Specific type 1 diabetes risk genes underpin age-at-diagnosis and indicate joint defects in immunity, beta-cell fragility and responses to viral infections in early-onset disease

Abstract

Background Immunohistological analyses of pancreata from patients with autoimmune type 1 diabetes (T1D) suggest a stratification of islet pathology of both B and T lymphocyte islet inflammation common in children diagnosed under age 7 years, whereas B cells are rare in those diagnosed age ≥13. Based on these observations, we would expect to see genetic susceptibility differences between these age-at-diagnosis groups at the population level. Moreover, these genetic susceptibility differences could inform us on the aetiology of this most aggressive form of T1D that initiates in the first years of life.

Methods Using multinomial logistic regression models we tested if the known T1D loci (17 within the human leucocyte antigen (HLA) region and 55 others, non HLA regions) had significantly stronger effect sizes in the <7 group compared to the ≥13 group, using genotype data from 26,991 individuals (18,400 controls, 3,111 T1D diagnosed <7 years of age, 3,759 at 7-13 and 1,721 at ≥13).

Findings Six associations of the HLA class II and I genes had stronger effects in the <7 group, and seven non-HLA regions, one of which functions specifically in beta cells (GLIS3), and the other six likely affecting key T cell (IL2RA, IL10, SIRPG), thymus (PTPRK) and B cell development/functions (IKZF3, IL10) or in both immune cells and beta cells (CTSH).

Interpretation In newborn children with the greatest load of certain HLA and non-HLA risk alleles, inherited variants in immune and beta cells, and their inherent disregulated response to environmental stresses such as virus infection, combine to cause a rapid loss of insulin production, thereby driving down the age at which T1D is diagnosed.

Introduction

Type 1 diabetes (T1D) is a multifactorial disease in which the insulin-producing beta cells of pancreatic islets are destroyed or rendered dysfunctional by an autoimmune process that often initiates in the first few months of life, causing a pre-diabetic, non-symptomatic state in approximately 0.4% of children ¹. The actual diagnosis could happen many years after this prodromal phase, the joint environmental and genetic mechanisms of which remain ill defined, with the median age-at-diagnosis being around age 11 years. Even after diagnosis there is still often sufficient endogenous insulin production to lower insulin treatment and reduce the later in life complications of early mortality, cardiovascular, kidney, eye and peripheral neuron disease ^2,. The exceptions to this are the children diagnosed with T1D under the age 10 years in whom there is little insulin production shortly after diagnosis, as measured by circulating C-peptide concentrations ^2,3. This subgroup represents the largest unmet clinical challenge, since they suffer the greatest complications of the disease ³. Yet any intervention of T1D autoimmunity in these young children must be as safe and precise as possible, modulating the causative molecules, cells, pathways and mechanisms. Hence we need to identify the specific mechanisms underlying early-diagnosed T1D.

Recent evidence suggests that children diagnosed under age 7 years may have a different, more aggressive form of islet inflammation (insulitis), characterised by a B lymphocyte infiltrate coincident with a T cell insulitis (CD4⁺ and CD8⁺ T cells), than children aged 13 years and over, who have reduced B cell participation ⁴. In cases diagnosed between 7 and 12 years there is a mixture of islet infiltrate phenotypes, some with the “under 7” B cell infiltrate and others with “13 and over” phenotype. There is already evidence that some genetic variants reduce age-at-diagnosis, which provides insight into the biology of this most beta-cell destructive form of the disease ^5–8. The autoantigen-presenting genes human leukocyte antigen (HLA) class II and class I are the major drivers of younger age-at-diagnosis. Class II molecules are recognised by CD4⁺ T cells which provide help for CD8⁺ beta-cell cytotoxic T cells and islet antigen-specific B cells. Class I molecules are expressed on beta cells, upregulated during viral infection or by immune cytokines, rendering them more susceptible to autoreactive CD8⁺ T cells. More recently, a genome-wide association genetic analysis of age-at-diagnosis of T1D identified a locus on chromosome 6q22.33 that acts almost exclusively in the cases of T1D diagnosed under age 5 years ⁹, encoding the protein tyrosine phosphatase receptor kappa (PTPRK) and thymocyte-expressed molecule involved in selection (THEMIS) genes. However, this approach has to meet the stringent genome-wide multiple testing correction criterion (p < 5 × 10⁻⁸) and informative, true signals were likely to have been missed. In the present study, we analysed the association of specified known T1D gene regions, thereby reducing the multiple testing burden. In addition, a biological or phenotypic prior could provide greater sensitivity in the search for age-at-diagnosis-associated genes. The stratification of patients into age-at-diagnosis categories according to their pancreatic histology, as opposed to treating age-at-diagnosis as a continuous phenotype provides us with just this opportunity.

Here, we analysed T1D-associated variants according to the proposed pancreatic infiltrate stratification of T1D, namely the age-at-diagnosis groups, the under 7’s versus the 13’s and over. If T1D has a particular pancreatic immunophenotype then it might be expected that it could have distinct genetic features, characterised by susceptibility genes with larger effects in the under 7’s. Moreover, the intermediate group, age-at-diagnosis 7-13 years, would have risk for these age-at-diagnosis-sensitive genes lying between the under 7’s and the 13’s and over. Six HLA haplotypes/alleles and seven non-HLA loci fulfil this risk profile informing the biology of the most aggressive form of T1D, revealing a mixture of predisposition in both the beta cell and immune cell compartments.

Methods

Study populations

Our dataset consists of 18,400 controls, 3,111 T1D cases diagnosed at <7 years (the <7 group), 3,759 at ≥7 to <13 years (the 7-13 group) and 1,721 at ≥13 years (the ≥13 group). The majority of individuals are from the UK, with others from central Europe, Asia-Pacific, Finland and the USA (Table 1), and comprises only unrelated individuals, since related individuals were removed (Supplementary methods).

Table 1:

Characteristics of individuals included in the analysis.

Loci studied

We examined eight HLA class II haplotypes and nine HLA class I classical alleles for their association with T1D diagnosed at each age group, where the haplotypes and classical alleles were a subset of the most protective and susceptible haplotypes identified for T1D to date ¹⁰ that we also found to be associated with T1D in our analysis after conditioning for the other associated HLA haplotypes (logistic regression Wald test p<0.01). Supplementary Table 1 summarises which haplotypes and classical alleles were examined, how they were defined and whether they were common enough to include in our analysis, defined as at least 5 individuals from each group with the classical allele/haplotype.

We also examined 55 loci outside the HLA, which have previously shown association with T1D (Supplementary Table 2). Each locus contains an ‘index’ variant, chosen to be the most strongly disease associated from a set of variants in linkage disequilibrium (LD) that constitute a single genetic signal. We have allocated locus names to each of these variants based on a candidate gene(s), but the named genes may not be causal for T1D.

Imputation

Classical HLA alleles as well as non-HLA variants that were excluded due to variant quality control filtering were imputed for analysis (Supplementary methods). Some individuals were genotyped for a subset of their classical HLA alleles ⁸ and therefore accuracy of imputation was assessed at those classical alleles for a proportion of individuals.

Multinomial logistic regression

In order to examine whether or not there was heterogeneity in effect size for each examined variant between the <7 and ≥13 groups, we fitted two multinomial logistic regressions per locus, one assuming identical effect sizes for the genetic variant in the <7 and ≥13 groups and the other allowing different effect sizes between groups. A comparison of how well these models fit the data allows us to test for heterogeneity in effect size between the two groups. Both models were adjusted for the ten largest principal components derived from the set of ImmunoChip variants passing quality control filters (Supplementary methods).

To test stability of our results at non-HLA loci, we did four sensitivity analyses. Firstly, we sampled without replacement 50% of cases and controls from each of the ancestry groups in our collection. In lieu of a valid replication dataset, to assess the possibility that age-at-diagnosis genetic heterogeneity was due to an unlikely chance distribution of genotypes between age strata, we repeated the heterogeneity test, for the 50% that were sampled and also on the remaining 50%. We performed this procedure 100 times, giving us 200 heterogeneity tests and noted the proportion of times the variant under consideration reached nominally significant heterogeneity (p<0.05). Secondly, to exclude the possibility of spurious associations due to population structure in our data, we repeated the analysis but only including individuals from the UK and Northern Ireland and adjusted for the five largest principal components derived from Immunochip data in these individuals only. Finally, to test sensitivity of our results to age-strata thresholds, we performed the same analysis but instead compared individuals diagnosed at <6 years to the ≥13 group and also individuals diagnosed at <5 years compared to the ≥13 group. We declared a locus differentially-associated if the heterogeneity p-value was associated to a False Discovery Rate (FDR) of <0.1. To explore whether there were more age-at-diagnosis associated variants which we cannot detect in the present analysis due to a lack of statistical power, we examined all loci which did not reach the association threshold (FDR<0.1) and counted how many loci had the largest effect in the <7 group, the intermediate effect in the 7-13 group and the smallest effect in the ≥13 group and compared this to the expected frequency of this ordering using a binomial test (Supplementary methods).

Fine mapping

For each non-HLA locus with strong evidence of heterogeneity between age-at-diagnosis groups, as determined by Bonferroni correction, a more conservative multiple-comparison correction than FDR, we fine mapped a 0.5 Mb region around the index variant to identify a list of potentially causal variants for T1D diagnosed at <7 years. Analysis was limited to individuals from the UK and Northern Ireland, amounting to 2,888 cases diagnosed at <7 years and 11,064 controls, in order to examine a homogeneous population, as fine mapping is sensitive to differences in LD structure between ancestrally divergent groups. We used the GUESSFM software, ¹¹ which carries out a Bayesian variable selection stochastic search to identify the combinations of variants constituting separate genetic susceptibility to T1D. We then examined whether the T1D-associated variants colocalised with expression quantitative trait loci (eQTL) associations in whole blood from a dataset of over 30,000 individuals, gauging which genes the variants are most likely to be regulating and in what direction the effects are on gene transcription and disease risk ¹² (eQTL statistics downloaded from http://www.eqtlgen.org/cis-eqtls.html) (Supplementary methods).

The scripts used to analyse these data are available at https://github.com/jinshaw16/AAD_t1d, commit 1727d18c3fe2559ac527681142155b83e8294165.

Funding

This work was funded by the JDRF (9-2011-253, 5-SRA-2015-130-A-N) and Wellcome (091157, 107212) to the Diabetes and Inflammation Laboratory, University of Oxford.

We use data generated by the Wellcome Trust Case Control Consortium (076113). The Northern Irish GRID, IDDMGEN, T1DGEN and Warren cohorts were genotyped using the T1DGC grants from the NIDDK, the NIAID, the NHGRI, the NICHD and the JDRF (U01 DK062418, JDRF 9-2011-530).

Results

Multinomial logistic regression: HLA

We found six HLA haplotypes to be differentially-associated between the <7 and ≥13 group (FDR<0.1). The strongest susceptible class II effect was for the DR3-DQ2/DR4-DQ8 diplotype, whilst the protective DRB1*15:01-DQB1*06:02 and DRB1*07:01-DQB1*03:03 haplotypes showed greater protection from T1D in the <7 group compared to the ≥13 group. Class I alleles A*24:02 and B39*06 showed more susceptibility to T1D in the <7 compared to and ≥13 group (Figure 1). Comparison of imputed classical 4 digit HLA alleles with directly genotyped 4 digit HLA alleles showed concordance of over 91% for each of gene examined (Supplementary Figure 1).

Figure 1:

Classical HLA alleles association with type 1 diabetes diagnosed at <7 years old (red circle; mean log-odds ratio age-at-diagnosis +/-95%CI), 7-13 years old (green circle; mean log-odds ratio age-at-diagnosis 7-13+/-95%CI) and ≥13 years old (blue circle mean log-odds ratio age-at-diagnosis >13+/-95%CI), from a multinomial logistic regression. Left panel shows the log-odds ratios with a dashed red line showing a log-odds ratio of 0. The right panel shows the association statistics from a likelihood ratio test comparing a multinomial logistic regression constraining the log-odds ratios from the <7 and ≥13 groups to be equal compared to an unconstrained model. Red dotted line shows nominal significance in heterogeneity (p<0.05), red dashed line show Bonferroni-corrected significance in heterogeneity.

Multinomial logistic regression: non-HLA regions

Outside the HLA, nine regions were differentially-associated between the <7 and ≥13 group (FDR<0.1), near Ikaros family zinc finger 3 (IKZF3), Cathepsin H (CTSH), GLIS family zinc finger 3 (GLIS3), Chymotrypsinogen B1 (CTRB1), the third index variant at interleukin 2 receptor alpha (IL2RA), interleukin 10 (IL10), Calmodulin-Regulated Spectrin-Associated Protein 2 (CAMSAP2), Signal Regulatory Protein Gamma (SIRPG) and PTPRK (Figure 2). Three of these (IKZF3, CTSH and GLIS3) survived Bonferroni correction (p<0.05/55=0.00091). At each locus associated with FDR<0.1, the 7-13 group had a larger effect size than the ≥13 group and smaller than the <7 group. Given the ≥13 group comprises just 1,721 individuals, it is probable that with increased sample size and hence statistical power, other T1D risk loci with sizeable estimated effect size differences between groups might reach statistical significance with regards to heterogeneity (Supplementary Figure 2). Of the 46 variants not satisfying an FDR<0.1, 20 have the strongest signal in <7s, weakest in ≥13s and intermediate in 7s-13s, compared to 8 occurrences in that order expected by chance (p=4.27×10⁻⁶, binomial test), suggesting the presence of substantial additional signal in variants that we are not able to declare show evidence individually.

Figure 2:

Non-HLA type 1 diabetes associated variants, showing log-odds ratios for those diagnosed at <7 years old (red circle; mean log-odds ratio age-at-diagnosis +/-95%CI), 7-13 years old (green circle; mean log-odds ratio age-at-diagnosis 7-13+/-95%CI) and ≥13 years old (blue circle mean log-odds ratio age-at-diagnosis ≥13+/-95%CI), from a multinomial logistic regression. Left panel shows the log-odds ratios with a dashed red line showing a log-odds ratio of 0. The right panel shows the association statistics from a likelihood ratio test comparing a multinomial logistic regression constraining the log-odds ratios from the <7 and ≥13 groups to be equal compared to an unconstrained model. Showing only loci with a false discovery rate of less than 0.1. Red dotted line shows threshold for false discovery rate of <0.1, red dashed line shows threshold for Bonferroni-corrected heterogeneity.

Stability of non-HLA results

After sampling half of the cases and controls 100 times, we found that nominal significance (p<0.05) was observed for the heterogeneity in effect size test between age-at-diagnosis groups >50% of the time for the IKZF3, CTSH, GLIS3, CTRB1 and IL2RA (3^rd index variant) loci (Supplementary Figure 3), and >44% of the time at the other FDR heterogeneous loci (IL10, CAMSAP2, SIRPG and PTPRK/THEMIS). In the UK-specific sensitivity analysis, six of the nine FDR heterogeneous loci from the primary analysis were heterogeneous between the <7 and ≥13 group (FDR<0.1) in this ancestry-homogeneous population, two of the loci showed no heterogeneity in effect size (CTRB1 p=0.310 and CAMSAP2 p=0.578) and were thus removed from our set of differentially-associated regions and the remaining locus, IL10, had a p-value of 0.06, which we considered differentially associated between the <7 and ≥13 groups, given the decrease in statistical power in this sensitivity analysis (Supplementary Figure 4).

When changing the threshold for the early-diagnosed group to <6 and <5, all seven associated loci from the primary analysis and UK-specific analysis were heterogeneous (FDR<0.1) (Supplementary Figures 5 and 6).

Minor allele frequency plots by age-at-diagnosis for the seven differentially associated loci that remained heterogeneous between the <7 and ≥13 group in all analyses are shown in Supplementary Figures 7-13, whilst Supplementary Tables 3 and 4 summarise the most likely causal genes at these loci.

Fine mapping

We fine mapped the three loci (IKZF3, CTSH and GLIS3) that reached Bonferroni-corrected heterogeneity between age-at-diagnosis groups. The posterior probability of there being one causal variant was >0.63 at each locus. All variants contained within a group that has a group posterior probability of causality of >0.9 are listed in Supplementary Tables 5-7, though our stringent post-imputation variant quality control filtering means variants in high LD with the listed variants could also be causal for T1D but were removed from the current analysis due to low quality imputation at that variant.

The IKZF3 locus results prioritise an LD block containing 34 variants, all of which could be causal, which also effects expression of at least three genes (p<5 × 10⁻¹⁵⁰), where the minor allele at the most likely causal variants decrease T1D risk and IKZF3 expression and also increase expression of GSDMB and ORMDL3. Colocalisation analyses support the hypothesis that the disease causal variant and the eQTL causal variant were the same for all three genes (posterior probability of colocalisation for T1D and eQTL with IKZF3=0.973, GSDMB=0.841 and ORMDL3=0.846). The CTSH locus showed evidence of colocalisation with the CTSH whole blood eQTL (posterior probability of colocalisation=0.655); the susceptibility allele for T1D is associated with more expression of CTSH (Figure 3).

Figure 3:

Results from fine mapping the IKZF3 and CTSH loci regions. Analysis includes individuals from the UK and Northern Ireland only and only controls and cases diagnosed at <7 years. Top panel: Gene positions (genome build 37), with arrows indicating direction of transcription. Second panel, univariable early-diagnosed (<7) type 1 diabetes log-odds ratios and 95% confidence intervals for each of the most likely causally associated variants as prioritised by GUESSFM. Third panel: log_e(absolute eQTL z score) if z score>0 and -log_e(absolute eQTL z score) if z score<0, so direction of effect can be compared, from publically available whole blood eQTL dataset of over 30,000 individuals, including only eQTLs with a p value of <5×10⁻¹⁵⁰ (IKZF3 locus) and <5×10⁻⁵⁰ (CTSH locus). The symbols are coloured red if contained in the set of most likely causal variants, as produced by GUESSFM and the shape corresponds to the gene that the variant is effecting transcription of.

There was no evidence of colocalisation between disease risk and GLIS3 whole blood eQTL (posterior probability of colocalisation=0.036), suggesting the variant might be acting elsewhere to alter T1D risk.

Discussion

The stratification of patients by age-at-diagnosis according to islet phenotypes has provided a rich source of genes, molecules and pathways with greater effects in children diagnosed with T1D under age 7 years. We expected to see strong differential associations with the HLA class II haplotypes, in particular the strongest single susceptibility determinant in the genome, the heterozygous diplotype DR3-DQ2/DR4-DQ8. Previously with smaller sample sizes and without dichotomising patients into biologically-defined discrete age categories, HLA class I alleles, A*24:02 and B*39:06 have been shown to be associated with younger age-at-diagnosis ^5–8. Here, we show for the first time that the protective HLA class II haplotypes DRB1*15:01-DQB1*06:02 and DRB1*07:01-DQB1*03:03 are less prevalent amongst individuals diagnosed at <7 years compared with controls and those diagnosed at ≥13 years. Therefore, the earliest and most aggressive phenotypic subtype of T1D results primarily from carriage of high risk alleles and haplotypes of the HLA class II and I genes, which probably act at four levels: (i) altering the T cell receptor repertoire in favour of anti-islet antigen reactivity, for example preproinsulin, and/or reducing the protective repertoire of T regulatory cells; (ii) providing a strong autoantigen presentation environment in the islets and pancreatic draining lymph nodes enabling the infiltration and cytolytic activity of CD8⁺ T cells but also by disrupting B cell anergy ¹³ permitting binding and presentation of autoantigen to provide potent help to T cells in a self-reinforcing spiral of autoreactivity; (iii) affecting the immune response to the viral infections that are involved in the disease; (iv) affecting how the gut microbiome develops in early life, a system that is known to affect T1D susceptibility ¹⁴.

In addition to the HLA heterogeneity, we obtained robust evidence of differences in effect size between the age-at-diagnosis groups at seven non-HLA loci. Of these loci, one plausible candidate gene, GLIS3, most likely perturbs disease risk in the islet beta cells, given the expression levels in the pancreas, lack of expression in immune cells, colocalisation with type 2 diabetes risk variants ¹⁵ and lack of association with other autoimmune diseases (https://genetics.opentargets.org). This finding supports a mechanism of beta-cell fragility, for example, susceptibility to apoptosis ¹⁶, in which increased risk of disease is encoded in the beta cell, not only in the immune system. The GLIS3 effect can be mimicked in a mouse model of non-immune diabetes by a high fat diet, linking obesity as a risk factor in T1D and type 2 diabetes ^16,17. Two of the loci, CTSH and IKZF3, could act in the islets or elsewhere, whilst all of the other candidate causal genes (IL2RA, IL10, SIRPG, PTPRK/THEMIS, as well as IKZF3/ORMDL3/GSDMB and CTSH) have known functions in T and/or B cell biology (Supplementary Table 4). This implies that in addition to HLA-susceptibility, risk of T1D in the very young is also impacted by particular malfunctions in the infiltrating T and B cells, leading to increased risk of autoreactivity, resulting in a perfect storm of immune infiltration, antigen recognition and a rapid destruction of beta cells.

Of the three non-HLA risk regions with the strongest evidence of heterogeneity between age-at-diagnosis groups, we focus on the IKZF3 and CTSH loci, which colocalise with whole blood eQTLs. The region containing IKZF3 has a complex structure with a large LD block, which is associated with multiple diseases, including asthma and paediatric asthma ^{18, 19} (Supplementary Table 4). However, the direction of effect of the risk variant is opposite in asthma to all associated autoimmune diseases, including T1D, where the C allele at a variant within the haplotype, rs921649 (C>T), increases susceptibility to autoimmunity, whereas the C allele is protective for asthma ¹⁸. Whole blood eQTL data shows the expression of 13 protein-coding genes is modulated by variants in the disease-associated haplotype, with IKZF3, ORMDL3 and GSDMB the most affected ¹². All three genes are expressed in lymphocytes and are up-(IKZF3) or down-regulated (ORMDL3, GSDMB) (https://dice-database.org/) following activation, with good biological candidacy for altering disease risk. IKZF3 is a transcriptional repressor with a key role in B-cell activation and differentiation ²⁰ and T cell differentiation ²¹. ORMDL3 is a central regulator of sphingolipid biosynthesis ²² and has also been proposed to negatively regulate store-operated calcium, lymphocyte activation and cytokine production ^18,23, while GSDMB can act as a pyroptotic protein ²⁴. Therefore one or more of these genes may be causal for T1D risk. Pertinent to the increased frequency of B-cell infiltration in the islets of the <7 group, there is evidence that carriers of the T1D risk allele have decreased anergic high affinity insulin-binding B cells in circulating blood, implying some of this population may have relocated to the pancreas ¹³. This loss of anergic circulating B cell frequencies is also associated with the most predisposing age-at-diagnosis diplotype HLA-DRB1*03:01-DQB1*02:01/DRB1*04:01-DQB1*03:02 compared to donors with the protective HLA class II haplotypes ¹³.

The candidate T1D risk variants at the CTSH locus, for example the C allele at rs2289702 (C>T), are associated with increased expression of CTSH RNA in multiple cell types and tissues (Supplementary Table 4). The locus has previously been implicated in T1D aetiology by altering sensitivity of beta cells to apoptosis ²⁵, where rs3825932 (C>T) was investigated, which is in low LD (r²=0.26) with the disease-associated variant reported here and the T1D risk allele counter-intuitively resulted in protection from beta-cell apoptosis. Thus, beta-cell apoptosis may not be the primary mechanism underlying disease aetiology in this region. CTSH functions as an endopeptidase and can cleave the N-terminus of the Toll-like receptor 3 (TLR3) protein, increasing its functionality ²⁶. Given TLR3 is expressed in islets ²⁷, it is possible that the increase in CTSH expression associated with the T1D susceptibility allele (the C allele of rs2289702) results in increased TLR3 N-terminus cleavage, heightened responses to viral infections and increased release of type 1 interferon (IFN). This may increase baseline risk of T1D and specifically the risk of early-diagnosed T1D in individuals carrying this allele, since viral infections are more frequent in childhood. There is mounting evidence that enteroviral infections predispose to T1D, a type 1 IFN transcriptional signature precedes anti-islet autoantibody appearance in children ²⁸, and another receptor for viral RNA, MDA5 encoded by IFIH1 is a proven T1D susceptibility gene with its higher IFN-inducing activity increasing risk of the disease ²⁹. Exposure of beta cells to type 1 IFN greatly increases their HLA class I expression and susceptibility to CD8⁺ cytotoxic killing, and heightened class I expression on beta cells is a hallmark phenotype of the T1D pancreas ³⁰.

Our genetic results imply a dynamic, fully integrated pathogenic collaboration between the immune system, the beta cells and viral infection in the initiation and rapid development of extreme insulin-deficiency starting in the first few weeks and months of life in those that carry the heaviest load of age-at-diagnosis alleles. Combinations of modulators of these pathways could be an effective way of preventing the cessation of endogenous insulin-production.

Supplementary Figure 1:

Concordance of HiBag imputed versus directly genotyped classical HLA alleles. Concordance is defined as identical 4 digit HLA classical allele at both chromosomes.

Supplementary Figure 2:

All 55 non-HLA type 1 diabetes associated variants, showing log-odds ratios for those diagnosed at <7 years old (red circle; mean log-odds ratio age-at-diagnosis +/-95%CI), 7-13 years old (green circle; mean log-odds ratio age-at-diagnosis 7-13+/-95%CI) and ≥13 years old (blue circle mean log-odds ratio age-at-diagnosis ≥13+/-95%CI), from a multinomial logistic regression. Left panel shows the log-odds ratios with a dashed red line showing a log-odds ratio of 0. The right panel shows the association statistics from a likelihood ratio test comparing a multinomial logistic regression constraining the log-odds ratios from the <7 and ≥13 groups to be equal compared to an unconstrained model. Red dotted line shows threshold for nominally significant heterogeneity between groups (p<0.05), red solid line shows threshold for false discovery rate of <0.1, red dashed line shows threshold for Bonferroni-corrected significant heterogeneity.

Supplementary Figure 3:

Proportion of times each locus had a nominally significant heterogeneity test (p<0.05) when sampling 50% of the entire collection 100 times, generating 200 analysis datasets.

Supplementary Figure 4:

All 55 non-HLA type 1 diabetes associated variants examined using only individuals from the UK or Northern Ireland, showing log-odds ratios for those diagnosed at <7 years old (red circle; mean log-odds ratio age-at-diagnosis +/-95%CI), 7-13 years old (green circle; mean log-odds ratio age-at-diagnosis 7-13+/-95%CI) and ≥13 years old (blue circle mean log-odds ratio age-at-diagnosis >13+/-95%CI), from a multinomial logistic regression. Left panel shows the log-odds ratios with a dashed red line showing a log-odds ratio of 0. The right panel shows the association statistics from a likelihood ratio test comparing a multinomial logistic regression constraining the log-odds ratios from the <7 and ≥13 groups to be equal compared to an unconstrained model. Red dotted line shows threshold for nominally significant heterogeneity between groups (p<0.05), red solid line shows threshold for false discovery rate of <0.1, red dashed line shows threshold for Bonferroni-corrected significant heterogeneity.

Supplementary Figure 5:

All 55 non-HLA type 1 diabetes associated variants, showing log-odds ratios for those diagnosed at <6 years old (red circle; mean log-odds ratio age-at-diagnosis +/-95%CI), 6-13 years old (green circle; mean log-odds ratio age-at-diagnosis 6-13+/-95%CI) and ≥13 years old (blue circle mean log-odds ratio age-at-diagnosis ≥13+/-95%CI) from a multinomial logistic regression. Left panel shows the log-odds ratios with a dashed red line showing a log-odds ratio of 0. The right panel shows the association statistics from a likelihood ratio test comparing a multinomial logistic regression constraining the log-odds ratios from the <6 and ≥13 groups to be equal compared to an unconstrained model. Red dotted line shows threshold for nominally significant heterogeneity between groups (p<0.05), red solid line shows threshold for false discovery rate of <0.1, red dashed line shows threshold for Bonferroni-corrected significant heterogeneity.

Supplementary Figure 6:

All 55 non-HLA type 1 diabetes associated variants, showing log-odds ratios for those diagnosed at <5 years old (red circle; mean log-odds ratio age-at-diagnosis +/-95%CI), 5-13 years old (green circle; mean log-odds ratio age-at-diagnosis 5-13+/-95%CI) and ≥13 years old (blue circle mean log-odds ratio age-at-diagnosis ≥13+/-95%CI) from a multinomial logistic regression. Left panel shows the log-odds ratios with a dashed red line showing a log-odds ratio of 0. The right panel shows the association statistics from a likelihood ratio test comparing a multinomial logistic regression constraining the log-odds ratios from the <7 and ≥13 groups to be equal compared to an unconstrained model. Red dotted line shows threshold for nominally significant heterogeneity between groups (p<0.05), red solid line shows threshold for false discovery rate of <0.1, red dashed line shows threshold for Bonferroni-corrected significant heterogeneity.

Supplementary Figure 7:

Minor allele frequency of the index variant near the IKZF3 gene for controls and individuals diagnosed at various ages.

Supplementary Figure 8:

Minor allele frequency of the index variant near the CTSH gene for controls and individuals diagnosed at various ages.

Supplementary Figure 9:

Minor allele frequency of the index variant near the GLIS3 gene for controls and individuals diagnosed at various ages.

Supplementary Figure 10:

Minor allele frequency of the index variant near the IL2RA gene (third index variant) for controls and individuals diagnosed at various ages.

Supplementary Figure 11:

Minor allele frequency of the index variant near the IL10 gene for controls and individuals diagnosed at various ages.

Supplementary Figure 12:

Minor allele frequency of the index variant near the SIRPG gene for controls and individuals diagnosed at various ages.

Supplementary Figure 13:

Minor allele frequency of the index variant near the PTPRK/THEMIS genes for controls and individuals diagnosed at various ages.

Supplementary Table 1: Classical HLA alleles/haplotypes examined in analysis.

Supplementary Table 2: Non-HLA variants examined in analysis.

Supplementary Table 3: Non-HLA region variants with evidence of heterogeneity in effect size between the <7 and ≥13 groups: Promoter Capture Hi-C (PCHi-C) candidate genes.

Supplementary Table 4: Details of non-HLA variants with evidence of heterogeneity in effect size between the <7 and ≥13 groups.

Supplementary Table 5: Most likely variants causally associated with T1D at the IKZF3 locus from GUESSFM fine mapping analysis.

Supplementary Table 6: Most likely variants causally associated with T1D at the CTSH locus from GUESSFM fine mapping analysis.

Supplementary Table 7: Most likely variants causally associated with T1D at the GLIS3 locus from GUESSFM fine mapping analysis.

Abbreviations

T1D

Type 1 diabetes

HLA

Human leukocyte antigen

FDR

False discovery rate

eQTL

Expression quantitative trait loci

NIDDK

The National Institute of Diabetes and Digestive and Kidney Diseases

NIAID

The National Institute of Allergy and Infectious Diseases

NHGRI

The National Human Genome Research Institute

NICHD

The National Institute of Child Health and Human Development

JDRF

The Juvenile Diabetes Research Foundation

GRID

Genetic resource investigating diabetes

IDDMGEN

Tyypin 1 Diabetekseen Sairastuneita Perheenjäsenineen□

T1DGEN

Tyypin 1 Diabetekseen Genetiikka

T1DGC

Type 1 diabetes genetics consortium

IFN

Interferon