After quality filtering, 61,427 participants (Supplementary Table 1 and Supplementary Fig. 1) and 140,333 genotyped ImmunoChip variants (Online Methods) were included in analyses, providing dense coverage in 188 autosomal regions (“ImmunoChip regions”)²⁴ and sparse genotyping in other regions (Supplementary Tables 2 and 3). Each participant was assigned to one of five ancestry groups using principal component analysis (Online Methods, Supplementary Fig. 2): European (EUR, n = 47,319), African Admixed (AFR, n = 4,290), Finnish (FIN, n = 6,991), East Asian (EAS, n = 588) and Other Admixed (AMR, n = 2,239). Association analyses included 16,159 T1D cases, 25,386 controls and 6,143 trio families (i.e., an affected child and both parents) (Supplementary Tables 4 and 5 and Supplementary Fig. 3). Genotypes at additional variants were imputed using the Trans-Omics for Precision Medicine (TOPMed)²³ multi-ethnic reference panel to improve discovery and fine-mapping resolution (Online Methods). After imputation, the number of variants in ImmunoChip regions with imputation R-squared > 0.8 and minor allele frequency (MAF) > 0.005 in each ancestry group was 166,274 (EUR), 322,084 (AFR), 163,612 (FIN), 137,730 (EAS), and 188,550 (AMR). We compared imputed genotypes to whole genome sequencing data from a subset of individuals and observed high concordance (Online Methods, Supplementary Note, Supplementary Figs. 4 and 5).

Initially, we analyzed unrelated cases and controls (n = 41,545), assuming an additive inheritance model. With minimal evidence of artificial inflation of association statistics due to population structure (Supplementary Note, Supplementary Fig. 6, and Supplementary Table 6), we identified 64 T1D-associated regions outside the major histocompatibility complex (MHC, including the HLA loci), including 24 regions associated with T1D at genome-wide significance (P < 5 × 10⁻⁸) for the first time. Following conditional analysis, 78 independent associations were identified (P < 5 × 10⁻⁸; Supplementary Table 7). On the X chromosome, the most T1D-associated variant was rs4326559 (A>C, C allele OR = 1.09, P = 4.5 × 10⁻⁷).

We extended the discovery analysis to incorporate T1D trio families (n = 6,143 trios, some trio families were multiplex and analyzed as multiple trios; Online Methods). Meta-analysis of case-control and trio results identified 78 chromosome regions associated with T1D (P < 5 × 10⁻⁸), including 42/43 chromosome regions previously identified in an ImmunoChip-based study³ (rs4849135 (G>T) was P = 2.93 × 10⁻⁷). When comparing these 78 regions to previous T1D studies^{1–3,15–21}, 36 novel regions associated with T1D at genome-wide significance for the first time (Table 1). In the remaining 42 regions, the lead variant was within 250 kb of the lead variant in a previous T1D study. The 1q21.3 region, which contains the gene encoding the interleukin-6 receptor (IL-6R), was among the regions associated with T1D at genome-wide significance for the first time. Interestingly, the lead variant in this region was rs2229238 (NC_000001.11:g.154465420T>C, P = 3.02 × 10⁻⁹), not the nonsynonymous variant rs2228145 (NC_000001.11:g.154454494A>C; NP_000556.1:p.Asp358Ala; P = 2.20 × 10⁻⁴), which was previously suggested to be causal for T1D in targeted analysis²⁵ and remains a candidate causal variant for rheumatoid arthritis²⁶.

Applying the Benjamini-Yekutieli false discovery rate (FDR) < 0.01²⁷ to assess statistical significance, 143 regions were associated with T1D (Supplementary Table 8). Their lead variants overlapped substantially with lead variants for 14 immune-mediated diseases from published studies, but the direction of effects frequently differed between traits (Supplementary Fig. 7). Associated variants with FDR < 0.01 but not meeting genome-wide significance (P < 5 × 10⁻⁸) had smaller absolute effect sizes but similar MAFs to those satisfying genome-wide significance (median (IQR) OR = 1.07 (1.06, 1.09) vs. 1.11 (1.09, 1.13); median (IQR) MAF = 0.301 (0.152, 0.397) vs. 0.306 (0.184, 0.374)). These results indicate that remaining regions associated with T1D may have increasingly smaller effect sizes (Supplementary Fig. 8), requiring genome-wide coverage and larger sample sizes for detection.

One exception underscores the need for inclusion of understudied populations to enhance biological insight, even with limited sample sizes, and suggests the potential value of considering alternative metrics for defining statistical significance in genetic studies²⁸. On chromosome 1p22.1 near the Metal Response Element Binding Transcription Factor 2 (MTF2) gene, rs190514104 (NC_000001.11:g.93145882G>A) had a large effect on T1D risk (OR (95% CI) = 2.9 (1.9–4.5); P = 6.6 × 10⁻⁷) in the AFR ancestry group. The minor allele (A) at rs190514104:G>A was common in the AFR ancestry group (> 1%) but rare in the others (< 0.1%). Considering the limited sample size, potential heterogeneity of the AFR cohort, and possible over-estimation of effect sizes due to “the winner’s curse”, this association requires replication in an independent cohort.

Use of recessive and dominant models of inheritance identified 35 regions (25 dominant, 10 recessive) with a better fit than the additive model (lower Akaike Information Criterion (AIC) in Europeans) at FDR < 0.01, including nine regions that did not reach FDR < 0.01 under the additive model (Supplementary Table 9). Thus, a total of 152 regions were associated with T1D at FDR < 0.01, 143 under an additive model and nine under recessive or dominant models.

To define the local architecture of T1D regions, we applied a Bayesian stochastic search method (GUESSFM²⁹) to the European ancestry case-control data (Online Methods, Statistical fine mapping). Of 52 ImmunoChip regions (Supplementary Table 2) associated with T1D, GUESSFM predicted 21 (40%) to contain more than one causal variant (Fig. 1a), compared to nine regions using stepwise conditional regression. In four regions, the lead variant in the discovery analysis was not prioritized by fine mapping (posterior probability < 0.5): 2q33.2 (CTLA4), 4q27 (IL2), 14q32.2 (MEG3) and 21q22.3 (UBASH3A). In these regions, the lead variant likely tags two or more T1D-associated haplotypes that can be identified using GUESSFM but not stepwise logistic regression, a phenomenon observed previously^29,30. For example, although stepwise regression analysis in the UBASH3A locus supported a single causal variant (Supplementary Table 7), GUESSFM fine mapping and haplotype analyses indicated that the lead variant in this region, rs11203203 (NC_000021.9:g.42416077G>A), is unlikely to be causal. GUESSFM fine mapping supported a three-variant model (rs9984852 (NC_000021.9:g.42408836T>C), rs13048049 (NC_000021.9:g.42418534G>A) and rs7276555 (NC_000021.9:g.42419803T>C)) (Fig. 1b), which had a better fit than the single variant model (AIC 45073 vs. 45138, Fig. 1c). Haplotype analysis (Online Methods) demonstrated that when rs11203203:G>A is present without the GUESSFM-prioritized variants, there is no effect of rs11203203:G>A on T1D risk (Fig. 1d). Resampling experiments consistently supported two or more causal variants in the region, with at least one of the three GUESSFM-prioritized variants more likely to be causal than rs11203203:G>A (Supplementary Table 10). Given the complexity of association in the UBASH3A region, and likely at many loci, statistical methods designed to use univariable summary statistics alone are not sufficient to explore the genetic architecture of T1D. We provide the comprehensive list of T1D credible variants and haplotype analyses for all 52 fine-mapped regions (Supplementary Table 11, https://github.com/ccrobertson/t1d-immunochip-2020).

Differences in linkage disequilibrium (LD) between ancestry groups can be advantageous in prioritizing causal variants³¹. In the 30 regions where analysis suggested a single causal variant, we performed multi-ethnic fine-mapping using PAINTOR³². Eight regions identified an associated variant (P < 5 × 10⁻⁴) in more than one ancestry group: five with associations in EUR and FIN, and three with associations in EUR and AFR. In three regions, the number of variants prioritized was markedly reduced by including multiple ancestry groups: 4p15.2 (RBPJ), 6q22.32 (CENPW) and 18q22.2 (CD226) (Fig. 2a, Extended Data Figs. 1 and 2, and Supplementary Table 12). In the chromosome 4p15.2 (RBPJ) region, the credible set from EUR ancestry contained 24 variants. In contrast, using PAINTOR with EUR and AFR summary statistics, only five variants were prioritized with a posterior probability > 0.1 (Fig. 2a). Among these prioritized variants, rs34185821 (NC_000004.12:g.26083858A>G) and rs35944082 (NC_000004.12:g.26093692A>G), both located in the non-coding transcript LINC02357, have the potential to disrupt multiple transcription factor binding motifs³³. rs35944082:A>G also overlaps open chromatin in multiple adaptive immune cell types (Fig. 2b) and resides in a FANTOM enhancer site³⁴. Further, rs34185821:A>G is one of three prioritized variants flanking an activation-dependent Assay for Transposase-Accessible Chromatin using sequencing (ATAC-seq) peak in lymphocytes and a stable response element in human islets³⁵, with potential to perturb an extended TATA box motif³⁶.

Only 34/2,732 (1.2%) credible variants (group posterior probability > 0.5) were protein-altering (nonsynonymous, frameshift, stop-gain, or splice-altering) with 12 having support for a role in T1D (Online Methods, Supplementary Table 13). We identified several previously unreported protein-altering variants as highly prioritized in T1D credible sets (posterior probability > 0.1): a protective missense variant in UBASH3A, rs13048049:G>A (NP_061834.1:p.Arg324Gln; OR = 0.84; AF_EUR = 0.051); two low-frequency splice donor variants in IFIH1, rs35732034 (NC_000002.12:g.162268086C>T; OR = 0.63; AF_EUR = 0.0089) and rs35337543 (NC_000002.12:g.162279995C>G; OR = 0.61; AF_EUR = 0.0099); and a missense variant in CTLA4, rs231775 (NC_000002.12:g.203867991A>G; NP_001032720.1:p.Thr17Ala; OR = 1.20; AF_EUR = 0.36).

ATAC-seq offers a high-resolution map of accessible chromatin with potential regulatory function³⁷. Using publicly available^38–40 and newly generated ATAC-seq data from healthy donors, we assessed enrichment (Online Methods) of 2,431 T1D credible variants (group posterior probability > 0.8) in accessible chromatin across diverse immune and non-immune cell types (including 25 primary immune cell types, pancreatic islets, and, as control cell types unlikely to be central to T1D etiology, fetal and adult cardiac fibroblasts). T1D credible variants were enriched in open chromatin in multiple primary immune cell types based on two complementary enrichment analysis approaches (Online Methods, Supplementary Fig. 9), with strong enrichment observed in stimulated CD4⁺ effector T cells (Supplementary Fig. 9b). There was no enrichment in pancreatic islets (P = 0.14), the primary target of autoimmunity in T1D, even after exposure to proinflammatory cytokines (P = 0.05) or in cardiac fibroblasts (P > 0.60) (Supplementary Fig. 9). We also examined enrichment for T1D credible variants in condition-specific accessible chromatin and observed the largest enrichment in stimulation-specific peaks from effector CD4⁺ T cells (Supplementary Note, Supplementary Table 14, and Supplementary Fig. 10).

Chromatin accessibility profiles were generated across 115 participants (n_EUR = 48, n_AFR = 67) in primary CD4⁺ T cells, the cell type in which accessible chromatin is most strongly enriched for T1D credible variants (Supplementary Figs. 9 and 10). We examined additive effects of genotype on local chromatin accessibility (cis window < 1 Mb), identifying 11 “peaks” of chromatin accessibility significantly (P < 5 × 10⁻⁵) associated with T1D credible variants. Colocalization analysis of T1D association and caQTLs (R package coloc⁴¹, Online Methods) identified five regions supporting a common causal variant underlying association with T1D and chromatin accessibility (PP.H4.abf > 0.8; Table 2). In all five regions, at least one T1D credible variant overlapped the caQTL-associated peak. Six of these “within-peak” credible variants were directly genotyped on the ImmunoChip, allowing us to examine allele-specific accessibility in heterozygous participants (Online Methods). At all six variants, the proportion of ATAC-seq reads from heterozygotes containing the alternative allele was consistent with the direction of the caQTL effect (Supplementary Table 15). When integrated with whole blood cis-eQTLs^41,42, colocalization identified T1D candidate genes in four of five T1D-caQTL regions (PP.H4.abf > 0.8; Table 2).

Fine mapping of the BACH2 locus refined the T1D association to two intronic variants, rs72928038 (NC_000006.12:g.90267049G>A) and rs6908626 (NC_000006.12:g.90296024G>T) (Fig. 3a). The EUR minor alleles of rs72928038:G>A and rs6908626:G>T are associated with increased T1D risk (OR = 1.18; P < 1 × 10⁻²⁰, MAF_EUR = 0.18). Chromatin-state annotations across cell types from the BLUEPRINT Consortium and NIH Roadmap Epigenomics Project annotate rs72928038:G>A as overlapping a T cell-specific active enhancer and rs6908626:G>T as lying in the ubiquitous BACH2 promoter (Fig. 3b). Promoter-capture Hi-C data from diverse immune cell types⁴³ indicates that the enhancer region containing rs72928038:G>A contacts the BACH2 promoter in T cells (Fig. 3c). Although weak interactions were observed in multiple T cell subtypes, only naïve CD4⁺ T cells had a significant interaction score.

In caQTL analysis, rs72928038:G>A is associated with decreased accessibility of the enhancer it overlaps (chr6:90266766–90267715) (Fig. 3d, left), while rs6908626:G>T does not affect accessibility at the BACH2 promoter (chr6:90294665–90297341) (Fig. 3d, right). Similarly, among 14 subjects heterozygous for rs72928038:G>A, only 4% (5/121) of ATAC-seq reads overlapping that site contain the T1D risk allele (A) (Fig. 3e, left, and Supplementary Table 15), suggesting it leads to restricted accessibility. In contrast, chromatin accessibility at rs6908626:G>T does not exhibit allelic bias in heterozygotes (Fig. 3e, right). These data help to prioritize rs72928038:G>A, rather than rs6908626:G>T, as functionally relevant in CD4⁺ T cells.

In eQTL studies, rs72928038:G>A is associated with decreased expression of BACH2 in whole blood⁴² and purified immune cell types⁴⁴. In the DICE consortium⁴⁴, rs72928038:G>A is associated with decreased expression of BACH2 in multiple cell types, with the strongest effects in naïve CD4⁺ and CD8⁺ T cells. This result is consistent with the observation that the enhancer region overlapping rs72928038:G>A is accessible specifically in unstimulated bulk CD4⁺, unstimulated bulk CD8⁺, and naïve CD4⁺ T effector cells (Fig. 3f). Both the enhancer caQTL and BACH2 eQTL colocalize with T1D association (Fig. 3g and Table 2).

The BACH2 rs72928038:G>A variant overlaps binding sites for STAT1 and the ETS family of transcription factors, based on canonical transcription factor binding motifs³³. We performed super-shift electrophoretic mobility shift assay (EMSA) experiments of the DNA sequence flanking rs72928038:G>A that demonstrated allele-specific ETS1 binding, but no STAT1 binding (Supplementary Fig. 11). This result builds on experiments demonstrating allele-specific nuclear protein binding of rs72928038:G>A in Jurkat cells⁴⁵. These data prioritize rs72928038:G>A as a likely functional variant in T cells and provide preliminary support for a candidate regulatory mechanism underlying the 6q15 region association with T1D. Specifically, we hypothesize that the rs72928038:G>A minor allele (A) disrupts ETS1 binding, which leads to decreased enhancer activity and BACH2 expression in naïve CD4⁺ T cells.

To identify potential T1D therapeutic targets with human genetic support, we used the Priority Index (Pi) algorithm⁴⁶, which integrates genetic association results with genome annotations, regulatory maps, and protein-protein networks (Online Methods). Using improved T1D association statistics and additional eQTL resources from whole blood⁴², we identified 50 highly-ranked gene targets (Supplementary Table 16). These targets include 26 “seed genes” (implicated by T1D-associated loci through proximity, eQTL effects, or chromatin looping) and 24 non-seed genes (not in T1D regions but highly connected to T1D seed genes in immune protein networks). Although we excluded variants in the MHC region from algorithm input, the networks implicated by non-MHC seed genes led to prioritization of HLA-DRB1, an established T1D risk factor. Among the top 50 gene targets, 13 were not previously implicated by Pi analyses (STAT4, RGS1, CXCR6, IL23A, PTPN22, NFKB1, MAPK3, EPOR, DGKQ, GALT, IL12RB1, IL12RB2, IL6R)⁴⁶, while 12 have been targeted in clinical trials for autoimmune diseases (IL2RA, IL6ST, IL6R, TYK2, IFNAR2, JAK2, IL12B, IL23A, IL2RG, JAK3, JAK1 and IL2RB). T1D susceptibility alleles may alter expression of gene targets in either direction, and gene regulatory effects may be seen across multiple major immune cell populations or be restricted to a single cell type (Supplementary Fig. 12). For example, T1D risk alleles are associated with increased expression of MAPK3 and DGKQ but decreased expression of TYK2 across multiple major immune cell populations. In contrast, risk alleles decrease expression of RGS1 across most immune cell types but increase expression specifically in CD8⁺ T cells. The directionality and cell type-specificity of gene regulatory effects associated with T1D risk alleles may inform therapeutic target considerations.

DNA samples were genotyped on the Illumina ImmunoChip at University of Virginia (UVA) Genome Sciences Laboratory (n = 52,219), Sanger Institute (n = 4,347), University of Cambridge (n = 2,941), and Feinstein Institute (n = 1,811). Raw genotyping files were assembled at UVA. Genotype clusters were generated using the Illumina GeneTrain2 algorithm. Stringent SNP- and sample-level quality control filtering and data cleaning was performed to ensure high quality genotypes and accurate pedigrees (Supplementary Fig. 1). The following variant filters were applied: (1) re-annotated ImmunoChip variant positions by aligning probe sequences to GRCh37 and removed any variants with <100% match or multiple matches at different positions in the genome; (2) removed variants with call rates <98%; (3) removed variants with any discordance between duplicate or monozygotic twin samples, as confirmed by genotype-inferred relationships; (4) removed variants with Mendelian inconsistencies in >1% of informative trios or parent-offspring pairs, based on genotype-inferred relationships.

For sample filtering, we used X chromosome heterozygosity and Y chromosome missingness to identify and exclude participants with apparent sex chromosome anomalies or resolve inconsistencies with reported sex. Pedigree-defined and genotype-inferred sample relationships were compared using KING version 2.1.3⁶⁴. Samples were excluded when inconsistencies could not be resolved, including relationships between families, within and across cohorts. For each pair of related families observed, we randomly selected one to remove from association analysis. After resolving sex and relationship issues, samples with genotype call rate < 98% were removed. Variants with genotype frequencies deviating from Hardy-Weinberg Equilibrium (P < 5 × 10⁻⁵) in unrelated European ancestry controls were excluded before imputation.

Principal components (PC) were generated in 1000 Genomes phase 3 individuals using 8,297 autosomal ImmunoChip variants selected by excluding regions of long-range linkage disequilibrium (LD)⁶⁵, pruning for short-range LD (r² < 0.2 in 50-kb windows), and filtering for minor allele frequency (MAF) > 0.05). Participant genotypes were projected onto the 1000 Genomes PC space using PLINK v1.9⁶⁶. The first ten PCs were used in k-means clustering to define five clusters of ancestrally similar participants, European (EUR), African-American (AFR), East Asian (EAS), Finnish (FIN), and Admixed (AMR), labeled according to their closest 1000 Genomes super-population. For case-control analyses to be performed within each ancestry cluster, affected trios were excluded and a set of unrelated individuals was selected from the remaining subjects using KING version 2.1.3 software (“--unrelated” option)⁶⁴. Cluster-specific PCs were calculated by performing PC analysis on unrelated controls and projecting the remaining subjects onto the resulting axes. Remaining population stratification within each ancestry cluster was assessed visually (Supplementary Fig. 3).

The ImmunoChip densely covered genetic variation in immune-associated genomic regions. Discovery analyses included all genotyped variants, as well as imputed variants from any 500-kb region that contained more than 50 genotyped variants (Supplementary Table 3). To define boundaries for fine-mapping regions, we mapped previously defined “ImmunoChip regions” (provided by the R package humarray) from GRCh36 to GRCh38 coordinates (Supplementary Table 2): for each region, we mapped all variants originally included in the region to GRCh38 to define boundaries as the lowest and highest observed GRCh38 positions among these variants (+/− 50 kb either side). Fine-mapping analyses were then restricted to densely genotyped regions overlapping these “ImmunoChip regions”.

Genotypes were imputed with the NHLBI Trans-Omics for Precision Medicine (TOPMed) Freeze 5 (Supplementary Note) reference panel. We analyzed association with type 1 diabetes (T1D) for all genotyped variants and high-confidence imputed variants separately in the five ancestry groups (Supplementary Tables 2 and 3 and Supplementary Note). Assuming an additive mode of inheritance, we used logistic regression for unrelated case-control analyses, adjusting for five ancestry-specific PCs and using genotype posterior probabilities to account for uncertainty in imputed genotypes using the SNPTEST version 2.5.4 software⁶⁷. Due to small sample size (38 cases and 106 controls), EAS subjects were excluded. We combined results using an inverse-variance weighted fixed-effects meta-analysis (METAL software version released on 2011-03-25)⁶⁸. Forward stepwise logistic regression was performed to identify loci with more than one independent association with T1D. All conditionally independent associations (P < 5 × 10⁻⁸) were reported. Case-control analyses were performed under recessive and dominant models of inheritance. To evaluate the relative fit of the three models, we compared the Akaike Information Criterion (AIC) in EUR ancestry and identified the model providing the lowest AIC (best fit). On the X chromosome, only genotyped variants were examined for their association with T1D. The Y chromosome was not examined.

Trio families (two parents and an affected offspring) were analyzed within ancestry group using the transmission disequilibrium test (TDT)⁶⁹. As TDT statistics are susceptible to substantial bias when applied to imputed genotypes⁷⁰, a stringent variant filter was applied to imputed genotypes, removing all variants with Mendelian inconsistencies in >1% of trios with heterozygous offspring or parent-offspring pairs with homozygous offspring. From TDT summary statistics, we derived effect sizes and standard error estimates⁷¹ and meta-analyzed with Phase I results.

Two complementary approaches were used to define credible variant sets within each T1D-associated ImmunoChip region. Fine mapping included high-confidence variants within 750 kb of the lead variant (1.5-Mb region total), usually consisting of imputed variants across the entire ImmunoChip region and genotyped variants adjacent to the ImmunoChip region.

Haplotype analyses were performed in the EUR ancestry cases and controls by taking “best-guess” genotype values for the variants included in the analysis and obtaining haplotype phase distribution estimates for each individual, using an expectation-maximization algorithm⁷⁵. Each individual’s haplotype was sampled ten times and a logistic regression was fitted estimating the effect size of the haplotype relative to the most common haplotype in the population, with T1D status as the outcome and adjusting for five PCs. The estimates and standard errors for each haplotype relative to the most common were averaged over the ten logistic regression models to obtain overall haplotype effect sizes on T1D risk.

The functional impacts of T1D credible variants (Supplementary Table 11) were annotated using ANNOVAR (version released on 16 April 2018)⁷⁶ and the Ensembl and refGene annotation databases.

We downloaded publicly available ATAC-seq data from diverse immune cell types³⁸, pancreatic islets³⁵, and cardiac fibroblasts⁴⁰ (see Data Availability statement).

We generated additional ATAC-seq data on CD4⁺ T cells (n = 6 donors) and CD19⁺ B cells (n = 4 donors), using different culture and stimulation conditions from Calderon et al.³⁸. CD4⁺ T cells were enriched and stimulated as previously described⁷⁷. B cells were positively selected from PBMCs using anti-CD19 beads (Miltenyi Biotec, GmbH) and cultured for 24 hours in X-VIVO 15 (Lonza, Switzerland) supplemented with 1% Human Ab Serum (Sigma) and penicillin/streptomycin (Thermo Fisher) and plated in 96-well CELLSTAR U–bottomed plates (Greiner Bio-One, Austria) at concentration of 2.5 × 10⁵ cells/well. Cells were left untreated or stimulated with 10 μg/ml goat anti-human IgM/IgG/IgA antibody (109‐006‐064, Jackson Immunoresearch), 0.15 μg/ml rhCD40L (ALX-522-110-C010, ENZO Lifesciences), 20 ng/ml rhIL-21 and rhIL-4 (200-21 and 200-04 respectively, Peprotech) for 24 hours. ATAC-seq data was generated from 50,000 cells from each cell type and culture condition following the Omni-ATAC protocol⁷⁸. ATAC-seq datasets were mapped to GRCh38.p12⁷⁹ with minimap2 (version 2.17)⁸⁰, except for GSE123404 (pancreatic islets dataset) where bowtie2 (version 2.3.5) was used. After mapping, the technical replicates (where available) were merged and PCR duplicated reads were detected with Picard tools (version 2.20.2). The percentage of detected duplicated reads was very low (mean value < 1%) in all datasets. bigWig files were generated with bamCoverage from the deeptools package (version 3.3.0), using reads per genome coverage (RPGC) normalization and ignoring allosomes and the mitochondrial chromosome. Peaks were called using macs2⁸¹ (version 2.1.2) with the params “--nomodel --shift 37 --extsize 73 --keep-dup all”.

The immune cell ATAC-seq dataset GSE118189³⁸ was used to create a consensus list of peaks. For each cell type, the donor contributing the fewest number of reads to that cell type was selected and the number of reads was divided by two. Reads were then randomly pooled by that number for each sample, creating a representative alignment file for that cell type. This procedure was performed twice in order to obtain two pseudo-replicates. Peaks were called with macs2 with the same parameters. Irreducible discovery rate (IDR) was calculated between the two pseudo replicates⁸², any peak with an IDR ≤ 0.05 were included in the consensus list of peaks. This list was then used as a feature reference and reads were counted per feature with featureCounts from the package subread⁸³ (version 1.6.4). A similar approach was used for the other datasets in the analysis. IDR was used to obtain a reliable list of peaks. In these datasets, no feature reference was derived from the IDR, and counting was performed directly from the list obtained from GSE118189. Workflows were implemented using conda and snakemake.

To examine enrichment of T1D credible variants (group marginal posterior probability > 0.8 from GUESSFM) in open chromatin, for each cell type, two complementary approaches—SNP-matching and GoShifter⁸⁴ (http://software.broadinstitute.org/mpg/goshifter/)—were employed. In the SNP-matching approach, variants were randomly sampled across the genome, matched on LD structure and gene density, to generate a null distribution of SNPs overlapping accessible chromatin (see below). GoShifter, in contrast, generates a null distribution within each locus (see Trynka et al.⁸⁴).

We profiled chromatin accessibility in 115 individuals (57 controls and 58 T1D cases; 67 AFR and 48 EUR) from the Type 1 Diabetes Genetics Consortium (T1DGC). CD4⁺ T cells were purified from viably frozen PMBC samples using magnetic cell separation according to the manufacturer protocol, using either negative selection (n = 42; STEMCELL Technologies EasySep Human CD4⁺ T Cell Isolation Kit) or positive selection (n = 73; MACS Miltenyl Biotec). The selection approach was incorporated in data processing and analysis. After CD4⁺ T cell purification, the “Omni-ATAC-seq” protocol⁷⁸ was followed for nuclei isolation, transposase incubation, and library preparation. Libraries were sequenced using 75 bp paired-end reads on an Illumina NextSeq and data were processed using the PEPATAC pipeline⁸⁵. Briefly, reads were trimmed using Skewer (version0.2.2)⁸⁶ and, after removing reads mapping to mitochondrial and human repeat regions, were mapped to GRCh38 using bowtie2⁸⁷. PCR duplicates were removed, enzymatic cut sites were inferred based on read alignment, and peaks called using macs2⁸¹. Libraries with transcription start site (TSS) enrichment scores less than 6 million or fewer than 10 million aligned reads were excluded from analyses. A set of consensus peaks was determined by merging peaks across all samples using bedops (version 2.4.35)⁸⁸. A matrix of peak counts was calculated by counting the number of cut sites within each consensus peak in each sample using the R package bigWig (https://github.com/andrelmartins/bigWig).

Peaks with low counts were excluded (required ≥ 10 reads in ≥ 50% of samples). Further peak quality filtering and normalization was performed using the R package edgeR⁸⁹. These steps included:

filtering for peaks with ≥ 10 counts-per-million (CPM) across samples within each batch;

We confirmed matching sample identity between ATAC-seq libraries and genotyped subjects using the “Match BAM to VCF” (MBV) command in the software tool set QTLtools⁹¹. Association between imputed genotype dosage and chromatin accessibility (caQTL analysis) was tested with a linear model, adjusting for the first two genotype principal components, age at sample collection, TSS enrichment score, and CD4⁺ T cell purification approach using the R package MatrixEQTL⁹². The caQTL discovery analyses were performed separately by ancestry group (EUR and AFR) and combined in an inverse-variance weighted fixed effect meta-analysis (R package meta). All variant-peak combinations were tested where the accessibility peak was within 1 Mb of a T1D credible variant.

We evaluated colocalization of T1D and caQTL for all peaks where at least one T1D credible variant (as defined by GUESSFM) was associated with peak accessibility (meta-analysis P < 5 × 10⁻⁵) using the R package coloc⁴¹, and visualized colocalized signals using the R package locuscomparer⁹³. Conditional summary statistics were used in regions predicted to have more than one causal variant underlying the T1D association or regions with multiple, conditionally independent variants associated with accessibility of the same peak. When running coloc for T1D-caQTL colocalization, we used a prior probability of colocalization of 5 × 10⁻⁶ and provided association betas and standard errors as input data. When running coloc for T1D-eQTL colocalization, we used the same priors and supplied association z-scores. We considered GWAS and QTL signals were considered to be significantly colocalized when the posterior probability of colocalization was greater than 0.8 (‘PP.H4.abf’ > 0.8).

For significant caQTLs that colocalized with T1D-associated variants, we tested for allele-specific accessibility of the caQTL peak. First, we identified individuals heterozygous for T1D credible variants overlapping the caQTL peak. Within each heterozygous individual, we then counted the number of reads overlapping the variant position containing the reference or alternative allele. We only performed this analysis if the T1D credible variant overlapping the caQTL peak was directly genotyped on the ImmunoChip, since uncertainty in the heterozygous status of an individual could lead to biased results. For peaks with at least 5 participants who had at least 5 reads overlapping the peak, we formally tested whether the proportion of reads containing an alternative allele significantly deviated from the expected null hypothesis proportion of 0.5. We calculated P-values for deviation from “allelic balance” (proportion = 0.5 for each read) by fitting a generalized linear mixed model where the dependent variable is the number of reads and follows a Poisson distribution and the independent variables include a fixed effect for the allele and a random effect for the participant.

Jurkat cell line (E6–1) was purchased from ATCC and grown in Roswell Park Memorial Institute, RPMI (RPMI-1640; Gibco) supplemented with 10% fetal bovine serum, 1% penicillin-streptomycin, 1% sodium pyruvate), at 37 °C and 5% CO₂.

Labeled (5’ IRDye 700) and unlabeled 31-bp, single-stranded oligonucleotides containing rs72928038 were obtained from Integrated DNA Technologies (Reference Allele strand: 5’AGGGACGGATTTCCTGTAAGCTGATCTTGAA 3’ and Alternative Allele strand: 5’ AGGGACGGATTTCCTATAAGCTGATCTTGAA 3’) along with complementary oligonucleotides. Double-stranded oligonucleotides were generated by annealing equal amount of labeled or unlabeled complementary oligonucleotides at 95 °C for 5 min, followed by gradual cooling with a ramp rate of −1.2 °C/min for 1 h (Bio-Rad C1000 Touch Thermal Cycler). Nuclear extract from Jurkat cells was obtained by following the manufacturer’s protocol for NE-PER^™ Nuclear and Cytoplasmic Extraction Reagents kit (Thermo Scientific) and the extracted nuclear protein was dialyzed with Slide-A-Lyzer MINI Dialysis Units, 10,000 MWCO (Thermo Scientific) against a 1 L buffer (10 mM Tris, pH 7.5, 50 mM KCl, 200 mM NaCl, 1 mM dithiothreitol, 1 mM phenylmethylsulfonyl fluoride, and 10% glycerol) for 16 h at 4 °C with slow stirring.

Binding reaction for the EMSA was carried out using 2 μL 10X binding buffer (100 mM Tris, 500 mM KCl, 10 mM DTT; pH 7.5), 2 μL 25 mM DTT (2.5% Tween 20), 1 μL Poly (dI-dC) (1 μg/μL in 10 mM Tris, 1 mM EDTA; pH 7.5), 1 μL 1% NP-40, 100 mM MgCl₂, 20 fmol IRDye double-stranded oligonucleotide probe, and 16 μg Jurkat nuclear extract in a final volume of 20 μL. For supershift lanes, tested transcription-factor-binding antibodies (ETS-1 Rabbit mAb and Stat1 Rabbit mAb) were diluted 1:50 with ddH₂O. Negative control Rabbit IgG was diluted to the same concentration as tested antibody. 1 μL of diluted antibody was added to the binding reaction mixture while maintaining a total volume of 20 ul. Binding reaction was incubated for 20 min at room temperature, after which 2 μL of 10X Orange Loading Dye was added. Electrophoresis was performed with binding reaction mixture on a pre-run 6% DNA retardation gel for 70 min at 70 V. To capture the image, the gel was placed directly on the Odyssey-CLx (Licor) scan bed. The gel was scanned with a thickness of 0.5 mm at 700 nm channel. The EMSA binding condition for rs72928038 was repeated three times to ensure reproducibility of the experiment.

To prioritize drug targets implicated by T1D genetic associations, we ran the Prioritiy Index (Pi) algorithm, as implemented in the R package Pi⁴⁶. Data used to identify eQTL colocalization (eGenes) included those from the initial publication (unstimulated monocytes⁹⁴, n = 414; LPS-stimulated monocytes after 2 hours⁹⁴, n = 261; LPS-stimulated monocytes after 24 hours⁹⁴, n = 322; interferon-gamma-stimulated monocytes after 24 hours⁹⁴, n = 367; unstimulated B cells⁹⁵, n = 286; unstimulated NK cells (unpublished), n = 245; unstimulated neutrophils⁹⁶, n = 114; unstimulated CD4⁺ T cells⁹⁷, n = 293; unstimulated CD8⁺ T cells⁹⁷, n = 283; and whole blood⁹⁸, n = 5,311), as well as a larger whole blood study (n = 31,684)⁴². Hi-C data from monocytes, fetal thymus, naïve CD4⁺ T cells, total CD4⁺ T cells, activated total CD4⁺ T cells, non-activated total CD4⁺ T cells, naïve CD8⁺ T cells, total CD8⁺ T cells, naïve B cells, and total B cells⁴³ were used to identify genes interacting with index variants (cGenes). Data used to define functional genes (fGenes, pGenes and dGenes) were those used in the initial publication. The STRING database⁹⁹ was used to define protein-protein interaction networks, where confidence scores ≥ 700 were considered.

Unless otherwise noted, all statistical analysis and data visualization was performed using R version 3.6¹⁰⁰. All statistical tests based on symmetrically distributed test statistics were two-sided. No repeated measures data were analyzed in this study. All genotyped and ATAC-seq samples analyzed in association tests represent distinct individuals. The R packages ggplot2, cowplot, ggbio, GenomicRanges, gridExtra, RColorBrewer, and rtracklayer were used for data visualization.

Code used to generate the results presented in this paper is available at https://github.com/ccrobertson/t1d-immunochip-2020. Pipelines for processing ATAC-seq data are available at https://github.com/dfloresDIL/MEGA and http://pepatac.databio.org.

Raw FASTQ files were obtained from Gene Expression Omnibus (GEO) accession number GSE118189. These data included four individuals and 25 immune cell types under resting conditions and after stimulation with anti-human CD3/CD28 dynabeads and human IL-2 (for 24 hours, T lymphocytes), F(ab)’2 anti-human IgG/IgM³⁸ and human IL-4 (for 24 hours, B lymphocytes), human IL-2 (for 48 hours, NK cells), or LPS (for 6 hours, monocytes)³⁷.

ATAC-seq data from pancreatic islets of five donors without glucose intolerance and five EndoCβH1 cell line replicates, under resting conditions and after stimulation with IFN-γ and IL-1β for 48 hours were downloaded from GEO, accession number GSE123404³⁵.

ATAC-seq data from cardiac fibroblasts (two fetal and three adult) were downloaded from the European Nucleotide Archive (https://www.ebi.ac.uk/ena/data/view/SRX2843570 and https://www.ebi.ac.uk/ena/data/view/SRX2843571), as a control cell type that we did not expect to be involved in the etiology of T1D⁴⁰.

Epigenome annotation tracks, chromHMM¹⁰¹ tracks from diverse primary human cells were obtained from the NIH Epigenome Roadmap, http://dcc.blueprint-epigenome.eu/#/md/secondary_analysis/Segmentation_of_ChIP-Seq_data_20140811 and additional immune-specific human primary and cell lines from the Blueprint consortium, https://egg2.wustl.edu/roadmap/data/byFileType/chromhmmSegmentations/ChmmModels/coreMarks/jointModel/final.

Summary statistics from whole blood cis eQTL analysis from 31,683 individuals⁴² were downloaded from https://eqtlgen.org.

Additional databases used in the Priority Index (Pi) drug target prioritization analysis were obtained through the relational database provided in the R package Pi (http://pi.well.ox.ac.uk:3010/download).