Family-based Genome Scan for Age at Onset of Late Onset Alzheimer’s Disease in Whole Exome Sequencing Data

Abstract

Alzheimer’s disease (AD) is a common and complex neurodegenerative disease. Age at onset (AAO) of AD is an important component phenotype with a genetic basis, and identification of genes in which variation affects AAO would contribute to identification of factors that affect timing of onset. Increase in AAO through prevention or therapeutic measures would have enormous benefits by delaying AD and its associated morbidities. In this paper, we performed a family-based genome wide association study for AAO of late-onset AD in whole exome sequence data generated in multigenerational families with multiple AD cases. We conducted single marker and gene-based burden tests for common and rare variants, respectively. We combined association analyses with variance component linkage analysis, and with reference to prior studies, in order to enhance evidence of the identified genes. For variants and genes implicated by the association study, we performed a gene-set enrichment analysis to identify potential novel pathways associated with AAO of AD. We found statistically significant association with AAO for three genes (WRN, NTN4, and LAMC3) with common associated variants, and for four genes (SLC8A3, SLC19A3, MADD, and LRRK2) with multiple rare associated variants that have a plausible biological function related to AD. The genes we have identified are in pathways that are strong candidates for involvement in the development of AD pathology and may lead to a better understanding of AD pathogenesis.

INTRODUCTION

Alzheimer’s disease (AD: MIM104300) is a common and complex neurodegenerative disease. In the US it is reported as the 6th leading cause of death (Murphy et al., 2013), with direct costs in caring for subjects estimated as >200 billion dollars annually (Thies et al., 2013). Unlike a number of other common diseases associated with aging for which death rates have declined, including heart disease, stroke, and some cancers, the death rate attributable to AD has substantially increased since 2000 (Thies et al., 2013). These issues lead to a projection of substantial rising costs in the near future, not only in the US, but in other countries as well (Banerjee, 2014, Thies et al., 2013). Even a modest increase in age-at-onset (AAO) of AD through prevention or therapeutic measures would have enormous benefits as it would delay the disease and associated morbidities. To date, there have been no successful pharmacological or other therapies that achieve this goal (Thies et al., 2013).

AD risk has a genetic basis, with evidence for familial aggregation first noted more than 60 years ago (Sjögren et al., 1952). Evidence for a genetic basis was later strengthened with results from twin (Gatz & Pedersen, 2013, Gatz et al., 1997, Pedersen et al., 2004) and family (Mayeux et al., 1991, Van Duijn et al., 1991) studies. Identification of four genes with variation that contributes to AD risk provided definitive confirmation (Corder et al., 1993, Goate et al., 1991, Levy-Lahad et al., 1995, Sherrington et al., 1995). Rare mutations in APP, PSEN1, and PSEN2 are typically characterized by highly-penetrant early-onset disease (EOAD, age < 65 yrs) (Bertram et al., 2008, Bird et al., 1996), while common variation in APOE is associated with altered AD risk in more typical and common late-onset AD (LOAD, age ≥ 65 yrs). Risk increases as a function of number of APOE ε4 alleles, and decreases as a function of number of APOE ε2 alleles, relative to the baseline ε3 allele (Corder et al., 1994, Corder et al., 1993). A recent study employing targeted high-throughput sequencing also implicates rare variation in APP, PSEN1, and PSEN2 in risk of LOAD (Cruchaga et al., 2012) and rare variants in TREM2 and PLD3 genes have been implicated in LOAD as well (Benitez et al., 2013, Cruchaga et al., 2014, Guerreiro et al., 2013, Jonsson et al., 2013). Finally, large genome wide studies (GWAS) have recently implicated multiple additional risk loci (Harold et al., 2009, Lambert et al., 2009, Lambert et al., 2013, Naj et al., 2011, Seshadri et al., 2010). Currently, efforts are under way to obtain additional evidence for involvement of the genes proposed in GWAS studies (Holton et al., 2013, Lord et al., 2014).

In AD, age is an important factor. While AD is very rare in younger individuals, especially below age 60 yrs, mutations in the known early onset genes are believed to account for ~50% of early-onset cases (Finckh et al., 2005, Ikeuchi et al., 2008, Lleo et al., 2002, Tandon & Fraser, 2002). Incidence of AD increases with age in all populations surveyed (Fratiglioni et al., 2000, Hebert et al., 2003, Lobo et al., 2000, Rocca et al., 1991, Sosa-Ortiz et al., 2012), with annual incidence in the US increasing from ~1% at ages 65–70 yrs, to 6–8% by age 85 yrs and up (Mayeux, 2003). As a result of this high annual incidence rate, the prevalence of AD among individuals of age 85 yrs and above is ~32% (Thies et al., 2013). In the study of AD, age may be used either as a covariate, or as a phenotype of direct interest. As a phenotype that is directly relevant to AD, AAO is correlated among family-members, with wide variability among families (Bird et al., 1996). Transmission models of AAO in family-based samples support a genetic basis (Daw et al., 2000). Also consistent with a genetic basis, AAO differs among APOE genotypes, with genotype-specific risk inversely proportional to genotype-specific AAO. Differences among APOE genotype-specific onset-distributions are consistent across studies, whether measured on a genetic background of an EOAD mutation in one of the presenilin genes (Pastor et al., 2003, Wijsman et al., 2005) or in more typical LOAD (Farrer et al., 1997). These observations all suggest that AAO may be a useful phenotype for study of the genetic basis of AD.

To date AAO as the phenotype of interest has been used in only a few genome scans of LOAD. Family-based linkage-analysis is the primary approach that has been used to identify regions of interest (Choi et al., 2011, Dickson et al., 2008, Holmans et al., 2005, Lee et al., 2008, Zhao et al., 2013b), with more limited recent use of GWAS in samples of unrelated, affected subjects (Kamboh et al., 2012, Naj et al., 2014). Together with regional analyses (Wijsman et al., 2004), these genome scans implicate regions containing AAO loci. In particular, regions on chromosome 6 and 19p replicated across independent sets of pedigrees. No inclusion of sequence data in AAO studies has yet been reported. There are two main issues with use of AAO. First, study subjects typically are identified through retrospective sampling designs. Analysis of AAO as a continuous variable from a GWAS study can give biased or misleading results when analyses are conducted as if a prospective sampling design had been used (Lin & Zeng, 2009). This contrasts with logistic regression typically used for analysis of case-control studies, which provides equivalent estimates of relative risk for both prospective and retrospective studies (Prentice & Pyke, 1979). Second, with AAO it is necessary to address age-censoring. As a result, analysis has either been carried out using only affected subjects (Holmans et al., 2005, Kamboh et al., 2012, Lee et al., 2008, Naj et al., 2014), or with a model that accounts for age-censoring (Choi et al., 2011, Dickson et al., 2008, Zhao et al., 2013b). Restricting analysis of AAO to AD cases leads to different interpretation of results than does analysis that includes unaffected subjects and incorporates age censoring. Analysis of AAO only in AD cases provides information about the genetic basis only of AAO modification, given predisposition to AD. Inclusion of unaffected subjects and incorporation of age-censoring allows a broader interpretation. In this context, reduced age-at-onset in cases compared to controls as inferred from age-censored data may lead to the inference of increased age-specific risk that is a function of genotype in a region of interest.

In our study here, we performed a family-based GWAS for AAO of LOAD in whole exome sequencing (WES) data in families with multiple AD cases. Our WES data consisted of only AD cases, so our analysis is a case-only analysis of AAO. While GWAS SNP chips provide a relatively high genomic coverage of the common (Minor Allele Frequency, MAF>0.05) genetic variation (>80% (Li et al., 2008)), a portion of common variation and all rare genetic variation is poorly covered. The sequence data allow access to both rare and common variants. Therefore, the use of WES family data allows the study of both common and rare variants, which are likely to be enriched in families (Wijsman, 2012). In order to identify new loci (both rare and common variants) and genes that may be a modifying factor of AAO, we conducted a classical single marker association analysis for common variants, and a gene-based burden association analysis for multiple rare variants in genes. We combined association analyses with variance component linkage analysis, in addition to referring to previous analyses of AAO in other family-based samples, in order to enhance evidence of the identified genes. We also performed a gene-set enrichment analysis on the list of identified genes and explored the biological function of these genes to help understanding AAO/AD pathogenesis.

MATERIALS AND METHODS

Subjects and phenotyping

Our data consisted of 77 subjects diagnosed with AD. These subjects were selected from pedigrees with multiple cases of late onset AD (> 60 yrs onset age) in multigenerational families from public repositories: 58 subjects from the NIA-LOAD/NCRAD collection (Wijsman et al., 2011) and 19 subjects from the NIMH (Blacker et al., 1997) collection. Single subjects, but not families, from the NIA-LOAD/NCRAD families have been incorporated into GWAS studies of AAO, and the NIMH families have previously been used for a linkage-analysis genome scan of AAO (Choi et al., 2011). Neither set of families has been used for evaluation of sequence-based variants as contributors to AAO, which was the goal of the current study. Pedigrees ranged in size from 10 to 25 subjects, and consisted of 3–4 generations/pedigree. The assumption was that use of families increases ability to detect effects of (potentially) rare alleles. Selection of specific families additionally required availability of DNA from at least two relatively distantly-related LOAD cases (e.g., avuncular to second cousin relationships), thus minimizing their relatedness relative to subjects from other available pedigrees, and reducing the sizes of regions of interest identified in the families. DNA samples from up to four cases per family were used when they were available, therefore also including additional, closer relationships. For inclusion in the current analysis, subjects also were required to be of European ancestry. Additional Hispanic pedigrees selected at the same time with the same criteria were only used for Principal Component Analysis in order to ensure a homogenous group of subjects. They were not used for association analysis because of analytical complications of joint analysis of an admixed set of pedigrees with European- and Hispanic-descent families. The AD affection status was defined as meeting NINCDS-ADRDA criteria for definite, probable, or possible AD (Mckhann et al., 1984, Wijsman et al., 2011). In both the NIMH and NIA-LOAD/NCRAD samples, AAO for AD cases was defined as the age at which first symptoms of AD were reported (Choi et al., 2011, Wijsman et al., 2011). In the subjects used for our analysis, the mean of AAO and its standard deviation are 70.6 and 9.24, respectively. Our study was approved by the University of Washington institutional review board. All samples used were collected with appropriate consent for this study.

Whole Exome Sequencing data

The gene coding sequences were captured using Nimblegen SeqCap EZ Human Exome Library v2.0 kit (Roche, Basel, Switzerland) following the manufacturer’s recommended instruction. The capture kit targets 28,858 genes with total size of the target regions 36.5 Mbp. The sequencing library clusters were generated on Illumina flowcells using cBlot (Illumina, Inc.) and pair-end 101bp sequencing was performed on the Illumina HIseq2000 sequencing platform at the Department of Genome Sciences, University of Washington. The raw base calling was performed with CASAVA (Illumina, Inc.). Sequenced reads were aligned to NCBI human reference genome GRCh37 (hg19) using the Burrows-Wheeler Aligner (Li & Durbin, 2010). BAM files were generated using SAMtools (Li et al., 2009). PCR duplicates were marked using Picard (http://broadinstitute.github.io/picard/). After base recalibration the sequence reads were realigned around indels and mapped. For single nucleotide variant calling, the Genome Analyzer Toolkit (GATK) was utilized (Mckenna et al., 2010). The average read depth for called positions was 70.89.

In all analyses, we considered autosomal di-allelic polymorphic variants (i.e. 102,603 SNPs). Quality control steps were performed to filter out possible sequencing errors. We used the following exclusion criteria: ABHet (Allele balance for heterozygotes) > 0.75, HRun (Largest Contiguous Homopolymer Run of Variant Allele In Either Direction) > 4.0, QUAL (Phred-scaled quality score) ≤ 50, QD (Variant Confidence/Quality by Depth) < 5, or SB (Strand Bias) ≥ 0.10. We also excluded SNPs with evidence of Hardy-Weinberg (HW) disequilibrium (p-value<10⁻³). HW testing was based on the 32 unrelated individuals in our WES dataset. QC procedures led to the exclusion of 15,937 SNPs (i.e. 14,948 due to sequence call quality filters and 989 due to HW disequilibrium). The number of remaining SNPs was 86,666.

Statistical Analysis

Principal Component Analysis

Families in the NIA-LOAD/NCRAD sample were recruited by multiple sites across the US (Wijsman et al., 2011) and families in the NIMH sample were recruited by three different sites (Blacker et al., 1997). As a result, subjects may be drawn from different genetic backgrounds even within subjects declared as of European ancestry. For the NIA-LOAD/NCRAD sample, several families are known to be Caribbean Hispanic. As described previously, there was strong evidence for population stratification in the NIA-LOAD/NCRAD European-American sample (Wijsman et al., 2011), and an indication of possible stratification in the NIMH sample (Choi et al., 2011). To account and correct for population stratification in the full sample used here, we performed a supervised principal component analysis (PCA) using EIGENSTRAT (Price et al., 2006) using SNPs from the WES data. For the PCA, we added to our AD cases the 1000 Genomes (1KG) project (Abecasis et al., 2010) subjects (release November 2010) of European (EUR), African (AFR), and Asian (ASN) descent. PCA was performed using common variants (Minor allele frequency (MAF) >0.05) with low pairwise linkage disequilibrium (LD), and with available genotypes in both our data and the 1KG data. We used PLINK (Purcell et al., 2007) to select SNPs as follows: in a window of 150 SNPs, we estimated LD for all pairs of SNPs and filtered out one of each pair having an r² > 0.2. We used overlapping sliding windows with a step-size of five SNPs. The procedure led to a set of 8,822 SNPs. Our PCA results identified 16 WES subjects who do not cluster with the remaining subjects of European descent (Fig. S1 in supplementary material). These subjects were all of known Caribbean Hispanic descent and were excluded from subsequent analysis, which led to exclusion of entire pedigrees, with no exclusion of any of the European-descent subjects. Among the remaining subjects, 56 subjects had available AAO information, and out of these 56 subjects, APOE genotypes were available for 47 subjects.

Imputation Analysis for APOE

As recently shown (Radmanesh et al., 2014), missing APOE genotypes can now be accurately imputed using the very dense 1000 Genomes data as a reference. To avoid decrease of the sample size due to missing APOE genotypes for nine subjects, we performed an imputation analysis to infer the missing APOE genotypes using the 1000 Genomes data (August 2010). Briefly, we used SHAPEIT2 (Delaneau et al., 2012) to phase the NIA-LOAD/NCRAD GWAS subjects and minimac (http://genome.sph.umich.edu/wiki/Minimac) to impute the missing genotypes. Then we extracted the allelic dosages for the two APOE SNPs (i.e., rs7412 and rs429358) in the case of missing APOE genotypes and we used them along with the known APOE genotypes in subsequent association analyses. With these procedures, all 56 subjects were used.

Association Analysis

We conducted two family-based genome wide association analyses for the log-transformed AAO. The first analysis used a single marker test where one SNP was tested at a time. The second analysis used a gene-based association burden test (Weighted Sum approach (Madsen & Browning, 2009)). The aim of the first analysis was to find evidence of association between AAO and SNPs with MAF>0.05. The aim of the second analysis was to find evidence of association between AAO and multiple rare variants in a gene (SNPs with MAF≤0.05). All association analyses were performed using the statistical package R (http://www.r-project.org/). To account for family relationship among subjects, we used a linear mixed model (LMM) implemented in the R-package “kinship”. In all models, we used the theoretical kinship matrix obtained from the pedigree structure information.

Single Marker Association Test for Common Variants

For the single marker association test, assuming an additive model, we considered 39,993 SNPs with MAF > 0.05 and with at least 50 non-missing genotypes. Among these SNPs, 30,384, 20,574, and 13,590 have minor allele frequencies greater than 0.1, 0.2, and 0.3, respectively. We used the model: log(AAO) = β × X + δZ + ε, where X is the vector of genotypes coded additively as 0, 1, or 2 copies of the minor allele, Z is the vector of observed covariates (e.g., APOE genotypes when used), β and δ are the marker and covariate fixed effect coefficients, respectively, and $\epsilon ~N(0,{\sigma }_g^2\mathit{\Phi }+{\sigma }_e^{2}I)$ where Φ is a matrix of twice the coefficient of kinship between pairs of subjects, I is an identity matrix, and ${\sigma }_g^2$ and ${\sigma }_e^2$ are the polygenic and residual variances, respectively. To test for association, we used the Wald test (H₁ : β ≠ 0 vs H₀ : β = 0).

Burden Association Test for Multiple Rare Variants

We performed the weighted sum burden association test (Madsen & Browning, 2009) for rare variants (MAF≤0.05). This approach collapses rare variants within genes by giving them weights that are inversely proportional to minor allele frequency. SNPs were annotated using ANNOVAR (Wang et al., 2010). A gene was tested if it had at least two non-synonymous SNPs with MAF < 0.05 and if the sum of the MAFs was greater than 0.05. The association model we used was: $log(\text{AAO})=\beta \times {\mathrm{\sum }}_{i=1}^p{w}_i\times X_i+\delta Z+\epsilon$ where √w_i ~ dBeta(5,25) (Wu et al., 2011), where Z and ε are defined above. Again, we used the Wald test to test for association.

Variance Component Linkage Analysis

In order to determine which of our association signals are also supported by evidence for linkage and to incorporate this evidence into the overall evaluation of the signals, we performed a variance component linkage analysis using SOLAR (Almasy & Blangero, 1998) for all 26 NIA-LOAD pedigrees, which have GWAS SNP data. The small number of NIMH pedigrees had a different, microsatellite marker scan, which could not be combined with the NIA-LOAD SNP markers for joint analysis. These pedigrees were therefore not used for the linkage analysis since the 19 available subjects is too small a sample to support estimation of the multiple parameters needed for a separate VC linkage analysis. For each chromosome, we first selected a set of equally-spaced SNPs (~0.5 centiMorgan) with relatively high MAF (>0.4) and in linkage equilibrium. Then, we estimated the Identity-By-Descent (IBD) distribution at each marker position in a full multipoint computation using a Markov Chain Monte Carlo (MCMC) method for pedigrees that have more than 15 transmitted meioses and exact computation for the remaining pedigrees. This analysis was performed using the program gl_auto in the MORGAN package (Thompson, 2011). IBD estimates at 5 cM intervals were converted to SOLAR input-format and a model with additive variance components only was fitted and compared to a model with polygenic variance, only, in a likelihood ratio test. We ran two versions of this model: the first one adjusted for APOE genotypes (VC wApoe) and the second one did not (VC sansApoe).

RESULTS

Single Marker Test

We first ran association analyses without any covariates. The QQ-plot and genomic control coefficient (Devlin & Roeder, 1999) (λ= 1.022) showed slight inflation in the statistical test distribution (Fig. 1). This excess of significant results may be driven by the effect of APOE on sample ascertainment, and thus on sample structure as suggested in (Wijsman et al., 2011). To reduce the observed inflation, and to also avoid detecting signals driven by APOE (a very well established associated factor with AD (Corder et al., 1994, Farrer et al., 1997)), we ran a second association analysis adjusting for APOE as covariate in the LMM. The statistical distribution obtained by this analysis was better controlled, as both the genomic control coefficient (λ=0.975) and QQ-plot showed (Fig. 1).

Fig. 1.

QQ-plot of the single marker test analysis that adjusts (red) and does not adjust (black) for APOE. The plot is based on all 39,993 tested SNPs.

As expected, we observed significant evidence of association for SNPs in the APOE region on chromosome 19 in the first analysis without adjustment for APOE (from 44 Mbp to 47 Mbp). None of the SNPs tested are the two SNPs that define the three critical APOE alleles, as those SNPs fail QC analysis because of low read depth. At a nominal threshold (α=0.05), 10 SNPs were significant in this region. The SNP rs11879355 provided the strongest evidence for association with a p-value of 6.2×10⁻³ (Table S1 in supplementary material). After adjusting for APOE, all 10 significant SNPs were no longer significant at a nominal threshold of α =0.05 (Table S1 in supplementary material).

Our main results from single SNP analyses are based on the second analysis, which adjusts for APOE, and are shown in the Manhattan plot in Fig. 2. Our strategy was to focus on the most significant non-synonymous SNPs with p-values less than 5×10⁻⁴ (i.e. 15 SNPs). The results of these SNPs are shown in Table 1. All these SNPs, except rs1800378, had negative effect sizes, indicating that they decrease AAO. The MAF of 13 of these SNPs was less than 0.1. The significance of association tests of the 15 SNPs ranged from 4.99×10⁻⁴ to 4.12×10⁻⁷. We identified one SNP on chromosome 19 (rs2291516, MAF=0.08, p-value=4.12×10⁻⁷) with Bonferroni-corrected significance (0.05/39993 = 1.25×10⁻⁶). This SNP is in the gene RGL3. The remaining 14 SNPs are located in 11 different genes on nine different chromosomes.

Fig. 2.

Manhattan plot of the single marker test analysis that adjusts for APOE. The plot is based on all 39,993 tested SNPs. The horizontal blue line is the threshold we used to decide the most significant SNPs. The horizontal red line is the genome-wide significance threshold.

Table 1.

Most significant SNPs obtained by the single marker test

Chr	Position	Gene	rsname	Minor	Major	MAF	F	β	P	βa	Pa
1	201,180,100	IGFN1	-	T	C	0.071	-	−0.147	1.18E-03	−0.144	3.12E-04
1	201,180,340	IGFN1	rs139390045	G	A	0.071	0.084	−0.147	1.18E-03	−0.144	3.12E-04
5	43,613,046	NNT	rs35201656	G	A	0.054	0.042	−0.116	2.59E-02	−0.157	4.08E-05
5	43,653,243	NNT	rs41271083	T	C	0.054	0.042	−0.116	2.59E-02	−0.157	4.08E-05
8	30,921,935	WRN	rs2230009	A	G	0.054	0.056	−0.346	1.07E-10	−0.296	5.88E-05
9	133,963,008	LAMC3	rs4740412	A	G	0.214	0.266	−0.077	1.31E-02	−0.089	4.17E-04
9	139,118,673	QSOX2	rs12380852	C	T	0.107	0.117	−0.132	1.71E-04	−0.113	2.17E-04
10	23,729,362	OTUD1	-	T	A	0.063	-	−0.102	8.68E-02	−0.203	1.63E-04
15	28,230,318	OCA2	rs1800407	T	C	0.054	0.078	−0.142	4.34E-03	−0.157	3.11E-04
16	615,048	C16orf11	rs113068385	A	G	0.055	0.036	−0.053	3.05E-01	−0.151	2.71E-04
17	7,324,788	SPEM1	rs33989543	A	G	0.054	0.070	−0.258	6.57E-06	−0.151	2.71E-04
17	7,735,063	DNAH2	rs57985356	G	T	0.071	0.085	−0.141	3.84E-03	−0.138	3.82E-04
17	7,735,934	DNAH2	rs61745181	A	G	0.071	0.076	−0.141	3.84E-03	−0.138	3.82E-04
19	11,508,177	RGL3	rs2291516	A	G	0.080	0.101	−0.155	3.60E-04	−0.168	4.12E-07
19	17,361,116	USHBP1	rs1043963	A	G	0.063	0.086	−0.133	8.46E-03	−0.157	2.17E-04

MAF= Minor Allele Frequency in the sample; F= Minor allele frequency in the ESP project data of European-American sample; b = Effect size not adjusted for APOE; P= P-value not adjusted for APOE;

^aAnalysis that adjusts for APOE. All analyses were based on N=56 subjects.

Burden Test

The number of tested genes with at least two non-synonymous SNPs with MAF < 0.05 and the sum of the MAFs greater than 0.05 was 1,949. The first quartile, median, mean, and third quartile of the number of rare variants in a gene were 3, 4, 4.96, and 6, respectively. We ran two versions of burden tests. The first one adjusts for APOE (wApoe) and the second one does not (sansApoe). Genomic control coefficients (0.776 for wApoe and 0.996 for sansApoe) and QQ-plots showed better properties of the burden test obtained by sansApoe. This trend was different from what we obtained in the single marker test. This is due to the fact that the weighted sum of variants in a gene is likely to be less correlated with APOE than the alleles of each marker alone. Therefore, our results were based on the sansApoe analysis. Manhattan and QQ-plots can be found in Fig. S2 and Fig. S3 in supplementary material. In the burden test analysis, we focused on the 10 most significant genes (Table 2) from our genome scan, with p-values ranging from 3.11×10⁻⁴ to 5.3×10⁻⁶. Again, to ensure that our most significant results were not driven by the effect of APOE, we explored the wApoe analysis of the 10 selected genes. The significance of all genes decreased slightly in the wApoe analysis.

Table 2.

Burden test results for the ten most significant genes

Chr	Gene	Pos first SNP	Pos last SNP	V0.05	S_MAF	β	P	βa	Pa
2	SLC19A3	228,552,234	228,563,911	3	0.054	−0.266	5.85E-06	−0.166	2.77E-03
11	MADD	47,296,533	47,315,499	2	0.054	−0.254	2.00E-05	−0.169	2.85E-03
11	C11orf82	82,625,814	82,644,904	4	0.080	−0.208	3.08E-05	−0.105	2.71E-02
11	KDELC2	108,352,777	108,357,137	2	0.054	−0.241	5.24E-05	−0.219	5.21E-04
12	LRRK2	40,619,082	40,758,652	4	0.054	−0.223	2.84E-04	−0.133	2.09E-02
14	SLC8A3	70,527,576	70,634,200	3	0.071	−0.196	2.56E-04	−0.049	3.68E-01
14	SLC24A4	92,909,807	92,959,940	5	0.063	−0.177	5.30E-06	−0.288	1.00E-04
15	STARD9	42,930,972	43,011,009	8	0.080	−0.109	3.11E-04	−0.053	3.45E-01
19	GRIN3B	1,000,785	1,009,585	3	0.063	−0.190	9.00E-05	−0.119	3.47E-02
19	LENG8	54,966,557	54,968,038	3	0.063	−0.202	2.35E-04	−0.112	3.52E-02

V0.05 = SNPs with MAF<0.05; S_MAF= Sum of the V0.05 minor allele frequencies; β= Effect size not adjusted for APOE; P= P-value not adjusted for APOE.

^aAnalysis that adjusts for APOE. All analyses were based on N=56 subjects.

Variance Component Linkage Analysis

In the VC sansApoe, we identified five regions with lod-scores ≥1.5 (Chromosome 2, 4, 9, 13, and 19) (Supplementary material). The region on chromosome 19 is the longest one and has the highest maximum lod-score, reaching ~2, out of these five regions (Fig. 3). The relevant gene in this region is most likely APOE since the lod-score maximizes at approximately the location of APOE on chromosome 19. Indeed, the VC wApoe analysis showed a drastic decrease of lod-scores in this region (i.e., lod-scores drop from ~2 for VC sansApoe to <0.2 for VC wApoe at the position of APOE at ~72 cM, Fig. 3). In the same analysis, most, but not all, lod-scores of other regions also decreased. The region on chromosome 9 was the most robust to APOE adjustment with its lod-score decreasing modestly from 1.47 to 1. New regions appeared when adjusting for APOE: 1) one around 210 cM on chromosome 2 (lod-score adjusted for APOE = 1.1) and one around 50 cM on chromosome 4 (lod-score adjusted for APOE = 1.4). Moreover, a modest signal on chromosome 19p around 30 – 40 cM (lod ~0.6) is effectively immune to whether or not there is adjustment for APOE.

Fig. 3.

Lod-score plot for chromosome 19. The red dashed line represents the lod-scores of VC analysis adjusting for APOE. The black line represents the lod-scores of VC analysis without adjustment for APOE.

Analysis of known genes from the literature

We explored our association results for the SNPs and genes reported in a previous AAO GWAS (Naj et al., 2014). The authors considered previously identified genes associated with AD (i.e.; CR1, BIN1, CD2AP, EPHA1, CLU, MS4A4A, PICALM, ABCA7, and CD33). They identified association between AAO and SNPs in CR1 (rs6701713, p-value=7.2×10⁻⁴), BIN1 (rs7561528, p-value=4.8×10⁻⁴), and PICALM (rs561655, p-value=2.2×10⁻³). From the list of SNPs reported in this previous study, only one SNP was found in our WES data (rs3752246, ABCA7). This SNP was not significant in both our and their study (p-value=0.69 and 0.064, respectively). Note that the remaining SNPs were located in introns. Nonetheless, several SNPs with MAF greater than 0.05, in CR1 and ABCA7, were nominally significant in our single marker test analysis of AAO. In CR1, two SNPs (rs2274567 and rs3811381) had p-values of 0.034. In ABCA7, two SNPs (rs3764645 and rs3752234) were nominally significant with p-values equal to 0.018 and 0.023, respectively. These results are shown in Table S2 in supplementary material. Using the burden test, three genes (i.e.; BIN1, EPHA1, and ABCA7) out of the nine considered in (Naj et al., 2014) had at least two SNPs with MAF less than 0.05, which means that they were tested. However, none of them was significant (Table S3 in supplementary material).

Bioinformatic enrichment analysis

We ran a gene-set enrichment analysis (GSEA) using a user-friendly web-based tool, WebGestalt (http://bioinfo.vanderbilt.edu/webgestalt/, (Wang et al., 2013)). We performed GSEA enrichment analysis using both GO and KEGG databases on the list of: 1) the 16 most significant genes obtained by the single marker test, 2) the ten most significant genes obtained by the burden test, and 3) the combined list of 26 genes from both tests. We focused on pathways with p-values less than 0.05 after a Bonferroni correction for multiple testing (Table 3). Only the burden test achieved this threshold, and using the list of “burden test genes”, six GO pathways containing five genes (SLC8A3, SLC24A4, SLC19A3, GRIN3B, and LRRK2) from the burden list showed significant results after Bonferroni correction (adjusted p-value range=[0.0096 – 0.0288]) (Table 3). It is notable that identified genes belong to multiple associated pathways.

Table 3.

Gene-set enrichement analysis using WebGestalt for the most significant genes obtained by the burden test

Pathway ID	Pathway name	Database	C	O	E	R	rawP	adjP
GO:0015291	Secondary active transmembrane transporter activity	GO molecular function	192	3 (SLC24A4, SLC8A3, SLC19A3 )	0.11	26.53	2.0E-04	9.6E-03
GO:0043025	Neuronal cell body	GO cellular component	291	3 (SLC8A3, GRIN3B, LRRK2 )	0.15	19.53	4.0E-04	1.5E-02
GO:0044297	Cell body	GO cellular component	312	3 (SLC8A3, GRIN3B, LRRK2 )	0.16	18.22	5.0E-04	1.9E-02
GO:0015297	Antiporter activity	GO molecular function	60	2 (SLC24A4, SLC8A3 )	0.04	56.61	5.0E-04	2.4E-02
GO:0006816	Calcium ion transport	GO biological process	216	3 (SLC24A4, SLC8A3, GRIN3B )	0.13	22.58	2.0E-04	2.7E-02
GO:0022804	Active transmembrane transporter activity	GO molecular function	309	3 (SLC24A4, SLC8A3, SLC19A3 )	0.18	16.49	6.0E-04	2.9E-02

C= The number of reference genes in the pathway; O= The number of genes in the gene set and also in the pathway; E= The expected number in the pathway; R= Ratio of enrichment; rawP= P-value from hypergeometric test; adjP= Bonferroni-corrected P-value.

DISCUSSION

In this paper, we presented the result of a (WES) family-based association study for AAO of LOAD subjects. Risk of AD and its AAO are related traits, as shown by overlap on both traits of effects of known genetic factors, such as APOE (Corder et al., 1993, Farrer et al., 1997). Nonetheless, the interpretation and implication of results focused on AD risk vs. AAO is different. A focus on AAO among AD cases, as in our study, may bring new insights into factors affecting onset of AD in those that are at risk. This allows a more nuanced and useful measure than simple risk of AD, and is highly pertinent to downstream investigation. Even a modest increase in AAO of AD through prevention or therapeutic measures would have enormous benefits by simply delaying the onset of disease.

In our study design, WES was carried on two to three affected subjects per family that has multiple affected subjects. This family-based design might be more efficient than a population-based design, especially for rare variants that are enriched in pedigrees (Wijsman, 2012). In addition, the rare variants that might be implicated in AD or the AAO of AD are likely to be shared by these affected subjects. Despite the modest size of our WES dataset, our results suggest that a good design that uses a carefully selected set of subjects can provide promising results. This is demonstrated by our replication of several signals from (Naj et al., 2014), a study based on more than 9,000 unrelated subjects, by our identification of several additional candidate genes for AD, and also finding that some of the identified genes are located in regions with evidence of linkage for AAO in this study, as well as in other studies.

There are also several aspects of both the design and the analysis that should contribute to robust results. First, as is common in similar studies, we took a number of steps to make sure that our results are statistically-robust and not explained by artifacts of confounders (e.g.; poor SNP quality and population stratification). Second, by focusing on AAO in affected subjects, only, we avoid the problem posed by the censored age data in unaffected subjects, for which there are not yet analytical methods that give statistically-robust results for variance-components analysis in pedigree samples. Third, our choices of analysis details were chosen to make the results robust to small sample size. By using direct sequence data and methods that allow for the possibility of multiple variants within relevant genes while also capitalizing on the increased information that can be obtained from a continuous trait, we eliminate many of the reasons that very large sample sizes became necessary during the era of GWAS case-control studies.

The use of WES data permits the evaluation of rare genetic variations in the functional parts of the genome (exons), which are not genotyped or well-tagged in classical GWAS SNP chips (Li et al., 2008). In addition, WES may provide a direct observation of common variants that are not well-tagged in GWAS SNP chips. Using a single marker test analysis for common variants and a burden test analysis for multiple rare variants have permitted us to detect novel candidate genes that may play a functional role in modifying AAO of AD. Based on literature review, among genes we have identified, three (WRN, OMIM 604611; NTN4, OMIM 610401; and LAMC3, OMIM 604349) with common associated SNPs and four with multiple rare variants associated with AAO (SLC8A3, OMIM 607991; SLC19A3, OMIM 606152; MADD, OMIM 603584; and LRRK2, OMIM 609007) have strong prior evidence for involvement in AD (http://www.genecards.org/).

Two genes, NTNA4 and LAMC3, belong to a family of proteins related to laminins. The gene Netrin 4 (NTN4) (rs17288108, p-value=3.48×10⁻⁵, MAF=0.125, p.(Y205H)) was originally found to have a role in neuronal axon migration and may play an important role in development (Cirulli & Yebra, 2007). Shen et al (Shen et al., 2012) recently showed that NTN4 expression is up-regulated during β-amyloid induced injury of neurite outgrowth, an effect that is reversed with addition of acetylcholinestarase inhibitors. This is consistent with the possibility that NTN4 might play a role in the development of AD pathology (Shen et al., 2012). The gene LAMC3 (rs4740412, p-value=1.44×10⁻⁴, MAF= 0.214, p.(R1459Q)) is a member of the Laminin family of heterotrimeric molecules that function in stabilization of epithelial structures. LAMC3 is strongly expressed in developing human fetal brain with highest expression in temporo-occipital regions. Recessive mutations in LAMC3 cause a syndrome with cortical malformations and seizures (Barak et al., 2011) further underlining its importance in brain development. Laminin interacts with β-amyloid supporting its role in Alzheimer’s disease (Morgan & Inestrosa, 2001). In addition, LAMC3 is a part of network that includes PICALM, a gene previously reported as associated with AD (Carrasquillo et al., 2010, Harold et al., 2009, Lambert et al., 2013).

Among genes identified by the burden test, there was LRRK2 (leucine-rich repeat serine/threonine-protein kinase 2, p-value=2.84×10⁻⁴), which is the most common cause of dominant inherited Parkinson’s disease (PARK8) (Lesage & Brice, 2009). Common variants in this gene have been also found to increase the susceptibility to Parkinson’s disease (MIM168600) (Gilks et al., 2005, Nalls et al., 2011, Simon-Sanchez et al., 2009). Several LRRK2 protein functions might lead to its effects in AD. LRRK2 regulates autophagy through a calcium-dependent activation of the CaMKK/AMPK signaling pathway (Gomez-Suaga et al., 2012) and mediates the synaptotoxic effects of Amyloid beta oligomers through tau phosphorylation (Mairet-Coello et al., 2013). LRRK2 might also contribute to Lewy Body pathology in Alzheimer’s disease (Linnertz et al., 2014). The remaining four genes that have functions related to AD can be found in the supplementary material.

Finally, the VC linkage analysis we performed gives strength to some genes identified in our single marker and burden association analyses. An interesting gene is RGL3, which both gave the strongest single-variant results in the current analysis, for SNP rs2291516, and is in a region with evidence of linkage on chromosome 19p located upstream of APOE. Even though the evidence for linkage in the current sample is moderate, it is interesting because at the same position, a strong signal was identified when adjusting for APOE in an earlier evaluation of regions containing AAO loci (Wijsman et al., 2004) with later confirmatory evidence provided by two other independent samples (Choi et al., 2011, Zhao et al., 2013a). The pedigrees used in these studies do not overlap with pedigrees used here. The region with evidence for linkage to AAO also includes SNP rs1043963 in USHBP1, which also gave positive results in the single marker association test performed here. Another gene, SLC19A3, which was identified by the burden test, is also located in a region with evidence of linkage on chromosome 2. This gene has a biological function that might be related to AD (supplementary material).

Current advances in sequencing allow for whole exome and genome sequencing in tens of thousands of samples. However, in the context of sequencing data, replication is not simple. Combining or replicating studies that use next-generation sequence data has new challenges. Read depth and analytical procedures for alignment and variant calling affect results, requiring both re-calling of the sequence data, and use of methods that include modeling read depth to avoid spurious results (Derkach et al., 2014). Such procedures do not yet include related subjects. In addition, even such large samples might not have sufficient power to account for multiple testing and large samples introduce additional complications of aggregating large datasets, including variation in phenotypic measurements, such as AAO across studies.

In conclusion, with our statistical approach that uses a family-based association study design, we have identified several candidate genes that have additional functional evidence for association with AD. Our family-based design and focus on coding regions of the genome attenuates the issues of multiple testing that complicate classic case-control designs. The identification of candidate genes is important from the perspective of a follow-up in larger case-control samples or in other family-based samples. Our approach that identifies a focused list of candidate genes, specific types of variants that are responsible for association (single low frequency variant vs. burden of rare variants) and the specific AAO phenotype allows for increase in power and replication study in smaller datasets.

Supplementary Material

Supp TableS1-S3 & FigureS1-S3

Fig. S1 Principal Component Analysis results. Our WES subjects and subjects from the three populations of the 1000 genomes project are plotted on the principal component 1 (PC1) vs. principal component 2 (PC2). Our WES subjects are represented by red crosses, and the EUR, AFR, and ASN subjects are represented by blue, cyan, and pink circles, respectively.

Fig. S2 Manhattan plot of the burden test analysis (Weighted Sum Approach) that does not adjust for APOE. The plot is based on all 1,949 tested genes. The horizontal blue line represents the least significant gene among the 10 most significant genes. The horizontal red line is the genome-wide significance threshold.

Fig. S3 QQ-plot of the burden test analysis (Weighted Sum Approach) that adjusts (red) or does not adjust (black) for APOE. The plot is based on all 1,949 tested genes.