Tag Polymorphisms at the A20 (TNFAIP3) Locus Are Associated With Lower Gene Expression and Increased Risk of Coronary Artery Disease in Type 2 Diabetes

Two populations of Caucasian individuals with type 2 diabetes were studied, one from Boston (n = 479) and the other from San Giovanni Rotondo, Italy (n = 563). The study protocol and informed consent procedures were approved by the local research ethic committees. Each population included a group of case subjects with CAD and a group of CAD-negative control subjects. The selection criteria have been previously described (17). In Boston, the CAD-positive case subjects, defined as subjects who had a stenosis >50% in at least one major coronary artery or their main branches, were recruited among type 2 diabetic patients who underwent cardiac catheterization at the Beth Israel Deaconess Medical Center. CAD-negative control subjects were Joslin Clinic patients (the Joslin Clinic serves as the Beth Israel Deaconess Medical Center Diabetes Clinic) aged 55 years or more, who had diabetes for 5 years or more but had a negative cardiovascular history and a normal exercise treadmill test (18). The Italian population consisted of type 2 diabetic patients who resided in the Gargano area in the central eastern region of Italy and attended the Endocrine Unit of the Institute Casa Sollievo della Sofferenza in San Giovanni Rotondo, Italy. Case subjects were patients who had angiographic evidence of stenosis >50% in at least one major coronary artery or their main branches (n = 189) or had a previous acute myocardial infarction without angiographic documentation (n = 42). Control subjects included diabetic patients without cardiovascular symptoms and with normal resting electrocardiogram and exercise treadmill test (n = 282) or with coronary stenosis (at angiography) ≤50% (n = 50) (17).

Tag SNPs selected on the basis of the linkage disequilibrium pattern as described below were genotyped in the study subjects by means of PCR followed by single base extension/fluorescence polarization (AcycloPrime-FP SNP Detection System) using a Wallac VICTOR2 Multilabel Plate Reader (Perkin-Elmer, Boston, MA). Genotyping quality was tested by including six blinded duplicate samples in each 96-well assay. The average agreement rate of duplicate samples was >99%. Sequences of the primers and probes used for typing are available from the authors.

Data generated by the SeattleSNP initiative in the European descent group were used to identify tag SNPs capturing all common variants at the A20 locus. Tag SNPs were determined by means of the LD Select software using an r² of 0.90 to define linkage disequilibrium bins (19).

Genotype distributions were tested at each polymorphic locus for departure from Hardy-Weinberg equilibrium. For each polymorphic locus, the association with CAD was evaluated by logistic regression analysis including age, sex, BMI, smoking status, hypertension status, and A1C as covariates. The significance of differences in the association of SNPs with CAD between A1C strata was determined by adding appropriate interaction terms (SNP [0,1] × A1C stratum [0,1]) to the logistic regression model. Analyses were conducted in each population separately and in the two population combined. In the latter analysis, an indicator variable for the study population (Boston versus Italy) was added to the model. The significance of differences in the association of SNPs with CAD between study populations was determined by adding an interaction term (SNP [0,1] × study population [0,1]) to the model. Finally, the presence of significant differences between populations in the interaction between SNPs and A1C strata was evaluated by adding a SNP × A1C stratum × population interaction term. The proportion of variability in CAD outcome explained by the SNPs was estimated by calculating the deviance r² (i.e., the proportional decrease in −2 log likelihood generated by the introduction of the SNPs into the logistic regression model). The association between CAD and common (≥0.05) A20 haplotypes was analyzed using Haplo.score in R software (http://www.biostat.wustl.edu/genetics/geneticssoft/manuals/haploscore/haplo.score.html) (20). This method allows adjustment for covariates (in this case, age, sex, BMI, smoking status, hypertension status, A1C, and study population) and provides a global test of association and haplotype-specific tests.

Levels of A20 mRNA were measured by quantitative, real-time PCR in peripheral blood mononuclear cells from 83 nondiabetic employees of the Joslin Diabetes Center (47% men, mean age 31.3 ± 10.2 years, and BMI 24.7 ± 3.7 kg/m²). Peripheral blood mononuclear cells were isolated from peripheral blood by density gradient (Vacutainer CPT; Becton Dickinson, Franklin Lakes, NJ). RNA was extracted using the RNeasy kit (Qiagen, Valencia, CA). A20 mRNA levels relative to β-actin mRNA levels were determined by quantitative RT-PCR using two predeveloped TaqMan Gene Expression Assays from Applied Biosystems (Foster City, CA) according to the manufacturer’s instructions in an ABI PRISM 7000 Sequence Detection System. A20–to–β-actin mRNA ratios were obtained from the equation 2^−ΔCT, where ΔCT is the difference in threshold cycles between A20 and β-actin. The significance of mRNA level differences across genotype groups was estimated by linear regression using log-transformed values to account for the skewed distribution of the A20–-to–β-actin ratios.

The human A20 (TNFAIP3) gene includes nine exons spanning 16 kb on chromosome 6q23. The entire locus has been sequenced by the SeattleSNP initiative in 23 Caucasian subjects, leading to the identification of 22 SNPs having minor allele frequencies >5% in the study population (http://pga.gs.washington.edu/). We used these data to select five tag SNPs representative of all the others on the basis of a pairwise r² ≥ 0.9 (Table 1). The association between these tags and CAD was evaluated in a population of 479 Caucasian individuals with type 2 diabetes from Boston. A second similar set of patients from Italy was used to replicate significant findings. Salient clinical characteristics of both study populations are described in Table 2.

Genotypes were in Hardy-Weinberg equilibrium at each polymorphic locus. Three of the tagging SNPs (rs583522, rs5029930, and rs610604) were significantly associated with CAD in the Boston population (nominal P = 0.042, 0.001, and 0.003, respectively, adjusted for age, sex, BMI, smoking status, hypertension status, and A1C) (Table 3). For both rs5029930 and rs610604, the association was consistent with a dominant model of inheritance, with minor allele carriers having about twice the odds of cardiovascular disease than major allele homozygotes (Table 3). In the case of rs583522, the increase risk of CAD concerned only heterozygotes (odds ratio [OR] 1.7 [95% CI 1.2–2.6]). The three SNPs, although tagging different linkage disequilibrium bins, showed some degree of disequilibrium with each other (r² = 0.67 for rs583522 and rs610604; 0.13 for rs5029930 and rs610604; and 0.08 for rs583522 and rs5029930 in the CAD control subjects). To determine the primary driver(s) of the association with CAD, we conducted a logistic regression analysis applying a backward elimination process to a model that initially included all three tags. Both rs5029930 and rs610604, but not rs583522, were retained in the final model (P = 0.012 and 0.013, respectively), indicating that these two SNPs were independently associated with CAD, whereas the effect of rs583522 was secondary to its linkage disequilibrium with rs610604.

Associations between rs5029930 and rs610604 and CAD were weaker and did not reach statistical significance in the Italian population (Table 4). The ORs for minor allele carriers were 1.4 for rs5029930 and 1.1 for rs610604, and the 95% CIs included one in both cases. We hypothesized that such lack of replication might be due to differences in some of the clinical characteristics of the two populations. The most prominent disparity was in the degree of glycemic control, which was significantly better in Boston than in the Italian population (mean A1C 7.3% in Boston vs. 8.6% in Italy, P < 0.0001) (Table 1). In a stratified analysis of the Boston population by median A1C value (7.0%), it became clear that most of the association between SNPs and CAD observed in this group originated from the lower A1C stratum. For instance, the OR of CAD for rs610604 minor allele carriers was 3.9 in subjects with A1C ≤7.0% compared with 1.2 in subjects with A1C >7.0% (P for interaction = 0.015) (Table 5). A similar interaction between rs610604 and A1C was observed in the Italian subjects (OR 2.2 for those with A1C ≤7.0 vs. 0.9% in the A1C >7.0% group, P for interaction = 0.05). Thus, the weaker association with CAD observed in the Italian group before stratification was explained by the smaller proportion of individuals with A1C ≤7.0% in this population (22%) compared with that from Boston (49%) (Table 5). If data were analyzed by A1C stratum, the ORs of CAD were not significantly different in Boston and Italy, and a significant association (P = 0.05) was also detected in the Italian population in the good glycemic control stratum (Table 5). In the two populations combined, the minor alleles of both SNPs were associated with a threefold increase in the risk of CAD among individuals with low A1C (P = 0.0012 and 8.3 × 10⁻⁶ for rs5029930 and rs610604, respectively). There was no significant effect in the high A1C stratum (Table 5). Together, the two SNPs explained ∼6% of the variability in CAD outcome in the good glycemic control stratum (deviance r² = 0.062) compared with only 0.4% in the poor glycemic control stratum. A similar interaction between SNPs and glycemic control was observed if A1C values of 6.5 or 7.5% rather than 7.0% were used to stratify the study populations (data not shown).

The two SNPs defined four haplotypes, the frequencies of which are shown in Table 6 along with their score statistics for association with CAD in the combined population. In the stratum with A1C ≤7%, the haplotype formed by the two major alleles (haplotype AA) was significantly more frequent in control subjects than case subjects (P = 4.8 × 10⁻⁷). Conversely, the three haplotypes carrying at least one of the minor alleles at rs5029930 or rs610604 (i.e., haplotypes AC, CA, or CC) were all more frequent in case subjects than control subjects, with P values ranging from 0.001 to 0.006. Associations were not seen between haplotypes and CAD in subjects with A1C >7% (Table 6).

None of the SNPs in the linkage disequilibrium bins tagged by rs5029930 or rs610604 encoded changes in A20 amino acid sequence. Thus, we hypothesized that the association between these two SNPs and CAD risk could be due to allelic differences in A20 expression. To test this hypothesis, we measured A20 mRNA levels by real-time, quantitative PCR in peripheral blood mononuclear cells from 83 healthy nondiabetic subjects (47% men, age 31 ± 10 years, and percent ideal body wt 112 ± 16). We chose to study nondiabetic individuals to avoid the potential confounding effects of differences in glycemic control and associated treatments that might have anti-inflammatory effects (e.g., thiazolidinediones). We used two different assays developed by Applied Biosystems, one targeted to exons 2 and 3 and capturing the isoforms corresponding to GenBank entries M59465, BC064689, and BC041790 and the other targeted to exons 1 and 2 and capturing only isoforms M59465 and BC064689 (Fig. 1A). With the first assay, minor allele homozygotes at rs5029930 and rs610604 had 45 and 30% lower A20 mRNA levels, respectively, compared with major allele homozygotes, with heterozygotes having intermediate values (P for trend = 0.028 and P = 0.04, respectively) (Fig. 1B). A similar pattern of association was observed with the second assay, although in this case, significance was reached only for rs610604 (P = 0.042) (Fig. 1C). A significant association between genotypes and mRNA levels was also observed if data were analyzed according to haplotype carrier status (AA homozygotes vs. AA heterozygotes versus carriers of any other genotype, P = 0.018 and 0.038 for the first and second assay, respectively).

A sequence variant in the A20 gene has been previously shown to account for differences in susceptibility to atherosclerosis between mouse strains (12,13). Our results indicate that genetic variants at the A20 locus have a similar effect on atherosclerosis in humans with type 2 diabetes. Whereas the effect in animals is due to an amino acid substitution (13), the effect in humans is mediated by variants affecting the expression of the A20 gene, with lower mRNA levels being associated with increased cardiovascular risk and higher levels with protection. We speculate that lower A20 expression may predispose to CAD by permitting NF-κB to remain more active, with resultant increases in the expression of downstream atherogenic mediators. NF-κB targets have been implicated at many stages in the atherogenic process, including P- and E-selectins and intracellular adhesion molecule–1 and vascular cell adhesion molecule–1, which localize circulating immune cells to the endothelium (5,6). Cytokines and chemokines produced locally in atherosclerotic lesions include MCP-1, macrophage inflammatory proteins (MIP-1α, MIP-1β, MIP-2, and MIP-3α), interferon-γ, lymphotoxin, TNF-α, and interleukin-6 (21). Even a small upregulation in NF-κB activity as might be induced by variants decreasing A20 activity would potentially increase the expression of these targets. Conversely, slightly lower NF-κB activity as presumably occurs in carriers of variants increasing A20 activity might decrease the expression of multiple target genes. It is easy to envision how the orchestrated regulation of numerous gene products could translate into clinically apparent phenotypes.

Other studies have examined genes involved in inflammatory processes as candidate loci for CAD. Among the most notable findings to date is the identification of a haplotype in the HLA region on 6p21 that is associated with myocardial infarction in Japanese and Europeans (22,23). Two of the three genes encompassed by this haplotype, LTA and NFKBIL1, code for modulators of NF-κB, as does the A20/TNFAIP3 gene. LTA encodes lymphotoxin-α, which stimulates NF-κB through its interaction with TNF receptors, whereas NFKBIL1 encodes a protein homologous to the NF-κB inhibitor, IκB (24,25). Thus, the variants that we identified in the A20 gene as being associated with CAD could well be part of a larger group of polymorphisms influencing susceptibility to atherosclerosis through a modulation of NF-κB activity.

The lack of significant association in the Italian study before stratification by A1C levels may raise the concern of a false-positive result in the Boston population. On the other hand, the fact that we could replicate the interaction between SNPs and degree of glycemic control in the Italian population argues in favor of a genuine effect of this locus on cardiovascular risk, even if this is limited to a subset of type 2 diabetic patients. Importantly, the same alleles that confer increased cardiovascular risk were found to be associated with decreased levels of A20 expression in vivo, reinforcing the evidence from the case-control study and providing a plausible biological explanation of the association with CAD.

We do not know why the genetic effect at the A20 locus is more visible in diabetic individuals under better glycemic control, but we envision two possible scenarios. One is that allelic differences in the inhibitory effect of A20 may be overpowered by a massive induction of NF-κB secondary to chronically severe hyperglycemia (26). Under these conditions, whether a person has 30–40% lower A20 expression compared with another may be irrelevant. The modulatory effect of allelic differences in A20 expression may instead be critical in the presence of a more moderate induction of NF-κB such as that elicited by lower (albeit abnormal) glycemic levels. The other possible scenario is that the cellular pathways mediating the increased risk of CAD associated with type 2 diabetes differ in the two glycemic strata. NF-κB–mediated mechanisms may be preponderant in the presence of moderate hyperglycemia, whereas other pathways (e.g., those secondary to PKC activation [4]), may prevail in the presence of more severe hyperglycemia. Under these circumstances, the effect of allelic differences in NF-κB inhibition may be more visible among individuals with relatively good glycemic control than in those with higher glycemic levels.

Two distinct polymorphisms at the A20 locus were independently associated with CAD in the logistic regression analysis. When the two SNPs were considered together as haplotypes, the two major alleles identified a haplotype that was highly protective, whereas all the three other haplotypes were associated with similar predisposition to CAD. One possible interpretation of these findings is that there are two independent genetic effects at this locus, one tagged by rs5029930, the other by rs610604. For both effects, the minor allele is associated with predisposition to CAD, the major allele with protection. The presence of a minor allele at one locus is sufficient to confer predisposition, and the addition of a minor allele at the other locus does not further increase the risk. Another possibility is that there is a single genetic effect due to an unidentified polymorphism having the major allele in linkage disequilibrium with the protective haplotype and the minor allele in linkage disequilibrium with the three predisposing haplotypes.

As far as the nature of the causal polymorphism(s) is concerned, both rs5029930 and rs610604 are within introns, as are the eight SNPs in the corresponding linkage disequilibrium bins. One of these (rs5029928, in the linkage disequilibrium bin of rs5029930) is in a conserved region, where it disrupts potential binding sites for the transcription factors forkhead box I1 and cut-like homeobox. Whether this affects the transcriptional activation of the A20 gene will need to be tested further. SNP rs5029940 is also interesting, consisting of a CCT insertion/deletion in intron 2 at a short distance from exon 3. It must be considered, however, that in the HapMap database, linkage disequilibrium with rs5029930 and rs610604 extends, although in a discontinuous fashion, for ∼50 kb on both sides of the gene (www.hapmap.org). Thus, the functional variants responsible for the observed associations might be in regulatory elements at some distance from the coding sequence.

There are limitations in our study. We based our SNP selection on the pattern of linkage disequilibrium determined by the SeattleSNP initiative through the resequencing of the DNA from 23 Caucasian subjects (27). This sample size had >99% power to detect SNPs with a minor allele frequency (MAF) ≥10% but only 90% power to detect SNPs with an MAF ≥5%, resulting in the possible exclusion of some SNPs with MAF between 5 and 10% from the linkage disequilibrium analysis. Also, some of the r² estimates obtained in this small sample of individuals might have been relatively imprecise, decreasing tagging efficiency. Thus, our study might not have been as comprehensive as one would assume on the basis of our SNP selection strategy. Another potential limitation concerns the selection of CAD control subjects. Because the vast majority of the subjects referred to the catheterization laboratory have significant CAD, we recruited the control group from diabetes clinics on the basis of a negative cardiovascular history, normal resting electrocardiogram, and a normal exercise treadmill test. It is possible that some of the individuals in this group actually had asymptomatic CAD that was missed by the exercise tolerance tests (28). Such misclassification, however, would have biased the results toward the null hypothesis, making our findings of association more notable. Finally, it should be emphasized that these results apply to type 2 diabetes, a condition characterized by insulin resistance and increased propensity to chronic, low-grade inflammation, even in the presence of relatively good glycemic control (15). Whether A20 polymorphisms influence CAD risk in the general population will have to be determined by future studies.

This research was supported by National Institutes of Health Grants HL73168, HL71981, and DK60837 (to A.D.); DK45943, DK51729, and HL83813 (to S.E.S.); DK36836 (Genetics Core of the Diabetes and Endocrinology Research Center at the Joslin Diabetes Center); a grant from the Donald W. Reynolds Foundation (to A.D.); and by the Italian Ministries of Health Grants RC0303ED19 and RC0201ED02 (to V.T.).

We acknowledge the invaluable contribution by the individuals who participated in this study.