Genetics of Alzheimer disease in the pre- and post-GWAS era

Since the 1990s, the genetics of Alzheimer disease (AD) has been an active area of research. The identification of deterministic mutations in the APP, PSEN1, and PSEN2 genes responsible for early-onset autosomal dominant familial forms of AD led to a better understanding of the pathophysiology of this disease. In the past decade, the plethora of candidate genes and regions emerging from genetic linkage and smaller-scale association studies yielded intriguing 'hits' that have often proven difficult to replicate consistently. In the last two years, 11 published genome-wide association studies (GWASs) in AD confirmed the universally accepted role of APOE as a genetic risk factor for late-onset AD as well as generating additional candidate genes that require confirmation. It is unclear whether GWASs, though a promising novel approach in the genetics of complex diseases, can help explain most of the underlying genetic risk for AD. This review provides a brief summary of the genetic studies in AD preceding the GWAS era, with the main focus on the findings from recent GWASs. Potential approaches that could provide further insight into the genetics of AD in the post-GWAS era are also discussed.

of presenilin function [26] as additional pathophysiologic mechanisms underlying AD.

Late-onset Alzheimer disease
Th e identifi cation of a chromosome 19 risk locus in LOAD families [27] was followed by the discovery of a higher frequency of the APOE ε4 allele in LOAD patients compared with controls [28,29]. APOE is a component of senile plaques [30], binds Aβ [31], can infl uence neuritic plaque formation in transgenic mouse models of AD in an isoform-specifi c fashion [32], and is thought to contribute to both Aβ clearance and deposition in the brain [33]. In vitro and in vivo studies also suggest a role for APOE in isoform-specifi c synaptogenesis and cognition, neurotoxicity, tau hyperphosphorylation, neuroinfl ammation, and brain metabolism, although these non-Aβ-related mechanisms require further investigation [34]. Unlike the EOFAD mutations that are fully penetrant, APOE ε4 is a genetic risk factor that is neither necessary nor suffi cient for the development of AD (see [35] for review of APOE in LOAD). Th e odds that APOE ε3/ε4 genotype carriers have AD is estimated to be two to four times greater than that of APOE ε3/ε3 carriers, according to population-based association studies in subjects of European origin. Th e odds ratio (OR) increases to approximately 6 to 30 in the APOE ε4/ε4 genotype carriers. Although there is evidence of a risk eff ect of APOE ε4 in non-Europeans, the estimated eff ect sizes are smaller with less consistent results in African-American and Hispanic subjects, which may suggest diff erent underlying genetic or environmental factors or both for these ethnic groups. Th e eff ect of APOE ε4 appears to be age-dependent, with the strongest eff ect observed before age 70. Th e use of APOE as a diagnostic [36] or predictive factor in clinical practice is not warranted. Population attributable risk (PAR) of AD due to a genetic factor describes the diff erence in rate of AD in the population between those who are carriers of the genetic risk versus those who are not. Another way to consider PAR is the amount of decrease that would be observed in the incidence of AD if the genetic factor could be eliminated. It is important to note that the PAR due to a combination of risk factors is usually smaller than the sum of PARs from multiple risk factors because a person with disease may have more than one risk factor (that is, the cause of disease in any given case may be attributable to more than one factor). Th e estimated LOAD PAR for APOE is 20% to 70% [4,35]. Th is relatively wide range of PAR estimates for APOE likely refl ects the diff erence in study types (population-based versus clinicbased, longitudinal versus cross-sectional), age range of the population, and other characteristics such as family history. PAR esti mates may have an upward bias in family-based studies and younger populations given that APOE has a stronger eff ect in younger ages (from younger than age 70 to eighties) and longitudinal studies may provide more accurate estimates of PAR of APOE where estimates of 20% were obtained [35,37,38]. Statistical approaches suggesting the presence of AD loci in addition to APOE [39] led to intensive research eff orts in the genetics of LOAD.

Linkage and smaller-scale association studies
Generation of the genetic and physical map of the human genome [40] largely based on polymorphic tandem repeat regions led to the fi rst-generation genome-wide scans in AD as well as many other diseases. Ten inde pendent whole-genome linkage and four association studies in AD were completed using microsatellite markers (reviewed in [35]) between 1997 and 2006. Th ese studies typically used approximately 200 to 400 micro satellite markers, covering the genome at every 5 to 16 centimorgans (1 cM is approximately equal to 1 million base pairs). Th e whole-genome linkage studies were conducted in AD families or sibships collectively composed of about 100 to 2,000 subjects. Whole-genome association studies were conducted on AD case-control series composed of a small number of subjects (n is approximately equal to 10 to 210, including an approximately equal number of AD patients and controls). Th ese studies were hypothesisfree (or hypothesis-generating) in that they provided a survey of the whole genome without requiring any prior knowledge or theory about the genes/genomic regions for their role in AD. All of the studies that were done in subjects of European origin identifi ed the APOE region on chromo some 19 as a risk region in AD. Importantly, however, they also detected other genetic regions, at times with a signal stronger than that of APOE. Some of these regions were detected in multiple independent studies, strongly suggesting the presence of non-APOE genetic factors underlying the risk of AD. Loci on chromosomes 6, 9, 10, and 12 [29,35] yielded multiple, independent, strong signals that led to subsequent fi nemapping eff orts, which have proven diffi cult due to the typically large regions identifi ed in these studies covering tens of millions of base pairs. Nonetheless, these fi ndings generated positional and functional candidate genes that were assessed in association studies.
Th e identifi cation of single-nucleotide polymorphisms (SNPs) in the human genome [41] and development of high-throughput SNP genotyping technologies galvanized such studies to understand the underlying genetics of common, complex diseases. Since the 1990s, more than 1,000 association studies have been carried out in nearly 600 AD candidate genes, regularly updated in the AlzGene website [42], which is a database of genetic linkage and association studies in AD [29]. Some of the AD candidate gene studies led to the identifi cation of genes with plausible functions in AD pathophysiology and evidence of association with AD risk in relatively large series. Sortilin-related receptor (SORL1) [43] and calcium homeostasis modulator 1 (CALHM1) [44] are two such genes. Variations in both genes were assessed in more than 3,000 to 4,000 subjects from multiple independent family-based or case control series in these studies. Furthermore, considerable functional analyses implicated SORL1 in APP sorting [43] and CALHM1 in calcium signaling that also infl uences Aβ levels [44]. Despite these initial results and some supportive followup studies from independent groups, there have also been reports of negative or weaker association for these genes summarized in AlzGene. Inconsistent replication of original association fi ndings has been the rule rather than the exception in AD, except for APOE, and even for candidate genes with convincing functional data and thorough genetic assessment with positive association results in the initial study.
Potential problems in candidate gene association studies of AD and proposed solutions are discussed elsewhere [29,45,46] and will be briefl y mentioned here. Th e potential causes for inconsistent replication of associa tion studies in AD (and other common, complex diseases) include initial false-positive results, falsenegative follow-up results, and heterogeneity in phenotype, genotype, or environment. Multiple testing, population stratifi cation, genotyping errors, and initial small sample size are potential reasons for false-positive fi ndings in the original study. Underpowered studies that are too small to detect a modest eff ect size can lead to false-negative follow-up studies. Meta-analyses of associa tion studies in AD [29] and other common diseases reveal modest estimated eff ect sizes with ORs of less than 2.0 (or greater than 0.5 for protective alleles) [47]. It is estimated that thousands to tens of thousands of subjects are required to have suffi cient power to detect such eff ect sizes, a prerequisite that has typically not been fulfi lled for many association studies in AD until recently.
Genetic heterogeneity in the form of both diff erent disease genes and diff erent alleles (in the same genes) underlying disease risk could account for failure of replication in association studies of common, complex diseases. Th e situation becomes even more complicated when potential gene-gene and gene-environment interactions are considered. Finally, given that association studies rely on the existing linkage disequilibrium (LD) between the genotyped markers and the actual functional disease susceptibility variant (typically not genotyped), heterogeneity in the extent of this LD between diff erent study populations could also account for inconsistent results between diff erent studies. Study of isolated populations or disease groups or both with distinct or extreme phenotypes could be used in an attempt to decrease genic, allelic, phenotypic, and environmental heterogeneity but may come at the expense of limited generalizability of study results. Use of multiple, informative, putative functional markers; testing for and correcting population substructure; and use of biologically relevant, quantitative phenotypes (endophenotypes) in addition to the disease phenotype are additional approaches to address heterogeneity and potentially to increase power in association studies. Testing multiple independent series internally prior to publication, supplementing genetic data with functional assays, and choice of candidate genes and variants with increased a priori probability of association based on biology and position of the gene can increase the probability of success in association studies of common diseases such as AD.

Genetics of Alzheimer disease in the GWAS era
Th e International HapMap Project, launched in October 2002, led to the generation of a database of the common variations (defi ned as minor allele frequency of greater than 0.05) and the underlying LD structure in the human genome [48,49] that provided the foundation for the genome-wide association studies (GWASs) that use highthroughput genotyping platforms composed of common SNP markers that tag a subset of the known common SNPs in the human genome. For example, the Aff ymetrix GeneChip 500K platform (Aff ymetrix, Santa Clara, CA, USA) has 68% coverage of the phase II HapMap SNPs at r 2 of at least 0.8 in the subjects of European ancestry (CEU), whereas the Illumina Hap300 platform (Illumina, Inc., San Diego, CA, USA) has 77% coverage [49]. Furthermore, coverage is less in non-European populations, up to 1% of the common variations are untaggable by other SNPs using LD, and there likely exist millions of 'rare' SNPs not covered by these platforms. Despite their limitations, GWASs allowed a higher-resolution screen of the human genome for common diseases and traits. In the last two years, 11 GWASs in LOAD have been published. In this section, brief summaries of each of these studies are provided, followed by collective conclusions.

Study-specifi c characteristics and results of the late-onset Alzheimer disease GWASs
Th e study designs and results of the 11 LOAD GWASs are depicted in detail in Tables 1 and 2, respectively [50][51][52][53][54][55][56][57][58][59][60][61]. Th e fi rst published LOAD GWAS used a select set of 17,343 SNPs from 11,211 genes that were chosen based on their likelihood of being functional polymorphisms [50]. Th ese SNPs were weighted heavily toward missense mutations, although there were also variants in transcription factor-binding sites, introns, intergenic regions as well as other putative functional SNPs. Authors used a multi-stage design, in which the SNPs were genotyped in pooled DNA from 380 LOAD and 396 control subjects in the fi rst stage. One thousand fi ve hundred forty-four SNPs with AD association at P values of less than 0.075 or less than 0.1 and in biological candidate genes were genotyped in a second pool of 376 LOAD and 344 control subjects. Of these SNPs, 119 showed association at a P value of less than 0.15 with the same risk allele associating in the second pool. Th ese 119 SNPs were genotyped individually in a total of four series, including the two initial pooled series. Eighteen of these SNPs that had a P value of less than 0.05 in the two pooled series and meta-analysis P value of less than 0.005 in the four series combined were also genotyped in a fi fth series. Altogether, nearly 4,000 subjects were genotyped in this study, although the multi-stage design with two pooled DNA series in the initial stages considerably reduced the cost of this study. Nonetheless, some associations may have been missed due to the reduced sensitivity of the pooling approach and the limited number of SNPs assessed. Despite these shortcomings, APOE-related SNPs (SNPs in APOE or in LD with SNPs in APOE) were identifi ed with study-wide signifi cance in this study. Fifteen additional non-APOE SNPs were identifi ed with nominal signifi cance after meta-analyses of all fi ve series, with modest ORs of 1.07 to 1.2 in all series combined. Of these SNPs, one was a missense mutation (rs3745833) in the galanin-like peptide pre cursor (GALP) gene on chromosome 19 which showed suggestive study-wide signifi cance. Th e galanin gene has been implicated in neuronal survival, regeneration, and neuroprotection and inhibition of learning and memory through reductions in glutamate release and reduction of long-term potentiation (reviewed in [62]). Galanin expression has been shown to be upregulated in AD brains, although it is unclear whether this is a cause or eff ect. In this study, the authors highlighted additional genes among the fi nal 15 -such as phosphoenolpyruvate carboxykinase 1 (PCK1), traffi cking protein, kinesin binding 2 (TRAK2), and tyrosine kinase, non-receptor 1 (TNK1) -with potential biological relevance in AD. Th e second LOAD GWAS used the Aff ymetrix 500K platform to genotype more than 500,000 SNPs in a histopathologically confi rmed series of 664 LOAD patients and 422 controls and identifi ed the APOE locus as the only region with SNPs that reached genome-wide signi ficance after Bonferroni correction for multiple testing [51]. Th e same group reassessed their data by analyzing APOE ε4-positive and -negative subjects separately and dividing their histopathologic series into discovery and replication series [52]. Th ey focused on 312,316 SNPs after quality control exclusions and also included a clinical replication cohort in their genotyping and analysis. Th ey identifi ed 10 SNPs in the GRB-associated binding protein 2 (GAB2) gene on chromosome 11q14.1 . GAB2 encodes a scaff olding protein, GRB-associated binding protein 2, which is involved in cell signaling pathways, especially in the immune system [63]. Its potential role in AD pathophysiology remains to be elucidated; however, preliminary functional studies accompanying this LOAD GWAS revealed diff erential expression of GAB2 in AD versus control brains, co-localization of GAB2 with dystrophic neurites, and variation of GAB2 expression infl uencing tau phosphorylation [52]. Th e third LOAD GWAS [53] analyzed more than 400,000 SNPs from the Aff ymetrix 500K platform in 753 AD patients and 736 controls from the discovery series and followed up the 120 top SNPs in the 418 AD patients and 249 controls from the replication series. Th ey identifi ed APOE-related SNPs with genome-wide significance in the combined series. Of the top 120 associations, 3 non-APOE SNPs passed their quality control thresholds and had nominal logistic regression P value of less than 0.05 in their replication series. Two of these SNPs, rs7019241 and rs10868366, reside in the Golgi phosphoprotein 2 gene (GOLPH2) on chromosome 9, which is in a suggestive linkage region [35] and encodes a protein functionally involved in Golgi transmembrane traffi cking, a subcellular localization where APP cleavage by the γsecretase complex occurs [64]. Th e other SNP, rs9886784, resides within a copy number deletion polymorphism on chromosome 9. A fourth SNP, rs10519262, which resides in an intergenic region on chromosome 15, was identifi ed with the nominal P cutoff of less than 0.05 using Cox proportional hazards regression analysis for age-at-onset phenotype.
Using a pooled DNA approach on 1,082 LOAD patients and 1,239 controls, the fourth GWAS tested more than 500,000 SNPs from two Illumina platforms [54]. Th ese SNPs were submitted to three diff erent analyses to test for allelic frequency diff erences in this pooled population, followed by individual genotyping in the same series. One hundred fourteen of the 237 chosen SNPs could be genotyped individually. In addition to the fi ve APOErelated SNPs with a high level of signifi cance (P = 8.97 × 10 −5 to 5.89 × 10 −9 ), 74 SNPs had nominal P values of not more than 0.05 and were genotyped in an additional control group of 1,400 subjects. Although this is not truly an independent test, the authors used this approach to detect fi ve non-APOE SNPs that became even more signifi cant, of which rs727153 had the best P value (P = 3.4 × 10 −6 ). Th is SNP resides in a haplotype block on chromosome 4, including the lecithin retinol acyltransferase gene (LRAT), which is implicated in the retinoid pathway [65]. Components of this pathway have been suggested to play a role in AD pathogenesis and proposed as potential drug targets. Genotyping additional SNPs in this study led to improved association signal (rs201825, P = 6.12 × 10 −7 , OR = 1.3) in the overall group, although none reached genome-wide signifi cance after Bonferroni correction.
Th e fi rst LOAD GWAS to use family-based analysis tested more than 400,000 SNPs in 941 AD patients versus 404 controls from 410 families of European descent [55]. Th e follow-up was done in three series composed of 1,767 AD patients versus 838 controls from 875 families, also of European descent. Th e authors used a familybased association approach that assesses disease status and age of onset jointly. To obtain the corrected levels of signifi cance, they used not the traditional Bonferroni method but a weighted-Bonferroni correction approach [66] that fi rst screens all of the markers and estimates the conditional power of each marker, followed by the familybased association tests that are corrected using the weights determined from the screening analysis. According to this paradigm, they detected four non-APOE SNPs with P values of 2.0 × 10 −3 to 4.0 × 10 −6 , which would not have been signifi cant after the traditional Bonferroni correction. Th ese four markers had nominal signifi cance of P = 0.3 to 2.0 × 10 −5 in the replication series. Th e two SNPs with nominal signifi cance in both stages, rs11159647 and rs3826656, achieved P values of 2.0 × 10 −6 and 6.0 × 10 −6 , respectively, for AD association in the combined series, which fell short of genome-wide significance after Bonferroni correction. rs11159647 appeared to have an eff ect on age of onset as well. Th ese SNPs had a nominal P value of less than 0.05 in another LOAD GWAS [52,53]. rs11159647 on 14q31.2 and rs3826656 on chromosome 19q13.33 reside in regions without any mapped RefSeq genes, although the latter is in a predicted gene, which overlaps with the CD33 antigen gene. Th is group also evaluated the top genes from the AlzGene database at the time and determined some variants at P = 0.03 and 0.002. Notably, GAB2 [52] showed one of the strongest nominal association signals. Th e sixth LOAD GWAS genotyped more than 500,000 SNPs from the Illumina HumanHap550 platform in 492 LOAD cases and 496 controls in addition to imputing genotypes using a previously published LOAD GWAS [52]. Th e authors used a modifi ed false discovery rate approach (false discovery rate-beta uniform mixture, or FDR-BUM) [67] to declare signifi cance at the genomewide level. Besides APOE-related SNPs, some of which would be signifi cant after Bonferroni correction, rs11610206 on 12q13, which does not reside in a RefSeq gene, was identifi ed at genome-wide signifi cance using the more relaxed FDR-BUM criteria (nominal P = 1.93 × 10 −6 ). When genotyped in the follow-up series of 238 LOAD and 220 control subjects, this SNP was nominally signifi cant at P = 0.0496 and the signifi cance in the joint series improved to P = 3.452 × 10 −7 , falling short of signifi cance after Bonferroni correction. Additionally, any SNP that had a P value of 0.0001 or less in this study or another LOAD GWAS [52], was imputed in the study that did not genotype that SNP. Th ey identifi ed eight SNPs with a nominal P value of less than 0.05 in the other study, in which the joint P value reached 1.51 × 10 −6 for the most signifi cant SNP in ZNF224 on chromosome 19q13. Th ey also evaluated the AlzGene candidates in these two LOAD GWASs by assessing the genotyped or imputed SNPs in these genes for association with AD. Twenty-fi ve SNPs from nine genes (ADAM12, CSF1, GBP2, KCNMA1, NOS2A, SORCS2, SORCS3, SORL1, and WWC1) had nominal P values of 0.003 to 0.05 in the individual LOAD GWAS with P values of 0.0001 to 0.01 in the joint analysis. In addition to providing suggestive evidence for a number of loci in the genome, this study drew attention to the challenges of comparing multiple GWASs via impu tation, in which lack of suffi cient marker infor mation could lead to false-negative results. In these two studies, this was the case with the APOE SNPs that were highly signifi cant in each study but could not be imputed in the other study, leading to missing the APOE signal in the joint analyses.
A single-stage LOAD GWAS from Germany analyzed 491 AD patients versus 479 younger controls and focused on the SNPs for the top 10 genes from AlzGene at the time of their study and SORL1. Th us, although 555,000 SNPs were genotyped, this study, due to its analytical approach, should be considered a follow-up to other association studies rather than a hypothesis-generating GWAS. Both single SNP and haplotype analyses were performed. In addition to APOE-related SNPs, some of which achieved genome-wide signifi cance after Bonferroni correction, SNPs in 8 of the 11 genes tested (MAPT, SORL1, CHRNB2, CH25H, GAB2, PGBD1, PCK1, and LMNA) revealed nominal signifi cance of P values of 0.05 to 6.8 × 10 −3 in this study. None of the non-APOE SNPs achieved genome-wide signifi cance in their study, but the authors provided follow-up analysis results that can be used in meta-analyses of all available data on these genes.
Th e second family-based LOAD GWAS assessed the smallest sample size to date, focusing on two extended LOAD families with 9 aff ected and 10 unaff ected family members versus 60 unrelated controls from the CEPH (Centre d'Etude du Polymorphisme Humain) collection. In stage I, they genotyped more than 400,000 SNPs and identifi ed association with 6 SNPs in the TRPC4AP (transient receptor potential cation channel, subfamily C, member 4 associated protein) gene at genome-wide signi fi cance (P = 5.63 × 10 −11 to 3.85 × 10 −10 ). Because this initial analysis compared ADs from families with unrelated controls outside of the families, it cannot be considered a family-based analysis. Th e authors subsequently genotyped 10 SNPs in TRPC4AP and identifi ed a common haplotype with increased frequency in the AD patients from the families compared with the control spouses. Th e same haplotype was also associated with LOAD in an unrelated case control series of 199 LOAD versus 85 control subjects. Th e authors suggest a functional role for this gene in AD through its interaction with proteins in the infl ammatory cascade and its role in calcium homeostasis. Despite the genome-wide signifi cant results in this study, the fi ndings need further replication given the small sample sizes and analytic approach in the fi rst stage, which used unrelated controls versus AD patients from the families. Indeed, this is the only GWAS that failed to identify APOE; this could be due to the fact that these families have risk factors that are distinct from the general population, but this could also suggest lack of suffi cient power.
Th e third largest LOAD GWAS to date assessed 844 LOAD subjects versus 1,255 controls in the fi rst stage, which assessed more than 300,000 SNPs on an Illumina platform. Th e authors followed the top 25 SNPs from this stage for association in the follow-up series of 1,547 LOAD subjects versus 1,209 controls. APOE-related SNPs were the only ones that were signifi cant at the genome-wide level in the fi rst stage. Upon combined analysis of all samples, rs5984894 in the protocadherin 11X (PCDH11X) gene on chromosome Xq21.31 was the only one that achieved genome-wide signifi cance (P = 3.8 × 10 −8 , OR = 1.29). Multi-variable logistic regression analysis using male sex as a covariate and comparing male hemizygotes, female heterozygotes, and female homozygotes with the female non-carriers yielded global a P value of 3.9 × 10 −12 for the global association. Male hemizygotes showed suggestive association (P = 0.07), whereas female heterozygotes and homozygotes had nominally signifi cant associations (P = 0.01 and 2.0 × 10 −7 , respectively), with OR estimates of 1.18, 1.26, and 1.75, respectively, for these groups. PCDH11X is the fi rst X-chromosomal candidate AD gene identifi ed in a LOAD GWAS. Protocadherins belong to the superfamily of cadherins that are involved in cell adhesion, cell signaling, and neural development. Protocadherins are expressed predominantly in the brain, suggesting their potential role in brain morphogenesis [68]. Although the functional role of this gene in AD needs to be established and the genetic eff ect confi rmed through additional studies, it is an intriguing hypothesis that this X-chromo somal gene could explain the increased risk of AD in women.
Th e two most recent LOAD GWASs are by far the largest ones published to date in this fi eld [60,61]. One of the two GWASs in LOAD case control series in which non-APOE SNPs reached genome-wide signifi cance after Bonferroni correction in the fi rst stage was performed in 3,941 LOAD and 7,848 control subjects from 13 diff erent centers in Europe and the US, where up to more than 500,000 SNPs were analyzed. In addition to the APOErelated SNPs that revealed genome-wide signifi cance (P = 4.9 × 10 −37 to 1.8 × 10 −157 ), rs11136000 in clusterin (CLU or ApoJ) on chromosome 8 and rs3851179 in the phosphatidylinositol-binding clathrin assembly protein (PICALM) gene on chromosome 11 yielded genomewide signifi cance with P = 1.4 × 10 −9 and 1.9 × 10 −8 , respectively. When genotyped in the follow-up series of 2,023 LOAD versus 2,340 control subjects from fi ve centers in Europe, these SNPs had nominal signifi cance at a P value of less than 0.05. Signifi cance for rs11136000 in CLU improved to P = 8.5 × 10 −10 (OR = 0.86) and that for PICALM improved to P = 1.3 × 10 −9 (OR = 0.90) in the combined series. In addition, the authors identifi ed more SNPs than would be expected by chance with P values of less than 1.0 × 10 −5 , including an SNP in the complement receptor 1 gene (rs1408077) that was identifi ed at genome-wide signifi cance in the other largest LOAD GWAS to date [61] (P = 8.3 × 10 −6 ). Th e authors tested more than 100 SNPs in their series that were identifi ed in other LOAD GWASs and identifi ed nominal signifi cance at a P value of less than 0.05 for a number of these, including PCDH11X and SORL1.
Th e other large LOAD GWAS analyzed more than 500,000 SNPs in 2,032 LOAD versus 5,328 control subjects from France [61]. Like the other large LOAD GWASs, this study identifi ed rs11136000 in CLU with genome-wide signifi cance in the fi rst stage (P = 9.0 × 10 −8 ) in their analysis, correcting for population stratification in addition to APOE-related markers (P = 5.06 × 10 −7 to less than 2 × 10 −16 ). Th ose SNPs with a P value of less than 1.0 × 10 −5 were genotyped or imputed in the follow-up series composed of 3,978 LOAD versus 3,297 controls from 15 centers in four countries. CLU SNP rs11136000 had nominal signifi cance in their second stage and enhanced signifi cance of P = 7.5 × 10 −9 (OR = 0.86) in their combined series. Another SNP, rs6656401, in complement component receptor 1 (CR1) also achieved genome-wide signifi cance in the combined series with P = 3.7 × 10 −9 (OR = 1.21). Th e authors also identifi ed nominal signifi cance for SNPs in PICALM (P = 1.0 × 10 −2 to 1.0 × 10 −3 ) and PCDH11X (0.01 < P < 0.05) in their series, thereby providing additional evidence for these genes identifi ed in other LOAD GWASs [59,60].
CLU encodes clusterin or ApoJ, which along with APOE is one of the most abundant apolipoproteins in the human brain. In vivo studies suggest that clusterin, like APOE, is involved in Aβ clearance from the brain (reviewed in [33]). Th ere are also studies that revealed a role for clusterin in Aβ fi brillogenesis and neurotoxicity. Th ese results raise the possibility that, in Aβ pathophysiology, clusterin may have a dual role similar to that of APOE. PICALM encodes a protein involved in clathrin-mediated endocytosis, a suggested pathway for traffi cking of APP that could also infl uence Aβ formation [69]. Th ere is also evidence that PICALM infl uences endocytosis of the synaptic vesicle protein VAMP2 from the plasma mem brane [70], suggesting a role in synaptic function. CR1 is a receptor for the complement component C3b, which has been suggested to be involved in the peripheral clearance of Aβ [71]. Th us, all three candidate genes that emerged from the two largest LOAD GWASs to date have putative functions in the Aβ cascade and synaptic machinery (for PICALM). Additional non-Aβmediated pathophysiologic mechanisms for the proteins encoded by these genes may exist and require further investigation.
It is important to note that a number of the 'hits' identifi ed from the LOAD GWASs reside in prior linkage regions (for example, GALP on chromosome 19, LMNA on chromosome 1, GOLPH2 on chromosome 9, and CLU on chromosome 8) and that others (such as MAPT and SORL1) have previously been implicated in candidate gene association studies. Multiple lines of evidence in support of the same gene provide additional support for their role in the risk of AD; however, in general, these results need to be interpreted with caution and special attention to any sample overlaps, eff ect size, and direction of eff ect in the diff erent studies. Th e AlzGene database facilitates the visualization and cross-checking of fi ndings from diff erent linkage and association studies [29].

Cumulative results from late-onset Alzheimer disease GWAS
GWASs in LOAD led to the generation of a multitude of novel candidate genes for follow-up. Table 3 depicts the list of genes and variants with the strongest level of association in these studies. Th e most consistent pattern that emerges from this table is that the eff ect size and strength of association for APOE-related variants are greater than the best non-APOE associations. Th ere are multiple explanations for this fi nding: It is possible that      Harold et al. [

3.3%
The minor allele frequencies (MAFs), odds ratios (ORs), and population attributable risks (PARs) of the best single-nucleotide polymorphisms (SNPs) in the top gene(s) from the 11 late-onset Alzheimer disease genomewide association studies are shown, where available from the original publications. MAFs of the rs157581 and rs4420638 SNPs were not shown in the original publications and were obtained from the dbSNP database for subjects of European origin. The formula used to calculate PAR was F(OR -1)/(F(OR -1) + 1), where F = frequency of the risk allele and OR = odds ratio of the risk allele. Where available, the MAFs, ORs, and PARs emerging from meta-analyses in the AlzGene website are shown for the non-APOE results.
all LOAD risk variants are common with modest eff ect sizes, as suggested by Table 3. However, the fi ndings from the GWASs are, for the most part, not based on the true susceptibility variants but are refl ective of their tagging markers, which may harbor greater heterogeneity than the former with respect to alleles and extent of LD. Th us, it remains a possibility that the actual functional susceptibility variants may have bigger eff ect sizes. Another possibility is the existence of rare variants with bigger eff ect sizes underlying risk of LOAD, particularly given the insuffi cient coverage of the rare variants in the human genome with the current GWAS platforms. Table 3 provides estimates of the PARs of the strongest APOE and non-APOE variants from the existing GWASs, where necessary information is available. PARs for the APOE-related variants vary between 19% and 35%. Th e PAR estimates for the three genes identifi ed from the two largest LOAD GWASs to date are approximately 9% for CLU and PICALM and 4% for CR1. Th e eff ect sizes from the original studies tend to be overestimates. A number of the GWAS hits were assessed in independent studies summarized in AlzGene, in which estimates from metaanalysis may be closer to the true eff ect sizes of genetic variants in the population. As seen in Table 3, PAR estimates of GALP, GAB2, PICALM, and CR1 decreased upon analyses of additional series, although that of CLU increased slightly. Replication studies with a thousand to tens of thousands of subjects are required to validate these fi ndings and obtain more accurate estimates of eff ect sizes. Th e current combined LOAD PARs for APOE, CLU, PICALM, and CR1 are approximately 56% at most. Given that genetic factors underlie up to 80% of the risk for LOAD [4], the remaining genetic risk for AD could be due to variants in other genes such as PCDH11X or GAB2, rare or structural variants, presence of diff erent genetic factors in non-European populations, and/or gene-gene and gene-environment interactions.

Conclusions and future directions in the post-GWAS era
Discoveries in the sequence and structure of the human genome, technical advances in high-throughput genotyping, and development of novel analytical approaches led to an explosion in the number and extent of genetic studies in common, complex diseases such as LOAD. Numerous candidate genes that emerged from these studies require further validation not only via genetic replication but also through functional assessment by in vitro and in vivo approaches. Experience from other common, complex diseases suggests that, despite their success in uncovering common genetic factors for complex diseases, GWASs may fall short of explaining all of the underlying genetics [72] because of their limitations discussed earlier. Th us, alternative approaches are necessary to identify the remaining genetic susceptibility factors. To identify the actual functional variations responsible for the current association and linkage signals as well as to detect rare and structural variations in candidate regions, next-generation sequencing is emerging as a promising approach. Additionally, novel analytical approaches that can mine available GWAS data may uncover structural variations (such as insertions, deletions, translocations, and inversions) that are associated with disease. Indeed, novel approaches that can re-assess GWAS results may provide an effi cient method to generate novel information. Combining biologically relevant quantitative phenotype (endophenotype) and GWAS data may be a powerful approach that can potentially lead to the identifi cation of genetic factors that infl uence disease risk and plausible biological pathways. It will also be important to jointly assess the results of GWASs and other large-scale association and linkage studies to capitalize on the cumulative knowledge that can be gained from such meta-analyses [73]. Recent advances in the genetics of LOAD provide promise in furthering our understanding of its pathophysiology. With the new discoveries in genomics, genetic studies in the post-GWAS era hold potential for uncovering the complex susceptibility factors for this disease which may translate into therapeutic potential for this devastating epidemic.