A multi-centre study of ACE and the risk of late-onset Alzheimer’s disease

Introduction

Late-onset Alzheimer’s disease (LOAD; MIM 104300) is the most common form of dementia accounting for almost two-thirds of all dementia cases. Its key pathological features include the formation of intracellular neurofibrillary tangles comprised of microtubule-associated tau, abnormal extracellular accumulations of amyloid beta (Aβ) peptide in the form of characteristic senile plaques and in most LOAD cases, deposition of intracerebrovascular Aβ in the form of cerebral amyloid angiopathy [1, 2]. It is increasingly recognised that altered degradation and clearance of Aβ is likely to be of importance in the development and progression of LOAD [3] and that this may be contributed to by both environmental and genetic factors. Cumulative evidence from in vitro, in vivo and ex vivo studies now strongly support the role of ACE (EC 3.4.15.1), a zinc metalloprotease widely expressed in the brain, as an Aβ degrading enzyme (reviewed in [4]). Taken together with the observation that increased ACE levels and activity are observed in LOAD brains (reviewed in [5]) and are associated with increased plasma levels of Aβ [6] and reduced levels of Aβ in CSF [7, 8], these all point to the likely involvement of ACE in Aβ-related pathology in AD. This is further supported by evidence that variation in the gene encoding ACE (ACE; OMIM 106180), may play a role in LOAD pathology and modify LOAD risk. For example, the insertion/deletion (indel) of a 287bp Alu repeat in intron 16 (rs1799752 Alu I/D) of ACE, and perhaps the most widely studied for LOAD association, is predicted to explain 29–47% of the variation in plasma ACE levels [9–11]. In their meta-analysis of 39 case-control series, comprising 6,037 LOAD cases and 12,099 controls, Lehmann et al reported that homozygotes for the Alu deletion were at reduced risk of LOAD (p=0.0004), while heterozygotes were at increased risk [12], thus supporting a genetic association of ACE with LOAD. The fact that the indel does not account for all of the observed variation in ACE levels suggests that other functional ACE variants may be present and in turn associated with ACE levels and/or LOAD risk.

The APOE ε4 allele (107741) remains the most widely studied and accepted susceptibility gene for LOAD since its first report as a candidate gene almost 20 years ago [13, 14]. The remaining genetic component of AD risk may involve many genes, each with individually small-to-moderate effect sizes that interact to produce greater effects on disease susceptibility and/or disease modification. However, detection and confirmation of the involvement of genes with these effect sizes requires very large sample sizes. By example, over the last two decades 664 different genes and almost 3,000 variants have been investigated as susceptibility factors for LOAD risk [15] and until recently, the majority of these studies have been relatively underpowered, often resulting in inconclusive or inconsistent results for the majority of putative candidate genes. AlzGene (www.Alzgene.org) [15] was designed and established to resolve this problem to some extent by regularly performing meta-analyses of published data as it emerged to continually compile a list of “Top LOAD genes” that show the strongest associations in LOAD. A relatively constant member of this list has been ACE for which two (rs4291 and rs1800764) of the six variants studied show significant association with LOAD risk following the AlzGene meta-analyses based on total sample sizes of n=10,588 and n=4,756, respectively. Notably, rs1800764 has also been associated with elevated CSF Aβ42/Aβ40 ratio [7].

Despite the large number of reported independent genetic associations between ACE variants and LOAD in the last decade (22 out of 55 populations published to-date; for details see AlzGene), few studies utilised more comprehensive haplotype approaches [7, 8, 16–19]. Keavney and colleagues identified seven haplotypes in a Caucasian British population derived from data from ten polymorphisms spanning 26kb of ACE. From these haplotypes, they constructed a cladogram that contained three main branches (clades A, B and C), which accounted for 90% of the observed haplotypes. Clade A has since been associated with low plasma ACE levels and increased risk of LOAD [16], clades B and C with higher ACE levels [19–21] and clade C with increased risk for LOAD in families [18]. Kehoe and colleagues also analysed seven variants within ACE (rs4363, rs4362, rs4343, rs4331, rs4309, rs4291, rs1800764) that formed ten haplotypes with an LD structure that enabled the selection of three ‘tagging’ variants (rs4291, rs4343 and rs1800764) [8]. The most frequent haplotype (H1) contained the previously reported AD-associated (‘risk’) ACE indel I allele [22] while the H2 haplotype contained the (‘protective’) D allele [8]. Some indication that the indel was not the only functional ACE variant involved in LOAD pathogenesis came from H5 (also containing the I allele) which was also associated with a reduced risk of LOAD ref [8].

In line with previous haplotype and cladistic approaches described we have used the three tagging variants rs4291, rs4343 and rs1800764 to investigate the association of ACE with LOAD in our large multi-centre cohort comprising ten case-control series, nine of which 3,930 LOAD cases and 4,282 controls. This represents the largest study to-date to investigate the effects of ACE haplotypes in LOAD.

Materials and Methods

Results

We tested three ACE variants for association with LOAD in our seven European and three North American case-control series (series details are shown in Table 1). Genotype and allele counts for each series are shown in Table 2. We first tested for association with LOAD in each case-control series by logistic regression using an additive/allelic dosage model correcting for sex, age-at-diagnosis and possession of at least one copy of the APOE ɛ4 allele as covariates (Table 2). None of the variants were associated with LOAD risk in any series (all p>0.09). We also tested for association the three variants in all ten series pooled (n=8,212) but again found no association (all p>0.18). Since some genetic variants may exert dominant or recessive effects we also performed logistic regression for the total dataset using these models but found no association with LOAD risk (all p>0.36). We also tested for association of all three variants with LOAD in individuals not possessing the APOE ε4 allele in all series (1,495 LOAD, 3,175 controls) but found no association (all p>0.13; data not shown). In order to determine whether the lack of observation could be attributed to population heterogeneity, we performed meta-analyses of each variant testing for association in the combined data from all ten series using a DerSimonion-Laird random effects model to estimate a pooled odds ratio and testing for heterogeneity between series using the Breslow-Day test (Fig. 1). We found no evidence for population heterogeneity (all p>0.38) or for association of rs4343 (OR=1.00, p=0.90), rs4291 (OR=0.97, p=0.47) or rs1800764 (OR=0.99, p=0.72) with LOAD risk.

In an attempt to replicate previous findings that the two most common haplotypes (H1 and H2) are significantly associated with opposing risk for LOAD [8], we constructed haplotypes using the three tagging variants. As shown in Table 3, the haplotype frequencies were comparable across all series and are consistent with previous studies [8]. In order to limit the number of tests used, global association of the six haplotypes in each series were tested and individual haplotypes only tested for association for populations with a global p-value <0.05. A significant global haplotypic p-value was observed in the MRC sample only (p=0.007), largely to the protective association of H3 (OR=0.72, p=0.02) and the trend towards a risky association of H4 (OR=1.38, p=0.09) and H7 (OR=2.07, p=0.08), associations that were not previously observed by Kehoe and colleagues [8]. The global haplotypic p-value for our complete dataset (n=7,557) was not significant (p=0.51) and the individual haplotypic ORs observed in our complete dataset for the two most common haplotypes gave weaker, ORs compared to those observed by Kehoe et al in their complete dataset (H1; OR=1.0 vs OR=1.2, H2; OR=0.96 vs 0.80 [8]). Therefore, these data do not replicate the previously reported association of the Alu indel [12], which tags H1.

Discussion

We have conducted a large case-control study of three haplotype tagging variants in the LOAD candidate gene, ACE, that has previously been associated with LOAD risk [8, 29–34]. Meta-analyses of ten case-control series totalling 3,930 LOAD and 4,282 controls showed no population heterogeneity (all p>0.38) or evidence for association of rs4343 (OR=1.00 p=0.90), rs4291 (OR=0.97 p=0.47) or rs1800764 (OR=0.99 p=0.72) with LOAD risk.

We also tested for association of six ACE haplotypes with LOAD but found no evidence for association in the total dataset (global p=0.51). However, we did observe a significant global haplotypic p-value of 0.007 in one of the series (the MRC population - 1,430 LOAD patients and 1,611 controls) from the UK, which was largely due to a novel protective association of H3, the haplotype containing the major allele (G-A-T) at all three variants (9% LOAD, 11% Controls, OR=0.72, p=0.02). However, this association was not observed in any of the other case-control populations (all p>0.07) or in the pooled dataset (p=0.58). Indeed the directionality of ORs observed in all other populations was opposite to that seen for H3 (OR=1.03–1.25) but none produced significant findings. The modest p-value (p=0.03) for association of H3 with LOAD, which would not survive Bonferroni correction for the six haplotypes studied (p<0.008) along with the lack of association of H3 in the other nine populations studied here or in previously published studies suggests that the significance of this association should be treated with caution.

The lack of association in any of our ten series for the two ACE variants that have been significantly associated with LOAD following AlzGene meta-analyses is perhaps not surprising. None of the eight Caucasian populations used previously to study rs4291 and included in AlzGene showed significant association for AD and only one [35] of the seven Caucasian populations used to study rs1800764 showed significant association (OR=0.86 95% CI 0.74–0.99). Despite this, AlzGene reported significant association for both rs4291 (OR=0.87, 95%CI 0.80–0.95) and rs1800764 (OR=0.84, 95%CI 0.77.0.92) prior to this study. This is largely due to the fact that there was a consistently similar directionality of the ORs in the majority of the Caucasian populations previously used to study rs4291 and rs1800764 and the resulting increase in sample size achieved by analyzing the studies together provides sufficient power to detect association. When the present data is eventually incorporated into Alzgene these overall associations will likely diminish further towards the null. However, the fact that we also observed same direction ORs in seven out of ten series for rs4291 and six out of ten series for rs1800764, raises the possibility that a true association of modest effect size (OR~0.90) is present, but which requires even larger studies to gain sufficient power for detection.

It is possible that the initial association may have, by chance, been the result of an over-estimation of the effect size of these variants that has since diminished in subsequent follow-up case-control studies. For example, in the case of rs4291, the initial study reported ORs ranging from 0.76–0.84 in four Caucasian populations each consisting of ~400 subjects [8]. In comparison, the four subsequent Caucasian studies (all of equal or larger size than the initial study populations) reported ORs ranging from 0.76–1.00 [31, 32, 36, 37] and here we report ORs ranging from 0.80–1.28 in ten populations of equal or larger size than the initial study thus further diminishing the effect size. The same appears true for rs1800764 where the initial study reported ORs ranging from 0.74–0.84 [8], compared to ORs ranging from 0.80–1.09 in subsequent follow-up studies [31, 38, 39] and in the ORs 0.83–1.07 reported here. This further supports the need for multiple, large follow-up studies and meta-analyses of all data to reduce the likelihood of an over-estimation of the effect size.

Failure to detect association could also be explained by these variants merely tagging one or more truly functional variants. In such instances the corresponding degree of linkage disequilibrium between these variants could differ between series leading to weaker and/or opposing effects. However, given the Caucasian background of all these series and the lack of significant population heterogeneity or association with LOAD risk for any variant in any individual population, this possibility seems unlikely.

If ACE variants modify LOAD risk this effect may be dependent on interactions with other environmental/phenotypic background. For example, ACE mediates hypertensive effects by its function on angiotensin I to convert it to the vasoactive angiotensin II [4]. Thus the co-occurrence of hypertension in sub-groups of AD patients and controls and this being treated to varying extents by drugs that target the pathway in which ACE is functional is likely to be a confounding variable in studies for ACE. Indeed studies of AD brain tissue have shown that while ACE genotypes did not influence levels or activity of brain ACE [40], rs4343 and rs1800764 have been associated with soluble Aβ levels [41] and exposure of neuronal SH-SY5Y cell lines to oligomeric Aβ1–42 for 24 hours resulted in significant increases in ACE protein level and activity [41]. These collectively suggest that along with previous evidence of elevated ACE activity in AD brain [40, 42] that there could be phenotypic-specific post-translational changes to ACE that contribute to AD pathogenesis. Further information supporting this are the findings of Ellul and colleagues [43]. They noted in a longitudinally assessed clinical cohort, that drugs affecting the Renin Angiotensin System, in which ACE is very important, can slow the rates of deterioration of AD and could serve as a confounder in clinical outcome measurement in clinical trials. The possibility of phenotype-specific pharmacogenetic considerations are also reinforced by findings from a recent GWAS [44] of two young-onset hypertension populations totalling 1,023 subjects that reported eight ACE variants were significantly associated with ACE activity and rs4343 showing the strongest association (p=3.0×10⁻²⁵). In the same study an association between blood type and ACE activity in an independent young-onset hypertension family study (n=428) was reported and showed a potential differential blood pressure response to anti-hypertensive treatment (ACE inhibitors) in subjects dependent on ACE genotype. This latter observation is particularly important in view of findings from a number of observational studies where anti-hypertensive treatments such as inhibitors of ACE activity (i.e. ACE-inhibitors) or of angiotensin II function (i.e. angiotensin II receptor antagonists – ARAs) appeared to be protective against the development of and/or progression of LOAD and Mild Cognitive Impairment (MCI) [45–51].

These data do not replicate previously reported haplotype associations with LOAD risk in a large case-control series of 3,930 LOAD patients and 4,282 controls. However, it is clear that when considered alongside data from other disciplines, these findings contribute just one important part of what appears to be a highly complex interaction between ACE genetics, phenotype and pharmacological effects in AD and which traditional case-control studies are not equipped to unpick. Larger studies which would include richer phenotypic data that would allow for more accurate adjustment for confounders and where possible incorporation of additional measurements specific to candidate gene function (e.g. qPCR, protein based measurements, etc), would likely increase the chances of unpicking the real genetic involvement of many current candidate genes.