Introduction

Studies from the World Health Organization have demonstrated that Alzheimer’s disease (AD) is a complex and common neurodegenerative disease in the elderly that causes more than 50–75 % of dementia types [1]. It is estimated that genetic factors underlie approximately 60–80 % of AD risk [2]. To identify common AD genetic variants, large-scale genome-wide association studies (GWAS) have been conducted in individuals of Caucasian ancestry, revealing nine AD susceptibility genes, namely CR1, BIN1, CLU, PICALM, MS4A4/MS4A6E, CD2AP, CD33, EPHA1, and ABCA7 [3–7]. Among these loci, a single nucleotide polymorphism (SNP), rs3764650 in ABCA7, was found to be significantly associated with AD (P = 5.00E − 21, odds ratio (OR) = 1.23, 95 % confidence interval (CI) 1.17–1.28, minor allele = C) [4]. This study indicated that European carriers of the rs3764650 risk variant (C) had an additional 23 % increased risk for AD [4].

In addition to the rs3764650 polymorphism, three other polymorphisms in ABCA7, rs115550680 (P = 2.20E − 09, African American) [8], rs3752246 (P = 5.00E − 07, European), [7] and rs4147929 (P = 1.10E − 15, European) [9], have been reported. Evidence from the HapMap Reference Sample (African Americans in southwest USA) showed that rs115550680 is in high linkage disequilibrium (LD) with rs3764650 and rs3752246 (0.8 < D′ < 0.9) [8]. Using 1,000 Genomes data (YRI) in SNP Annotation and Proxy Search (SNAP) [10], a web-based tool for the identification and annotation of proxy SNPs, we found that rs3764650 was in LD with rs115550680 (D′ = 1). rs3764650 was in LD with rs3752246 (D′ = 0.70 for CEU; D′ = 0.80 for CHB + JPT) and rs4147929 (D′ = 0.81 for CEU; D′ = 0.83 for CHB + JPT). LD also exists between rs3752246 and rs4147929 (D′ = 0.94 for CEU; D′ = 0.96 for CHB + JPT).

Recent studies that investigated the genetic mechanisms of ABCA7 in AD pathogenesis have indicated that ABCA7 plays an important role in AD risk in individuals of European ancestry [11–13]. AD GWAS and candidate gene studies have also investigated ABCA7 polymorphisms (especially rs3764650) in other populations (Chinese, Japanese, Korean, African American, and Canadian) [14–19]. However, these studies reported a weak or negligible association between rs3764650 and AD. We hypothesized that these negative results were due to relatively small sample sizes compared with those used in previous GWAS studies in individuals of European ancestry or the genetic heterogeneity of the rs3764650 polymorphism in different populations. Here, we reevaluated the association between the rs3764650 polymorphism (or its proxies, namely rs115550680, rs3752246, and rs4147929) and AD using relatively large-scale samples from previous studies (N = 79,381—30,590 cases and 48,791 controls), which may be helpful for revealing significant disease associations.

Methods and Materials

Literature Search

Guiyou Liu searched the PubMed (http://www.ncbi.nlm.nih.gov/pubmed) and AlzGene (http://www.alzgene.org/) databases to select all possible studies with the key words “Alzheimer’s disease,” “ATP-binding cassette, sub-family A (ABC1), member 7,” or “ABCA7.” The literature search was updated on January 18, 2014. To compile additional studies, we used Google Scholar (http://scholar.google.com/) to query articles citing the initial studies and all references in these studies identified by PubMed and AlzGene. We only selected published articles written in English.

Inclusion Criteria

We selected studies meeting the following criteria: (1) the study was conducted by a case-control design, (2) the study evaluated the association between the rs3764650 polymorphism (or one of its proxies) and AD, (3) the study provided the number of rs3764650 (or its proxies) genotypes or sufficient data to calculate the number of rs3764650 (or its proxies) genotypes, and (4) the study provided an OR with a 95 % CI or sufficient data to calculate the OR and a 95 % CI.

Data Extraction

Guiyou Liu and Yongshuai Jiang extracted the following information from each study: (1) the name of the first author; (2) the year of publication; (3) the population and ethnicity; (4) the numbers of AD cases and controls; (5) the genotyping platform; (6) the number of rs3764650 (or its proxies) genotypes or if not provided, calculated the number of rs3764650 (or its proxies) genotypes; and (7) the OR with 95 % CI or if not provided, calculated the OR and 95 % CI. All relevant calculations were completed using R (http://www.r-project.org/).

Quality Evaluation

Here, the criteria proposed by Clark et al. were used to evaluate the quality of selected genetic association studies [20]. This scoring system consists of ten components. A component of the criteria was scored as 1 if present or 0 if absent. A final quality score was obtained by summing each component, resulting in a scoring range of 0–10 for case-control association studies [20]. Selected studies were scored as “good” if the score was greater than or equal to 8, “mediocre” if the score was 5–7, and “poor” if the score was less than 4 [21]. Two authors performed the quality evaluation independently using the criteria proposed by Clark et al. A third author adjudicated any differences between the two authors.

Genetic Models

We used allele, dominant, recessive, and additive models for our meta-analysis. The ABCA7 rs3764650 polymorphism has two alleles, C and A. C is the minor allele. We assumed that C was the high-risk allele and that A was the lower risk allele. The four models can be described as follows: the allele model (C allele versus A allele), the dominant model (CC + CA versus AA), the recessive model (CC versus CA + AA), and the additive model (CC versus AA) [22].

Heterogeneity Test

We evaluated genetic heterogeneity among the included studies using Cochran’s Q test, which approximately follows a χ 2 distribution with k − 1 degrees of freedom (where k is the number of studies analyzed). Another statistic, \( {I}^2=\raisebox{1ex}{$\left(Q-\left(k-1\right)\right)$}\!\left/ \!\raisebox{-1ex}{$Q$}\right.\times 100\% \), was also used. I 2 ranges from 0 to 100 % [23]. Low, moderate, large, and extreme heterogeneity corresponded to I 2 values of 0–25, 25–50, 50–75, and 75–100 %, respectively [23]. Heterogeneity was considered significant when P < 0.01 and I 2 > 50 %.

Meta-analysis

If there was no significant heterogeneity among the selected studies, the pooled OR was calculated using the fixed effect model (Mantel-Haenszel). For all other cases, the pooled OR was calculated using the random-effect model (DerSimonian-Laird). A Z test was used to determine the significance of the OR. All statistical tests for heterogeneity and the meta-analysis were computed using the R Package (meta: Meta-Analysis with R, http://cran.r-project.org/web/packages/meta/index.html) or RevMan (v.5.1) software (http://ims.cochrane.org/revman/download).

Sensitivity Analyses and Publication Bias Analyses

For sensitivity analyses, we evaluated the relative influence of each study by omitting each study, one at a time. In addition, we used funnel plots to evaluate potential publication bias [24]. Begg’s and Egger’s tests were used to evaluate the asymmetry of the funnel plots [24].

Results

Literature Search and Quality Evaluation

A total of 48 articles were identified using PubMed, AlzGene, and Google Scholar. In the end, nine articles describing 18 independent studies were included for analysis. More detailed information about the inclusion or exclusion of selected studies and quality evaluation is provided in the Supplementary materials. The main characteristics of the included studies are described in Table 1.

Table 1 Main characteristics of the individual studies included in the meta-analysis
Full size table

Heterogeneity Test and Meta-analysis

We did not identify significant heterogeneity among the included studies using the allele model (P = 0.02 and I 2 = 45.9 %). We then calculated the overall OR with the fixed effect model and found a significant association between the rs3764650 polymorphism and AD (P = 1.76E − 26, OR = 1.21, 95 % CI 1.17–1.26, minor allele = C). More detailed results are provided in Fig. 1.

Fig. 1
figure 1

Forest plot for the meta-analysis of the rs3764650 polymorphism using the allele model. M-H Mantel-Haenszel, Fixed fixed effect model, CI confidence interval

Full size image

In addition to the allele model, we used dominant, recessive, and additive models to investigate the association between the rs3764650 polymorphism and AD. Because all three models require the exact number of original genotypes, we selected only 6 of 18 studies for our further analysis. There was no significant heterogeneity among selected studies using the dominant model (P = 0.09 and I 2 = 47 %), recessive model (P = 0.18 and I 2 = 34 %), or additive model (P = 0.21 and I 2 = 30 %). Furthermore, a meta-analysis revealed significant associations using the dominant model (P = 4.00E − 04, OR = 1.17, 95 % CI 1.07–1.28), recessive model (P = 3.00E − 03, OR = 1.43, 95 % CI 1.13–1.81), and additive model (P = 3.00E − 03, OR = 1.49, 95 % CI 1.16–1.91). More detailed results are provided in Fig. 2.

Fig. 2
figure 2

Forest plot for the meta-analysis of the rs3764650 polymorphism using recessive, dominant, and additive models. M-H Mantel-Haenszel, Fixed fixed effect model, CI confidence interval

Full size image

Subgroup Analysis

The frequency of rs3764650 is variable among different populations. Only 6 of 18 studies were selected for further analysis using the dominant, recessive, and additive models. We then conducted a subgroup analysis using the allele model in East Asian and Caucasian populations. There was no significant heterogeneity in the East Asian (P = 0.85 and I 2 = 0 %) or Caucasian (P = 0.65 and I 2 = 0 %) populations. A subsequent meta-analysis revealed significant associations in the East Asian (P = 0.029, OR = 1.09, 95 % CI 1.01–1.19) and Caucasian (P = 6.95E − 22, OR = 1.22, 95 % CI 1.17–1.28) populations.

Sensitivity Analyses and Publication Bias Analyses

We found that the association between rs3764650 and AD did not vary substantially when excluding any one study using the four models (data not shown). However, by excluding the study from Logue et al. (African American) [19] using the allele model, the heterogeneity decreased from P = 0.02 and I 2 = 45.9 % to P = 0.46 and I 2 = 0 %. The four plots were symmetrical inverted funnels (Figs. 3 and 4), suggesting that there was no significant publication bias for the allele model (Begg’s test, P = 0.65, and Egger’s test, P = 0.65), dominant model (Begg’s test, P = 0.25, and Egger’s test, P = 0.25), recessive model (Begg’s test, P = 0.99, and Egger’s test, P = 0.99), or additive model (Begg’s test, P = 1, and Egger’s test, P = 1). The findings from the sensitivity analyses and publication bias analyses indicated that our results were robust.

Fig. 3
figure 3

Funnel plot for publication bias analysis of the rs3764650 polymorphism in AD using the allele model

Full size image
Fig. 4
figure 4

Funnel plot for publication bias analysis of rs3764650 polymorphism in AD using the recessive, dominant, and additive models

Full size image

Discussion

A previous GWAS showed that rs3764650 was significantly associated with AD [4]. Recent studies have investigated the genetic mechanisms of ABCA7 in AD pathogenesis and indicated that ABCA7 plays an important role in AD risk. Vasquez et al. demonstrated that the rs3764650 polymorphism T allele, which is associated with an increased risk of AD, is correlated with increased ABCA7 expression in human brain samples and that ABCA7 expression is increased in AD individuals [11]. Shulman et al. conducted an investigation of 725 subjects, followed by a targeted validation of 114 subjects, and found that ABCA7 (rs3764650; P = 0.02) was associated with neuritic plaque burden [12]. Karch et al. measured the influence of the rs3764650 polymorphism and ABCA7 expression levels on clinical and pathological measures of AD in brain tissue from the parietal lobe of both AD cases and age-matched, cognitively normal controls [13]. Their results showed that the minor allele of rs3764650 C was associated with age at onset and disease duration. ABCA7 expression levels were associated with a clinical dementia rating, with higher expression being correlated with more advanced cognitive decline [13].

Here, due to the weak or negligible association between rs3764650 and AD reported by previous studies in Chinese, Japanese, Korean, African American, and Canadian populations, we reevaluated the association using large-scale samples. We observed that there was no significant heterogeneity among the selected studies using our four models. Our results revealed a more significant association between rs3764650 and AD (P = 1.76E − 26) compared with a previous study (P = 5.00E − 21) [4]. Based on these findings, we consider that the weak and negligible associations between rs3764650 and AD may have been caused by relatively small sample sizes compared with the previous GWAS in populations of European ancestry [4].

In addition to the rs3764650 polymorphism (or its proxies, namely rs115550680, rs3752246, and rs4147929), common AD variants of another eight AD susceptibility genes (CR1, BIN1, PICALM, MS4A4/MS4A6E, CD2AP, CD33, EPHA1, and CLU) were also reported in the Caucasian population. Subsequent studies successfully corroborated the significance of these AD variants by analyzing a large-scale dataset of an Asian population (e.g., PICALM gene rs3851179 polymorphism, BIN1 gene rs744373 polymorphism, CR1 gene rs6656401 polymorphism, and CLU gene rs11136000 variant [23, 25–28]). These results indicated that analyzing relatively large-scale samples is effective in identifying significant associations between these common variants and AD.

Before our submission (January 27, 2014), we accessed the AlzGene database [29]. Among nine articles selected for our analysis, only two articles were included in AlzGene. Neither of the articles provided exact genotype numbers. Lambert et al. recently conducted a large, two-stage meta-analysis of AD GWAS in individuals of European ancestry [9]. In addition to the APOE locus, they identified 19 loci with genome-wide significance (P < 5.00E − 08), of which 11 were newly associated with AD. However, the rs3764650 polymorphism was not included in the 19 loci.

Despite these interesting results, our study has a limitation. Here, we investigated the association between rs3764650 (or its proxies, namely rs115550680, rs3752246, and rs4147929) and AD using allele, dominant, recessive, and additive models. The dominant, recessive, and additive models required exact genotype numbers of all studies used in our analysis. We attempted to obtain these genotype numbers but were not successful. The Alzheimer’s Disease Genetic Consortium (ADGC) study from Naj et al. contained 15 discrete datasets and had a total sample number that was much greater than all of the other included studies combined (after removing datasets that were counted twice). However, the ADGC study could not provide exact genotype numbers for all of the studies. Thus, we divided the selected studies into two classifications: those that could provide exact genotype numbers and those that could not, as described in Table 1. Six of the 18 studies with exact genotype numbers were selected for meta-analyses using dominant, recessive, and additive models. It was important to analyze the dominant, recessive, and additive models; however, the results from these models cannot be compared with those from the allelic model due to discrepancies between the numbers of studies included. Future studies using genotype data are required to corroborate our findings.

To our knowledge, this is the first meta-analysis to investigate the association between the rs3764650 polymorphism (or its proxies, namely rs115550680, rs3752246, and rs4147929) and AD in East Asian, African American, Canadian, and Caucasian populations. Our analysis supports previous findings that the ABCA7 rs3764650 polymorphism contributes to AD susceptibility. We believe that our findings will be very useful for future genetic studies on AD. The more detailed Supplementary Data can also be found from the website http://www.bioapp.org/research/ADrs3764650/.