|Home | About | Journals | Submit | Contact Us | Français|
Elevated serum soluble E-selectin levels have been associated with a number of diseases. Although E-selectin levels are heritable, little is known about the specific genetic factors involved. E-selectin levels have been associated with the ABO blood group phenotype.
We performed a high-resolution genome-wide association study of serum soluble E-selectin levels in 685 white individuals with type 1 diabetes from the Diabetes Control and Complications Trial (DCCT)/Epidemiology of Diabetes Intervention and Complications (EDIC) study to identify major loci influencing levels. Highly significant evidence for association (P=10−29) was observed for rs579459 near the ABO blood group gene, accounting for 19% of the variance in E-selectin levels. Levels of E-selectin were higher in O/O than O/A heterozygotes, which were likewise higher than A/A genotypes. Analysis of subgroups of A alleles reveals heterogeneity in the association, and even after this was accounted for, an intron 1 SNP remained significantly associated. We replicate the ABO association in nondiabetic individuals.
ABO is a major locus for serum soluble E-selectin levels. We excluded population stratification, fine-mapped the association to sub-A alleles, and also document association with additional variation in the ABO region.
Endothelial cells express leukocyte-specific cell adhesion molecules both constitutively and in response to cytokines and other mediators.1,2 Cellular adhesion molecules mediate the attachment and transmigration of leukocytes across the endothelial surface and are hypothesized to play an important role in the initiation of atherosclerosis.3,4 Binding of leukocytes to the endothelium requires the interaction of integrins on the surface of leukocytes with intracellular adhesion molecule-1 (ICAM-1), vascular cell adhesion molecule-1 (VCAM-1), and the selectins on the endothelial cells. Soluble forms of these adhesion molecules are present in endothelial cell culture supernatants and human sera and there is correlation of levels of soluble cell adhesion molecules5 with their membrane-bound expression.
Serum soluble E-selectin levels have been variably associated with a number of common diseases, including atherosclerosis,3,4,6,7 Type 1,7–9 and Type 2 diabetes.10 Investigations into the relationship of E-selectin levels to long-term diabetic complications have however produced inconsistent results. In a nested case-control analysis of the EURODIAB Prospective Complications Study (n=543), patients with macro- or microvascular complications had higher E-selectin levels than those without long-term complications.11 In another report from EURODIAB (n=540), soluble E-selectin levels were positively associated with retinopathy, albuminuria, and cardiovascular disease in univariate analysis.12 However, after adjustment for other risk factors, including HbA1c, the association between E-selectin levels and outcomes did not remain significant at the 5% level. A nested case-control study found E-selectin levels higher among probands with type 1 diabetes who subsequently developed coronary artery disease than among diabetics who did not.13 In a recent investigation of a subset of DCCT/EDIC participants, serum soluble E-selectin was associated with microalbuminuria (P<0.01), and the association remained significant even after adjustment for conventional risk factors (P<0.01).14 Most of these studies were cross-sectional, so it is not possible to separate causal association from potential confounding, due perhaps to a common risk factor such as glycemic control.
Family studies have demonstrated that serum soluble E-selectin levels (SELE, endothelial adhesion molecule 1, CD62E, chr1:167,958,406 to 167,969,803, build 36.3 position) are heritable (additive heritability of 0.50±0.11).15,16 To date, however, no genome-wide linkage or association studies of E-selectin levels have been reported, so the exact genetic basis for the heritability is unknown. A few reports of association of E-selectin levels with specific polymorphisms in SELE7 are available.
One study suggested that the classical ABO blood group phenotype is associated with E-selectin (P<0.001), with the following ranking of mean trait values by blood group phenotype: O≈B>AB≈A.17 Although all subjects were from Manchester, UK, no information was provided regarding the ethnicity of the subjects. The association could therefore be a false-positive because of confounding by population stratification, which is present for ABO in the UK.18–20 There was no attempt to fine-map the association—thus it is possible that the observed association was due to variation in another gene that is in linkage disequilibrium with ABO. Further, because blood group phenotyping cannot distinguish between certain genotypes (eg, A/A and A/O, etc) the effects of all ABO genotypes was not described. The only other report of association came from a case-control study of peripheral arterial disease, where ABO—E-selectin association was detected in both 200 cases and 213 controls (P<0.01).21 However, the controls from this study mostly overlapped with individuals from the earlier study (Andrew Blann, personal communication),17 and the specific direction of the association in the cases was not specified. These authors estimated that ABO blood group and peripheral arterial disease case status accounted for 20.7% of the variance in E-selectin levels.21
Using data from a high-resolution genome-wide association study, the purpose of the study reported herein was to identify and fine-map major loci for serum soluble E-selectin levels and estimate their effect sizes in individuals with type 1 diabetes, and confirm the associations in nondiabetic individuals.
We performed a genome-wide association study of soluble E-selectin in 685 DCCT/EDIC subjects. In an initial screen, we tested for association between E-selectin levels and the genotypes at each SNP using the Kruskal—Wallis test, which does not assume a particular genetic model and allows the detection of differences among any of the 3 genotype groups. This approach is robust, not requiring that the biomarker data are normally distributed, and we report asymptotic probability values. Generally, probability values <10−7 are considered to be genome-wide significant given the multiple testing performed.22The maximum sample size (individuals with both biomarker and genotype data) was 685. Linear regression was used to investigate the relationship of E-selectin level with multiple SNPs, covariates, and principal components from population stratification analyses. Covariates known to affect E-selectin levels were measured at the time closest to the collection of samples for E-selectin. For this analysis, E-selectin levels were winsorized below the 1st and above the 99th percentiles (a standard statistical approach to make outliers less influential) and log transformed to produce an approximate normal distribution. We present results based on Type 3 sum of squares from PROC GENMOD (SAS). All analysis was performed using SAS v9.1.3.
The mean (±SD) serum soluble E-selectin levels in the 685 white EDIC participants was 58.4±36.1 ng/mL (median=51.8, range 8.2 to 390.4) with a right-skewed distribution, which was approximately normalized by winsorization and ln transformation (Figure 1A and 1B). Descriptive information of subjects used for the analysis, as well as other variables measured at the time of collection of the samples for the E-selectin assay, are provided in supplemental Table 1. With the available data, there was good power (1-β=0.8, α=5×10−8) to identify a locus that accounts for 6% of the trait variance, assuming an additive genetic model.
In genome-wide association analysis, several SNPs in a region of chromosome 9 had highly significant association with E-selectin levels. No other SNPs in the remainder of the genome met genome-wide statistical significance (P<10−7)—see results at Phs000086.v1.p1 at dbGaP (www.ncbi.nlm.nih.gov) for the results for all SNPs analysed on the 1M assay. Specifically, the statistically significant SNPs surrounded the ABO blood group gene on chromosome 9 (ABO, Figure 2). The most striking result was for rs579459, located about 3.5 kb 5′ of ABO, probability value=1×10−29, clearly meeting the criteria for genome-wide significance.22 Twenty-six other SNPs in a 230-kb interval around the ABO gene on chromosome 9 all had probability values <10−6 (Table and Figure 3). The quality of the genotype data for each of the associated SNPs met all the predefined quality control criteria, as indicated by the low number of individuals with missing genotype data at any SNP, and the absence of marked deviation from Hardy-Weinberg equilibrium (Table).
It is evident from examination of the mean E-selectin values by genotype at rs579459 (Table) that homozygotes for the common allele (T/T) have the highest values, followed by the heterozygotes (C/T) and then the rare homozygotes (C/C). The trend at this SNP is consistent with an additive genetic model. Using linear regression, we estimated that this locus accounts for 19% of the variance in E-selectin levels in this cohort, however this estimate may be inflated by the Winner's curse.23 No significant evidence (probability value <0.01) of association was observed with SNPs in or near the E-selectin (SELE) gene.
Given the strength of findings at the ABO region, we implemented a specific assay to genotype the variant (rs8176719) that distinguishes the O blood group from A and B, because it was not present on the Illumina 1M chip. This variant was also highly significantly associated with E-selectin levels (P=10−19, Table). Further, using the combination of genotypes at rs8176719 and rs8176746 (a B-specific SNP) we inferred ABO blood group genotypes for all individuals and plotted the trait values by genotype (Figure 4). Specifically, at this marker, individuals who are homozygous for the O allele (O/O genotype) have the highest E-selectin levels, followed by A/O heterozygotes, and by the A/A genotypes. The B/O and B/B genotypes do not differ markedly from O/O whereas the A/B heterozygotes have mean values similar to A/O.
To evaluate the association of E-selectin levels with various covariates in addition to SNPs, we performed univariate and multivariate linear regression analysis (supplemental Table 2). In univariate analysis, E-selectin levels were significantly higher in males than females, higher in the secondary intervention cohort than the primary prevention cohort, higher in those with higher HbA1c at both the time of collection and prior the sample collection updated mean HbA1c during DCCT and EDIC, but were borderline significant for age, age2, and for DCCT treatment group (supplemental Table 2). Multivariate analysis including sex, age terms, cohort, and treatment revealed that sex (P=6×10−8), age2 (P=0.029), and cohort (P=2×10−3) remained significant covariates, DCCT treatment group became more significant (P=2.0×10−3), but current HbA1c and updated prior HbA1c were both less significant (P=0.014 and P=0.043, supplemental Table 2). In univariate analysis of the 6 classical ABO blood group genotypes (using 5 indicators) there was highly significant association (P=9.8×10−22, Figure 4, Model M1 in supplemental Table 3). Similar differences were observed among the ABO genotype groups when the covariates listed in supplemental Table 2 were added to the model.
To address the issue of population stratification, we performed a series of analyses using the first 3 principal components (PC1, PC2, PC3) from the Eigenvector analysis of the genome-wide genotype data designed to capture geographic variation in allele frequencies. Using scree plots we found no evidence that more than 3 principal components account for variation in allele frequencies between individuals in this cohort. Investigation of the association between E-selectin levels and each of the 3 PCs revealed no significant associations, either singly, or jointly (P>0.05) in a linear model, with or without other covariates (supplemental Table 2). Although PC2 (but neither PC1 nor PC3) was associated with the O SNP genotypes (rs8176719, P=3.3×10−4), and with the B SNP genotypes (rs8176746, P=5.2×10−3), inclusion of the PCs in the E-selectin multivariate model with covariates had little effect on the association with ABO blood group, reducing the probability value from P=1.4×10−21 to P=1.1×10−21. This argues strongly that the association between E-selectin levels and ABO blood group cannot be explained by population stratification as captured by these three PCs.
Because of the strong linkage disequilibrium between some SNPs in the ABO region (Figure 3 and supplementary Figure), we attempted to fine-map the association in the ABO region using multivariate linear regression including multiple SNPs (supplemental Tables 3 and 4). We began with a model that included the classical ABO blood group genotype indicator (6 categories), and added individual SNPs (one at a time) that had P<10−6 in the single-SNP analysis (Table). This analysis was done both with and without the other covariates (see above) as well as with and without PCs for population stratification—the results were similar for both approaches. One SNP, rs579459, that had the most significant univariate result (Table, P=10−29), remained highly statistically significant when it was included with the ABO blood group genotype (P=1.2×10−10, Model M2 in supplemental Table 3), and the result for this SNP did not differ markedly when other covariates were included (P=6.9×10−12). In this case, the joint test of the ABO genotype indicators had reduced significance (P=0.11, compared to P=10−21 for ABO alone) suggesting that rs579459 was capturing additional information beyond that provided by the ABO blood group. Within each of the AA, AB, and AO subgroups, the rs579459 C allele was associated with lower E-selectin levels, whereas nearly all BO, BB, and OO individuals had the TT genotype at rs579459 (supplemental Table 5).
Only 2 other SNPs retained significance at P<10−4 when included with the ABO genotype: rs505922 and rs739468 (Models M3 and M4 in supplemental Table 3), but in these cases the ABO indicator remained significant (P<10−5) when either of these 2 SNPs were included, with rs505922 describing heterogeneity within A alleles (supplemental Table 7).
Additional analysis was undertaken using variants that distinguish subgroups of A alleles24,`25 defined by rs8176704. It is important to note that this SNP had a univariate probability value=0.58 and tags A2 alleles. This SNP allowed us to distinguish a total of 9 ABO blood group genotypes, including separating the common A1/A1 homozygote, from A1/A2; A1/O from A2/O, and A1/B from A2/B (there were only 2 individuals homozygous for A2/A2 who were subsequently excluded). Comparison of E-selectin levels among these groups indicated strong overall evidence for association with ABO (P=3.7×10−29). As evident in Figure 5 and from supplemental Table 4, individuals with 1 or more A1 alleles had significantly lower E-Selectin levels than individuals with O/O genotypes. Addition of the C allele of rs579459, which tags the A1 allele in our sample (supplemental Table 6), provided little additional information in the extended ABO blood group analysis (Model M6 in supplemental Table 4).
When additional SNPs from supplemental Table 1 were included in this extended ABO blood group model both ABO group (P=5×10−16) and rs505922 were retained (P=2×10−7, M#7 in supplemental Table 4). However, when both SNPs (rs579459 and rs505992) were included in the model (M#8, rs579459 was no longer significantly associated (P=0.75), while rs505922 remained significant (P=3×10−6), and the ABO blood group with the sub-A alleles defined was also nominally significant (P=4×10−4) with the 3 non-A1 blood group genotypes (A2O, BB, BO) having significantly higher E-selectin levels than the OO reference genotype group. Rs505922 thus captured additional heterogeneity within the A1 alleles (supplemental Table 8).
To demonstrate that the association between ABO SNPs and E-selectin was not limited only to individuals with type 1 diabetes, we measured E-selectin and genotyped selected ABO SNPs in 477 nondiabetic siblings of the DCCT probands (supplemental Table 9). E-selectin levels (mean=38, SD=21) were significantly lower in the nondiabetic siblings compared to the DCCT probands (P=1×10−32). Comparing the probands and their paired nondiabetic sibs from the 283 families where both were measured, the nondiabetic sibs E-selectin levels were on average 21 U lower (SD=38, P=2×10−22). In addition, there was significant familial correlation in levels between probands and their nondiabetic siblings (Spearman rank r=0.20, P=9×10−4) consistent with genetic effects. Neither age, age2, nor sex were significantly associated with E-selectin levels in nondiabetic sibs (all P>0.3). rs579459 and rs8176719 were both significantly associated with E-selectin levels in the nondiabetic sibs (P=2×10−8 and 2×10−4 respectively, supplemental Table 10), in the same direction as in the DCCT probands, indicating that the genetic association is not diabetes-specific, but present also in the nondiabetic population.
The physiological role of the ABO blood group still remains enigmatic.26 Explanations for the association between E-selectin and ABO blood group17 include that A and B alleles downregulate E-selectin levels, alter the cleavage of molecules from the cell surface, or alter clearance from serum. It is of interest that a native ligand for E-selectin (fucosylated tetrasaccharide isalyl Lewis X) is structurally related to blood group antigens.27–30 This raises the possibility that the association reported here may be explained by interference between soluble E-selectin, the monoclonal antibody that detects it and soluble blood group modification. Use of other monoclonal or polyclonal antibodies to E-selectin, or analytic interference assays24 could allow this possibility to be addressed. Although the individuals used in the GWAS part of the current study all had type 1 diabetes, data from the nondiabetic individuals is consistent with 2 smaller studies in nondiabetic individuals supporting the association of ABO with E-selectin levels,17,21 with similar estimates of the effect sizes. No other major locus (accounting for >6% of trait variance) were identified elsewhere in the genome.
We have fine-mapped the association to the region of the ABO blood locus, and for most SNPs, we show that there was little residual association after inclusion of ABO blood group genotypes in the model. This has been facilitated in part by the gene-centric SNP coverage afforded by the 510S part of the Illumina 1M assay. However, we cannot conclude that all of the association signal can be accounted for solely by the 6 classical ABO blood group genotypes, because in models that included rs579459 and ABO blood group, this SNP was still highly statistically significant, suggesting loci that are independent of ABO blood group differences. However, when the 2 major subgroups of the A allele25,31 were included in the model, the effect of rs579459 was reduced, suggesting that it is predominantly tagging heterogeneity within the A alleles. Remarkably, it has been shown that the A2 allele has 38-fold lower A transferase activity than the A1 allele.32 The A101 (A1) and A201 (A2) alleles differ at multiple SNPs, however the primary difference is thought to be a single nucleotide deletion (1060delC) in exon 7 in A201 alleles which results in a protein with 21 additional amino acids (also differing between A101 and A201 is rs1053878 at nt467, Pro156Leu).25,32,33 However, when rs505922, in intron 1, was included in a model with the ABO genotypes subdivided by A1/A2 alleles, it remained significant suggesting additional heterogeneity within the sub-A alleles. However, direct evidence to support this hypothesis as opposed to the effect of additional independent variants is lacking because of the limited sample size in the largest resequencing study of the genomic region of ABO (SeattleSNPs, http://pga.gs.washington.edu/data/abo/).34 In addition, despite our large overall sample size, certain genotype subgroups become small after stratification on multiple variations in the region (supplemental Tables 7 and 8). Previous studies of E-selectin levels were limited to blood group phenotypes, and therefore were unable to distinguish certain blood group genotypes, subclasses of alleles, or whether the association was attributable to linkage disequilibrium with variations in genes close to ABO. It is of interest to note that rs579459 is ≈3.8 kb 5′ of ABO and within 100 nucleotides of a VNTR in a regulatory region of ABO which binds the transcription factor CBF/NF-Y.35,36 It is possible that rs579459 alters the transcription of A alleles, but strong linkage disequilibrium between it and the A1 alleles does not allow us to distinguish these possibilities.
There is growing evidence that the ABO blood group is highly significantly associated with variation in the levels of a number of biomarkers. Individuals with the O pheno-type (O/O genotype) have lower von Willebrand factor (vWF) levels than those with non-O phenotypes.37 It has been shown that O heterozygotes have intermediate vWF levels between the O/O and non-O genotypes.37 Multiple lines of evidence indicate that this association is predominantly attributable to the increased clearance of vWF in individuals with group O, as a result of differences in terminal carbohydrate moiety expression. Multiple studies also report that ABO is associated with Factor VIII levels,38 with lower levels in group O compared to non-O groups. Blann and colleagues also reported that thrombomodulin levels differed by ABO blood group17 with O phenotypes having the lowest levels. In a recent study, a highly significant association between tumor necrosis factor α (TNFα) levels and ABO blood group was observed (P=6×10−44) with levels highest in the O blood group, and levels similar in A, B, and A/B phenotypes.39 In that study, the use of 2 additional assays for TNFα did not show strong correlation with the measurements from the initial assay, and there was no association of levels from these assays with ABO blood group. This led to suggestions that either the assays are measuring different parts of the TNFα molecule, different fractions of multimeric TNFα molecules, or that the initial assay cross-reacts with ABO antigens. Potentially, the TNFα-ABO association is mechanistically related to the association between E-selectin and ABO because it is well known that TNFα induces E-selectin expression.5 Furthermore, in the EURODIAB study, E-selectin levels were positively associated with TNFα levels even after conventional risk factors were included in the model.11 It will be of interest to determine whether the association between levels of TNFα and E-selectin can be explained solely by their respective associations with ABO. Most recently, ABO blood group was reported to account for 1.5% of the variance in ICAM1 levels.24 Although it has been shown that there is phenotypic correlation between E-selectin and ICAM1 levels there was no significant evidence for genetic correlation between their levels.15
It is important to note that with the exception of the ICAM1 association, no other ABO trans effects have been fine-mapped, leaving open the question as to whether they are attributable to the classical ABO blood group genotype alone, or if other variations in the region contribute to the association signal.
Historically, the ABO phenotype was one of the first markers to be typed, and there are numerous reports of association with various diseases and traits. However, small sample sizes, the potential for population stratification, and modest statistical significance have led to skepticism regarding these associations. Nevertheless, some observed associations are striking, including an increased risk of severe Plasmodium falciparum malaria in those with non-O alleles,40–42 and the association of O blood group phenotype with peptic/duodenal ulcer.43 Importantly, a recent systematic review and meta-analysis has demonstrated that non-O blood group phenotypes are significantly associated with increased risk of myocardial infarction, peripheral vascular disease as well as venous thromboembolism.44 These observations have led a number of authors to postulate a mechanism whereby some of these associations with ABO operate through biomarkers known to be associated with ABO, in particular levels of vWF. However, few of these disease associations have been subject to fine-mapping.
In studies attempting to relate E-selectin levels with clinical outcomes, it may be advantageous to include ABO genotypes as covariates in a multivariate model because they account for a sizeable proportion of variance (19%) in levels—analogous to including gender in a study of HDL cholesterol. This would be expected to improve the power to detect association between E-selectin and clinical outcomes, which has typically been nonsignificant or modest. Inclusion of ABO genotypes may be particularly important when studying the association between E-selectin and diseases that are themselves associated with ABO are undertaken (see above for a list) because this may improve validity in the presence of confounding. Moreover, when the levels of a biomarker such as E-selectin are influenced by a gene whose allele frequencies are known to vary across populations, then association between such a biomarker and disease could be confounded by population stratification. Because genotyping assays to detect population stratification are now available,45 adjusting for it may become important in future biomarker studies.
In conclusion, we have confirmed and extended the previous study that reported an association between serum soluble E-selectin levels and variation in the ABO locus.17 In comparison to the previous study we provide evidence that the association is unlikely to be attributable to chance, because the association is highly significant. Secondly, by the use of high-resolution genotype data from the Illumina 1M chip we show that for the majority of SNPs in the ABO region, the association signal can be accounted for by a subgroup of the classical A alleles making this the most likely etiologic variant to explain the association. However, an intron 1 SNP, rs505922, remains strongly associated even after heterogeneity of the A alleles is included in the model suggesting additional allelic heterogeneity at this locus. Moreover, we provide evidence that the association cannot be explained by population stratification and is also present in nondiabetic individuals. Finally, in a European-derived population with type 1 diabetes, we find no other loci in the genome where common alleles have an effect comparable to that of the ABO gene on E-selectin levels.
The excellent technical assistance of Charlyne Chassereau, Andrea Semler, Karina Moller, Hasnae Elouardighi, Jean Bucksa, and Kandy Klump and efforts of study coordinators Jenny Smith, Leslie Nicholson, Marlene Brabham, and Erica Hood are gratefully acknowledged. The authors are grateful to the subjects in the DCCT/EDIC cohort for their long-term participation. Clinical data and DNA from the DCCT/EDIC study will be made available through the National Institute of Diabetes and Digestive and Kidney Diseases repository at https://www.niddkrepository.org/niddk/home.do. The content of this article is solely the responsibility of the authors and does not necessarily represent the official views of the National Institute of Diabetes and Digestive and Kidney Diseases or the National Institutes of Health.
Sources of Funding This research was supported by funding of the Merit Review Program of the Department of Veterans Affairs, by joint funding of the National Institutes of Health and Juvenile Diabetes Research Foundation (PO1 HL55782), National Institute of Diabetes and Digestive and Kidney Diseases Contract N01-DK-6-2204, National Institute of Diabetes and Digestive and Kidney Diseases Grant R01-DK-077510, and support from the Canadian Network of Centers of Excellence in Mathematics. The DCCT/EDIC Research Group is sponsored through research contracts from the National Institute of Diabetes, Endocrinology, and Metabolic Diseases of the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) and the National Institutes of Health. A.D.P. holds a Canada Research Chair in the Genetics of Complex Diseases. S.B.B. was a recipient of a Canadian Institutes of Health Research (CIHR) Senior Investigator award (2002–7). L.M. is a recipient of a CIHR Research Scholarship.
ABO Blood group database: http://www.ncbi.nlm.nih.gov/projects/gv/mhc/xslcgi.cgi?cmd=bgmut/systems_info&system=abo.