and non-activated CD4+ T cells (Fig. 6a). These three SNPs are also in LD with rs12722495 (r2>0.86) that has previously been associated with differential surface expression of CD25 on memory T cells39 and differential responses to IL-2 in activated Tregs and memory T cells50. We measured the relative expression of IL2RA mRNA in five individuals heterozygous across all group “A” SNPs and homozygous across most other associated SNP groups (Supplementary Table 9), in a four-hour activation time-course of CD4+ T cells. Allelic imbalance was observed consistently for two reporter SNPs in intron 1 and in the 3’ UTR in non-activated CD4+ T cells in each individual (Fig. 6b; Supplementary Fig. 14a), validating a functional effect of the PCHi-C-derived interaction between this PIR and the IL2RA promoter in non-activated CD4+ T cells. While the allelic imbalance was maintained in non-activated cells cultured for 2-4 hours, the imbalance was lost in cells activated under our in vitro conditions. Since increased CD25 expression with rare alleles at group "A" SNPs has previously been observed on memory CD4+ T cells but not the naive or Treg subsets that are also present in the total CD4+ T cell population39, we purified memory cells from 8 group "A" heterozygous individuals and confirmed activation-induced loss of allelic imbalance of IL2RA mRNA expression in this more homogeneous population (Fig. 6c, Supplementary Fig. 14b; Wilcoxon p=0.007). IL2RA is one of the most strongly upregulated genes upon CD4+ T cell activation, showing a 65-fold change in expression in our RNA-seq data. Concordant with the genome-wide pattern (Fig. 3), the IL2RA promoter fragment gains PIRs that accumulate H3K27ac modifications upon activation and these, as well as potentially other changes marked by an increase in H3K27ac modification at rs61839660 and across the group A SNPs in intron 1, could account for the loss of allelic imbalance. These results emphasise the importance of steady-state CD25 levels on CD4+ T cells for the disease association mediated by the group A SNPs, levels which will make the different subsets of CD4+ T cells more or less sensitive to the differentiation and activation events caused by IL-2 exposure in vivo51.
Discussion
Our results illustrate the dramatic global changes in chromosome conformation in a single cell type in response to a single activation condition, in addition to providing support for the candidacy of certain genes and sequences in GWAS regions as causal for disease. Recent attempts to link GWAS signals to variation in gene expression in primary human cells have sometimes found only limited overlap52–54. One explanation may be that these experiments miss effects in specific cell subsets or states, especially given the transcriptional diversity between the many subsets of memory CD4+ T cells.55 We highlight the complex nature of disease association at the IL2RA region where additional PIRs for IL2RA gained upon activation overlap other fine-mapped disease-causal variants (Fig. 6a), suggesting that other allelically-imbalanced states may exist in activated cells, which may also correspond to altered disease risk. For example, the PIR containing rs61839660, a group A SNP, also contains an activation eQTL for IL2RA expression in CD4+ T effectors56 marked by rs12251836, which is unlinked to the group A variants and was not associated with T1D56. Furthermore, rs61839660 itself has recently been reported as a QTL for methylation of the IL2RA promoter as well as an eQTL for IL2RA expression in whole blood57. The differences between CD25 expression in different T cell subsets58,59, and the rapid activation-induced changes in gene and regulatory expression, chromatin marks and chromosome interactions we observe, imply that a large diversity of cell types and states will need to be assayed to fully understand the identity and effects of autoimmune disease causal variants.
It will be challenging to assay this diversity of cell types and states in large numbers of individuals for traditional eQTL studies, particularly for cell-type or condition-specific eQTLs that have been shown to generally have weaker effects60,61. Allele-specific expression (ASE) is a more powerful design to quantify the effects of genetic variation on gene expression with modest sample sizes62 and the targeted ASE we adopt enables testing individual variants or haplotypes at which donors are selected to be heterozygous, while controlling for other potentially related variants at which donors are selected to be homozygous. By using statistical fine mapping of GWAS data, integrated with PCHi-C, to highlight both likely disease causal variants and their potential target genes, we enable the design of such targeted ASE analyses. This systematic experimental framework offers an alternative approach to candidate causal gene identification for variants with cell state-specific functional effects, with achievable sample sizes.
Online Methods
CD4+ T cell purification and activation, preparation for genomics assays
CD4+ T cells were isolated from whole blood using RosetteSep (STEMCELL technologies, Canada) according to the manufacturer’s instructions. Purified CD4+ T cells (average =96.5% pure, range 92.9 - 98.7%) were washed in X-VIVO 15 supplemented with 1% AB serum (Lonza, Switzerland) and penicillin/streptomycin (Invitrogen, UK) and plated in 96-well CELLSTAR U-bottomed plates (Greiner Bio-One, Austria) at a concentration of 2.5 × 105 cells / well. Cells were left untreated or stimulated with Dynabeads human T activator CD3/CD28 beads (Invitrogen, UK) at a ratio of 1 bead: 3 cells for 2-21 hours at 37°C and 5% C02. Cells were harvested, centrifuged, supernatant removed and either, (i) resuspended in RLT buffer (RNeasy micro kit, Qiagen, Germany) for RNA-seq (0.75-1 × 106 CD4+ T cells / pool and activation state) or microarray (1 × 106 CD4+ T cells / donor / timepoint and activation state) (ii) fixed in formaldehyde for capture Hi-C (44-101 × 106 CD4+ T cells / pool and activation state) or ChIP-seq (16-26 × 106 CD4+ T cells / pool and activation state) as detailed in34.
ChIP-seq was carried out according to BLUEPRINT protocols63. Formaldehyde fixed cells were lysed, sheared and DNA sonicated using a Bioruptor Pico (Diagenode). Sonicated DNA was pre-cleared (Dynabeads Protein A, Thermo Fisher) and ChIP performed using BLUEPRINT validated antibodies and the IP-Star automated platform (Diagenode). Libraries were prepared and indexed using the iDeal library preparation kit (Diagenode) and sequenced (Illumina HiSeq, paired-end).
For PCHi-C34, DNA was digested overnight with HindIII, end labeled with biotin-14-dATP and ligated in preserved nuclei. De-crosslinked DNA was sheared to an average size of 400 bp, end-repaired and adenine-tailed. Following size selection (250-550 bp fragments), biotinylated ligation fragments were immobilized, ligated to paired-end adaptors and libraries amplified (7-8 PCR amplification rounds). Biotinylated 120-mer RNA baits targeting both ends of HindIII restriction fragments that overlap Ensembl-annotated promoters of protein-coding, noncoding, antisense, snRNA, miRNA and snoRNA transcripts were used to capture targets. After enrichment, the library was further amplified (4 PCR cycles) and sequenced on the Illumina HiSeq 2500 platform.
PCHi-C interaction calls
Raw sequencing reads were processed using the HiCUP pipeline64 and interaction confidence scores were computed using the CHiCAGO pipeline12 as previously described34. We considered the set of interactions with high confidence scores (> 5) in this paper.
Raw PCHi-C read counts from 3 replicates and 2 conditions were transformed into a matrix, and a trimmed mean of M-values normalization was applied to account for library size differences. Subsequently, a voom normalization was applied to log-transformed counts in order to estimate precision weights per contact, and differential interaction estimates were obtained after fitting a linear model on a paired design, using the limma Bioconductor R package65.
Microarray measurement of gene expression
We recruited 20 healthy volunteers from the Cambridge BioResource. Total CD4+ T cells were isolated from whole blood within 2 hours of venepuncture by RosetteSep (StemCell technologies). To assess the transcriptional variation in response to TCR stimulation, 106 CD4+ T cells were cultured in U-bottom 96-well plates in the presence or absence of human T activator CD3/CD28 beads at a ratio of 1 bead: 3 cells. Cells were harvested at 2, 4, 6 or 21 hours post-stimulation, or after 0, 6 or 21 hours in the absence of stimulation. Three samples from the 6 hour unstimulated time point were omitted from the study due to insufficient cell numbers, and a further four samples were dropped after quality control, resulting in a total of 133 samples that were included in the final analysis. RNA was isolated using the RNAeasy kit (Qiagen) according to the manufacturer’s instructions.
cDNA libraries were synthesized from 200 ng total RNA using a whole-transcript expression kit (Ambion) according to the manufacturer’s instructions and hybridized to Human Gene 1.1 ST arrays (Affymetrix). Microarray data were normalized using a variance stabilizing transformation66 and differential expression was analysed in a paired design using limma65. Genes were clustered into modules using WGCNA67. Clustering code is available at https://github.com/chr1swallace/cd4-pchic/blob/master/make_modules.R.
ChIP sequencing and regulatory annotation
ChIP sequencing reads for all histone modification assays and control experiments were mapped to the reference genome using BWA-MEM68, a Burrows-Wheeler genome aligner. Samtools69 was employed to filter secondary and low-quality alignments (we retained all read pair alignments with PHRED score > 40 that matched all bits in SAM octal flag 3, and did not match any bits in SAM octal flag 3840). The remaining alignments were sorted, indexed and a whole-genome pileup was produced for each histone modification, sample and condition triple.
We used ChromHMM17, a multivariate hidden Markov model, to perform a whole-genome segmentation of chromatin states for each activation condition (Supplementary Table 8). First, we binarized read pileups for each chromatin mark pileup using the corresponding control experiment as a background model. Second, we estimated the parameters of a 15-state hidden Markov model (a larger state model resulted in redundant states) using chromosome 1 data from both conditions. Parameter learning was re-run five times using different random seeds to assess convergence. Third, a whole-genome segmentation was produced for each condition by running the obtained model on the remaining chromosomes. Each state from the obtained model was manually annotated, and states indicating the presence of promoter or enhancer chromatin tags were selected (E4-E11, Supplementary Fig. 6). Overlapping promoter or enhancer regions in non-activated and activated genome segmentations were merged to create a CD4+ T cell regulatory annotation. Thus, we defined 53,534 regulatory regions (Supplementary Table 8).
RNA sequencing
Total RNA was isolated using the RNeasy kit (Qiagen) and the concentrations and integrity were quantified using Bioanalyzer (Agilent); all samples reached RINs > 9.8. Two pools of RNA were generated from three and four donors and for each experimental condition. cDNA libraries were prepared from 1ug total RNA using the stranded NEBNext Ultra Directional RNA kit (New England Biolabs), and sequenced on HiSeq (Illumina) at an average coverage of 38 million paired-end reads/sample. RNA sequencing reads were trimmed to remove traces of library adapters by matching each read with a library of contaminants using Cutadapt70, a semi-global alignment algorithm. Owing to our interest in detecting functional enhancers, which constitute transcription units on their own right, we mapped reads to the human genome using STAR71, a splicing-aware aligner. This frees us from relying on a transcriptome annotation which would require exact boundaries and strand information for all features of interest, something not available in case of promoters and enhancers.
After alignment, we employed Samtools69 to discard reads with an unmapped pair, secondary alignments and low-quality alignments. The resulting read dataset, with an average of 33 million paired-end reads/sample, was sorted and indexed. We used FastQC (v0.11.3, http://www.bioinformatics.babraham.ac.uk/projects/fastqc/) to ensure all samples had regular GC content (sum of deviations from normal includes less than 15% of reads), base content per position (difference A vs T and G vs C less than 10% at all positions) and kmer counts (no imbalance of kmers with p < 0.01) as defined by the tool. We augmented Ensembl 75 gene annotations with regulatory region definitions obtained from our ChIP-seq analysis described above, and defined them as present in both genome strands due to their bidirectional transcription potential. For each RNA-seq sample, we quantified expression of genomic and regulatory features in a two-step strand-aware process using HTSeq72. For each gene we counted the number of reads that fell exactly within its exonic regions, and did not map to other genomic elements. For each regulatory feature we counted the number of reads that fell exactly within its defined boundaries, and did not map to other genomic or regulatory elements.
By construction, this quantification scheme counts each read at most once towards at most one feature. Furthermore, strand information is essential to be able to assign reads to features in regions with overlapping annotations. For example, distinguishing intronic eRNAs from pre-mRNA requires reads originating from regulatory activity in the opposite strand from the gene.
Feature counts were transformed into a matrix, and a trimmed mean of M-values normalization was applied to account for library size differences, plus a filter to discard features below an expression threshold of < 0.4 counts per million mapped reads in at least two samples, a rather low cutoff, to allow for regulatory RNAs to enter differential expression calculations. This threshold equates to approximately 15 reads, given our mapped library sizes of ~35 million paired-end reads. A voom normalization was applied to log-transformed counts in order to estimate precision weights per gene, and differential expression estimates were obtained after fitting a linear model on a paired design, using the limma Bioconductor R package65. There was a strong correlation (rho=0.81) between microarray and RNA-seq fold change estimates at 4 hours.
Comparison of regRNAs to FANTOM CAGE data
We compared expressed regRNA regions detected in our non-activated CD4+ T cell samples versus those found using CAGE-seq by the FANTOM5 Consortium. RNA-seq, using a regulatory reference obtained from chromatin states, yields 17,175 features expressed with at least 0.4 counts per million in both non-activated CD4+ T cell samples. Among those, 3,897 correspond to regulatory regions. Unstimulated CD4+ samples from FANTOM5 (http://fantom.gsc.riken.jp/5/datafiles/latest/basic/human.primary_cell.hCAGE/, samples 10853, 11955 and 11998) contain 266,710 loci expressed (with at least one read) in all 3 samples.
We found 13,178 of our 17,175 expressed CD4+ T cell features overlap expressed loci in CAGE data (77%). Conversely, 243,596/266,710 CAGE loci overlap CD4+ T cell features (91%). Similarly 2,888/3,897 expressed regRNAs overlap expressed loci in CAGE data (74%).
Comparison of PCHi-C and ChlA-PET interactions
We downloaded supplementary table 1 from http://www.nature.com/cr/ioumal/v22/n3/extref/cr201215xl.xlsx13 and counted the overlaps of PCHi-C interactions from CD4+ T cells and comparitor cells (megakaryoctyes and erythroblasts) in distance bins. R code to replicate the analysis is at https://github.com/chr1swallace/cd4-pchic/blob/master/chepelev.R. Calling interactions requires correction for the expected higher density of random collisions at shorter distances73 which are explicitly modelled by CHICAGO12 used in this study but not in the ChlA-PET study13. As a result, we expected a higher false positive rate from the ChlA-PET data at shorter distances.
Regression of gene expression against PIR count and eRNA expression
We related measures of gene expression (absolute log2 counts or log2 fold change) to numbers of PIRs or numbers of PIRs overlapping specific features using linear regression. We used logistic regression to relate agreement between fold change direction at PCHi-C linked protein coding genes and eRNAs. We used robust clustered variance estimates to account for the shared baits for some interactions across genes with the same prey. Enrichment of chromatin marks in interacting baits and prey were assessed by logistic regression modelling of a binary outcome variable (fragment overlapped specific chromatin peak) against a fragment width and a categorical explanatory variable (whether the HindIII fragment was a bait or prey and the cell state the interaction was identified in), using block bootstrapping of baited fragments (https://github.com/chr1swallace/genomic.autocorr) to account for spatial correlation between neighbouring fragments.
GWAS summary statistics
We used a compendium of 31 GWAS datasets 34 (Supplementary Table 5). Briefly we downloaded publicly available GWAS statistics for 31 traits. Where necessary we used the liftOver utility to convert these to GRCh37 coordinates. To remove spurious association signals, we removed variants with P< 5 × 10 -8 for which there were no variants in LD (r2>0.6 using 1000 genomes EUR cohort as a reference genotype panel) or within 50 kb with P<10−5. We masked the MHC region (GRCh37:chr6:25-35Mb) from all downstream analysis due to its extended LD and known strong and complex associations with autoimmune diseases.
Comparison of GWAS data and PIRs requires dense genotyping coverage. For GWAS which did not include summary statistics imputed for non-genotyped SNPs, we used a poor man’s imputation (PMI) method34 to impute. We imputed p values at ungenotyped variants from 1000 Genomes EUR phase 3 by replacing missing values with those of their nearest proxy variant with r2>0.6, if one existed. Variants that were included in the study but did not map to the reference genotype set were also discarded.
To calculate posterior probabilities that each SNP is causal under a single causal variant assumption, we divided the genome into linkage disequilibrium blocks of 1cM based on data from the HapMap project (http://hapmap.ncbi.nlm.nih.gov/downloads/recombination/2011-01_phaseII_B37/). For each region excluding the MHC we used code modified from Giambartolomei et al.74 to compute approximate Bayes factors for each variant using the Wakefield approximation75, and thus posterior probabilities that each variant was causal as previously proposed76.
Testing of the enrichment of GWAS summary statistics in PIRs using blockshifter
We used the blockshifter method34 (https://github.com/ollyburren/CHIGP) to test for a difference between variant posterior probability distributions in HindIII fragments with interactions identified in test and control cell types using the mean posterior probability as a measure of central location. Blockshifter controls for correlation within the GWAS data due to LD and interaction restriction fragment block structure by employing a rotating label technique similar to that described in GoShifter77 to generate an empirical distribution of the difference in means under the null hypothesis of equal means in the test and control set. Runs of one or more PIRs (separated by at most one HindIII fragment) are combined into ‘blocks’, that are labeled unmixed (either test or control PIRs) or mixed (block contains both test and control PIRs). Unmixed blocks are permuted in a standard fashion by reassigning either test or control labels randomly, taking into account the number of blocks in the observed sets. Mixed blocks are permuted by conceptually circularising each block and rotating the labels. A key parameter is the gap size - the number of non-interacting HindIII fragments allowed within a single block, with larger gaps allowing for more extended correlation.
We used simulation to characterise the type 1 error and power of blockshifter under different conditions and to select an optimal gap size. Firstly, from the Javierre et al. dataset34 we selected a test (Activated or Non Activated CD4+ T Cells) and control (Megakaryocyte or Erythroblast) set of PIRs with a CHiCAGO score > 5, as a reference set for blockshifter input.
Using the European 1000 genomes reference panel, we simulated GWAS summary statistics, under different scenarios of GWAS/PIR enrichment. We split chromosome 1 into 1cM LD blocks and used reference genotypes to compute a covariance matrix for variants with minor allele frequency above 1%, Σ. GWAS Z scores can be simulated as multivariate normal with mean μ and variance Σ78. Each block may contain no causal variants (GWASnull, μ = 0) or one (GWASalt). For GWASalt blocks, we pick a single causal variant, i, and calculate the expected non-centrality parameter (NCP) for a 1 degree of freedom chi-square test of association at this variant and its neighbours. This framework is natural because the NCP at any variant j can be expressed as the NCP at the causal variant multiplied by the r2 between variants i and j79. In each case we set the NCP at the causal variant to 80 to ensure that each causal variant was genome-wide significant (P < 5 × 10−8). μ is defined as the square root of this constructed NCP vector.
For all scenarios we randomly chose 50 GWASalt blocks leaving the remaining 219 GWASnull. Enrichment is determined by the preferential location of simulated causal variants within test PIRs. In all scenarios, each causal variant has a 50% chance of lying within a PIR, to mirror a real GWAS in which we expect only a proportion of causal variants to be regulatory in any given cell type. Under the enrichment-null scenario, used to confirm control of type 1 error rate, the remaining variants were assigned to PIRs without regard for whether they were identified in test or control tissues. To examine power, we considered two different scenarios with PIR-localised causal variants chosen to be located specifically in test PIRs with either 50% probability, scenario power (1), or 100%, scenario power (2). Note that a PIR from the test set may also be in the control set, thus, as with a real GWAS, not all causal variants will be informative for this test of enrichment.
For each scenario we further considered variable levels of genotyping density, corresponding to full genotyping (everything in 1000 Genomes), HapMap imputation (the subset of SNPs also in Stahl et al. REF dataset) or genotyping array (the subset of SNPs also on the Illumina 550k array). Where genotyping density is less than full, we used our proposed poor man’s imputation (PMI) strategy to fill in Z scores for missing SNPs.
We ran blockshifter, with 1000 null permutations, for each scenario and PMI condition for 4000 simulated GWAS, with a blockshifter superblock gap size parameter (the number of contiguous non-PIR HindIII fragments allowed within one superblock) of between 1 and 20 and supplying numbers of cases and controls from the RA dataset48.
For comparison we also investigated the behaviour of a naive test for enrichment for the null scenario. We computed a 2x2 table variants according to test and control PIR overlap, and whether a variant’s posterior probability of causality exceeded an arbitrary threshold of 0.01, and Fisher’s exact test to test for enrichment.
Enrichment of GWAS summary statistics in CD4+ and activated CD4+ PIRs
We compared the following sets using all GWAS summary statistics, with a superblock gap size of 5 (obtained from simulations above) and 10,000 permutations under the null:-
Total CD4+ Activated + Total CD4+ NonActivated (test) versus Endothelial precursors + Megakaryocytes (control)
Total CD4+ Activated (test) versus Total CD4= NonActivated (control).
Variant posterior probabilities of inclusion, full genotype data (ImmunoChip)
We carried out formal imputation to 1000 Genomes Project EUR data using IMPUTE2 80 and fine-mapped causal variants in each of the 179 regions where a minimum p < 0.0001 was observed using a stochastic search method which allows for multiple causal variants in a region, (https://github.com/chr1swallace/GUESSFM)22. The posterior probabilities for models that contained variants which overlapped PIRs for each gene were aggregated to compute PIR-level marginal posterior probabilities of inclusion.
Variant posterior probabilities of inclusion, summary statistics
Where we have only summary statistics of GWAS data already imputed to 1000 Genomes, we divided the genome into linkage disequilibrium blocks of 0.1cM based on data from the HapMap project (http://hapmap.ncbi.nlm.nih.gov/downloads/recombination/2011-01_phaseII_B37/). For each region excluding the MHC we use code modified from Giambartolomei et al.74 to compute approximate Bayes factors for each variant using the Wakefield approximation75, and thus posterior probabilities that each variant was causal assuming at most one causal variant per region as previously proposed76.
Computation of gene prioritisation scores
We used the COGS method34 (https://github.com/ollyburren/CHIGP) to prioritise genes for further analysis. We assign variants to the first of the following three categories it overlaps for each annotated gene, if any
coding variant: the variant overlaps the location of a coding variant for the target gene.
promoter variant: the variant lies in a region baited for the target gene or adjacent restriction fragment.
PIR variant: the variant lies in a region overlapping any PIR interacting with the target gene. We produced combined gene/category scores by aggregating, within LD blocks, over models with a variant in a given set of PIRs (interacting regions), or over HindIII fragments baited for the gene promoter and immediate neighbours (promoter regions), or over coding variants to generate marginal probabilities of inclusion (MPPI) for each hypothesised group. We combine these probabilities across LD blocks, i, using standard rules of probability to approximate the posterior probability that at least one LD block contains a causal variant:
Thus the score takes a value between 0 and 1, with 1 indicating the strongest support. We report all results with score > 0.01 in Supplementary Table 7, but focus in this manuscript on the subset with scores > 0.5.
Because COGS aggregates over multiple signals, a gene may be prioritised because of many weak signals or few strong signals in interacting regions. To predict the expected information for future users of this method, we considered the subset of 76 input regions with genome-wide significant signals (p<5×10−8) in ImmunoChip datasets. We prioritised at least one gene with a COGS score > 0.5 in 35 regions, with a median of three genes/region (interquartile range, IQR = 1.5-4). Equivalent analysis of the genome-wide significant GWAS signals prioritised a median of two genes/region (interquartile range = 1-3). This suggests that this algorithm might be expected to prioritise at least one gene in about half the genomewide significant regions input when run on a relevant cell type.
Whilst components 1 and 2 are fixed for a given gene and trait the contribution of variants overlapping PIRs varies depending on the tissue context being examined. We developed a hierarchical heuristic method to ascertain for each target gene which was the mostly likely component and cell state. Firstly for each gene we compute the gene score due to genic effects (components 1 + 2) and interactions (component 3) using all available tissue interactions for that gene. We use the ratio of gene effects score to interactions score in a similar manner to a Bayes factor to decide whether one is more likely. If gene effect is more likely (gene.score ratio >3) we iterate and compare if the gene score due to coding variants (component 1) is more likely than for promoter variants (component 2). Similarly if an interaction is more likely we compare interaction gene scores for activated vs non-activated cells. If at any stage no branch is substantially preferred over its competitor (ratio of gene scores < 3) we return the previous set as most likely, otherwise we continue until a single cell state/set is chosen. In this way we can prioritize genes based on the overall score and label as to a likely mechanism for candidate causal variants.
Allele-specific expression assays
Total CD4+ T cells were isolated from five donors and activated as described above and were harvested after 0, 2 and 4 hours in RLT Plus buffer. Selected donors were heterozygous at all eight group A SNP and, homozygous for group C and F SNPs. Two and three of the donors were homozygous for the group D and E SNP groups, respectively (Supplementary Table 9). Memory CD4+ T cells were sorted from cryopreserved PBMC as viable, αβ TCR+, CD4+, CD45RA−, CD127+. CD27+ cells using a FACSAria III cell sorter (BD Biosciences). Sorted cells were either activated for 4 hours in culture as described above or resuspended directly in RLT plus buffer post-sort. Total RNA was extracted using Qiagen RNeasy Micro plus kit and cDNA was synthesised using Superscript III reverse transcriptase (Thermo Fisher) according to manufacturer’s instructions. To perform allele-expression experiments we used a modified version of a previously described method for quantifying methylation in bisulfite sequence data81. A two-stage PCR was used, the first round primers were designed to flank the variant of interest using Primer3 (http://bioinfo.ut.ee/primer3-0.4.0/primer3/) and adaptor sequences were added to the primers (Sigma), shown as lowercase letters (rs61839660_ASE_F tgtaaaacgacggccagtGCACACACCTATCCTAGCCT, rs61839660_ASE_R caggaaacagctatgaccCCCACAGAATCACCCACTCT, product size 114bp; rs12244380_ASE_F tgtaaaacgacggccagtTTCGTGGGAGTTGAGAGTGG, rs12244380_ASE_R caggaaacagctatgaccTTAAAAGAGTTCGCTGGGCC, product size 180bp; rs12722495_ASE_F tgtaaaacgacggccagtGTGAGTTTCAATCCTAAGTGCGA, rs12722495_ASE_R caggaaacagctatgaccATTAAGCGGACTCTCTGGGG, product size 97bp). The first round PCR contains 10 μl of Qiagen multiplex PCR mastermix, 0.5 μl of 10 nmol forward primer, 0.5 μl of 10 nmol reverse primer, 4 μl of cDNA and made up to 20 μl with ultra-pure water. The PCR cycling conditions were 95°C for 15 minutes hot start, followed by 30 cycles of the following steps: 95°C for 30 seconds, 60°C for 90 seconds and 72°C for 60 seconds, finishing with a 72°C for 10 minutes cycle. The first round PCR product was cleaned using AmpureXP beads (Beckman Coulter) according to manufacturer’s instructions. To add Illumina sequence compatible ends to the individual first round PCR amplicons, additional primers were designed to incorporate P1 and A sequences plus sample-specific index sequences in the A primer, through hybridisation to adapter sequence present on the first round gene-specific primers. Index sequences are as published81. The second-round PCR contained 8 μl of Qiagen multiplex PCR mastermix, 2.0 μl of ultra-pure water, 0.35 μl of each forward and reverse index primer, 5.3 μl of Ampure XP-cleaned first-round PCR product. The PCR cycling conditions were 95°C for 15 minutes hotstart, followed by 7 cycles of the following steps: 95°C for 30 seconds, 56°C for 90 seconds, 72°C for 60 seconds, finishing with 72°C for 10 minutes cycle. All PCR products were pooled at equimolar concentrations based on quantification on the Shimadzu Multina. AmpureXP beads were used to remove unincorporated primers from the product pool. We used the Kapa Bioscience library quantification kit to accurately quantify the library according to manufacturer’s instructions before sequencing on an Illumina MiSeq v3 reagents (2 x 300 bp reads).
Statistical analysis of allele-specific expression data
Sequence data was processed using the Methpup package (https://github.com/ollyburren/Methpup) to extract counts of each allele at rs12722495, and rs12244380 (Supplementary Table 10). Individuals were part of a larger cohort genotyped on the ImmunoChip and were phased using snphap (https://github.com/chr1swallace/snphap") to confirm which allele at each SNP was carried on the same chromosome as A2=rs12722495:C or A1=rs12722495:T. Allelic imbalance was quantified as the ratio A2/A1 and was averaged across replicates within individuals using a geometric mean. Allelic ratios in cDNA and gDNA were compared using Wilcoxon rank sum tests. P values are shown in Fig. 6b and Supplementary Fig. 13. Full details are in https://github.com/chr1swallace/cd4-pchic/blob/master/IL2RA-ASE.R.
Supplementary Information is linked to the online version of the paper at www.nature.com/nature.
Author Contributions
Study conceived by: CW, JAT, LSW, PF, and led by CW. Interpreted the data: CW, OSB, AJC, ARG, DR, LSW, JAT. Sequence data analysis: ARG. HiCUP analysis: SW. ASE experiments and analysis: DR. Microarray experiments and analysis: XCD,RCF, RC, CW. Statistical analysis: OSB, JC, NJC, CW, ARG. Laboratory experiments: AJC, BJ, DR, JJL, FB, SPR, KD. Wrote the paper: CW, OSB, AJC, ARG and contributed to writing: JAT, LSW, MS. Revised the paper: all authors. Genetic association data processing: CW, OSB, ES. Supervised capture Hi-C experiments: MS and PF. Supervised cell experiments: AJC, MF, WO, PF, JAT and LW.
Data availability
The following datasets were generated:
PCHi-C data are available as raw sequencing reads (submission to EGA in progress) or CHICAGO-called interactions (Supplementary Table 4) and are available for interactive exploration via http://www.chicp.org
RNA-seq and ChIPseq data are available as raw sequencing reads (submission to EGA in progress).
Microarray data are available at ArrayExpress, https://www.ebi.ac.uk/arrayexpress, accession number E-MTAB-4832
Processed datasets are available as Supplementary Tables
Code used to analyse the data are available from https://github.com/chr1swallace/cd4-pchic except where other URLs are indicated in Methods
Reprints and permissions information is available at www.nature.com/reprints
Competing financial interests
The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to cew54@cam.ac.uk.
Acknowledgements
This work was funded by the JDRF (9-2011-253), the Wellcome Trust (089989, 091157, 107881), the UK Medical Research Council (MR/L007150/1, MC_UP_1302/5), the UK Biotechnology and Biological Sciences Research Council (BB/J004480/1) and the National Institute for Health Research (NIHR) Cambridge Biomedical Research Centre. The research leading to these results has received funding from the European Union’s 7th Framework Programme (FP7/2007-2013) under grant agreement no.241447 (NAIMIT). The Cambridge Institute for Medical Research (CIMR) is in receipt of a Wellcome Trust Strategic Award (100140).
We thank all study participants and family members.
We thank the Wellcome Trust for funding the AITD UK national collection; all doctors and nurses in Birmingham, Bournemouth, Cambridge, Cardiff, Exeter, Leeds, Newcastle and Sheffield for recruitment of patients and J. Franklyn, S. Pearce (Newcastle) and P. Newby (Birmingham) for preparing and providing DNA samples on Graves’ disease patients.
This research utilizes resources provided by the Type 1 Diabetes Genetics Consortium, a collaborative clinical study sponsored by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK), National Institute of Allergy and Infectious Diseases (NIAID), National Human Genome Research Institute (NHGRI), National Institute of Child Health and Human Development (NICHD), and JDRF and supported by U01 DK062418.
We gratefully acknowledge the participation of all Cambridge NIHR BioResource volunteers, and thank the Cambridge BioResource staff for their help with volunteer recruitment. We thank the National Institute for Health Research Cambridge Biomedical Research Centre and NHS Blood and Transplant for funding. Further information can be found at www.cambridgebioresource.org.uk
We thank the High-Throughput Genomics Group at the Wellcome Trust Centre for Human Genetics (funded by Wellcome Trust grant reference 090532/Z/09/Z) for the generation of the sequencing data.
We thank Stephen Eyre for helpful comments on the manuscript, and N. Soranzo and the HaemGen consortium for sharing blood trait GWAS summary statistics.
The authors acknowledge the assistance and support of the National Institute for Health Research (NIHR) Cambridge Biomedical Research Centre. Helen Stevens, Meeta Maisuria-Armer, Pamela Clarke, Gillian Coleman, Sarah Dawson, Simon Duley, Jennifer Denesha and Trupti Mistry for sample processing. Judy Brown, Lynne Adshead, Amie Ashley, Anna Simpson and Niall Taylor for laboratory administration and procurement support. Vin Everett and Sundeep Nanuwa for logistical and web development.
We thank investigators of published ImmunoChip studies for making available their raw genotyping data (David van Heel, celiac disease; Stephen Eyre, rheumatoid arthritis; Matthew Simmonds, Stephen Gough, Jayne Franklyn, and Oliver Brand, autoimmune thyroid disease).
References
- 1.↵
- 2.↵
- 3.↵
- 4.↵
- 5.↵
- 6.↵
- 7.↵
- 8.↵
- 9.↵
- 10.↵
- 11.↵
- 12.↵
- 13.↵
- 14.↵
- 15.↵
- 16.↵
- 17.↵
- 18.↵
- 19.↵
- 20.↵
- 21.↵
- 22.↵
- 23.↵
- 24.↵
- 25.↵
- 26.↵
- 27.↵
- 28.↵
- 29.↵
- 30.↵
- 31.↵
- 32.↵
- 33.↵
- 34.↵
- 35.↵
- 36.
- 37.↵
- 38.↵
- 39.↵
- 40.↵
- 41.↵
- 42.↵
- 43.↵
- 44.↵
- 45.↵
- 46.↵
- 47.↵
- 48.↵
- 49.↵
- 50.↵
- 51.↵
- 52.↵
- 53.
- 54.↵
- 55.↵
- 56.↵
- 57.↵
- 58.↵
- 59.↵
- 60.↵
- 61.↵
- 62.↵
- 63.↵
- 64.↵
- 65.↵
- 66.↵
- 67.↵
- 68.↵
- 69.↵
- 70.↵
- 71.↵
- 72.↵
- 73.↵
- 74.↵
- 75.↵
- 76.↵
- 77.↵
- 78.↵
- 79.↵
- 80.↵
- 81.↵