Hippocampal Atrophy as a Quantitative Trait in a Genome-Wide Association Study Identifying Novel Susceptibility Genes for Alzheimer's Disease Authors: Steven G. Potkin, Guia Guffanti, Anita Lakatos, Jessica A. Turner, Frithjof Kruggel, James H. Fallon, Andrew J. Saykin, Alessandro Orro, Sara Lupoli, Erika Salvi, Michael Weiner, Fabio Macciardi Published: 2009, PLOS ONE DOI: 10.1371/journal.pone.0006501 ABSTRACT With the exception of APOE ε4 allele, the common genetic risk factors for sporadic Alzheimer's Disease (AD) are unknown. We completed a genome-wide association study on 381 participants in the ADNI (Alzheimer's Disease Neuroimaging Initiative) study. Samples were genotyped using the Illumina Human610-Quad BeadChip. 516,645 unique Single Nucleotide Polymorphisms (SNPs) were included in the analysis following quality control measures. The genotype data and raw genetic data are freely available for download (LONI, http://www.loni.ucla.edu/ADNI/Data/ ). Two analyses were completed: a standard case-control analysis, and a novel approach using hippocampal atrophy measured on MRI as an objectively defined, quantitative phenotype. A General Linear Model was applied to identify SNPs for which there was an interaction between the genotype and diagnosis on the quantitative trait. The case-control analysis identified APOE and a new risk gene, TOMM40 (translocase of outer mitochondrial membrane 40), at a genome-wide significance level of≤10 −6 (10 −11 for a haplotype). TOMM40 risk alleles were approximately twice as frequent in AD subjects as controls. The quantitative trait analysis identified 21 genes or chromosomal areas with at least one SNP with a p-value≤10 −6 , which can be considered potential “new” candidate loci to explore in the etiology of sporadic AD. These candidates included EFNA5, CAND1, MAGI2, ARSB, and PRUNE2, genes involved in the regulation of protein degradation, apoptosis, neuronal loss and neurodevelopment. Thus, we identified common genetic variants associated with the increased risk of developing AD in the ADNI cohort, and present publicly available genome-wide data. Supportive evidence based on case-control studies and biological plausibility by gene annotation is provided. Currently no available sample with both imaging and genetic data is available for replication. Using hippocampal atrophy as a quantitative phenotype in a genome-wide scan, we have identified candidate risk genes for sporadic Alzheimer's disease that merit further investigation. INTRODUCTION Age is a major risk factor for Alzheimer's disease (AD), and with both life expectancy and population increases, its incidence is expected to double in the next two decades [1] . Genetic investigations have discovered some putative pathophysiological mechanisms related to the accumulation of senile plaques and neurofibrillary tangles, the histopathological hallmarks of AD. The high incidence of dementia in Down's syndrome and analysis of the amino acid sequence of beta-amyloid has led to identification of a series of mutations associated with increased amyloid production. However, such mutations only account for a very small percentage of Alzheimer cases. Risk factors have been identified for the sporadic cases, which are the majority of AD patients, although these factors collectively explain a relatively modest amount of the genetic risk for AD [2] . The most significant of these is the presence of apolipoprotein E4 (APOE ε4). The most commonly used strategies to find disease-related genes such as APOE have been linkage studies and candidate gene association studies. The late onset of Alzheimer's disease and the sporadic nature of most cases complicate family studies, although more than a dozen have been performed [2] . These family studies and the case-control studies strongly implicate multiple genetic loci in susceptibility to AD [2] . Candidate gene approaches targeting amyloid, tau, and inflammatory cascades have proved useful, yet the identification of other major risk factors have proved elusive, with few associations being replicated. The recent availability of high-throughput genotyping has allowed the simultaneous measurement of hundreds of thousands to more than one million SNPs (Single Nucleotide Polymorphisms), which has made genome-wide association studies (GWAS) feasible. A major challenge in case-control GWAS designs in which allele and genotype frequencies are compared between AD and control patients is achieving sufficient statistical power. Such categorical approaches require approximately 6,000 cases and controls to obtain 85% power to detect a 30% difference (odds-ratio of 1.3) with a minor allele frequency of 0.15 [3] . Quantitative trait association studies offer several advantages over case-control studies, since the dependent measures are quantitative and more objective than diagnostic categorization, and can increase the statistical power four to eight fold, thus greatly decreasing the required sample size to achieve sufficient statistical power [4] . We present the results of a GWAS in the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort, a longitudinal multi-site observational study including AD, mild cognitive impairment (MCI), and elderly individuals with normal cognition [5] assessing clinical and cognitive measures, MRI and PET scans (FDG and 11 C PIB) and blood and CNS biomarkers. Major goals of ADNI are to make this wealth of data available to the general scientific community and to improve methods for clinical trials by optimizing methods for imaging that are associated with clinical symptoms and the conversion of MCI to AD. The clinical and imaging measures from this longitudinal study have previously been made publicly available; the genome-wide genotyping results for 793 subjects from the Illumina Human610-Quad BeadChip (Illumina Corporation, San Diego, CA) as performed by The Translational Genomics Research Institute (TGen, Phoenix, AZ), are now also publicly available. In this initial analysis of the AD and the healthy control subsets of the data, we first performed a case control analysis and found APOE as well as an associated gene TOMM40, that provides additional risk for developing AD. To complement the case-control/candidate gene approach to discover new risk genes for AD we performed a QT (quantitative trait) analysis. Our method uses the hippocampal grey matter density from neuroimaging as the quantitative trait, and examines which SNPs (as proxies for genes) influence the quantitative trait differently for AD and healthy controls. The concentration of grey matter in the hippocampus [6] was chosen as the QT phenotype as it is affected early in AD, is implicated in the conversion to and progress of AD [7] , [8] and is associated with many of the cardinal symptoms, and it can be reliably measured in vivo . We used differences in the total grey matter concentration of the hippocampal area, (the four CA fields of the hippocampus proper, plus dentate gyrus, subiculum, entorhinal cortex, and parahippocampal gyrus) in AD subjects vs normal aged-matched controls in our model, since MRI studies have suggested that reductions in hippocampus over time can be particularly useful in predicting AD before the onset of clinical symptoms, and in assessing the efficacy of pharmacological treatment in clinical trials [9] – [12] . Therefore, in our GWA study we used hippocampal concentration as an imaging phenotype to reveal genes that potentially influence hippocampal atrophy and dementia in the context of AD. The genes which influence hippocampal grey matter concentration differentially in AD and healthy subjects may provide important information regarding the mechanisms of disease-related atrophy. The power of this approach is demonstrated in the results implicating novel genes which have been implicated in candidate neurodegenerative mechanisms. METHODS For the purpose of this analysis we used ADNI subject data that was previously collected across 50 sites. Study subjects gave written informed consent at the time of enrollment for imaging and genetic sample collection and completed questionnaires approved by each participating sites' Intitutional Review Board (IRB). ADNI Data used in the preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database ( www.loni.ucla.edu/ADNI ). The ADNI was launched in 2003 by the National Institute on Aging (NIA), the National Institute of Biomedical Imaging and Bioengineering (NIBIB), the Food and Drug Administration (FDA), private pharmaceutical companies and non-profit organizations, as a $60 million, 5-year public-private partnership. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer's disease (AD). Determination of sensitive and specific markers of very early AD progression is intended to aid researchers and clinicians to develop new treatments and monitor their effectiveness, as well as lessen the time and cost of clinical trials. The Principle Investigator of this initiative is Michael W. Weiner, M.D., VA Medical Center and University of California – San Francisco. ADNI is the result of efforts of many co-investigators from a broad range of academic institutions and private corporations, and subjects have been recruited from over 50 sites across the U.S. and Canada. The initial goal of ADNI was to recruit 800 adults, ages 55 to 90, to participate in the research—approximately 200 cognitively normal older individuals to be followed for 3 years, 400 people with MCI to be followed for 3 years, and 200 people with early AD to be followed for 2 years.” For up-to-date information see www.adni-info.org . PARTICIPANTS All subjects were part of the Alzheimer's disease Neuroimaging Initiative (ADNI), a longitudinal multi-site observational study including AD, mild cognitive impairment (MCI), and elderly individuals with normal cognition assessing clinical and cognitive measures, MRI and PET scans (FDG and 11 C PIB) and blood and CNS biomarkers. Brain imaging, biological samples, and clinical assessments are longitudinally collected for a target of 200 healthy controls, 400 MCI, and 200 AD subjects. For the AD sample, the study focuses on identification of sporadic cases of mild AD, subjects between the ages of 55–90, with an MMSE score of 20–26 inclusive and meeting NINCDS/ADRDA criteria for probable AD [13] – [15] , and having an MRI consistent with the diagnosis of AD. Major goals of ADNI are to make this wealth of data available to the general scientific community and to improve methods for clinical trials by optimizing methods for imaging that are associated with diagnosis and clinical symptoms. In this analysis we contrast the AD subjects with healthy controls. This analysis is confined to baseline diagnosis and MRI scans for each subject. Following volumetric analysis and genotyping quality control measures (described below), data from a total of 172 AD subjects and 209 healthy controls were included in this analysis. PROCEDURES The data collection methods have been described in more detail in Jack et al., 2008 [16] , and Mueller et al. 2005 [5] . The images analyzed here were the baseline or screening scans at the entry to the study; they were collected using the standard ADNI protocol on 1.5T GE, Siemens, or Philips MR scanners [16] . The analyzed datasets included the T1-weighted structural MRI scans from all available subjects. Following the segmentation and quality assurance defined below, a final dataset of 229 healthy controls and 194 AD subjects provided grey matter concentration estimates of the hippocampus, using a voxel-based morphometry approach. The MRI images were converted from DICOM format into NIFTI format and re-oriented into MNI coordinate conventions using FSL tools [17] . They were then registered with the MNI brain atlas, corrected for intensity inhomogeneities and segmented into 3 classes (grey matter (GM), white matter (WM), and CSF) using the VBM toolbox integrated with the SPM5 package [6] , [18] . Datasets with an extreme value of computed total brain density or the ratio of grey/white matter density were inspected, leading to the exclusion of several datasets in which segmentation was unsuccessful. Often, these were cases with large ventricles and/or diffuse white matter hypointensities. When there were multiple datasets per subject, the best segmentation for each subject was selected, blind to diagnosis or gender. The datasets of all subjects whose segmentation was successful were used to construct custom templates for the class prior probabilities. These custom templates were used in a second run to segment all datasets. To allow a comparison with the previous run, the same datasets were selected for further evaluation. Again, a quality check was performed using the total brain density and the ratio of grey/white matter density. The total grey matter was generally higher in this segmentation, which is very much expected with a better-matched template. The right and left hippocampal regions were defined using the AAL brain atlas [19] , and the mean of the grey-matter fraction over voxels within those regions was computed for the final imaging phenotypes on each subject. GENOTYPING Blood samples were obtained from each participant and sent to Pfizer for DNA extraction. All DNA extraction and genotyping was done blinded to group assignment. Genotyping on the blinded DNA was performed by TGen using the Illumina Human610-Quad BeadChip. Approximately 200 ng of DNA was used to genotype each subject sample according to the manufacturer's protocol (Illumina, San Diego, CA). After amplification, fragmentation and hybridization, the specifically hybridized DNA was fluorescently labeled by a single base extension reaction and detected using a BeadArray scanner. Non-specifically hybridized fragments were removed by washing while remaining specifically hybridized DNA were processed for the single base extension reaction, stained and imaged on an Illumina Bead Array Reader. Normalized bead intensity data obtained for each sample were loaded into the Illumina BeadStudio 3.2 software which generated SNP genotypes from fluorescent intensities using the manufacturer's default cluster settings. The data analyzed with BeadStudio are publicly available on the LONI website ( www.loni.ucla.edu/ADNI/Data ). The raw genotypic data were imported in a genome-wide data management system, i.e. SNPLims [20] , to allow the export of user-defined formats compatible with the genetic programs used for the statistical analysis. The following Quality Control (QC) procedures were implemented on the genome-wide data. We performed all data analysis and QC using the PLINK software package ( http://pngu.mgh.harvard.edu/purcell/plink/ ), release v1.05. DNA samples with a call rate<95% were excluded. The average call rate in the remaining samples was 99.4%. From a total of 600,470 SNPs in autosomal chromosomes (sex chromosomes, mitochondrial, and pseudo-autosomal SNPs were excluded), the intensity-only probes from chromosomes 1–22 were excluded (17,879), and 5,592 SNPs were excluded by the PLINK algorithm. SNPs were included in the imaging genetics analysis if their minor allele frequency (MAF) was>.05, were present for>90% of the subjects, and did not depart from Hardy-Weinberg equilibrium for the controls using a threshold of 10 −3 . (Some SNPs were excluded for more than one category). The final number of SNPs included in the analyses is 516,645. Since the Illumina chip does not include the SNPs associated with APOE alleles, a separate analysis was performed in which DNA was extracted from 3 ml blood samples. APOE genotyping was carried out on these samples by PCR amplification and HhaI restriction enzyme digestion. The genotype was resolved on 4% Metaphor Gel and visualized by ethidium bromide staining [21] . STATISTICAL ANALYSIS Differences in clinical and demographic values between these groups were analyzed by t test, chi-square or Fisher exact test using STATA10 [22] . Population stratification was examined with PLINK v1.5 and EIGENSTRAT [23] , [24] and a lambda of 1.06 was observed, indicating no stratification. CASE-CONTROL ANALYSIS Because APOE ε4 is a well documented risk factor for late onset AD, the Cochran-Armitage trend test was used to analyze the APOE ε4 allele (which are results of “diplotypes” constructed on SNPs rs429358 and rs7412) distribution between healthy controls and AD patients. Based on the significance of our finding regarding APOE ε4 in our sample and recent interest in this region we expanded our genotyping in the APOE region (∼300-Kb on chromosome 19). To increase the map density in the APOE region, we 1) merged the results of the genotyping (performed with the APOE-specific genotyping) of SNPs rs429358 and rs7412 and 2) performed imputation. To impute the region, the set “hapmap_r23” was downloaded from the HapMap site. This dataset includes 270 subjects from each one of the 4 populations studied by HapMap, i.e. CEU, YRB and CHB+JPT for 267 SNPs [25] . The imputation was considered successful for 159 SNPs with an in-silico call rate>90% in this region. After imputation, the new dataset was used to calculate case-control association tests: 1) single marker association and 2) haplotypes proxy association. The proxy association test was calculated using the flanking markers in strong LD with the “reference” SNP (i.e. proxies). QUANTITATIVE TRAIT ANALYSIS All autosomal SNPs that passed quality control checks were tested for quantitative trait association interaction using the general linear model as implemented in PLINK v1.5. The statistical model is based on comparing the differential effects of SNP association by diagnosis on the brain imaging quantitative trait, i.e. the interaction term. Based on the demographic analysis of our sample and the case-control allelic association test for APOE ε4, we covaried for age and gender and for the number of APOE ε4 alleles in the model. Out of the possible 4 models to choose from (i.e. additive, codominant, dominant and recessive) we a priori chose the additive model as it is the most general, both reflecting the additive contribution to risks for complex diseases [26] , and detecting strong non-additive effects. The general linear model used to determine the interaction term was: in which b i represents the coefficients, SNP represents the genotype of each marker tested, APOE ε4 represents the allele dose of ε4, gender and diagnosis were coded as dichotomous variables, and age is the age of the subject in years. Our focus is on the SNP×diagnosis interaction. There is no accepted method for determining a statistical threshold for a quantitative trait interaction in a context of a genome-wide association study (GWAS). Given 516,645 SNPs and the sample size of 381 subjects, any result at 10 −6 or smaller provides statistical evidence for a main effect associating a given SNP with a quantitative trait. There is a lack of consensus on how to identify a genome-wide significance threshold. Reaching a consensus definition for an appropriate statistical threshold for the interaction term in a GWAS is even more complex. For the purposes of presenting the initial analyses of our data and considering the small sample size, the number of variables analyzed and the complexity of the model, we chose a conservative hypothesis for a threshold of 10 −6 for our interaction term. In keeping with recommendations for presenting a GWAS analysis [27] the results of all analyses are provided in the supplemental material. The Linkage Disequilibrium (LD) maps of each chromosomal region harboring the most significant SNPs were created using a window size of±200 Kb. D' values were calculated using Haploview 4.1 [28] and visualized with WGAViewer software, Version 1.25W; 2009 ( http://www.genome.duke.edu/centers/pg2/downloads/wgaviewer.php ) [29] . The genetic annotation was performed using WGAViewer: SNPs location and gene names were derived from Ensembl release 52 - Dec 2008 ( http://www.ensembl.org/ ) [30] . RESULTS Table 1 represents the demographic characteristics of the sample analyzed. In this population the healthy control sample had significantly more education than the AD sample, as well as the expected differences in measures of cognitive performance. The APOE ε4 alleles were overrepresented in the AD cases as expected (Cochran-Armitage test χ 2 (1 d.f.) = 62.3, p-value = 2.98×10 −15 ). Of the two SNPs comprising the APOE genotype, rs429358 SNP was significant at p-value = 2.30×10 −16 and rs7412 at p-value = 5.00×10 −4 (see Table 2 ). The second most significant finding in the APOE region is the SNP rs2075650 (p-value = 7.47×10 −7 ) located in intron 2 of TOMM40 gene. The haplotype association via in silico reconstruction of proximal SNPs increases the importance of the findings at this locus. One of the haplotypes constructed with rs2075650 and the two flanking SNPs rs11556505 (TOMM40) and rs429358 (APOE) has greater significance (p-value = 1.50×10 −11 ) than the univariate p-value for the TOMM40 SNP. Case control association analysis in the APOE region with special focus on TOMM40, APOE and APOC1 genes. Gene name, SNP name (rs number), SNP base-pair position (build 36.3), genotype type (either obtained through actual genotyping or through in silico imputation), physical location or SNP function (consequence type), case-control allele frequencies and p-values ordered by base-pair position. TABLE: Baseline characteristic of mild AD subjects and healthy controls at baseline assessment. Category Control AD p -Value Number of Subjects 209 172 Gender (Male/Female) 108/101 91/81 ≥0.81 Chi Square Age 75.87±5.10 75.31±7.32 ≥0.39 MMSE 29.09±1.00 23.33±2.06 <0.0001 ADASCog 9.33±4.14 27.63±9.58 <0.0001 Years of Education 15.99±2.90 14.74±3.15 <0.0001 Handedness (right/left) 191/18 162/10 ≥0.29 Chi Square Ethnicity (Hispanic/non-Hispanic/unknown) 1/205/3 4/166/2 ≥0.327 Fisher's exact Race (African American/Asian/More than one race/White) 14/3/0/192 7/1/2/162 ≥0.256 Fisher's exact APOE (ε2/ε3/ε4) 34/322/62 8/194/142 <0.0001 Chi Square TABLE: Case – control association results in the region of interest harboring APOE gene. Gene SNP Coordinate (Build 36.3) Type Consequence Type Allele frequency [%] cases/controls p -Value TOMM40 rs8102977 50085313 imputed 5′UTR 2.3/1.3 0.3 TOMM40 rs157580 50087106 genotyped intronic 73.8/62.7 1.10×10 −3 TOMM40 rs2075650 50087459 genotyped intronic 29.9/15.1 7.48×10 −7 TOMM40 rs11556505 50087984 imputed synonymous-coding 26.5/14.3 2.96×10 −5 TOMM40 rs8106922 50093506 genotyped intronic 70.6/64.8 0.08 TOMM40 rs1160985 50095252 imputed intronic 58.7/52.0 0.07 APOE rs405509 50100676 genotyped 5′UTR 57.6/49.3 0.02 APOE rs769451 50102751 genotyped intronic 98.8/97.6 0.21 APOE rs429358 50103781 genotyped non-synonymous-coding 41.3/14.9 2.30×10 −16 APOE rs7412 50103919 genotyped non-synonymous-coding 97.7/91.9 5.00×10 −4 LOC100129500 rs439401 50106291 genotyped intronic 75.0/62.9 4.00×10 −4 LOC100129500-APOC1 rs5114 50110286 imputed intronic 1.00/99.5 0.19 LOC100129500-APOC1 rs389261 50112183 imputed intronic 98.3/96.4 0.15 LOC100129500-APOC1 rs10424339 50112333 imputed intronic 99.0/98.9 0.92 LOC100129500-APOC1 rs12721054 50114427 imputed intronic 1.00/99.5 0.2 QT ANALYSIS The QT analysis identified 25 SNPs significant at the level of≤10 −6 associated with either the left or right hippocampal grey matter volume (see Figure 1 ). A Q-Q plot of the association results is presented in Figure 2 . P-values from the observed distribution of the test of interest (i.e., the SNPxDiagnosis interaction) are plotted for all SNPs against the null distribution: p-values show a deviation in excess in the upper tail distribution, suggesting the presence of significant genetic association. Such deviation becomes evident at a p<1×10 −6 supporting the hypothesis of a threshold for a claim of genetic association at this significance value. The significant SNPs (see Table 3 ) represent 15 loci and 6 desert regions. The QT analysis for all SNPs is provided in the supplementary materials (see Supplementary files). Of the 25 significant SNPs, thirteen map directly in genes and pseudogenes or are linked based on the LD structure of the region. Four SNPs map in regions characterized by a high level of genomic complexity given the overlap of several different genes and isoforms or pseudogenes in these areas. Eight SNPs map to 6 different “desert” regions where no specific gene annotation could be found including the LD block structure of the region. Details of these 25 significant SNPs follow. Supporting figures ( Figures S1 – Figure S20 ) and tables ( Table S1 – Table S20 ) are available for each of the 25 significant SNPs. Manhattan plot of the p-values (-Log 10 (Observed p-value)) from the QT (right hippocampus) genome-wide association analysis. The spacing between the SNPs does not reflect the actual distances between SNPs in the genome. Each color identifies an autosomal chromosome (from chromosome 1 to chromosome 22). The horizontal lines display the cutoffs for 3 significance levels: green line for mild significance (10 −3