Ordered Subset Analysis of Copy Number Variation Association with Age at Onset of Alzheimer’s Disease

Abstract

Genetic heterogeneity is a common problem for genome-wide association studies of complex human diseases. Ordered-subset analysis (OSA) reduces genetic heterogeneity and optimizes the use of phenotypic information, thus improving power under some disease models. We hypothesized that in a genetically heterogeneous disorder such as Alzheimer’s disease (AD), utilizing OSA by age at onset (AAO) of AD may increase the power to detect relevant loci. Using this approach, 8 loci were detected, including the chr15: 30,44 region harboring CHRFAM7A. The association was replicated in the NIA-LOAD Familial Study dataset. CHRFAM7A is a dominant negative regulator of CHRNA7 function, the receptor that facilitates amyloid-β_1–42 internalization through endocytosis and has been implicated in AD. OSA, using AAO as a quantitative trait, optimized power and detected replicable signals suggesting that AD is genetically heterogeneous between AAO subsets.

INTRODUCTION

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder affecting approximately 5.4 million individuals in the US [1–3]. AD and age at onset (AAO) of AD have high heritability [4, 5]. Rare Mendelian forms confirmed and elucidated pathways involved in amyloid accumulation, but only account for a small percentage of AD [6]. Genome-wide association studies (GWAS) using single-nucleotide polymorphisms (SNPs) elucidated 10 loci contributing to the heritability of AD [7, 8].

Copy number variation (CNV) is a specific form of genetic variation representing extra or missing copies of normal sequence, such as deletions, duplications, and inversions [9]. CNVs have properties that are distinct from SNPs: they are often multiallelic and have a higher de novo mutation rate, and therefore they are often not adequately tagged by SNPs [10]. CNVs influence gene expression and phenotypic variation by altering either gene dosage or genome organization [11, 12]. As the mutational load by variant type is locus specific, CNVs serve as an alternative genetic marker map representing a supplementary approach to SNP association [13].

Genetic heterogeneity is a common challenge for GWAS of complex human diseases. Genetic heterogeneity at the population level may manifest as variable risk allele frequencies in subsets of patients. Ordered subset analysis (OSA) is a statistical method that is able to identify a more homogeneous subset of patients [14]. OSA is designed to test genotypic differences between cases and controls in the case-only quantitative trait situation [15]. If the quantitative trait is important in the genotype-phenotype relationship, then one end of the distribution of trait values will contribute disproportionately to the association signal and offset the penalty for multiple testing. The power of OSA for case-control studies was particularly superior to the Cochran-Armitage trend test (OSA power 83.4% versus trend test power 39.45 at α = 0.3) when the subset of interest was about 15–30% of the available dataset. Furthermore OSA allows the selection of the most informative subset of individuals for deeper genotyping [15].

This genome-wide CNV association study applying ordered subset by AAO analysis was undertaken to identify loci that confer risk of AD in a subset of AD subjects defined by AAO of the disease.

MATERIALS AND METHODS

Discovery set

781 subjects participated in the study. Probable AD was diagnosed based on NINCDS-ADRDA criteria [16]. The methodology of the Texas Alzheimer Research and Care Consortium project has been described [17]. Exclusion criteria included a Hachinski score >4 and clinical or imaging evidence of a stroke.

The control group consisted of non-demented subjects; inclusion criteria were the following: unrelated to cases, age over 55 years, normal performance on activities of daily living, and Clinical Dementia Rating (CDR) global score 0 (by surrogate historian). Control subjects underwent neuropsychological testing after enrollment in the study which included the assessment of global cognitive functioning/status (Mini-Mental State Exam and CDR), attention (Digit Span and Trails A), executive function (Trails B and Clock Drawing; Texas Card Sorting), memory (Wechsler Memory Scale (WMS) Logical Memory I and WMS Logical Memory II), language (Boston Naming and FAS Verbal Fluency), premorbid IQ (American National Adult Reading Test), visuospatial memory (WMS-Visual Reproduction I and II), psychiatric symptoms (Geriatric Depression Scale; Neuropsychiatric Inventory- Questionnaire), and functional assessment (Lawton-Brody Activities of Daily Living: PSMS, IADL). Subjects with impairment (Z-score < −1.5 on any measure) were excluded from the control cohort following consensus review.

IRB at each site approved the study. Informed consent was obtained. Genomic DNA was extracted from whole blood with the Puregene DNA isolation kit (Qiagen).

Replication set

For replication of the results we used the NIA-LOAD Familial Study dataset probands (n = 866) and unrelated controls (n = 906) deposited in dbGAP (http://www.ncbi.nlm.nih.gov/gap). Selection criteria for cases were probands only, Caucasian, non-Hispanic, and AAO data available. For controls, the selection criteria included unrelated to cases, Caucasian, and non-Hispanic.

Gene expression set

We obtained 22 pathologically confirmed AD temporal lobe tissue samples (mean age at death: 80 years, range 61–93) and 15 control temporal lobes (mean age at death: 65 years, range 41–93) from two institutions (Alzheimer’s Disease and Memory Disorders Center (ADMDC) tissue collection, New York Brain Bank). Board-certified neuropathologists assigned the diagnosis based on plaque and tangle assessment and Braak staging. De-identified brain samples were exempt from IRB.

AAO phenotyping

AAO was determined with two standardized methods in the TARC cohort according to previously published data [18, 19]: i) caregiver estimate of onset of symptoms and ii) physician estimate of duration of illness using a structured interview with landmark event to facilitate recall [20]. AAO in the LOAD cohort was determined by a standard question inquiring when the symptoms started.

APOE genotyping

APOE genotyping was performed using real-time PCR according to the manufacturer’s instructions. The amplifications were done on an ABI 7900HT thermal cycler (Applied Biosystems, Inc.). The custom Taq-Man probes (Applied Biosystems, Inc) were unique to SNPs at nucleotide positions 112 (rs7412) and 158 (rs429358) of the APOE gene, respectively. The combination of alleles at the two polymorphisms determined the APOE genotype. APOE status for the LOAD was not available and imputation was not possible due to lack of linkage disequilibrium between SNPs on the array and the APOE polymorphisms.

CNV genotyping by the genome-wide human SNP array 6.0

Array based genotyping was performed on the Genome-Wide Human SNP Array 6.0 (Affymetrix) according to the manufacturer’s instructions. Standard QC measures included contrast QC (>0.4) and Median of the Absolute values of all Pairwise Differences (MAPD) <0.4 and number of calls more than 2 SD from the mean.

CNV genotyping in LOAD

The NIA-LOAD Familial Study genotyping was performed on the Illumina Human610 Quad_array. The logR ratio was calculated in the GenomeStudio (Illumina) software. Candidate locus specific data was extracted and OSA was performed on the logR numeric data.

Detection of copy number variation and test of association

The Affymetrix intensity data was corrected for 16 principal components to avoid inflation of the test statistics (Supplementary Figure 1). The PCA corrected data was segmented by the CNAM algorithm implemented in GoldenHelix (GoldenHelix).

The univariate method is sensitive for larger, rare CNVs by scanning the data subject by subject. Algorithm settings included a moving window of 20000 probes, could combine maximum number of segments and minimum number of markers per segment set at 1, maximum pairwise permuted p-value 0.005, applying 2000 permutations per pair to achieve high sensitivity. The multivariate method is sensitive for common CNVs with superior resolution. Algorithm settings were similar to the univariate method, except we did not apply a moving window. Both algorithms reduce the dataset to regions where CNV events occur, thus limiting the number of tests of association.

For test of association, OSA was performed using the covariates from the univariate and multivariate segmentation algorithms. The subsets were added sequentially by 20 subjects from youngest AAO to oldest. The OSA script was created in R (script available upon request). We used the false discovery rate (FDR) approach considering all comparisons within the univariate or multivariate tests and the criteria for significance was set at p-value of <0.05. As the univariate and multivariate segments have substantial overlap, considering both sets would have been overly conservative. Ultimately, we required locus specific replication in the replication set. To account for the APOE effect on AAO in a way that does not change the fundamental concept of the ordered subset analysis, the following strategy was applied: a derived AAO matrix was generated where the APOE effect was accounted for by adding 5 years for one APOE4 allele, and 10 years for two APOE4 alleles. The statistics are added to Tables 2 and 3. All except one association remained significant.

Table 2.

Ordered subset analysis adding multivariate segmentation data by AAO groups in ascending order with bins of 20

Start probe	Chr	Position	traitval	Pval	lowerCI	upperCI	meanCtrl	meanCase	nCtrl	nCase	qval	APOE pval	Nearest gene
CN_857286	2	4213852	60	0.000281	−0.189	−0.058	−0.030	0.093	190	50	0.021	3.84E-05	None
CN_847338	2	208355166	86	0.000523	0.045	0.160	−0.135	−0.237	190	364	0.032	0.0006221	CREB1
SNP_A-8602968	5	97048466	55	0.000415	−0.072	−0.021	−0.011	0.035	190	24	0.029		None
CN_1196516	7	109437080	56.47	0.000679	−0.180	−0.051	−0.022	0.094	190	29	0.037	0.0152776	EIF3IP1
SNP_A-8363651	13	69248299	56	0.000148	−0.080	−0.026	−0.017	0.036	190	28	0.014		LINC00550
CN_119094	15	30444264	74.77	0.000726	0.016	0.060	−0.014	−0.052	190	226	0.038	0.00085	CHRFAM7A

Chr, chromosome; traitval, age at onset when maximum p value was detected; pval, p value; lowerCI, lower confidence Interval; upperCI, upper confidence interval; meanCtrl, mean logR of controls; meanCase, mean logR of cases; nCtrl, number of controls; nCase, number of cases; qval, False Discovery Rate (FDR) corrected p value; APOE pval, APO-adjusted p value; Affy N, number of probes covering the region on the Genome-Wide Human SNP Array 6.0 array; Illumina N, number of probes covering the region on the Illumina Human610_Quad Array.

Table 3.

Ordered subset analysis adding univariate segmentation data by AAO groups in ascending order with bins of 20

Start probe	Chr	Position	traival	pval	lowerCI	upperCI	meanCtrl	meanCase	nCtrl	nCase	qval	APOE pval	Nearest gene
CN_847338	2	208355166	84	6.55E-06	0.069	0.174	−0.117	−0.238	190	355	0.030	7.42E-06	CREB1
CN_1066295	4	173424935	55	4.33E-08	0.074	0.152	0.105	−0.008	190	24	0.001	0.1047305	GALNTL6
CN_641648	13	69248299	55	1.37E-05	−0.054	−0.021	−0.050	−0.012	190	24	0.042	7.30E-06	LINC00550
CN_690430	14	41607622	55	1.97E-06	−0.136	−0.058	−0.108	−0.011	190	24	0.013	0.0003186	BX248273

Chr, chromosome; traitval, age at onset when maximum p value was detected; pval, p value; lowerCI, lower confidence interval; upperCI, upper confidence interval; meanCtrl, mean logR of controls; meanCase, mean logR of cases; nCtrl, number of controls; nCase, number of cases; qval, False Discovery Rate (FDR) corrected p value; APOE pval, APOE-adjusted p value.

TaqMan assay for the CNV encompassing the CHRFAM7A fusion gene

TaqMan Copy Number Assay was performed to validate the copy number calls of CHRFAM7A. Primers (forward primer: GTAATAGTGTAATACTGTAACTTTAAAATGTGTTACTTGT, reverse primer: AGCCGGGATGGTCTCGAT) and probe (TCCTGACTGTACACATAAAA) were supplied by FAM dye-labeled assay targeted to CHRFAM7A and the VIC dye-labeled RNaseP Applied Biosystems. The duplex real-time PCR assays were performed using a FAM dye-labeled assay targeted to CHRFAM7A and the VIC dye-labeled RNaseP (TaqMan copy number reference assay, part #4403326) as a reference gene. Each sample was assayed in quadruplicate by using 10 ng DNA in each reaction. Real-time PCR was performed using the CFX384 Real-time PCR Detection System (Bio-Rad). Threshold cycle (Ct) values were determined for CHRFAM7A and compared with Ct values for RNase P. Relative quantity was determined by the DD Ct method [21].

Gene expression analysis of CHRFAM7A and CHRNA7 in postmortem human brain tissue

Microarray expression profiling was performed by the Microarray Core Facility of The University of Texas Health Science Center at Houston, Houston, TX. RNA QC included 260/280 nm 1.9–2.0, 260/230 nm >1.5 for RNA by Nanodrop 1000. RNA quality was further assessed by calculating RNA integrity number (RIN) with Agilent 2100 Bioanalyzer (Microarray core facility, Baylor College Medicine). Samples were entered to an expression array experiment if RIN > 4 (range 4.3–6.9) [22]. The RNA was amplified into cRNA and biotinylated by in vitro transcription using the Illumina^® TotalPrep RNA Amplification Kit (Ambion, Applied Biosystems, Foster City, CA) according to the manufacturer’s protocol. Biotinylated cRNAs were purified, fragmented, and subsequently hybridized to an Illumina Human-6 V3 BeadChip (Illumina, San Diego, CA).

RESULTS

756 samples passed the contrast QC and were adequate for copy number analysis. Forty samples failed MAPD cutoff of 0.4 and 143 samples failed due to number of CNV calls more than 2 SD of the mean. 573 samples passed QC, 381 AD subjects and 192 normal controls. Six AD subjects had missing AAO, thus 375 AD subjects entered the OSA AAO analysis. Demographics of the 375 subjects are depicted in Table 1.

Table 1.

Cohort characteristics

	N	% Female	AAO mean	AAO SD	AAO range	Age mean	Age SD	Age range	APOE 4/4	APOE N/4	APOEN/N
AD	375	59	71.4	8.6	45–90	79.3	8.5	56-1-5	52 (13.9%)	180 (48%)	143 (38.1%)
Control	192	74	NA	NA	NA	73.3	8.8	57–96	4 (2.1%)	56 (29.1%)	130 (67.7%)

AAO, age at onset; SD, age at onset standard deviation.

The OSA using the multivariate covariates identified six chromosomal regions (Table 2), while the univariate segmented data resulted in four chromosomal regions (Table 3). Two chromosomal regions (chr2: 208,35 and chr13: 69,24) were detected by both segmentation methods. The association of the chr15: 30,44 chromosomal region was replicated in the NIA-LOAD Familial Study dataset (Bonferroni corrected p-value <0.05, Supplementary Table 1). A subset of samples (n = 499) was validated by TaqMan assay for the chr15: 30,44 region (genotyping accuracy: 93.3%) (Fig. 1). We detected 0.73% homozygous deletion, 18.12% heterozygous deletion, and 79.21% diploid genotypes. In addition, we detected rare three copies of the fusion gene (1.87%) confirmed by Taq-Man assay, consistent with previous studies using locus specific genotyping (in contrast to array based genome wide genotyping) [23]. Expression of CHRFAM7A correlated with CNV state in the postmortem human temporal lobe tissue (Fig. 2).

Fig. 1.

Genotyping, validation, and genomic context of the CHRFAM7A fusion gene. (a) Genome-Wide Human SNP Array 6.0 is presented as a plot of normalized probewise log2ratio data in the four panels for duplication, diploid, heterozygous and homozygous deletion carriers (CN = 3,2,1,0). (b) Validation of the 4 CN states using TaqMan Copy Number Assay. (c) Genomic context of the structure of the CHRFAM7A region. Top: Chromosomal location and orientation of genes based on the human reference genome. The horizontal lines indicate the 9 CNV probes located in the CHRNA7 exons of the fusion gene (grey line: duplicated sequence, *: the site of the TaqMan probe). Bottom: The three structures of the CHRFAM7A allele. The ancestral allele contains both the CHRNA7 and FAM7A but no hybrid gene. Partial duplication and fusion of CHRNA7 and FAM7A generated the hybrid gene, CHRFAM7A (CHRNA7 and CHRFAM7A in direct orientation). Inversion of the fusion gene results in opposite orientation of CHRNA7 and CHRFAM7A.

Fig. 2.

CHRFAM7A and CHRNA7 gene expression. Expression level of CHRFAM7A correlates with the (A) disease state (p = 0.018) and (B) CNV state (p = 10⁻⁵) in human temporal lobe samples. Expression of CHRNA7 is similar in AD versus controls (A) or in deletion carriers and non-carriers (B) consistent with in vitro experiments suggesting that the presence of the fusion gene decreases the number of functional receptors reaching the cell membrane without changes in CHRNA7 expression levels.

DISCUSSION

All six regions associated with AD from the multivariate segmented dataset are known CNV loci and are significant after accounting for the APOE effect on AAO. The chr2: 4,21 and chr5: 97,04 regions do not harbor genes within 500 kb. Two of the CNVs are near noncoding RNA sequences. The chr7: 109,43 CNV is in close proximity of EIF3IP1, a pseudogene encoding Eukaryotic Translation Initiation Factor 3, Subunit I Pseudogene 1. The chr13:69,24 region encompasses Long Intergenic Non-Protein Coding RNA 550 (LINC00550). LINC00550 is of unknown function; however, it is overexpressed in neurofibrillary tangles as compared to neurons in subjects with AD (GEO: GDS2795). The chr2: 208,35 deletion upstream from CREB1 has been reported previously from the same dataset using alternative statistical approach incorporating gene expression as a quantitative trait and CNV into the model [24]. The chr15: 30,44 region was reported to be associated with AD [23]. Inspection of the kernel distribution of the logR data suggested that the previous reports grossly underestimated the deletion frequency on chr15: 30,44 likely due to over-stringent algorithm settings and detected only the homozygous deletion. We confirmed the CNV calls with TaqMan assay and demonstrated that the deletion allele frequency is 18.85%. Locus specific association studies in schizophrenia and bipolar disorder reported similar allele frequencies [25–27].

The four associations detected by the univariate segmented data included the chr2: 208,35 and the chr13:69,24 regions also detected by the multivariate segmentation. The univariate and multivariate segmentation methods are sensitive to larger, rarer and smaller, more common CNV events, respectively, with overlap between the two. These two regions were large enough for the univariate segmentation and frequent enough for the multivariate method, thus detected by both. Due to this overlap, correcting for multiple testing using Bonferroni correction for the two methods would be overly conservative. The chr4: 173,42 region is intragenic in GALNTL6. GALNTL6 has been associated with lipid metabolism [28], body mass index [29], and hypertension, raising the possibility of its association with AD through a vascular mechanism; this association lost its significance after correcting for the APOE effect on AAO. The nearest gene to the chr14: 41,60 region is BX248273 of unknown function.

The association of the chr15: 30,44 chromosomal region was replicated in the NIA-LOAD Familial Study dataset. The lack of replication for the other regions with the NIA-LOAD Familial Study dataset is inconclusive due to limited coverage of these regions on the Illumina Human610 Quad array (Table 2 and 3). Gene expression of CHRFAM7A was lower in deletion carrier status versus non-carriers in postmortem human temporal lobes. CHRFAM7A is a fusion gene derived through a presumptive series of duplication, deletion, and inversion events [25–27]. The ancestral allele does not possess the fusion gene; as the most frequent genotype is one copy of the fusion gene on each chromosome (2 copy of the fusion gene), and the CNV calls are relative to the most frequent variant, thus the ancestral allele is the homozygous deletion (0 copy of the fusion gene) (Fig. 1). CHRFAM7A association has been reported previously; however the segmentation algorithm failed to detect the heterozygous deletion in this study [30]. It has been implicated in schizophrenia [31, 32], bipolar disorder [33], and episodic memory impairment [34]. Large deletions of approximately 2 MB affecting the chromosomal region between two low copy repeats from CHRFAM7A to CHRNA7 are more frequent in idiopathic epilepsies and developmental disorders [31, 32, 35, 36].

The fusion gene is derived from the neuronal acetylcholine receptor subunit alpha-7 (CHRNA7) and family with sequence similarity 7A (FAM7A or ULK4) genes. The presence of the fusion gene reduces the nicotine-elicited alpha 7 current in Xenopus laevis oocytes [37, 38]. The reduction is mainly caused by the decreased number of functional receptors reaching the oocyte membrane [37, 38]. These findings suggest that CHRFAM7A acts as a dominant negative modulator of CHRNA7 function and is critical for receptor regulation in humans.

Amyloid-β (Aβ)₁₋₄₂ and α7 nicotinic acetylcholine receptor are co-expressed in AD brains [39]. Aβ_1–42 binds with high affinity to the α7 nicotinic acetylcholine receptor and the receptor facilitates internalization of Aβ_1–42 through endocytosis in cell culture [39] supporting the observation that neurons expressing the α7 nicotinic acetylcholine receptor are selectively vulnerable in AD. α7 nicotinic acetylcholine receptor expressed in vascular smooth muscle cells is implicated in cerebral amyloid angiopathy, one of the pathological hallmarks of AD [40].

Memantine, one of the drugs used in the treatment of AD, is a low molecular weight antagonist of α7 nicotinic acetylcholine receptor. As lower copy number and lower expression levels of the fusion gene (the negative regulator of α7 nicotinic acetylcholine receptor) are associated with AD, memantine could be particularly effective in this subgroup of AD, allowing molecular based therapy.

This genome-wide CNV association study applying OSA by AAO analysis was undertaken to identify loci that confer risk of AD in a subset of AD subjects defined by AAO of the disease. AAO was determined with two methods with high correlation as reported previously [18, 19]. The power of OSA for case-control studies is superior to trend tests when the subset of interest is about 15–30% of the available dataset, corresponding to the allele frequencies we detected for the replicated locus. Additional advantages include the method’s nonparametric nature: through the ranking procedure it is robust to outliers; and the identification of a case subset that is enriched for a particular susceptibility allele. The major limitation of OSA is the evaluation of a single covariate in the model.

CNV GWAS studies face multiple challenges that impact the power of traditional case control design: i) CNVs are often multiallelic in contrast with the biallelic SNPs, ii) CNVs cannot be computed, iii) compiling datasets is inefficient due to batch effects and variable coverage, thus limiting sample size and power, and iv) CNV calls cannot be optimized at the whole genome level as the frequency and contribution of deletions and duplications at each locus is different resulting in variable references and dynamic ranges which are locus specific. OSA as an alternative statistical approach with increased power is useful as it can detect a signal in a smaller, homogeneous dataset and subsequently these signals can be followed up from other datasets in a locus specific manner allowing optimization of CNV calls for that specific locus.

The contribution of CNVs to the heritability of AD has not been fully explored due to these limitations and larger scale studies with high resolution and high dynamic range assays are needed.