Alzheimer's disease is the most common form of dementia, with both environmental and genetic factors contributing to risk. Alzheimer's disease is genetically complex and shows heritability of up to 79% (ref. 1). Rare variants in three genes (APP, PSEN1 and PSEN2)1 cause disease in a minority of cases, but until recently, APOE (encoding apolipoprotein E) was the only gene known to increase disease risk for the common form of Alzheimer's disease with late onset2. In 2009, we published a genome-wide association study (GWAS) of Alzheimer's disease in a sample designated GERAD1 (Genetic and Environmental Risk in Alzheimer's Disease Consortium 1), in which we identified two new genome-wide significant susceptibility loci: clusterin (CLU: P = 8.5 × 10−10) and the phosphatidylinositol-binding clathrin assembly protein gene (PICALM: P = 1.3 × 10−9). We also observed more variants with P < 1 × 10−5 than were expected by chance (P = 7.5 × 10−6)3. These included variants in CR1 (the complement receptor 1 gene), BIN1 (the bridging integrator 1 gene) and the MS4A (membrane-spanning 4A gene) cluster. A second independent Alzheimer's disease GWAS4 using the EADI1 (European Alzheimer's Disease Initiative 1) sample showed genome-wide significant evidence for association with CLU (P = 7.5 × 10−9) and CR1 (P = 3.7 × 10−9) and support for association with PICALM (P = 3 × 10−3). A combined analysis of the GERAD1 and EADI1 data yielded highly significant support for all three loci (CLU meta P = 6.7 × 10−16; PICALM meta P = 6.3 × 10−9; and CR1 meta P = 3.2 × 10−12). The associations in CLU, PICALM and CRI have since been replicated in several independent datasets5,6,7,8, have shown trends in another dataset9 and have shown relationships with the neurodegenerative processes underlying disease10. In addition, members of this consortium have since reported genome-wide significant association for BIN1 (P = 1.6 × 10−11) and support for EPHA1 (encoding ephrin receptor A1) (P = 1.7 × 10−6)11.

This study sought to identify new common susceptibility variants for Alzheimer's disease by first undertaking a three-stage association study based upon predominantly European samples (GERAD+; Fig. 1) and then by testing these samples for loci showing suggestive evidence for association in the ADGC GWAS12.

Figure 1: GERAD+ study design. *Data for rs744373 and rs3818361 in the CHARGE consortium have been presented elsewhere15, as has data for rs381861 in the EADI2 samples4; as such these SNPs were not included in stage 3. Full size image

The first stage of this study comprised a meta-analysis of four Alzheimer's disease GWAS datasets (6,688 affected individuals (cases) and 13,685 controls) including: the GERAD1 (ref. 3), EADI1 (ref. 4), Translational Genomics Research Institute (TGEN1)13 and the Alzheimer's Disease Neuroimaging Initiative (ADNI)14 datasets. SNPs which remained significant at P ≤ 1 × 10−5 were then tested for replication in the second stage of this study, comprising 4,896 cases and 4,903 controls, including genotyping of the GERAD2 sample and in silico replication in the deCODE and German Alzheimer's Disease Integrated Genome Research Network (AD-IG) GWAS datasets. In stage 3, previously unidentified SNPs showing significant evidence of replication in stage 2 were then tested for association in a sample comprising 8,286 cases and 21,258 controls, which included new genotyping in the EADI2 (ref. 4) and Mayo2 samples and in silico replication in the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) sample11. Sample descriptions and characteristics can be found in the Supplementary Note and Supplementary Table 1.

In stage 1, we identified 61 SNPs associated with Alzheimer's disease at P ≤ 1 × 10−5 following meta-analysis of 496,763 SNPs in the GERAD1, TGEN1, ADNI and EADI1 samples (Supplementary Table 2 and Supplementary Note). Ten SNPs at newly associated loci and two SNPs at previously identified susceptibility loci that surpassed the P ≤ 1 × 10−5 threshold were selected for further analysis (see below). One SNP, rs610932 (stage 1 P = 1.8 × 10−8) at the MS4A (encoding membrane spanning 4A) gene cluster, surpassed the threshold for genome-wide significance (P < 5.0 × 10−8)15. We also observed strong evidence for association at ABCA7 (encoding ATP-binding cassette, sub-family A, member 7; rs3764650; stage 1 P = 2.6 × 10−7).

When selecting SNPs for testing in stage 2, we excluded known susceptibility loci that had previously been tested in GERAD2, and we limited our analysis of BIN1 and CR1, which had not been tested in GERAD2, to the most significant SNPs at each locus (Supplementary Table 2). Following pruning for linkage disequilibrium (LD), 12 SNPs were taken forward for replication in stage 2 (10 SNPs, excluding BIN1 and CR1).

Five of the 12 SNPs tested in stage 2 showed significant evidence for replication using a Bonferroni-adjusted threshold for significance of P = 4.2 × 10−3 (Table 1 and Supplementary Table 3). In addition to SNPs at BIN1 and CR1, one SNP within ABCA7 (rs3764650, stage 2 P = 1.9 × 10−5) and two SNPS at the MS4A gene cluster (rs610932, stage 2 P = 1.6 × 10−3; and rs670139, stage 2 P = 1.1 × 10−3) showed evidence of replication in stage 2. We tested the three SNPs implicating new risk loci for association in the stage 3 sample and showed further evidence of replication (rs3764650, stage 3 P = 2.9 × 10−7; rs610932, stage 3 P = 2.1 × 10−5; and rs670139, stage 3 P = 3.2 × 10−3; Table 1 and Supplementary Table 3).

Table 1: Results of the GERAD+ study Full size table

We conducted an inverse variance weighted meta-analysis of data from stages 1, 2 and 3 (Table 1 and Supplementary Table 3). This provided strong evidence for association with rs3764650 at ABCA7 (meta P = 4.5 × 10−17) and two SNPs at the MS4A gene cluster: rs610932 (meta P = 1.8 × 10−14) and rs670139 (meta P = 1.4 × 10−9). When combining GERAD+ and ADGC results (after removing overlapping samples), ABCA7 had P = 5.0 × 10−21 (odds ratio (OR) = 1.22). The two SNPs at the MS4A gene cluster, rs610932 and rs670139, showed P = 1.2 × 10−16 (OR = 0.91) and P = 1.1 × 10−10 (OR = 1.08), respectively, in the combined analysis of the GERAD+ and ADGC results. It is noteworthy that the most significant ADGC SNP at the MS4A locus is in LD with our top SNP (rs4938933 with rs610932, r2 = 0.62, D′ = 0.86), and thus both datasets may be detecting the same underlying signal.

This study also provides additional independent support for association with CR1 (stage 2 P = 1.4 × 10−3) and BIN1 (stage 2 P = 3.8 × 10−5; see Table 1 for meta-analysis.) We did not observe interaction between APOE and the new variants identified in this study, and indeed, we did not find evidence of epistasis between any of the genome-wide significant variants identified to date (ABCA7, MS4A, BIN1, CR1, PICALM, CLU or APOE) (Supplementary Table 4a). Likewise, adjusting for the presence of at least one APOE ɛ4 allele had little effect on the results of the analysis of the three newly associated variants (Supplementary Table 4b). We also found no evidence for association between these loci and age at onset of Alzheimer's disease (rs3764650, P = 0.17; rs670139, P = 0.38; rs610932, P = 0.95; rs744373, P = 0.87; and rs3818361, P = 0.58).

This study therefore identifies two new Alzheimer's disease susceptibility loci, which replicate over a number of independent case-control samples. The first of these is the ABCA7 (encoding ATP-binding cassette, sub-family A, member 7) locus (Fig. 2a). The associated marker is rs3764650, which is located in intron 13. This SNP was the only variant in the gene that passed our stage 1 criterion, which is not unexpected given the low levels of LD between this SNP and others included in the GWAS. However, in a preliminary attempt to identify an associated functional variant at the ABCA7 locus, we genotyped the GERAD2 sample for rs3752246, a non-synonymous SNP in exon 32 of the gene, which showed the highest LD with rs3764650 out of all HapMap ABCA7-coding variants based on r2 values (r2 = 0.36, D′ = 0.89). This variant (which was not genotyped in stage 1) was also associated with Alzheimer's disease (GERAD2 P = 1 × 10−3, OR = 1.17). rs3752246 encodes a glycine to alanine substitution at position 1527 of the protein, which is predicted to be a benign change16 and is unlikely to be the relevant functional variant. We used data from two published expression quantitative trait loci (eQTL) datasets (derived from lymphoblastoid cell lines17 and brain18) to determine if rs3764650 is associated with the expression of ABCA7. However, we observed no association (Supplementary Table 5). Further work will be required to identify the causal variant(s) at this locus.

Figure 2: Schematic of the associated variants reported in reference to (a) ABCA7 and (b) chromosomal region chr11: 59.81Mb–60.1Mb harboring members of the MS4A gene cluster. Chromosome positions are shown at the top of the schematics (UCSC Feb 2009). Gene schematic: horizontal arrows indicate directions of transcription, black boxes indicate gene exons and the untranslated region. The −log 10 P of the SNPs analyzed in stage 1 are shown in the chart and graph. The GERAD+ stage 1, 2 and 3 meta-analysis P values for rs3764650 (ABCA7), rs610932 (MS4A6A) and rs670139 (MS4A4E) are indicated by the red lines. The D′ LD block structure of ABCA7 plus the surrounding region and chr11: 59.81Mb–60.1Mb according to the CEPH HapMap data are provided at the bottom of each schematic with lines indicating where each SNP genotyped on the Illumina 610-quad chip is represented. Full size image

Second, we implicate the MS4A (encoding membrane-spanning 4A) gene cluster (Fig. 2b). The association spans an LD block of 293 kb (chr11: 59,814,287–60,107,105) and includes 6 of 16 known genes comprising the membrane-spanning 4-domains subfamily A (encoded by MS4A). These are MS4A2, MS4A3, MS4A4A, MS4A4E, MS4A6A and MS4A6E. The associated SNPs are found in the 3′ untranslated region of MS4A6A (rs610932) and the intergenic region between MS4A4E and MS4A6A (rs670139). rs610932 showed nominally significant association with expression levels of MS4A6A in cerebellum and temporal cortex (0.01 < P < 0.05; Supplementary Table 5) but not in the frontal cortex, pons or lymphoblastoid cell lines. The non-synonymous SNP that was most strongly associated with the genome-wide significant variants was rs2304933. This SNP was analyzed in stage 1 but showed weaker evidence for association (P = 0.006) than the genome-wide significant variant at this locus in the same sample. Figure 3 shows forest plots depicting association in the different datasets for SNPs at the ABCA7 (rs3764650) and MS4A (rs610932 and rs670139) loci.

Figure 3: Forest plots showing association in the different datasets for SNPs at the ABCA7 (rs3764650) and MS4A (rs610932 and rs670139) loci. Full size image

We also sought to follow up four additional loci showing suggestive evidence for association with Alzheimer's disease (1 × 10−6 ≥ P > 5 × 10−8) from the ADGC GWAS12. These loci included those in CD33, EPHA1, CD2AP and ARID5B. It should be noted that evidence for suggestive association with EPHA1 and CD33 has been reported previously. Members of this collaboration were the first to report EPHA1 as showing suggestive evidence of association with Alzheimer's disease (rs11771145, P = 1.7 × 10−6; LD with ADGC rs11767557, r2 = 0.28, D′ = 0.75)11, which included the GERAD1 and EADI1 samples reported on here. Similarly, researchers from another study were the first to show suggestive evidence for CD33 (rs3826656, P = 4.0 × 10−6; LD with ADGC rs3865444, r2 = 0.13, D′ = 1.0)19.

We combined data from the GERAD+ dataset, comprising the GERAD1, EADI1, deCODE and AD-IG GWAS datasets (up to 6,992 cases and 13,472 controls), using inverse variance meta-analysis. We included the TGEN1, ADNI and Mayo1 datasets in the ADGC discovery set and were thus excluded from these particular analyses. We observed support for association with CD2AP (rs9349407, P = 8.0 × 10−4, OR = 1.11), CD33 (rs3865444, P = 2.2 × 10−4, OR = 0.89) and EPHA1 (rs11767557, P = 3.4 × 10−4, OR = 0.90).

When these data were combined with the ADGC data, we observed genome-wide evidence for association with Alzheimer's disease (rs9349407, GERAD+ and ADGC meta P = 8.6 × 10−9, OR = 1.11; rs3865444, GERAD+ and ADGC meta P = 1.6 × 10−9, OR = 0.91; rs11767557, GERAD+ and ADGC meta P = 6.0 × 10−10, OR = 0.90). We observed nominally significant evidence of association with ARID5B (rs2588969, P = 3.3 × 10−2, OR = 1.06); however, the direction of effect was opposite to that reported by ADGC12 and was not significant overall (GERAD+ and ADGC meta P = 3.6 × 10−1, OR = 0.99). See Table 2 for results of the GERAD+ and ADGC combined analyses and Supplementary Table 6 for results of additional SNPs at these loci.

Table 2: Results of the combined analysis of the GERAD+ consortia with the ADGC GWAS × Full size table

Taken together, these results show compelling evidence for an additional five Alzheimer's disease susceptibility loci. ABCA7 encodes an ATP-binding cassette (ABC) transporter. The ABC transporter superfamily has roles in transporting a wide range of substrates across cell membranes20. ABCA7 is highly expressed in brain, particularly in the hippocampal CA1 neurons21 and microglia22. ABCA7 is involved in the efflux of lipids from cells to lipoprotein particles. Notably, the main lipoproteins in brain are APOE followed by CLU. We observed no evidence for epistatic interactions between the three genetic loci (Supplementary Table 4a), however, this is not a prerequisite for biological interaction between these molecules. In addition, ABCA7 has been shown to regulate APP processing and inhibit β-amyloid secretion in cultured cells overexpressing APP (ref. 23). ABCA7 also modulates phagocytosis of apoptotic cells by macrophages mediated through the C1q complement receptor protein on the apoptotic cell surface23. ABCA7 is an ortholog of Caenorhabditis elegans ced-7, the product of which is known to clear apoptotic cells, and the high levels of expression of ABCA7 in microglia are consistent with such a role.

The genes in the MS4A cluster on chromosome 11 have a common genomic structure with all other members of the family, including transmembrane domains, indicating that they are likely to be part of a family of cell-surface proteins24. MS4A2 encodes the β subunit of high affinity IgE receptors25. The remaining genes in the LD block have no known specific functions.

CD33 is a member of the sialic-acid-binding immunoglobulin-like lectins (Siglec) family, which is thought to promote cell-cell interactions and regulate functions of cells in the innate and adaptive immune systems26. Most members of the Siglec family, including CD33, act as endocytic receptors, mediating endocytosis through a mechanism independent of clathrin27. CD2AP (CD2-associated protein) is a scaffold adaptor protein28 that associates with cortactin, a protein also involved in the regulation of receptor-mediated endocytosis29. It is striking that these two new susceptibility genes for Alzheimer's disease, and the recently established susceptibility genes PICALM and BIN1, are all implicated in cell-cell communication and transduction of molecules across the membrane. EPHA1 is a member of the ephrin receptor subfamily. Ephrins and Eph receptors are membrane-bound proteins which play roles in cell and axon guidance30 and in synaptic development and plasticity31. However, EPHA1 is expressed mainly in epithelial tissues32 where it regulates cell morphology and motility33. Additional roles in apoptosis34 and inflammation35 have also been proposed.

Our study has identified variants at ABCA7 and the MS4A gene cluster associated with susceptibility to Alzheimer's disease with replication over a number of independent case-control samples. We also provide independent support for three loci showing suggestive evidence in a companion paper12: loci in CD33, CD2AP and EPHA1, which, when the data were combined, showed genome-wide levels of significance. Finally, we provide further replication evidence for BIN1 and CR1 loci as Alzheimer's disease susceptibility loci. What is striking about our findings is the emerging consistency in putative function of the genes identified. Five of the recently identified Alzheimer's disease susceptibility loci in CLU, CR1, ABCA7, CD33 and EPHA1 have putative functions in the immune system; PICALM, BIN1, CD33 and CD2AP are involved in processes at the cell membrane, including endocytosis, and APOE, CLU and ABCA7 are involved in lipid processing. It is conceivable that these processes would play strong roles in neurodegeneration and Aβ clearance from the brain. These findings therefore provide new impetus for focused studies aimed at understanding the pathogenesis of Alzheimer's disease.