Common variants at ABCA7, MS4A6A/MS4A4E, EPHA1, CD33 and CD2AP are associated with Alzheimer's disease.

We sought to identify new susceptibility loci for Alzheimer's disease through a staged association study (GERAD+) and by testing suggestive loci reported by the Alzheimer's Disease Genetic Consortium (ADGC) in a companion paper. We undertook a combined analysis of four genome-wide association datasets (stage 1) and identified ten newly associated variants with P ≤ 1 × 10(-5). We tested these variants for association in an independent sample (stage 2). Three SNPs at two loci replicated and showed evidence for association in a further sample (stage 3). Meta-analyses of all data provided compelling evidence that ABCA7 (rs3764650, meta P = 4.5 × 10(-17); including ADGC data, meta P = 5.0 × 10(-21)) and the MS4A gene cluster (rs610932, meta P = 1.8 × 10(-14); including ADGC data, meta P = 1.2 × 10(-16)) are new Alzheimer's disease susceptibility loci. We also found independent evidence for association for three loci reported by the ADGC, which, when combined, showed genome-wide significance: CD2AP (GERAD+, P = 8.0 × 10(-4); including ADGC data, meta P = 8.6 × 10(-9)), CD33 (GERAD+, P = 2.2 × 10(-4); including ADGC data, meta P = 1.6 × 10(-9)) and EPHA1 (GERAD+, P = 3.4 × 10(-4); including ADGC data, meta P = 6.0 × 10(-10)).

Alzheimer’s disease (AD) is the most common form of dementia, with both environmental and genetic factors contributing to risk. AD is genetically complex and shows heritability up to 79%1. Rare variants in three genes (APP, PSEN1 & PSEN2)1 cause disease in a minority of cases, but until recently the Apolipoprotein E gene (APOE), was the only gene known to increase disease risk for the common form of AD with late-onset2. In 2009 we published a genome-wide association study (GWAS) of AD in a sample designated GERAD1 (Genetic and Environmental Risk in AD Consortium 1), which identified two new genome-wide significant susceptibility loci: clusterin (CLU: P=8.5×10−10) and phosphatidylinositol-binding clathrin assembly protein gene (PICALM: P=1.3×10−9). We also observed more variants with P-values<1×10−5 than were expected by chance (P=7.5×10−6)3. These included variants in the complement receptor 1 (CR1) gene, the bridging integrator 1 (BIN1) gene and the membrane-spanning 4A gene cluster (MS4A gene cluster). A second independent AD GWAS by Lambert and colleagues4 using the EADI1 sample (European Alzheimer’s Disease Initiative 1) showed genome-wide significant evidence for association with CLU (P=7.5×10−9) and CR1 (P=3.7×10−9), and support for PICALM (P=3×10−3). Combined analysis of the GERAD1 and EADI1 data yield highly significant support for all three loci (CLU meta-P=6.7×10−16, PICALM meta-P=6.3×10−9, CR1 meta-P=3.2×10−12). The associations in CLU, PICALM and CRI have since been replicated in several independent datasets5-8, shown trends in another9 and relationships with 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 ephrin receptor A1 (EPHA1; P=1.7×10−6)11..

This study sought to identify new common susceptibility variants for AD by first undertaking a three-stage association study based upon predominantly European samples (GERAD+, see Figure 1) and second, by testing these samples for loci showing suggestive evidence for association in the American Alzheimer’s Disease Genetics Consortium (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.

The first stage of this study comprised a meta-analysis of four AD GWAS datasets (6688 cases, 13685 controls), including: GERAD13, EADI14, Translational Genomics Research Institute (TGEN1)13 and Alzheimer’s Disease Neuroimaging Initiative (ADNI)14. Single nucleotide polymorphisms (SNPs) which remained significant at P≤1×10−5 were then tested for replication in the second stage of this study, comprising 4896 cases and 4903 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, novel SNPs showing significant evidence of replication in Stage 2 were then tested for association in a sample comprising 8286 cases and 21258 controls, which included new genotyping in the EADI24 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 AD at P≤1×10−5 following meta-analysis of 496763 SNPs in the GERAD1, TGEN1, ADNI and EADI1 (see Supplementary Table 2 and the Supplementary Note). Ten SNPs at novel loci and two 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 (membrane spanning 4A) gene cluster, surpassed the threshold (P<5.0×10−8)15 for genome-wide significance. We also observed strong evidence for association at ABCA7 (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 limited analysis of BIN1 and CR1, which had not been tested in GERAD2, to the most significant SNPs at each locus (See Supplementary Table 2). Following pruning for linkage disequilibrium, twelve SNPs were taken forward for replication in Stage 2 (10 excluding BIN1 and CR1).

Five of the twelve SNPs tested in Stage 2 showed significant evidence for replication using a Bonferroni adjusted threshold for significance of P=4.2×10−3 (see 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; rs670139 Stage 2 P=1.1×10−3) showed evidence of replication in Stage 2. The three SNPs implicating novel risk loci were tested 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; rs670139, Stage 3 P=3.2×10−3; see Table 1 and Supplementary Table 3).

Table 1

Results of the GERAD+ study.

SNP	ClosestGene	CHR	MAF	Stage 1*			Stage 2†			Stage 3 ‡			Meta-analysis of GERAD+Stage 1, 2 and 3 §			Meta-analysis of GERAD+& ADGC
				P	OR	95% CI	P	OR	95% CI	P	OR	95% CI	P	OR	95% CI	P	OR	95% CI
rs3764650	ABCA7	19	0.10	2.6×10−7	1.22	1.13-1.32	1.9×10−5	1.28	1.14-1.44	2.9×10−7	1.22	1.13-1.32	4.5×10−17	1.23	1.18-1.30	5.0×10−21	1.23	1.17-1.28
rs610932	MS4A6A	11	0.42	1.8×10−8	0.88	0.85-0.92	1.6×10−3	0.90	0.84-0.96	2.1×10−5	0.91	0.87-0.95	1.8×10−14	0.90	0.87-0.92	1.2×10−16	0.91	0.88-0.93
rs670139	MS4A4E	11	0.41	1.0×10−5	1.11	1.06-1.16	1.1×10−3	1.11	1.04-1.19	3.2×10−3	1.06	1.02-1.11	1.4×10−9	1.09	1.06-1.12	1.1×10−10	1.08	1.06-1.11
rs3818361	CR1	1	0.19	3.2×10−12	1.21	1.14-1.27	1.4×10−3	1.14	1.05-1.23	NA	NA	NA	3.7×10−14	1.18	1.13-1.24	NA	NA	NA
rs744373	BIN1	2	0.29	1.5×10−10	1.17	1.11-1.22	3.8×10−5	1.17	1.08-1.25	NA	NA	NA	2.6×10−14	1.17	1.12-1.21	NA	NA	NA

CHR=Chromosome, MAF=Minor Allele Frequency in cases and controls.

GERAD1, EADI1, ADNI, & TGEN1 <6688 Cases, 13685 Controls.

GERAD2, deCODE, AD-IG: 4896 AD Cases, 4903 Controls.

EADI2, CHARGE, Mayo2 <8286 AD Cases, 21258 Controls,

GERAD1&2, EADI1&2, ADNI, TGEN1, Decode, AD-IG, CHARGE, Mayo2 <19870 AD Cases and 39846 Controls

We conducted an inverse variance weighted meta-analysis of data from Stages 1, 2 and 3 (See 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 has a P-value of 5.0×10−21 (OR=1.22). The two SNPs at the MS4A gene cluster, rs610932 and rs670139, showed P-values of 1.2×10−16 (OR=0.91) and 1.1×10−10 (OR=1.08), respectively, in the combined analysis of 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), 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 novel variants identified in this study, 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) (see Supplementary Table 4a). Likewise, adjusting for the presence of at least one APOE ε4 allele had little effect on the results of analysis of the three novel variants (see Supplementary Table 4b). We also found no evidence for association between these loci and age at onset of AD (rs3764650: P=0.17; rs670139: P=0.38; rs610932: P=0.95; rs744373: P=0.87; rs3818361: P=0.58).

This study therefore shows strong statistical support for two novel AD risk loci, which replicate over a number of independent case-control samples. The first of these is the ATP-binding cassette, sub-family A, member 7 (ABCA7) locus (Figure 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 linkage disequilibrium (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 (r2=0.36, D’=0.89). This variant (which was not genotyped in Stage 1) was also associated with AD (GERAD2 P=1×10−3, OR=1.17). Rs3752246 encodes a glycine to alanine substitution at position 1527 of the protein (accession number NP_061985.2) 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, no association was observed (see 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) the ABCA7 gene 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/UTR. The −Log10(P) of the SNPs analyzed in Stage 1 are shown in chart graph. The GERAD+ Stage 1, 2 and 3meta-analysis P-values for SNPs rs3764650 (ABCA7), rs610932 (MS4A6A) and rs670139 (MS4A4E) are indicated by the red lines. The D’ LD block structure of the ABCA7 gene plus 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.

Second, we implicate the membrane-spanning 4A (MS4A) gene cluster (Figure 2B). The association spans an LD block of 293 kb (chr11: 59,814,28760,107,105) and includes 6 of 16 known genes comprising the membrane-spanning 4-domains, subfamily A (MS4A). These are MS4A2, MS4A3, MS4A4A, MS4A4E, MS4A6A and MS4A6E. The associated SNPs are found in the 3′ UTR of MS4A6A (rs610932) and the intergenic region between MS4A4E and MS4A6A (rs670139). rs610932 shows nominally significant association with expression levels of MS4A6A in cerebellum and temporal cortex (0.01

We also sought to follow up four additional loci showing suggestive evidence for association with AD (1×10−6>=P>5×10−8) from the ADGC GWAS12. These loci included 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 AD (rs11771145, P=1.7×10−6; LD with ADGC SNP rs11767557: r2 = 0.28, D’=0.75)11, which included GERAD1 and EADI1 samples reported on here. Similarly, Bertram and colleagues were the first to show suggestive evidence for CD33 (rs3826656, P=4.0×10−6; LD with ADGC SNP rs3865444: r2 = 0.13, D’=1.0)19.

We combined data from the GERAD+ dataset comprising GERAD1, EADI1, deCODE and AD-IG GWAS datasets (up to 6992 cases and 13472 controls) using inverse variance meta-analysis. The TGEN1, ADNI and Mayo1 datasets were included 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 ADGC we observed genome-wide evidence for association with AD (rs9349407, GERAD+ & ADGC meta-P=8.6×10−9, OR=1.11; rs3865444, GERAD+ & ADGC meta-P=1.6×10−9, OR=0.91; rs11767557, GERAD+ & 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+ & ADGC meta-P=3.6×10−1, OR=0.99). See Table 2 for results and Supplementary Table 6 for results of additional SNPs at these loci.

Table 2

Results of the combined analysis of the ADGC and GERAD+ consortia.

SNP	ClosestGene	CHR	MAF	LinkageDisequilibriumwith the topADGC SNP ateach loci		GERAD+ Consortia *					GERAD+ & ADGC Metaanalysis
				r2	D’	Cases	Controls	P	OR	95% CI	P	OR	95% CI
rs9349407†	CD2AP	6	0.29	N/A	N/A	6283	7165	8.0×10−4	1.11	1.04-1.18	8.6×10−9	1.11	1.07-1.15
rs9296559‡	CD2AP	6	0.29	0.71	0.95	6283	7165	1.5×10−3	1.10	1.04-1.17	NA	NA	NA
rs11767557	EPHA1	7	0.21	N/A	N/A	6283	12935	3.4×10−4	0.90	0.85-0.95	6.0×10−10	0.90	0.86-0.93
rs2588969†	ARID5B	10	0.40	N/A	N/A	6283	7165	3.3×10−2	1.06‡	1.01-1.13	3.6 × 10−1	0.99	0.95-1.02
rs4948288	ARID5B	10	0.26	0.55	0.78	6992	13472	3.6×10−3	1.07‡	1.03-1.15	NA	NA	NA
rs3865444§	CD33	19	0.31	N/A	N/A	6283	7165	2.2×10−4	0.89	0.84-0.95	1.6 × 10−9	0.91	0.88-0.93

CHR=Chromosome, MAF=Minor Allele Frequency in cases and controls.

GERAD1, EADI1, deCODE, AD-IG.

results generated from imputed data. The results from the top genotyped SNP are also shown. See Supplementary Table 6 for full details.

opposite direction of effect to that reported by Naj et al.

data imputed in the deCODE dataset.

Taken together, these results show compelling evidence for an additional five novel AD 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 hippocampal CA1 neurons21 and in 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. Although no evidence for epistasitic interactions between the three genetic loci was observed (see 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 APP23. ABCA7 also modulates phagocytosis of apoptotic cells by macrophages mediated through the C1q complement receptor protein on the apoptotic cell surface23. ABCA7 is an orthologue of C. 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 beta 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 are 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 which associates with cortactin, a protein also involved in the regulation of receptor mediated endocytosis29. It is striking that these two new susceptibility genes for AD, 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 generated strong statistical evidence that variants at ABCA7 and the MS4A gene cluster confer susceptibility to AD, which replicates over a number of independent case control samples. We also provide independent support for three loci showing suggestive evidence in a companion paper12, CD33, CD2AP and EPHA1,which when the data are combined show genome-wide levels of significance. Finally, we provide further evidence for BIN1 and CR1 loci as susceptibility loci. What is striking about our findings is the emerging consistency in putative function of the genes identified. Five of the recently identified AD susceptibility loci CLU, CR1, ABCA7, CD33 and EPHA1 have putative functions in the immune system; PICALM, BIN1, CD33, CD2AP are involved in processes at the cell membrane, including endocytosis and APOE, CLU and ABCA7 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 AD.