Whole genome association analysis shows that ACE is a risk factor for Alzheimer's disease and fails to replicate most candidates from Meta-analysis

Abstract

For late onset Alzheimer's disease (LOAD), the only confirmed, genetic association is with the apolipoprotein E (APOE) locus on chromosome 19. Meta-analysis is often employed to sort the true associations from the false positives. LOAD research has the advantage of a continuously updated meta-analysis of candidate gene association studies in the web-based AlzGene database. The top 30 AlzGene loci on May 1^st, 2007 were investigated in our whole genome association data set consisting of 1411 LOAD cases and neuropathoiogicaiiy verified controls genotyped at 312,316 SNPs using the Affymetrix 500K Mapping Platform. Of the 30 “top AlzGenes", 32 SNPs in 24 genes had odds ratios (OR) whose 95% confidence intervals that did not include 1. Of these 32 SNPs, six were part of the Affymetrix 500K Mapping panel and another ten had proxies on the Affymetrix array that had >80% power to detect an association with α=0.001. Two of these 16 SNPs showed significant association with LOAD in our sample series. One was rs4420638 at the APOE locus (uncorrected p-value=4.58E-37) and the other was rs4293, located in the angiotensin converting enzyme (ACE) locus (uncorrected p-value=0.014). Since this result was nominally significant, but did not survive multiple testing correction for 16 independent tests, this association at rs4293 was verified in a geographically distinct German cohort (p-value=0.03). We present the results of our ACE replication aiongwith a discussion of the statistical limitations of multiple test corrections in whole genome studies.

Introduction

Until recently, most genetic analyses of complex traits involved candidate gene analysis. Frequently, in such analyses, an initial positive report of a genetic association is followed by a mixture of positive and negative reports, leading to an unclear outcome after many studies. After such studies, it is often not clear whether there is an association, albeit weaker than that reported in the original study, or whether there is simply no association. For late-onset Alzheimer's disease (LOAD) (OMIM #104300), the only consistently replicated association locus is apolipoprotein E (APOE) [1,2]. Meta-analysis, with and without the inclusion of the original report, is commonly used to determine what are true and what are false positives. However, the original reports which are used to compile these meta-analyses may be subject to publication bias and it is not clear if positive reports of association are marginally more likely to be reported. Opinions are thus divided between those that believe that positive meta-analyses have considerable value (“glass half full”) and those that believe these studies are too flawed by reporting bias to be genuinely useful (“glass half empty”). There has been little systematic analysis the utility of meta-analyses in any particular disorder in independent cohorts because the sample sizes and genotyping costs have been prohibitive.

In LOAD analysis we have the benefit of the AlzGene database [3], a continuously updated meta-analysis of all candidate gene association studies. This database allows us to assess large numbers of positive and negative meta-analyses simultaneously. With these resources we sought to assess whether the May 1^st freeze of the top 30 Alzheimer meta-analyses (see Table 1 for the list), 24 of which had 95% confidence intervals which did not include 1.0 could be translated into positive findings in our whole genome analyses using the 500K Affymetrix Mapping Array on 861 Alzheimer cases and 550 relevant controls. As previously reported, the most significant allele in this whole genome screen is rs4420638, which is located 14kb distal to the APOE locus [4]. We have also reported the association of a common haplotype of GAB2 with LOAD in carriers of the APOE ɛ4 locus [5].

Table 1.

SNPs from Alzgene Freeze May 2008

AlzGene	AlzGene Polymorphism	Risk Allele	Freq. Cases	Freq. Controls	AlzGene OR (95% CI)	Affymetrix 500K SNP	r2	D′	Power (alpha=0.05)	Power (alpha=0.001)
APOE	ɛ3/ɛ4	ɛ4	0.37	0.14	3.81 (3.38-4.29)	rs4420638	0.720	0.910	99.74%	92.92%
CHRNB2	rs48453782	G	0.93	0.91	1.45 (1.06-1.96)
CH25H	rs13500	T	0.12	0.1	1.38 (1.01-1.88)	rs11594137	0.014	1.000	100.00%	100.00%
PGBD1	rs3800324	A	0.05	0.03	1.42 (1.13-1.80)	rs742107	1.000	1.000	99.94%	97.26%
LMNA	rs505058	C	0.08	0.06	1.35 (1.12-1.63)	rs2485668	1.000	1.000	100.00%	100.00%
SOAT1	rs1044925	C	0.43	0.36	1.35 (1.13-1.60)	rs6666455	0.300	1.000	97.44%	73.19%
MAPT	rs2471738	T	0.22	0.19	1.30 (1.01-1.67)	rs1078268	0.107	1.000	100.00%	100.00%
SORL1	rs1010159	C	0.36	0.34	1.14 (1.02-1.29)	rs1010159	1.000	1.000	100.00%	100.00%
	rs1699102	C	0.35	0.33	1.13 (1.02-1.25)	rs7116734	0.689	1.000	100.00%	99.99%
	rs2070045	G	0.26	0.23	1.26 (1.08-1.46)	rs11218347	0.015	1.000	99.28%	86.74%
	rs2276346	T	0.38	0.33	1.28 (1.07-1.54)	rs2276346	1.000	1.000	100.00%	99.89%
	rs2282649	T	0.3	0.28	1.16 (1.04-1.30)	rs2282649	1.000	1.000	100.00%	100.00%
	rs3824968	T	0.31	0.28	1.30 (1.07-1.58)	rs1629493	0.891	1.000	100.00%	100.00%
	rs661057	T	0.59	0.56	1.14 (1.03-1.25)	rs610188	0.000	0.000	0.00%	0.00%
PCK1	rs8192708	A	0.13	0.11	1.29 (1.09-1.52)	rs1023049	0.066	0.782	77.33%	28.07%
CST3	rsl0640392	A	0.21	0.19	1.16 (1.00-1.33)
ACE	rs1800764	T	0.59	0.54	1.27 (1.09-1.47)	rs4293	0.964	1.000	99.85%	94.85%
	rs42912	T	0.65	0.62	1.22 (1.04-1.43)
SORCS1	rs6008792	min	0.11	0.09	1.24 (1.04-1.48)
TF	rs1049296	C2	0.19	0.17	1.24 (1.06-1.25)	rs1049296	1.000	1.000	100.00%	100.00%
hCG2039140	rs19039082	T	0.15	0.12	1.23 (1.06-1.44)
IDE1
GALP	rs3745833	C	0.39	0.35	1.21 (1.10-1.33)	rs4801296	0.187	0.598	94.68%	61.20%
CTSD	rs17571	T	0.09	0.08	1.20 (1.01-1.42)	rs17834326	0.786	1.000	100.00%	100.00%
TNK1	rs1554948	T	0.55	0.5	1.19 (1.08-1.32)	rs7219773	0.967	1.000	83.86%	36.62%
GWA_14q32.13	rs11622883	T	0.58	0.54	1.19 (1.08-1.30)	rs17091290	0.008	1.000	7.58%	0.25%
IL1B	rsll43634	T	0.28	0.26	1.18 (1.02-1.37)	rs3917365	0.013	1.000	97.75%	75.01%
LOC651924	rs69071752	G	0.49	0.45	1.16 (1.04-1.30)
PON11
GWA_7p15.21
LOC439999	rs498055	G	0.52	0.48	1.15 (1.03-1.29)	rs526928	0.887	1.000	15.89%	0.97%
DAPK1	rs4877365	G	0.74	0.69	1.25 (1.09-1.45)	rs4877365	1.000	1.000	100.00%	100.00%
	rs4878104	C	0.66	0.62	1.15 (1.05-1.27)	rs4877365
PRNP	rs1799990	A	0.69	0.66	1.14 (1.03-1.27)	rs7274758	0.048	1.000	16.73%	1.07%
MYH131
HMGCS21
BDNF	rs6265	A	0.19	0.18	1.10 (1.01-1.19)	rs6265	1.000	1.000	100.00%	100.00%
PSEN11

¹These genes had no polymorphisms with 95% confidence intervals that did not include 1.

²These SNPs were notgenotyped by the HapMap Consortium and therefore we were unable to determine whether there was a sufficiently powerful proxy on the Affymetrix 500K Mapping Array with which to test the association. Bolded SNPs (>80% power to detect α=0.001) were examined in our whole genome association data set.

Materials and methods

Genotyping

A cohort of 1411 subjects including 643 brain donors who have been clinically diagnosed and neuropathologically verified to have late-onset LOAD, 404 clinically and neuropathologically classified non-demented controls, 218 clinically diagnosed LOAD cases and 146 clinically characterized non-demented controls was utilized. The mean age of the AD cases was 82 8+/−7.7 years and of controls was 79.7+/−6.3 years. The 500K GeneChip (Affymetrix, Santa Clara, CA) was used to survey 502,267 SNPs in each subject as recently described [4, 5]. Genotypes were extracted using both SNiPer-HD<⁶) and BRLMM (Affymetrix) software. 312,316 SNPs were analyzed after excluding those that were monomorphic, clustered poorly, had Hardy Weinberg equilibrium p-values less than 0.01, had minor allele frequencies less than 2%, or exhibited less than 98% concordance between the SNiPer-HD and BRLMM calls. The software program STRUCTURE [7] was employed to test for underlying genetic stratification, using 5,000 randomly selected SNPs and including at least 100 SNPs per chromosome. The initial analysis yielded empirical evidence of three populations. Since fourteen subjects belonged to a population far removed from the rest of the study population, they were eliminated from further analyses. STRUCTURE then was used to demonstrate a comparable admixture of the two populations in the cases and controls. These data are at http://www.tgen.org/research/index.cfm?pageid=1065 (see reference 5). ABI Taqman genotyping of rs4293 was carried out according to manufacturer's instructions, using the off-the-shelf kit, and read using an ABI 7000.

Statistical Analysis

The PLINK analysis toolset (http://pngu.mgh.harvard.edu/~purcell/plink/index.shtml) was used for whole genome analysis. JMP was used to perform a logistic regression analysis of rs4293 vs. LOAD diagnosis with APOE ɛ4 gene dose as a covariate. LD mapping was performed by importing genotypes into the HaploView program version 3.32. Pair-wise LD values (as measured by D'), reflect the likelihood that two genetic markers are inherited together. Power calculations were performed using the Genetic Power Calculator http://pngu.mgh.harvard.edu/~purcell/gpc/) with a disease prevalence of 13%, and using AlzGene odds ratio (OR) as an estimator of relative risk [8].

Follow up genotyping on an independent sample series

We used a follow up sample of LOAD patients recruited from the Department of Psychiatry of the University of Bonn Patients. These were diagnosed according to DSM IV criterua, supported by clinical examination and detailed. Healthy controls from the general population were recruited with the support of the local Census Bureau and the regional Board of Data protection (Nordrhein-Westfalen, Germany). Neuropsychological testing, detailed structured interviews and clinical examination were performed. All patients and control subjects gave informed consent for participation in the study. The study protocol was approved by the Ethics Committee of the Faculty of Medicine of the University of Bonn. The LOAD patients (n=200) had a mean age of 76±6.5 years and a mean age at onset of 73±7.1 years and 72.5% were women and the controls (n=126) had a mean age of 73±6.4 years and 50.8 % were women. With the exception of their national origin, these subjects are similar to those we used in our whole genome study [5].

Results

Our first step was to determine the extent to which our study had 80% power to replicate the reported associations with an alpha of 0.001 and allow for replication after multiple testing correction. This relied on two factors other than the size of our sample: the size of the reported effect and the extent to which a proxy SNP on the Affymetrix array is in linkage disequilbrium with the SNP used in the candidate gene study. These parameters are illustrated in Table 1. As these data show, the SNP on the array was the same as the candidate SNP in six cases (rs1010159, rs2276346 and rs2282649 in SORL1, rs1049296 in TF, rs4877365 in DAPK1, and rs6265 in BDNF) and was in sufficient LD with that SNP (r² 0.014 to 1.00, D′ 0.91 to 1.00, power to detect association at α=0.001 80% to 100.00%) in a further ten cases. The uncorrected contingency test p-values for these sixteen SNPs are listed in Table 2. In eight cases the Affymetrix platform does not capture the variability of the reported SNP with sufficient power for our study to be a useful proxy.

Table 2.

Data from ALzgene-related SNPs in Whole Genome Analysis

AlzGene	Affymetrix 500KSNP	Risk Allele	Freq. Cases	Freq. Controls	P-value	OR (95% C.I.)
APOE	rs4420638	C	0.40	0.17	4.58E-37	3.2 (2.68-3.9)
CH25H	rs11594137	A	0.86	0.85	0.7159
PGBD1	rs742107	C	0.061	0.050	0.2319
LMNA	rs2485668	A	0.95	0.95	0.2833
ACE	rs4293	A	0.57	0.52	0.014	1.22 (1.04-1.41)
TF	rs1049296	C	0.82	0.82	0.82
DAPK1	rs4877365	C	0.71	0.70	0.76
BDNF	rs6265	T	0.21	0.21	0.65
MAPT	rs1078268	A	0.79	0.78	0.72
SORL1	rs1010159	T	0.67	0.64	0.12
SORL1	rs7116734	G	0.60	0.58	0.30
SORL1	rs11218347	T	0.92	0.93	0.91
SORL1	rs2276346	G	0.64	0.63	0.26
SORL1	rs2282649	C	0.71	0.70	0.39
SORL1	rs1629493	A	0.64	0.61	0.092
CTSD	rs17834326	T	0.099	0.092	0.56

Uncorrected, two-tailed contingency test p-values for the sixteen Affymetrix 500K SNPs that are good proxies for significant polymorphisms in the AlzGene database. The “Risk allele” is that allele reported to be associated with disease in previous studies or to be the Affymetrix surrogate for that allele.

As we have previously reported, and unsurprisingly, this analysis confirms that APOE is a risk locus for disease, since the genetic association between the ɛ4 allele is captured by rs4420638 (r²=0.72). Of more interest, is the fact that in only one of the other 15 cases was the association confirmed, and this was at the ACE locus (also known as DCP1). The Affymetrix 500K SNP at the ACE locus that we tested, rs4293, was nominally significant (uncorrected contingency test p-value=0.014) in our study population.

According to the HapMap CEPH data set (9), rs4293 is part of the same haplotype block as rs1800764, the significant polymorphism identified in the AlzGene database (Figure 1). Additionally, the risk allele for rs1800764 (T) is part of the same haplotype as the risk allele we identified in rs4293 (A), confirming that the risk allele for rs4293 displayed the same direction of effect as the identified risk allele of rs1800764.

Figure 1.

Linkage disequilibrium structure of HapMap genotyped SNPs in the CEPH population at the distal end of ACE. The plot follows the standard HAPLOVIEW color scheme: white boxes, D′ < 1, LOD < 2; shades of pink/red, D′ < 1, LOD ≥ 2; blue, D′=l, LOD < 2; bright red, D′=l, LOD ≥ 2. AizGene SNP rs1800764 (black box) is part of the same hapiotype block as Affymetrix SNP rs4293 (yellow box). The AD risk allele (T, in black box) associated with rs1800764 is transmitted with the AD risk allele (A, in yellow box) that was detected in rs4293 by this screen.

Given the nominal significance of the ACE replication, a second cohort was examined to determine the true significance of ACE in LOAD. A replication cohort of clinically assessed patient samples (see Methods) from Bonn, Germany was genotyped at rs4293 using an ABI Taqman (Applied Biosystems, Foster City, CA) custom genotyping assay. An association with LOAD was confirmed in this population at rs4293 with the same direction of effect observed in our whole genome association series (Table 3).

Table 3.

Association of rs4293 with AD in the Bonn Alzheimer Series

	Genotypes (AA/AG/GG)	A Allele Frequency	G Allele Frequency
Cases	75/89/36	0.60	0.40
Controls	34/61/31	0.51	0.49
Total	109/150/67	0.56	0.44

Association of rs4293 with AD in a clinically characterized cohort from Bonn, Germany. The odds ratio for the G allele was 0.71 (95% confidence interval, 0.51-0.97). The ‘p’ value (two tailed t test) was 0.03.

A logistic regression analysis was used to test for an interaction between rs4293 and APOE ɛ4 gene dose in the determination of LOAD risk. The p-value of this interaction term was 0.3567, indicating that there is not a significant interaction between ACE and APOE in determining disease risk in this cohort.

Discussion

These data are of interest for three reasons. First, they show that the ACE locus is indeed an Alzheimer risk factor. Second, they illustrate that Bonferroni correction leads to false negative findings in the context of whole genome associations since this risk factor was missed in our initial analysis. Lastly, they illustrate that meta analyses lead to reports of positive associations which fail to replicate even when samples series of sufficient power and correct ethnicity are used, presumably because of bias in the reporting of data.

ACE as an Alzheimer Risk factor

Kehoe and colleagues (1999) reported an association of an insertion/deletion polymerphism in intron 16 of ACE with LOAD [10] that has been shown to account for about half of ACE serum expression levels [11]. Subsequently, seven polymorphisms in ACE were examined, five of which are included in the AlzGene analysis. Of these five, two (rs4291 and rs1800764) were significant according to our criteria. Since rs4291 was not genotyped as a part of the HapMap initiative, we were unable to determine if it had a good proxy on the Affymetrix 500K Mapping Array. However, there is a good proxy for rs1800764 in rs4293, for which we have 95% power to detect the observed association with LOAD. A nominally significant association was detected in the whole genome association cohort (uncorrected two-tailed p-value=0.014) with the same direction of effect observed in the AlzGene meta-analysis. Since this association did not survive a Bonferroni multiple testing correction, we confirmed the association of rs4293 with LOAD in a second, geographically distinct population (two-tailed p-value=0.03). This association had the same direction of effect as our whole genome association series and the AlzGene analysis.

ACE is hypothesized to be a risk factor for atherosclerosis [12] and hypertension [13]. There is now epidemiological [14-16] and pathological [17-19] evidence pointing to vascular risk factors for LOAD. A genetic association with the allele of the ACE locus which is associated with risk for hypertension [13, 20], fits with this pattern and with the hypothesis that microhaemorrhages may play a part in Alzheimer pathogenesis [21].

Bonferroni correction and whole genome studies

These data show, unsurprisingly, that Bonferroni correction for multiple tests, leads to false negative findings in whole genome studies. The ACE polymorphism was the 4398th out of 312,316 in terms of the significance of the allelic ‘p’ values. Here we show that ACE is a risk factor for LOAD, however using a traditional whole genome association analysis approach, including a Bonferroni correction, we did not identify ACE [5]. A great deal of attention has been paid, with good reason, to limiting type I error in genome-wide scans [22, 23]. However, clearly and unsurprisingly the use of the Bonferroni correction greatly increases the type II error and decreased the power of our whole genome association dataset [24, 25]. This, and the recent success of the whole genome scans for diabetes [26-29] illustrates that a productive way forward for identifying loci which do not appear to stand out considerably from the mass of data, such as APOE does in this data set for LOAD [4], is to have large initial cohorts for whole genome analysis followed by large and independent replication cohorts in which replication of several thousand SNPs (at least 5000) is attempted, followed by third cohorts etc.

The (Limited) Utility of Meta-Analyses

Finally, this study suggests that meta-analyses have modest utility in distinguishing true from false positives in genetic associations. ACE was the 11^th highest ranked by allelic effect size (http://www.alzforum.org/res/com/gen/alzgene/methods.asp#topresults). The most likely reason for the failure to replicate the other candidates from meta analysis is that even subtle publication biases are likely to be enough to lead to the false assignment of statistical significance to odds ratios in the range of 1.1 to 1.3. Even the most robust meta-analysis methods are limited by the literature available. Meta-analyses assume that the literature represents a random sampling of studies for each association marker. However, if negative findings are underreported in comparison to positive association findings, meta-analyses may yield false positives or inflated odds ratios [30, 31]. A recent meta-analysis of twenty-five associations across multiple complex disorders found statistically significant replication for eight of the reported associations, or 32% of reported associations [32], and concluded that publication bias was not responsible for the observed associations. However, that analysis did not attempt external testing of these associations as we have done here and made the assumption that, following a positive report of association, associations reporting replication and failure to replicate were equally likely to be published. In contrast, we only identified sixteen SNPs from the AlzGene meta-analysis for which we had greater than eighty percent power to detect the reported association in our whole genome association dataset. Of these, only two (13%), APOE ɛ4 and rs4293 in ACE, were significantly associated with AD in our large case-control population. Clearly, one could argue that meta-analysis is not required for confirmation of the APOE association and this would reduce the predictive success rate for meta-analyses to be 1/15 (7%) in this disorder at least. While this may seem rather pessimistic, it suggests that systematic meta-analyses of the literature, such as AlzGene, are useful adjuncts to whole genome studies in the identification of risk alleles. It is notable that, subsequent to these analyses, ACE has moved to number two in the Alzgene list (September 1^st 2009) and that others have more recently suggest ACE is an important disease risk factor, in part because of their analysis of whole genome data [33-35].

Summary

As data from other diseases have indicated, whole genome association studies have the potential to identify genetic risk variants which contribute to disease. The data we present here show that genetic variability in ACE, does, as others had suggested [10, 30] contribute to the risk of LOAD and indicate that, while previous meta-analyses cannot be unequivocal determinants of whether genetic variability at a certain locus contributes to risk, they are a useful step as hypothesis generating analyses which can then be tested in a whole genome study.