Relative predispositional effects of HLA class II DRB1-DQB1 haplotypes and genotypes on type 1 diabetes: a meta-analysis

Study populations

Notation: Because of the varying definition of HLA typing in various studies where four-digit resolution was not available for a given DRB1 or DQB1 allele in every study we have used the notation DRB1*xx00, for example DRB1*0100 to indicate that DR1 subtyping was available for some but not for all samples.

A total of 38 nonoverlapping studies of HLA DRB1-DQB1 haplotypes and T1D were analyzed from 30 countries from the following geographic regions: Sub-Saharan Africa (2 studies), North Africa (2 studies), Europe (18 studies), Middle East (3 studies), North East Asia (4 studies), South East Asia (5 studies), Oceania (1 study), and three admixed populations (Table 1). An additional seven studies supplied data only for analysis of the DR4 subsets of DRB1*0400 DQB1*0302 haplotypes. For the genotype analyses, a total of 15 populations from 14 countries were studied. In addition, for the individual genotype analyses, combined European data from Koeleman et al. (17) were studied.

Sample sizes and the numbers of HLA DRB1-DQB1 haplotypes and genotypes studied from each population are listed in Table 1. Each study had on average 362 T1D patients (range, 2n= 74–2104, where n is the number of chromosomes) and 492 controls (range, 2n= 78–3834). For inclusion in the study, high-resolution typing was preferred and was required for DR4 subtypes. Some relatively older data sets are included especially if they were from a region of particular interest, for example Sub-Saharan Africa. For description of the individual studies, see the papers referenced in Table 1, although much of the data analyzed were supplied by the authors as the complete haplotype and/or genotype data were an extension of that listed in the papers. For a haplotype or genotype to be included in our analyses, the combined patient and control count had to be four or more.

For family data, the control numbers represent the total number of Affected Family BAsed Controls (AFBACs) (parental nontransmitted haplotypes) except for transmitted/nontransmitted data where it is the number of nontransmitted haplotypes (Table 1).

For countries with multiple studies (Table 1), a combined P/C ratio is listed for that country, resulting in a total of 31 data sets from 30 countries from the 38 studies. Sardinia was considered distinct from mainland Italy because of known differences between these regions. Where more than one study was available for a country, the P/C ratio was the average of the various studies regardless of the sample size in each study. Where the genotyping resolution of the various studies was not consistent (e.g. DRB1*0100 or DRB1*0101), we have used the lowest resolution so that the specificity of many alleles is limited to two-digit resolution. However, in individual studies, analyses can of course be performed at the four-digit level when this information is available.

Methods

The P/C ratio

To compare populations, we introduce a new measure, the P/C ratio, under the assumption that the HLA DR-DQ genes are directly involved in disease:

where freq(DR-DQ) denotes the frequency of the haplotype or genotype under consideration (or allele or amino acid as appropriate). This equation can also be appropriately modified for any other genetic system where a disease-predisposing locus has been identified and may include one amino acid site, or gene, or multiple genes in a region.

HLA DR-DQ genotype frequencies in the general (control) population are denoted by f(A_iA_j) and absolute penetrance values by w_ij (i≤j), where i, j= 1,2, …, k (Table 2). (The DR-DQ locus is considered as one multi-locus site denoted by locus A.) Note that these w_ij values represent penetrance values averaged over the effects of other T1D loci, both HLA and non-HLA. In the simplest case, the HLA DR-DQ effects would be independent of these other effects and would affect the penetrance values equally by a constant multiplicative factor. Haplotype DR-DQ frequencies in the control population are denoted by f(A_i), and their marginal (averaged over genotypes) absolute penetrance values by w_i·. The genotype and haplotype frequencies in the general (control) and patient populations are given in Table 2, along with the P/C ratios. The analogy between disease models and selection models without recombination has previously been noted (6) and is illustrated here.

Table 2. Patient and control frequencies and P/C ratios with genes directly involved in disease

	Genotypes	Haplotypes
	AiAj	Ai
Controls	f(AiAj)	f(Ai)
Patient population	wij f(AiAj)/T	wi · f(Ai)/T
P/C	wij/T	wi ·/T

• P/C, patient/control.T=Σw_ij f(A_iA_j) is the disease prevalence.

The P/C ratios give a maximum likelihood estimate of the relative penetrance values for each genotype (f_ij=w_ij/T) and haplotype (f_i.=w_i·/T), i.e. the absolute penetrance values normalized by the disease prevalence in that population (proof not shown) (Table 2). Unfortunately, the P/C ratios per se cannot be directly compared across populations and particularly across ethnic groups. However, within a population, the ratio for two different genotypes or haplotypes of their P/C ratios is a function only of the absolute penetrance values, allowing comparisons across studies and ethnic groups, i.e.

The population prevalence parameter T cancels from the relative penetrance estimates in this case, allowing direct comparisons between risk estimates from populations with different prevalences.

Hence, the ratio of two P/C ratios estimates the ratios of the absolute penetrance values for the two haplotypes or genotypes under consideration and is equivalent to the odds ratio for the comparison of these two haplotypes or genotypes. Koeleman et al. (17) applied a similar approach in that they used a consistent base of the DRB1*0101 DQB1*0501 (DR1) haplotype in their analyses of DRB1 DQA1 DQB1 haplotype effects in five European populations. The absolute penetrance values may vary between populations because they include the averaged effects of non-HLA genes, environmental factors, and other HLA genes. As mentioned above, in the simplest case, the ratio of the absolute penetrance values for the DR-DQ genotypes and haplotypes would be the same across all studies; and this is our null hypothesis.

Ranking

The rank of a haplotype or genotype was assigned 1 = most predisposing = highest P/C ratio, so that a larger ranking number corresponds to a more protective haplotype or genotype. If two haplotypes or genotypes had the exact P/C ratio, they were given the same rank. The relative ranks of haplotypes and genotypes within a study can be used in across study comparisons.

Pairwise tests

Within each population where both haplotypes were present [e.g. DR3 (DRB1*0301-DQB1*0201) and DRB1*0401 DQB1*0302], the haplotype counts in patients and controls were compared in a 2 × 2 contingency table. The P value from the corresponding Pearson’s chi-square was computed. Standard linear regression methods were carried out to test for correlation and significance level between pairs of populations or between populations and the mean P/C ratio for a given haplotype using S-plus 6.0 (Insightful Corp, Seattle, WA).

P/C ratios and relative ranks of DR-DQ haplotypes

In Table 3, the P/C ratios are listed by country, with the DRB1-DQB1 haplotypes listed in order of their mean P/C ratios, from most predisposing DRB1*0405 DQB1*0302 to most protective DRB1*1500 DQB1*0602. The average rank is also considered, and there is good agreement between ordering by the mean P/C ratios and the average rank. The same general pattern was also seen using all studies (data not shown). We must note that neither measure is perfect; P/C ratios will vary across populations and particularly across ethnic regions as they are a function of disease prevalence, and most haplotypes do not occur in all studies; a factor that also makes the average ranking not a perfect measure. However, combined with the pairwise analyses, the ordering from most predisposing to least predisposing T1D haplotypes is in general consistent.

Table 3. DRB1-DQB1 haplotype patient/control ratios and ranking with regards to susceptibility to type 1 diabetes

For all pairs of populations that had at least five haplotypes in common, we assessed how correlated were the P/C ratios. Over 255 such comparisons, we found an average Pearson’s correlation coefficient of 0.804 (R²= 0.65) and an average statistical significance of P < 0.0002.

In addition, for the 21 haplotypes present in at least 11 countries, we explored the correlation between the observed P/C ratio in individual populations and the mean PC ratios over all populations. By linear regression, we computed the square of the correlation coefficient and the P value for the four ethnic groupings (Figure 1). The overall correlation between mean P/C ratio and individual population P/C ratio yielded an R²= 0.49 and P < 5 × 10⁻⁶⁶. Interestingly, the best correlation between the overall mean P/C ratio and the individual ones for populations was seen in the East Asian countries. The worst fit was seen in the three countries from the Americas.

In this study, we use the mean P/C ratios to give a relative initial ordering of RPEs. Statistical differences can be inferred when specific combinations are compared within and across studies as shown below, using either 2 × 2 contingency table comparisons or the relative rank orders in each study.

General trends are as previously well documented. There is obviously a continuous gradient from highly predisposing to very protective haplotypes. However, we will continue the tradition, which is reasonable, of often referring to common haplotypes in categories of highly predisposing through to predisposing [the DRB1*0405, 0401, and 0402 haplotypes with DQB1*0302, the DR3 haplotype (DRB1*0301 DQB1*0201) and the DRB1*0404 DQB1*0302 haplotype], then a ‘neutral’ (intermediate) category including DRB1*0800 DQB1*0402, DRB1*0901 DQB1*0303 (DR9), and DRB1*0100 DQB1*0501 (DR1), followed by the protective to very protective category culminating in the most protective haplotypes such as DRB1*1400 DQB1*0503 and DRB1*1500 DQB1*0602.

Meta-analyses comparing relative ranks of pairs of haplotypes can be carried out from this table, for example, comparing DR3 and DR9 haplotypes, the P/C ratio for DR3 is greater than that for DR9 in all 15 data sets where they both occur (Table 3), and in fact considerably larger in all cases. Similarly, as is well known, DRB1*0401 DQB1*0302 is strongly and consistently more predisposing than the DRB1*0401 DQB1*0301 haplotype; this original observation led of course to the Asp at DQB1 position 57 protective vs non-Asp predisposing hypothesis. For these and other comparisons, no distinct ethnic group differences are seen. Other comparisons will be considered below.

Pairwise tests including single Amino acid site comparisons

In combination with the meta-analysis across populations (Table 3), we performed pairwise tests for a number of haplotype combinations (Table 4), in particular those that were present in a number of countries and at high frequency in some countries. We subdivided the haplotypes into sets of interest for studying their relative predisposition (Table 4A,B, DR3, and DRB1*0401, 0402, 0405, and 0404 with DQB1*0302), neutral (Table 4C,D), and protective (Table 4E,F). The number of populations where both haplotypes tested were present is shown, as is the percent where one or the other had a higher P/C ratio and the percent (over the total number of populations with this haplotype) where this difference was statistically significant with P < 0.05. When significant pairwise comparisons involve only one allele of DRB1 or DQB1 but not both being different, then the amino acids will be mentioned when these are of interest, i.e. if only one or a few amino acids are involved.

With these tests, we show that overall the DR3 haplotype is significantly less predisposing than the DRB1*0405, *0401 and *0402 with DQB1*0302 haplotypes (11.5%, 46.5%, and 5.3%, respectively, of countries with significant effects), but the DR3 haplotype is significantly more predisposing than DRB1*0404 DQB1*0302 (17.9% of countries with significant effects) (Table 4A). Within these DR4 haplotypes, the pairwise comparisons show no significant differences between DRB1*0405, *0401 and *0402 with DQB1*0302, but as expected from the DR3 results, these three DR4 haplotypes are significantly more predisposing than DRB1*0404 DQB1*0302 (Table 4B) (30.2% of countries show significant differences between DRB1*0401 and *0404).

For the intermediate-risk or neutral haplotypes (Table 4C,D), DRB1*0800 DQB1*0402 and DRB1*0901 DQB1*0303 (DR9) while not significantly different from each other are both significantly slightly more predisposing than DRB1*0100 DQB1*0501 (DR1) (10.7% and 11.8% of countries, respectively, show significant differences).

For protective haplotypes (Table 4E,F), DRB1*1500 DQB1*0602 is slightly more protective than DRB1*1400 DQB1*0503 (in 62.5% vs 37.5% of populations, but significant effects are divided equally between 4.2% of countries), and both these haplotypes are significantly more protective than DRB1*1100 DQB1*0301, DRB1*0701 DQB1*0303, and DRB1*1300 DQB1*0603.

Of interest also is further comparison of the DRB1*0400 DQB1*0302 haplotypes (4, 5, 12, 14). In accord with previous observations, the protective DRB1*0403 haplotype has a lower P/C ratio than the predisposing DRB1*0404 haplotype in 17 vs 3 populations (and these three cases involve at least one rare haplotype) and very highly significantly differences are seen in four populations. Comparisons involving DRB1*0406, 0407, and 0408 are more difficult because of their lower frequencies. DRB1*0407 is more predisposing than DRB1*0403 in six vs two populations (the latter two involve rare haplotypes) and is significantly more predisposing in three populations. A slightly higher predisposing effect for DRB1*0408 over DRB1*0404 as indicated in Table 3 has no statistical support; similarly for DRB1*0407 vs DRB1*0404, except that in the Mexican mestizos sample, there is highly significant (P < 0.00001) support for DRB1*0404 being more predisposing than DRB1*0407. In comparison of the protective DRB1*0403 and DRB1*0406 haplotypes, there is only a very slight trend (6:2) toward greater protection of the latter. Thus, our results for DRB1*0400 DQB1*0302 haplotypes can be summarized as follows, with the placement of those in parentheses lacking statistical support: 0405 = 0401 = 0402 > 0404 (0407, 0408) > 0403 (0406).

Amino acid site differences at DRB1 involving one amino acid are as follows (with the more predisposing allele listed first): *0405 vs*0408 [#57 S(Ser) vs D(Asp)], *0401 vs*0408 [#71 K(Lys) vs R(Arg)], *0407 vs*0403 [#86 G(Gly) vs V(Val)]. Note, however, that both *0405 and *0401 differ from *0402 at position 86 with again G vs V (as well as at other sites #57, 67, 70, and 71), and because these three DRB1*0400 alleles are not significantly different in risk, these results implicate combinations of sites as risk factors rather than single amino acids.

The protective effect of the DRB1*0403 allele with the predisposing DQB1*0302 allele balances out to be equivalent to the overall protective effect of the DQB1*0301 allele on DRB1*0401 haplotypes (no significant overall or individual comparisons).

For DRB1*0405 haplotypes, in three populations those with DQB1*0302 were more predisposing than those with DQB1*0201, and this difference was very highly significant in the Sardinian sample (P < 0.0000001). DQB1*0302 was more predisposing than DQB1*0401 in four of four samples, with significance in two comparisons. A number of amino acid differences separate these pairwise contrasting allele predisposition effects.

Amino acid sites

We have selected a set of residues in DRB1 and DQB1 that have either been implicated in T1D susceptibility by various authors or are in linkage disequilibrium with DQ-alpha residues that have, as DQA1 genotyping was not available in most studies (6, 12, 16–25) (Table 5). We do not claim that the residues selected can explain susceptibility to T1D, but rather we wanted to explore how consistent such combinations would be across haplotypes and ethnic boundaries.

Table 5. Patient/control ratios and ranking of DRB1-DQB1 haplotypes as defined by 10 amino acid residues with regards to susceptibility to type 1 diabetes

Where the genotype resolution available does not allow us to establish allele subtype (e.g. DQB1*0400), an ‘x’ has been used at any residue that is undetermined. If two or more alleles correspond to a given residue combination (e.g. ‘DLQRAG’ can be DR1 or DRB1*0408) but the haplotype amino acid combination does not occur in any population (e.g. DRB1*0408 DQB1*0501 was not seen in any of the populations at least once), then only the actually observed haplotype combination corresponding to a given residue sequence is presented. ‘Inconsistent’ refers to haplotypes or genotypes that have a P/C ratio > 1 in one or more populations and, also, <1 in one or more populations. The P/C ratios that are discordant with the mean ratio for that haplotype or genotype have been highlighted in gray in Table 5 and Table 7.

Table 7. Patient/control ratios and ranking of DRB1-DQB1 genotypes as defined by 10 amino acid residues with regards to susceptibility to type 1 diabetes

Although these amino acids have been shown by the haplotype method to explain much of the HLA DR-DQ risk (6), nonetheless there is not a consistent hierarchical pattern of risk involving these amino acids per se. Combinations of these amino acids in specific allelic combinations appear to confer predisposing vs protective properties on the HLA DR-DQ molecules, and various combinations of amino acids can give equivalent risks.

P/C ratios and relative ranks of DR-DQ genotypes

In Table 6, the P/C genotype ratios are listed by country, with the DRB1-DQB1 genotypes listed in order of their mean P/C ratios, from most predisposing to most protective. Note that with the genotype analyses, less confidence can be placed on the P/C values because in many cases, the control genotype counts were small. We have investigated various hypotheses using both the overall rankings in Table 6 and pairwise comparisons in individual populations. Table 7 includes amino acid data as well the genotype data.