Risk of Celiac Disease in Children With Type 1 Diabetes Is Modified by Positivity for HLA-DQB1*02-DQA1*05 andTNF −308A

We performed a case-control study where the case subjects were children with both CD and type 1 diabetes and the control subjects were children with type 1 diabetes only, matched to the case subjects by center, age at type 1 diabetes onset, and duration of diabetes. Optimally, two control subjects should have been identified for each case. Ten centers from seven countries participated in the study. All participating children should have manifested with type 1 diabetes by their 15th birthday, and case subjects should have manifested with CD by their 18th birthday. All available diabetic children with CD cared for by the center were approached and asked to participate. The nature of the study was described to the child, parents, or guardian before a written consent was obtained, in accordance with the procedures approved by the ethics committee at the particular center. Basic clinical and laboratory data were collected retrospectively from the patient’s records. Two matched control subjects were identified for each case. First, the case was assigned to a 5-year band of age at type 1 diabetes onset. Then, two control children were selected at the same center from the same age-at-onset band, one being the child manifesting with type 1 diabetes just before and the other just next to the case. If the selected control subjects were not available or did not agree to participate, the selection went further to the next eligible control subject.

The characteristics of subjects are given in Table 1. The participating centers recruited 130 eligible case subjects with both type 1 diabetes and CD. The case subjects had been identified by endomysium antibody screening that was positive on two separate occasions and confirmed by small bowel biopsy in most instances. The case subjects were matched to 245 control subjects (for single case subjects, two matched counterparts could not be found). The control subjects were deemed free of CD based on the annual endomysium antibody testing. The case and control subjects were European Caucasians except for four subjects of Asian origin (one control subject from Vienna, one case subject from Berlin, and one case and one control subject from Hannover). Several exceptions were allowed from the matching criteria because the center had no better control subjects available: 25 (10%) control subjects were selected from the neighboring lower or higher band of age at type 1 diabetes onset, and seven control subjects were selected although they had manifested with type 1 diabetes after their 15th birthday, with the maximum age at onset being 17 years and 1 month.

Either genetic material was obtained from the center’s DNA archives or EDTA-anticoagulated blood was drawn from the child, sent by mail to the central laboratory, and processed using the QiaAmp Blood DNA extraction kits (Qiagen, Germany). All DNA samples were quantified using a real-time assay for the human albumin gene, and 83 of 375 (22%) samples, which had an insufficient total DNA amount, were pre-amplified using the Genomi-Phi whole-genome amplification kit (Amersham) according to the manufacturer’s instructions. Because all downstream assays were based on PCR with sequence-specific primers (PCR-SSP), the quantity of the DNA was equalized among all samples by diluting, using data from real-time quantification of the human albumin gene.

The genotypes of HLA-DQB1, -DQA1 were determined using an in-house assay based on PCR-SSP. The performance of the assays is controlled by an exchange of samples with the Tissue Typing Laboratory IKEM in Prague, a laboratory accredited by the European Federation of Immunogenetics. The assay for DQA1 and DQB1 typing includes 24 reactions distinguishing the DQA1*01–06 alleles and the DQB1*02, 0301, 0302, 0303, 0304, 0401, 0402, 0501, 0502, 0503, 0601, 0602, 0603, and 0604–9 alleles. Full-typing data were necessary to determine the dose of DQB1*02-DQA1*05 and DQB1*0302 in the individual. The CTLA4 +49A/G was determined using a previously reported PCR-SSP technique (11). The cytokine single nucleotide polymorphism (SNP) PCR-SSP typing was performed using a protocol with the sequence-specific primers developed for the cytokine component of the 13th International Histocompatibility Workshop and Conference (IHWC) (12). The laboratory was controlled within the 13th IHWC proficiency testing and achieved the accuracy of 98%. Seventeen SNPs were tested within 10 different genes for cytokines (interleukin [IL]-1α −889 T/C; IL-1β −511 C/T and +3962 C/T; IL-2 −330 T/G and +160 G/T; IL-6 −174 G/C and +565 G/A; IL-10 haplotypes of three SNPs at −1082, −819, and −590, ACC/GCC/ATA; TNF −308 A/G and −238 G/A; TGFβ1 codon 10 C/T and 25 G/C; IFN-γ UTR 5644 A/T) and for IL-1–related molecules (IL-1R 1970 C/T, IL-1RA 11100 T/C). The design of the sequence-specific primers also allowed us to directly determine the phase of SNPs in haplotypes of TGF-β1, TNF, IL-2, IL-6, and IL-10.

The association of the variants with CD was first tested by conditional logistic regression for matched data. The outcome variable identified whether the individual was a case subject (patient with type 1 diabetes and CD, encoded as 1) or a control subject (patient with type 1 diabetes only, encoded as 0). The identifier of the matching group was used as the conditioning variable. The independent variables were the phenotype positivity for the examined alleles or haplotypes in a univariate model. Then, the variables achieving a P value of the Wald statistics at <0.05 in the univariate analysis were used for step-up building of a conditional logistic regression model. The best-fitted model was determined using the likelihood ratio test statistics. All models were—apart from conditioning for the matching identifier—adjusted for the matching variables. Analyses were performed using the Stata 9.0 program (StataCorp, College Station, TX). The sample size provided an 80% power at α = 0.05 to detect an effect size with an odds ratio (OR) of ≥1.83 for a genetic factor present in 50% of the control subjects (e.g., the HLA-DQB1*02-DQA1*05): this OR corresponds to a ≥65% prevalence in case subjects. Similarly, the study had an 80% power to detect an OR >1.86 for a genetic factor with a 30% prevalence among control subjects or OR >2.07 for a genetic factor with a 70% prevalence.

The results of the univariate analysis of the case-control status versus phenotypic positivity of the studied factors are shown in Table 2. HLA-DQB1*02-DQA1*05 was found in 77% of the case subjects and 53% of the control subjects, being strongly associated with CD among type 1 diabetic subjects (P < 10⁻³). The effect of DQB1*02-DQA1*05 is adequately described by a codominant model taking the gene dose into account; there is a proportional rise in risk with gene dose. HLA-DQB1*0302-DQA1*03 shows a weak negative association with CD. Among the cytokine genes, TNF −308A is strongly associated with CD (P < 10⁻³), whereas the weak negative association of IL1α −889 disappears after correction for multiple testing.

The best-fitted regression model is shown in Table 3. Although HLA-DQB1*0302-DQA1*03 completely lost its effect after being adjusted for HLA-DQB1*02-DQA1*05, both cytokine SNP alleles retained the association observed in the univariate analysis. Homozygozity for HLA-DQB1*02-DQA1*05 increased the risk of CD roughly fourfold, whereas the presence of a single dose of DQB1*02-DQA1*05 doubled the risk relative to individuals carrying no HLA-DQB1*02-DQA1*05. In the final model, both HLA-DQB1*02-DQA1*05 and TNF −308A retained the statistical significance, indicating an independent effect: the risk was roughly doubled also by the presence of the −308A variant of the TNF. While the association of HLA-DQB1*02-DQA1*05 was dose dependent, such an effect could neither be confirmed nor refuted for the TNF −308A dose.IL-1α −889T was associated with a marginal decrease in CD risk. No interaction terms between the variables could further improve the fit.

The results did not change after exclusion of four subjects having other than European Caucasian descent (e.g., the OR for a double dose of DQB1*02-DQA1*05 was 4.45 [95% CI 1.75–11.3], in the whole dataset vs. 4.27 [1.67–10.9], after exclusion of the subjects of non-European descent). Including sex into the analysis as one of the predictors neither improved the fit nor changed the results at the first two decimal positions of the risk estimate.

The present study clearly demonstrates that factors inside HLA significantly modify the risk of CD in children with type 1 diabetes. The HLA-DQB1*02-DQA1*05 increases the risk in a dose-dependent manner, with a significant contribution of TNF −308A.

The observed effect of HLA-DQB1*02-DQA1*05 may have been expected in populations where DQB1*02-DQA1*05 confers a lower risk of type 1 diabetes than CD but has seldom been documented. The first study to indicate that DQB1*02 is overrepresented among diabetic children with CD was published 10 years ago by Saukkonen et al. (13), based on a longitudinal observation of 775 diabetic children that identified 18 case subjects of CD diagnosed after type 1 diabetes onset. A later Czech study suggested that HLA-DQB1*0201-DQA1*05 is associated with a fourfold increased risk of CD in children with type 1 diabetes (14), based on data from 15 children with CD identified from a cohort of 345 diabetic children, of whom 186 were genotyped. However, the CI of the risk estimate was wide. The association was not replicated in an Italian study of 25 children with CD and type 1 diabetes compared with 79 children with type 1 diabetes only (15): the DQB1*0201-DQA1*05 frequency was 68% in the former group and 62% in the latter group. The authors concluded that the previously seen association may have been population specific, which is likely particularly for the Italian population where DQB1*0201-DQA1*05 is frequent among diabetic subjects. Another study comes from Australia (16): an increased frequency of HLA-DQB1*0201-DQA1*05 was seen in diabetic subjects with CD (10/13, 77%) compared with subjects with type 1 diabetes but not CD (70/118, 59%), but this difference was not statistically significant. A clear limitation to these studies is the low number of identified case subjects (diabetic children also having CD): a single-center study therefore has a very substantial risk of type II error, i.e., not refuting the false null hypothesis due to low statistical power. Negative results reported by some of the studies should therefore be interpreted with caution. Two articles were published by a group from Denver, Colorado, who used positivity for the transglutaminase antibody as an end point, thus increasing the number of case subjects: the studies indicate a significant contribution of genetic factors to CD risk in diabetic individuals, since an increased prevalence of transglutaminase antibodies was reported in diabetic children carrying DQB1*02-DQA1*05 or being DQB1*02-DQA1*05 homozygous (17,18). Possible weak points of these two reports are that the antibodies were measured only one time and that analysis on ethnicity is not reported in a population that is likely to contain ethnic minorities.

In the present report, we show a significant effect of HLA-DQB1*02-DQA1*05 positivity that is probably a compound of the effects of DQB1*02-DQA1*05 itself and other genetic factors carried on the DQB1*02-DQA1*05 haplotypes, namely the TNF −308A. Whereas the pathogenic effect of DQB1*02-DQA1*05 in CD is well established and the association is primary, the association of the TNF −308A may be due either to its functional significance as a primary risk modifier or to a linkage disequilibrium to another variant on the haplotype. The former alternative is plausible for TNF-α, a potent proinflammatory cytokine involved in T-cell immunity. However, the data on the functional significance of its polymorphisms are conflicting (rev. in 19), and the TNF −308A may be just a passive marker traveling on a proinflammatory haplotype. The TNF-α gene lies between HLA class II and class I in what is termed the class III region. TNF −308A is known to be associated with the proinflammatory 8.1 ancestral haplotype, so other components of the haplotype may significantly contribute as well. There are many genes with known inflammatory roles in the close vicinity of TNF-α: lymphotoxin α, the heat-shock protein-70 complex, the complement components 4A and 4B, and others.

Although the TNF −308A remains associated when adjusted for DQB1*02-DQA1*05 positivity, this fact does not allow us to deem the genetic association truly independent. Adjusting or matching only for DQB1*02-DQA1*05 phenotypic positivity is not enough, because we do not control the other chromosome in heterozygotes carrying one dose of DQB1*02-DQA1*05 (20). An alternative approach to dissecting the association may be to analyze only subjects positive for a fixed HLA-DQ genotype, e.g., the frequent HLA-DQB1*02-DQA1*05/DQB1*0302-DQA1*03. However, the power of this type of analysis of our dataset is low, because there are only 23 matching groups where a case and at least one control are positive for this genotype. In this small subset, we observed indications that the TNF −308A is associated with an increased risk of CD (OR 2.4 [95% CI 0.63–9.1], P = 0.20).

An optimal design for a future study verifying the independence of the TNF −308A effect in determining the CD risk among diabetic individuals would involve matching for the primarily associated DQB1*02-DQA1*05 on both chromosomes. As seen in Table 2, there are not many DQB1*02-DQA1*05 homozygotes, only 17% among the case subjects and 7% among the control subjects. Thus, for such a relatively rare combination of diseases, it would be especially difficult to recruit a sufficient number of HLA-DQB1*02-DQA1*05 homozygous diabetic individuals with CD and to match them adequately with control subjects with type 1 diabetes but without CD.

The present study was the first to use matching for age at diabetes onset and duration of diabetes. This is because the age at onset is known to be influenced by the child’s HLA genotype and because the cumulative incidence of CD rises with duration of exposure to type 1 diabetes. This careful matching, together with matching for study center, should eliminate possible bias due to the recent shift toward a younger age at type 1 diabetes onset observed in some populations. Another important reason for matching is the heterogeneity in the CD prevalence in diabetic children: the prevalences previously reported from populations in this study range from 3 to 12%, moreover with significant variations over time (14,21–25). Consequently, reliable estimations of absolute risks attributable to particular genotypes are impossible, although knowledge of them might be of interest for the clinician.

An important source of technical error was eliminated by equalizing the DNA content across the samples using real-time quantification. All the genotyping reactions we used were based on PCR with sequence-specific primers, a technique known to be sensitive to uneven DNA content. The adequacy of this approach is well documented by the high success rate in IHWC proficiency testing.

In conclusion, at least in some European populations, the risk of CD in children with type 1 diabetes is substantially modified by the child’s genotype, with a contribution from both HLA-DQB1*02-DQA1*05 and the TNF −308A variant. Although there had been previous indications that children with diabetes differ in their risk of CD based on their HLA genotype, our study has conferred the strongest evidence so far, having collected the largest adequately matched case-control dataset. Although an immediate transition of the knowledge into clinical use is unlikely, in the future, genetic testing may be useful for stratifying the frequency of testing diabetic children for antibodies indicative of CD.

This study was supported by the Czech Ministry of Health, grant no. 7413-3.

We acknowledge Catarina Limbert (Lisbon, Portugal) and Magdalena Paskova (Kosice, Slovakia) for providing the study with DNA samples from their patients.