Gene-Wide Analysis Detects Two New Susceptibility Genes for Alzheimer's Disease

Background Alzheimer's disease is a common debilitating dementia with known heritability, for which 20 late onset susceptibility loci have been identified, but more remain to be discovered. This study sought to identify new susceptibility genes, using an alternative gene-wide analytical approach which tests for patterns of association within genes, in the powerful genome-wide association dataset of the International Genomics of Alzheimer's Project Consortium, comprising over 7 m genotypes from 25,580 Alzheimer's cases and 48,466 controls. Principal Findings In addition to earlier reported genes, we detected genome-wide significant loci on chromosomes 8 (TP53INP1, p = 1.4×10−6) and 14 (IGHV1-67 p = 7.9×10−8) which indexed novel susceptibility loci. Significance The additional genes identified in this study, have an array of functions previously implicated in Alzheimer's disease, including aspects of energy metabolism, protein degradation and the immune system and add further weight to these pathways as potential therapeutic targets in Alzheimer's disease.


Introduction
The prevalence of Alzheimer's disease (AD) is increasing as more people live into old age. Hope for finding preventative and clinical therapies lies in the ability to gain a better understanding of the underlying biology of the disease, and genetics will provide a valuable starting point for advancement. Rare monogenic forms of AD, the majority of which are attributable to mutations in one of three genes, APP, PSEN1 and PSEN2, exist, but common, lateonset AD is genetically complex with heritability estimated to be between 56-79% [1,2]. Along with the APOE polymorphism [3], 20 common susceptibility loci have been identified associated with AD [4][5][6][7][8][9]. (This figure does not include CD33 as it did not show genome-wide significance in the original report [9].) Recently, a moderately rare variant in TREM2 has also shown evidence for association [10]. However, new variants remain to be found. This study sought to identify new susceptibility genes, using an alternative gene-wide analytical approach, which focuses on the pattern of association within gene regions.
Genome-wide association (GWA) studies to date have focused on single nucleotide polymorphisms (SNPs) as the unit of analysis. Single locus tests are the simplest to generate and to interpret, but have limitations. For example, if susceptibility is conferred by multiple variants within a locus [11,12], this gives rise to complex patterns of association that might not be reflected by association to the same SNPs in different samples, despite apparently reasonably powered tests [13,14]. In addition, rare risk-increasing variants may not be tagged by single SNPs, as is e.g. the case for CLU in which significant enrichment of rare variants in patients was observed independent of the single locus GWA signal [15]. It is therefore likely that the power to detect association might be enhanced by exploiting information from multiple signals within genes encompassed by gene-wide statistical approaches [12]. Disease risk may reflect the co-action of several loci but the number of loci involved at the individual or the population levels are unknown, as is the spectrum of allele frequencies and effect sizes [16]. The observations of multiple genome-wide significant or suggestive linkage signals for disorders, that do not readily replicate between studies but which are not randomly distributed across the genome [17,18] is compatible with the existence of multiple risk alleles of moderate effect that would implicate a locus in disease risk, when analysed together. Thus the first aim of this study is to test for gene-wide association with AD, using a powerful mega-meta analysis of genome-wide datasets as part of the International Genomics of Alzheimer's Project (IGAP) Consortium comprising four AD genetic consortia (see the full list of consortia members in Materials S1): Genetic and Environmental Risk in Alzheimer's Disease (GERAD), European Alzheimer's Disease Initiative (EADI), Cohorts for Heart and Aging in Genomic Epidemiology (CHARGE) and Alzheimer's Disease Genetics Consortium (ADGC) (see full IGAP datasets description in Materials S2). A two stage study was undertaken. In Stage 1 the combined sample included 17,008 AD cases and 37,154 controls. In Stage 2 loci with p-values (combined over all SNPs at the locus) less than 10 24 were selected for replication for 8,572 AD cases and 11,312 controls of European ancestry. We observed evidence for gene-wide association at loci which implicate genes which already show genome-wide significant association from single SNP analysis (CR1, BIN1, HLA-DRB5/HLA-DRB1, CD2AP, EPHA1, PTK2B, CLU, MS4A6A, PICALM, SORL1, SLC24A4, ABCA7, APOE), three new genes in the vicinity of lately reported single SNP hits [9] (ZNF3, NDUFS3, MTCH2) and two novel loci (TP53INP1, combined p = 1.4610 26 and IGHV1-67 combined p = 7.9610 28 ).

Results
Initially, we tested for excess genetic signal revealed by the Stage 1 IGAP SNP GWAS study. We observed more SNPs at all significance intervals, and more genes at multiple significance thresholds, than expected by chance (Table S1). This is unlikely to be due to uncorrected stratification, since each of the individual GWAS samples in the IGAP Stage 1 analysis was corrected for ethnic variation. Thus it is likely that the sample contains novel genetic signals, in addition to those detected by the primary analysis [9,19].
Next, we looked at overrepresentation of significant genes in the Stage 1 data. Table 1 gives the observed and expected numbers of significant genes at significance levels 10 24 , 10 25 , 10 26 when all genes are counted in the analyses and when the known genes (Table S1) and genes within 500kb of them are excluded, the observed numbers of genes are much larger than expected at all significance levels (all p#0.001). Thus there are more loci associated with AD to find.
Furthermore, the number of independent nominally significant loci at Stage 2 (N = 60, (13.5%)) was significantly greater than expected by chance (p = 4.6610 212 ). The percentage of replicated loci increased with the decrease of the gene-wise significance threshold at Stage 1 (see Table 2 for details).
Combining the gene-wide p-values in both stages 1 and 2, using Fisher's method revealed two new gene-based genome-wide significant (p,2.5610 26 ) loci TP53INP1 and IGHV1-67. The TP53INP1 gene is located on chromosome 8:95,938,200-95,961,615 and its combined gene-based p-value = 1.4610 26 (Table 3). Table S3 provides details for each SNP contributing to the gene-based result. Out of 45 SNPs in the gene, three SNPs (rs4735333, rs1713669, rs896855) have p-value#10 24 . Figure 1 shows the LD plot of this gene and suggests that there are at least two partially independent signals in the TP53INP1 gene (r 2 between the pairs of most significant SNPs rs4735333-rs1713669 and rs1713669-rs896855 are 0.65 and 0.6 respectively).
To look at the gene expression patterns in these novel genes, we used the Webster-Myers expression dataset [21], available at http://labs.med.miami.edu/myers/LFuN/data%20ajhg.html. Comparing 137 AD vs 176 controls with temporal or frontal cortex expression values by t-test, t showed significantly higher TP53INP1 expression in cases compared to controls (p = 0.0128). Further examination in the BRAINEAC database [22] (www. braineac.org) from the UK Brain Expression Consortium showed TP53INP1 to have a best cis-eQTL p-value of 6.8610 26 (for rs4582532 SNP, which is about 7.6 kb upstream of the gene). The three SNPs with association p#10 24 mentioned above (rs4735333, rs1713669, rs896855) had significant cis-eQTL pvalues of 8.2610 26 , 7.8610 25 and 1.1610 25 respectively in BRAINEAC brain expression data. The r 2 between the cis-eQTL and the three associated SNPs were 0.80, 0.65, and 0.81, respectively). Further analysis of additional independent brain expression and methylation datasets (see Methods S1) indicated significant cis eQTLs and meQTLs for TP53INP1 (Tables S10 and S11). The probe for the meQTL is in a CpG island region that corresponds well with ENCODE DNAse/ChIP-seq/Histone marks and is located upstream (,1.5 kb) of the TP53INP1 Table 1. Overrepresentation of replication of significant genes/loci available at Stage 2, excluding all loci of 0.5 Mb around genes previously reported [4][5][6][7][8] and Stage 1 IGAP genes [9,19] containing genome-wide significant SNPs. transcription start site. In combination these results suggest a possible epigenetic mechanism whereby the associated variants in the region influence TP53INP1 expression in several brain regions. These expression data provide further evidence supporting the functional relevance of TP53INP1 to AD susceptibility. The IGHV1-67 gene was not found in those databases. In addition we detected two genome-wide significant loci 1) ZNF3 (chr7: 99,661,653-99,679,371; p = 8.6610 27 ) and 2) two closely located genes on chromosome 11 MTCH2 (47,638,858-47,664,206, combined p = 2.5610 26 ) and NDUFS3 (47,600,632-47,606,114, combined p = 4.8610 27 ) ( Table 4). None of these genes harbour genome-wide significant SNPs in the SNP GWAS analysis on its own (see Tables S5-S7). Figures S1-S3 show LD plots of these additional genes.
ZNF3 and NDUFS3, MTCH2 genes on chromosomes 7 and 11, respectively, lie close to rs1476679 (chr7:100,004,446; ZCWPW1) and rs1083872 (chr11:47,557,871; CELF1) SNPs, which are shown to be genome-wide significant in the IGAP study, when combining Stage 1 and Stage 2 data. Figures S1-S3 show LD structure of these genes in relation to the IGAP singe genome-wide significant hits. (Note that the NDUFS3 gene on chromosome 11 was genebased genome-wide significant already at Stage 1.) Although none of these SNPs actually lie within the genes mentioned above, it is possible that they may account for the gene-based signals through linkage disequilibrium. In order to test whether the gene-based signals are independent of these strongly-associated SNPs, we performed single-SNP association for each SNP annotated to these genes by regression, adjusting for the significant SNPs mentioned above, along with the other study covariates. The resulting pvalues were combined into gene-based tests, as described previously. Under this conditional analysis ZNF3 gene does not show significant association, however NDUFS3 still shows a trend towards significance (p = 0.081) (see Table S8 for details).
Furthermore, five genes in chr11:47,593,749-47,615,961 (KBTBD4, NDUFS3, LOC100287127, FAM180B, C1QTNF4) all have p,0.05 with gene-based analysis 610 kb, when conditioning by the genome-wide significant hit rs10838725 in this region. This may partially be explained by the SNP rs10838731 (p = 1.2610 23 after conditioning by rs10838725) which is shared by all latter five genes.
Gene-based analysis with 610 kb around genes did not reveal additional genome-wide significant loci in the Stage 1 data set. Moreover, the significance of the genes identified above did not improve in general, indicating that adding 10 kb flanking regions to genes introduces more noise to the gene-based signal. The combined Stage 1 and Stage 2 gene-based analysis provided further evidence for significant signals in the loci on chr 11 with 8 genes (SPI1, SLC39A13, LOC100287086, PTPMT1, KBTBD4, NDUFS3, LOC100287127, FAM180B) and on chr 7 with 6 genes (LOC100128334, MCM7, PILRB, PILRA, LOC100289298, C7orf51), all reaching genome-wide significance. This is likely to be due to the fact that including genes' flanking regions captures a greater number of the same SNPs or SNPs in high LD showing significant association.
The Manhattan plot of the gene-based p-values ( Figure 3) gives a general overview of the gene-based results and shows the new loci in relation to previously reported genes (see also QQ-plots in Figure S4). The results of gene-wide analysis for the genes, which were previously reported as associated with AD [4][5][6][7][8] and those which are GWAS significant in the Stage 1 analysis are presented in Table S9. Out of 16 reported susceptibility genes, 15 are nominally significant with gene-wide analysis (almost all p-values are smaller than 10 24 ), however not all of them reach the genebased genome-wide significance level (2.5610 26 ) when the number of SNPs per gene and LD structure of the gene is taken into account. We did not observe genome-wide significance for CD33 gene. This gene was genome-wide significant in Stage 1 (p = 1.9610 26 ), but the association was attenuated when combining Stage 1 and Stage 2 data (p = 1.79610 25 ), similar to the single SNP association result in the SNP GWAS study [9,19].

Discussion
In this study we show that there are more signals in the GWAS imputed data at SNP-and gene-based levels than revealed by single SNP analysis. A gene-based analysis is a next logical step after the single SNP analyses in any attempt to combine possible several signals in genes and thus enhance the power of the association analyses.
The first new gene TP53INP1 (chromosome 8) encodes a protein that is involved in mediating autophagy-dependent cell death via apoptosis through altering the phosphorylation state of p53 [23] and in modulating cell-extracellular matrix adhesion and cell migration [24]. TP53INP1 encodes a pro-apoptotic tumor suppressor and its antisense oligonucleotide has been used as potential treatment for castration-resistant prostate cancer [25]. This association is notable, given the potential inverse association between cancer and AD that has previously been reported [26,27].
The second new gene IGHV1-67 (chromosome 14) is a pseudogene in the immunoglobulin (IgG) variable heavy chain region of chromosome 14: its function is unknown but all genes in this region are most likely to be involved in IgG heavy chain VDJ recombinations that lead to the full repertoire of antigen-detecting immune cell clones [28].
The gene-based analysis in this study has shown its utility to enhance the information provided by single SNP analysis (i.e. NDUFS3 gene was genome-wide significant from Stage 1 using gene-based analysis whereas this gene was only genome-wide significant after combining the two stages of single SNP analysis).
ZNF3 is a zinc-finger protein at the same locus on chromosome 7 as ZCWPW1 thus rendering it a candidate as the gene that contains the functional signal in this region. Although we can not identify which gene actually confers the risk to AD, it is interesting that ZNF3 function is unknown though it interacts with BAG3 which is involved in ubiquitin/proteasomal functions in protein degradation [29] and ZNF3 is regulated by upstream binding of BACH1 whose target genes have roles in the oxidative stress response and control of the cell cycle [30].
In the cluster of genes on chromosome 11, MTCH2 encodes one of the large family of inner mitochondrial membrane transporters [31] which is associated with mitochondrially-mediated cell death [32], adipocyte differentiation [33], insulin sensitivity [34] and has a genetic association with increased BMI [35]. NDUFS3 also has functions in the mitochondria as it encodes an iron-sulphur component of complex 1 (mitochondrial NADH:ubiquinone oxidoreductase) of the electron transport chain. A deficiency causes a form of Leigh syndrome [36] an early-onset progressive neurodegenerative disorder with a characteristic neuropathology consisting of focal lesions including areas of demyelination and gliosis [37].
In summary, we report two novel genes TP53INP1 (chr8: 95,938,200-95,961,615; combined p = 1.4610 26 ) and IGHV1-67 (chr14: 107,136,620-107,137,059; combined p = 7.9610 28 ), which were not reported as genome-wide significant before. We also report ZNF3 gene on chromosome 7 and a cluster of genes on chromosome 11 (SPI1-MTCH2), showing gene-based genomewide significant association with Alzheimer's disease. These genes are in proximity with, but not the same as, those detected by genome-wide significant SNPs, demonstrating support for the Table 3. New genome-wide significant genes associated with AD. signals identified by IGAP [9,19]. They have an array of functions previously implicated in AD including aspects of energy metabolism, protein degradation and the immune system and add further weight to these pathways as potential therapeutic targets in AD.

Stage 1 data
The main dataset was reported by the IGAP consortium [9,19] and consists in total of 17,008 cases and 37,154 controls. This sample of AD cases and controls comprises 4 data sets taken from genome-wide association studies performed by GERAD, EADI, CHARGE and ADGC (see primary IGAP manuscript [9,19] for more details). The full details of the samples and methods for conduct of the GWA studies are provided in the respective manuscripts [4][5][6][7][8].
Each of these datasets was imputed with Impute2 [38] or MACH [39] software using the 1000 genomes data (release Dec2010) as a reference panel. In total 11,863,202 SNPs were included in the SNPs allelic association result file. To make our analysis as conservative as possible, we only included autosomal SNPs which passed stringent quality control criteria, i.e. we included only SNPs with minor allele frequencies (MAF) $0.01 and imputation quality score greater than or equal to 0.3 in each individual study, resulting in 7,055,881 SNPs which are present in at least 40% of the AD cases and 40% of the controls in the analysis. The summary statistics across datasets were combined using fixed-effects inverse variance-weighted meta-analysis. We corrected all individual SNPs p-values for genomic control (GC) l = 1.087. These SNPs are well imputed on a large proportion of the sample, which increases confidence in the accuracy of the association analysis upon which gene-wide analysis is based.

Stage 2 data
11,632 SNPs with p-values ,10 23 in the IGAP meta-analysis were successfully genotyped in a Stage 2 sample comprising 8,572 cases and 11,312 controls (see primary IGAP manuscript [9,19] for more details). An additional 771 SNPs were successfully genotyped to test all genes with gene-wide p-values ,10 -4 in the IGAP Stage 1 analysis, excluding genes reported prior to IGAP [4][5][6][7][8], the four loci reaching genome-wide significance in the Stage 1 IGAP meta-analysis [9,19] and the 0.5Mb regions around them (Table S2). These SNPs cover 887 genes and correspond to 444 independent loci where all genes within 0.5 Mb are counted as one locus.

Assignment of SNPs to genes
SNPs were assigned to genes if they were located within the genomic sequence lying between the start of the first and the end of the last exon of any transcript corresponding to that gene. The chromosome and location for all currently known human SNPs were taken from the dbSNP132 database, as was their assignment to genes (using build 37.1). In total, we retained 2,804,431 (39.7% of the total) SNPs which annotated 28,636 unique genes with 1-16,514 SNPs per gene. For the gene-wide analysis we have excluded genes which contain only one SNP in the IGAP Stage 1 analysis, leaving a total of 25,310 genes. If a SNP belongs to more than one gene, it was assigned to each of these genes. In order to account for possible signals which are correlated with those in a gene, gene-wide analysis was also performed using a 10 kb window around genes to assign SNPs to genes. Gene-wide analysis The gene-wide analysis was performed based on the summary p-values while controlling for LD and different number of markers per gene using an approximate statistical approach [40] adopted for set-based analysis of genetic data [41]. This is a method for calculating the significance of a set of SNPs in the absence of individual genotype data based on a theoretical approximation to Fisher's statistic for combining p-values. Fisher's statistic (-gln(p i )) combines probabilities and under the null hypothesis has a chisquare distribution with 2N degrees of freedom, where N is the number of markers, and the summation above is for i = 1,…,N). If Fisher's statistic combines the results of several tests when the tests are independent, the approximate method combines non-independent tests and requires only the list of p-values for each SNP and knowledge of correlations between SNPs. Then the value of Fisher's statistic and the number of degrees of freedom is corrected by the coefficient which depends upon the number of SNPs and correlations (LD) between them. This approximation was applied to the Stage 1 and Stage 2 samples separately, and the resulting gene-wide p-values combined using Fisher's method (since these are independent). LD between markers was computed using 1000 genomes data. The gene-based genome-wide significant level was set to 2.5610 26 to account for the number of tested genes [42].

Test for excess of associated SNPs/loci
The effective number N of independent SNPs in the whole genome (excluding genes with SNPs that are genome-wide significant in the Stage 1 IGAP dataset 6 0.5 Mb was estimated by the method described in [43] taking LD into account, as were the observed number of independent SNPs significant at each pvalue criterion (adjusting individual SNP p-values for genomic control l = 1.087 before hand). LD was computed from the 1000 Genomes database (http://www.1000genomes.org/). In the absence of excess association, the expected number of independent SNPs significant at significance level a is a normally distributed random variable whose mean and standard deviation (SD) can be calculated as aN and !Na(1-a) (mean and SD for a binomial distribution). The number of independent SNPs (and thus statistical tests) in the whole genome were estimated as ,3.7610 6 , ,3.6610 6 and ,3.5610 6 at significance levels below 0.1, between 0.05 and 0.1, and 0.2 and above respectively (see [43] for details on the dependence between the significance levels and the estimated number of independent tests). We then calculated mean of the expected number of significant SNPs in intervals a 1 , p # a 2 , (a 1 , a 2 = 0, 10 26 , 10 25 , …, 0.5) as difference between the expected numbers of independent SNPs at a 2 and a 1 significance levels and SD as the square root of sum of the corresponding variances.
We calculated the significance of the excess number of genes attaining the specified thresholds based upon the assumption that, under the null hypothesis of no association, the number of significant genes at a significance level of a in a scan is distributed as a binomial (N,a), where N is the total number of genes, assuming that genes are independent. Genes within 0.5 Mb of each other are counted as one signal when calculating the observed number of significant genes. This prevents significance being inflated by LD between genes, where a single association signal gives rise to several significantly-associated genes. The total number of genes was not corrected for LD in this way, making the estimate of significance of the excess number of genes conservative. Table 4. New genome-wide significant genes associated with AD in the vicinity of recently reported single SNP genome-wide significant hits [9,19]. Supporting Information Table S1 Overrepresentation of significant SNPs excluding previously reported [4][5][6][7][8] genes ±0.5Mb and the APOE region as above.
(DOCX) Table S2 List of genes that are genome-wide significant in the IGAP stage 1 dataset and the flanking regions which included SNPs either in r 2 $0.3 or association p-value#10 -3 whichever covers the largest region. (DOCX)