Genetic Control of Alternative Splicing in the TAP2 Gene: Possible Implication in the Genetics of Type 1 Diabetes

Genomic DNA (gDNA) was obtained after informed consent from type 1 diabetes–affected subjects and their two parents (889 trios or 2,667 individuals after removing families with Mendelian discrepancies at multiple independent single-nucleotide polymorphisms [SNPs]). The Research Ethics Board of the Montreal Children’s Hospital and other participating centers approved the study. Ethnic backgrounds were of mixed European descent, with the largest single group being of Quebec French-Canadian origin (40% of the total collection). All patients were diagnosed as aged <18 years and required insulin treatment continuously from the time of diagnosis.

Six SNPs were genotyped by AcycloPrime-FP SNP Detection kit (PerkinElmer, Boston, MA). PCR primers and fluorescence polarization probes are available from the authors on request. PCR was performed in a Dual 384-Well GeneAmp PCR system 9700 in clear 384-well microplates (Axygen Scientific, Union City, CA). Unincorporated primers and dNTPs were removed according to the PE AcycloPrime PCR Clean-Up protocol. The final extension reaction was performed according to the PE AcycloPrime-FP protocol in Dual 384-well GeneAmp PCR system 9700 in black microplates (Axygen Scientific). Final detection of the SNP was done by the Criterion Analyst HT system (Molecular Devices, Sunnyvale, CA). All six SNPs had unambiguous clusters corresponding to the three genotypes. The average Mendelian error rate is <0.1%.

The sequencing of the HLA-DQB1 exon 2 was based on the reported method (23) with a small modification. The exonic sense primer, 5′-CATGTGCTACTTCACCAACGG-3′, and antisense primer, 5′-GGCGACGACGCTCACCTC-3′, are the same as reported. Considering the polymorphic site in the intron 1/exon 2 primer, a degenerate sense primer spanning the intron 1/exon 2 border, 5′-ATTCCYCGCAGAGGATTTCG-3′, was adapted. To identify the HLA-DQB1 alleles from the sequencing results, Visual Basics for Applications scripts were written to align the sequence results with the allele sequences from the international ImMunoGeneTics project/HLA database (24) (http://www.ebi.ac.uk/imgt/hla). This algorithm permits the identification of the HLA-DQB1*02 alleles from all other DQB1 alleles unambiguously. Of HLA-DQB1*02 alleles, the vast majority (99.8%) are *0201. This allele marks a remarkably conserved haplotype that removes most of the complexities of the DR-DQ region.

Transmission disequilibrium test, LD analysis of SNPs were performed by Haploview software (available at www.broad.mit.edu/personal/jcbarret/haploview) (25). Haplotype association was tested by the TRANSMIT software (available at http://www-gene.cimr.cam.ac.uk/clayton/software/) (26). By logistic regression, the odds ratio (OR) was estimated based on the methods described by Lohmueller et al. (27).

RNA was prepared from lymphoblastoid cell lines immortalized from our type 1 diabetes family collection and used to determine the relative abundance of each SNP allele in RNA representing each of the two splicing isoforms by single-nucleotide primer extension. After RT-PCR with primers specific for each isoform (Fig. 1), an internal oligonucleotide probe immediately upstream of the polymorphic base was extended using ddNTPs corresponding to the two alleles and labeled with different fluorochromes (SnapShot kit; Applied Biosystems, Foster City, CA). Labeled extension products were visualized by capillary electrophoresis. Sequences of primers and probes are given in online appendix Table 1 (available at http://diabetes.diabetesjournals.org).

Allelic proportions in heterozygous gDNA PCR products were used to establish 1:1 stoichiometry and correct for different efficiencies of the two ddNTPs and fluorochromes. To assess the relative expression levels of each allele, the ratio of two alleles in the RT-PCR product was corrected by the average allelic proportion of the gDNA PCR product. Allelic proportions in RT-PCR products outside the 95% CI for gDNA were taken as showing allelic imbalance (AI), i.e., different representation of two alleles in the given isoform. The significance of allelic expression differences was tested by paired t test. AI association with a given SNP was assessed by χ² test (28). Haplotypes of all five SNPs in all subjects used in the AI studies was unambiguously phased by genotyping family members.

General-population allele frequencies of the SNPs examined are given in online appendix Table 2. By the transmission disequilibrium test, we confirm previous observations that, except for the relatively rare SNP rs2228396, all the other five SNPs show a highly significant type 1 diabetes association (Table 1). The five coding SNPs are in tight LD (Fig. 2). Six haplotypes with frequency >1% composed of these five SNPs account for 99.5% of chromosomes in our type 1 diabetes nuclear family collection (Table 2). Each haplotype corresponds to one TAP2 type defined by the 12th HLA Workshop (29). Dramatic type 1 diabetes association can be seen in the four common haplotypes with a frequency >5%, while type 1 diabetes association of the two more rare haplotypes cannot be excluded because of the low frequencies. Both individual SNP analysis and haplotype analysis suggest that two SNPs in tight LD (r² = 0.990), 665 Thr>Ala (rs241447) and 687 Gln>Ter (rs241448), are the most highly associated. Their position in an alignment with the known TAP1 tertiary structure is shown in online appendix Fig. 1. Because r² = 0.99 between these SNPs, two of the four possible haplotypes accounted for >99% of chromosomes; AT was overtransmitted and GC undertransmitted.

For this reason, we proceeded to study the effect of these two haplotypes on splicing. Because of the high LD in the HLA region, dissecting this genetic effect from that of the adjacent DR-DQ locus requires regression analysis of large datasets. Evaluation of the functional significance of the polymorphisms involved may help focus such analysis.

The relative abundance of each SNP haplotype in RT-PCR products specific for each isoform, in RNA from six unrelated heterozygous subjects, was determined by sinlge-nucleotide primer extension using the synonymous G/A SNP at 604 Gly (rs241441). The six subjects used for the allelic expression study were selected for being heterozygous for the two most highly type 1 diabetes–associated SNPs, 665 Thr>Ala and 687Gln>Ter. Because these SNPs do not exist in the transcribed region of NM_018833, we used the synonymous SNP rs241441 as the allelic expression marker. This SNP is known to be in complete LD with the other two (29) and unambiguously marks the haplotypes. In all six individuals, the G allele at this SNP marks the GC haplotype at 665 Thr>Ala and 687Gln>Ter, while the A marks AT (Table 3). For isoform NM_000544, the results were also confirmed using 687 Gln>Ter, with nearly identical results. A representative electrophoresis profile is shown in Fig. 3, and individual results detailed in Table 3. At NM_000544, the average ratio of GC over AT is 2.23 ± 0.44 (mean ± SD) (P = 1.55 × 10⁻⁴). Thus, in heterozygotes, approximately twice as much of that transcript is derived from the GC haplotype as from AT.

Conversely, transcript NM_018833 is almost exclusively derived from chromosomes containing the AT haplotype at the two associated SNPs (AT-to-GC ratio 18.06 ± 5.58, P = 2.04 × 10⁻⁷). The effect is clearly due to these haplotypes (χ² = 8.3, P = 0.004). Other common exonic SNPs are not involved, as differential splicing is clearly seen in individuals homozygous for them (Table 3). Based on these results, it can be calculated that ∼92% of mRNA carrying GC is spliced into NM_000544, while 75% of mRNA with AT gives NM_000544.

As a first approach to dealing with the complexities of the HLA locus, we looked for the presence of a distortion in the transmission of haplotypes carrying each TAP2 allele, in cis to a DQB1*0201 allele. DQB1*0201 is common in the general population and much more so in individuals with type 1 diabetes. It is in tight LD with DRB1*0301, DRA*0102, and DQA1*0501 as part of a remarkably conserved, type 1 diabetes–predisposing haplotype (HLA-DR3) (30). Thus, if only chromosomes carrying this haplotype are examined, distortion in the transmission of TAP2 alleles must be independent of DR-DQ.

Table 4 shows the results of transmission analysis of those haplotypes. Of the 742 parental chromosomes carrying DQB1*0201, irrespective of TAP2 allele, 472 were transmitted to the offspring, whereas 270 were not (P = 1.1 × 10⁻¹³), confirming the well-known predisposing effect of this allele. Most or all of this distortion is attributable to the common (because of LD) haplotype DQB1*0201-TAP2*Thr. The less common DQB1*02-rs241447*Ala is transmitted at a rate no different from the expected 50%. This is consistent with the protective effect of the Ala allele (Table 1) that, in these haplotypes, appears to counterbalance the predisposing effect of DQB1*0201.The transmission ratios of the two haplotypes are significantly different (χ² = 6.5, P = 0.011). Since the two haplotypes are largely identical at DR-DQ, the difference must be driven by the TAP2 polymorphism (or a genetic variant in LD with it).

It has long been known that the HLA locus on chromosome 6p21 accounts for approximately half of the familial clustering of type 1 diabetes (1). However the multiplicity of genes, their allelic complexity, and extensive LD at this locus have made it difficult to define the contribution of the various genes mapping to this area and the mechanisms involved. The strongest associations detected to date map to the DR-DQ region, but it is equally clear that the LD block containing these two crucial genes does not explain all of the linkage to the region (31). Recent high-density SNP genotyping in the region revealed complex patterns of LD over an extended region of >4 Mb (20) and evidence of strong, recent positive selection (16,17,21). Whether polymorphisms in any of a number of genes in the region may be independently contributing to type 1 diabetes risk can only be determined with regression analysis of densely genotyped large DNA sets. It is a reasonable expectation that such complex analysis may be greatly aided by functional clues.

Functionally, the TAP genes are prime candidates for association with autoimmune disease because they encode proteins that directly interact with antigenic epitopes. TAP2, specifically, has an unusually large number of nsSNPs with minor allele frequency >0.05. These may have been driven by the same evolutionary forces that have created the coding hypervariability in HLA molecules through positive selection of new mutations that broaden the spectrum of epitopes to accommodate emerging pathogens. Indeed, differential antigen selectivity has been reported with allelic TAP2 variants between strains of rats (18), and there are also clear differential selectivities across species (32). One report of differential selectivity among human protein–coding alleles (19) has not been substantiated, but differences between isoforms have been shown: The model peptide IYLGPFSPNVTL has a 30-fold lower affinity for isoform NM_018833 than for isoform NM_000544, while, conversely, TVDNKTRYE had 7-fold lower affinity for NM_000544 (22). It is worth noting that TVDNKTRYE also shows marked selectivity for rat TAP2 alleles (33).

Such isoform selectivity would be irrelevant to genetic association with the locus, unless alternative splicing generating each isoform is genetically determined. The data presented in this study clearly show that the diabetes-predisposing haplotype favors formation of NM_018833. An alternative possibility for the AI seen in NM_000544 is nonsense-mediated decay of the allele creating a termination codon at position 687 (rs241448). However, nonsense-mediated decay is typically seen when a premature termination codon occurs >50–55 bp upstream from the last exon-exon junction (34,35), while rs241448 is well downstream of this position. In addition, nonsense-mediated decay cannot explain the AI seen in NM_018833, which does not contain this SNP.

In addition to demonstrating the functional effect of the TAP2 polymorphisms, we show that part of the previously known association of these variants with type 1 diabetes is clearly independent of DR-DQ, at least when in occurs on DQ*02 chromosomes. On the basis of our current genetic data, we cannot distinguish whether such independent contribution can be attributed to the polymorphisms responsible for this functional effect or due to LD with a remote variant of another gene in this genetically complex region. The two possibilities are, of course, not mutually exclusive, as multiple contributions to the linkage of type 1 diabetes to 6p21 are quite plausible. Conclusively dissecting genetic effects in this region of strong and extensive LD reqires the power of large sample sizes. It is to be hoped that the several thousand affected sibling pairs being assembled by the type 1 diabetes genetics consortium (http://t1dgc.org) will provide a conclusive answer, including confirmation of the finding reported here. Regardless, we propose that knowledge of the functional effects of key associated variants, such as the one we describe in this study, will be a crucial adjunct to genetic studies toward understanding the role of the locus in the pathogenesis of type 1 diabetes and, possibly, other autoimmune diseases.

This study was funded by the Juvenile Diabetes Research Foundation and Genome Canada. H.-Q.Q. is supported by a fellowship from the Montreal Children’s Hospital Foundation.

We thank all participating families and the personnel of collaborating clinics.