Genome-wide association study for variants that modulate relationships between cerebrospinal fluid amyloid-beta 42, tau, and p-tau levels

For the rQTL and G×G analyses, the dataset for CSF biomarker analysis consisted of 3146 participants from nine different studies and is described in detail in our recent publication [9]. Included were 805 individuals (29.34% cases) enrolled in studies at the Knight ADRC, 787 individuals (more than 71% cases) from the ADNI (390 from ADNI1 and 397 from ADNI2), 184 individuals (5.43% cases) from Predictors of Cognitive Decline Among Normal Individuals (BIOCARD), 105 individuals (no AD status) from Saarland University in Homburg/Saar, Germany, 433 individuals (22.17% cases) from the Mayo Clinic, 293 individuals (all cases) from Skåne University Hospital, Sweden, 164 (62.8% cases) from studies at Perelman School of Medicine at the University of Pennsylvania, and 375 (33.33% cases) from studies at the University of Washington. Table 1 shows the demographic data for each study. Clinical assessments, CSF collection, and proteins were measured by each site. Prior to combining data for analyses, CSF levels of tau, p-tau, and Aβ42 were log₁₀-transformed to approximate a normal distribution and the mean from each dataset was standardized to zero to account for the different platforms used by the different studies to measure protein levels. There were no significant differences in the transformed and standardized values for the different studies. Study, age, sex, and the first two principal components were identified as confounding factors by stepwise regression analyses for each protein and corrected for in applicable analyses [9]; other papers have been published that have combined data in similar ways [10, 11].

Five million single nucleotide polymorphisms (SNPs) (5,088,365) were included for analyses from the 12 million available imputed SNPs after excluding those with a minor allele frequency (MAF) less than 0.01. Data was imputed using the 1000 Genomes Phase 3 data (October 2014) as the reference panel [9]. For the rQTL screens we used a standard 5 × 10⁻⁸ genome-wide significance threshold and for the G×G screens we corrected for 5 × 10⁻⁸ divided by the number of independent and significant rQTL from each screen (tau/Aβ42 and ptau/Aβ42; see Table 3). Most analyses were performed using the Julia programming language (v0.3.12) [12, 13] and some plots were made in R [14].

As a strategy to detect G×G interactions, we performed genome-wide single-locus screens for rQTL [4] with the CSF biomarker data. All models used the same covariates: age, sex, series, APOε2, APOε4, and the first two principle components for ancestry. Significant rQTL were used as independent a-priori hypotheses to identify G×G interactions. Two rQTL models were screened:

where the SNP genotypes are treated as factors and only genotypes with a count of 20 or more were included, leading to models with sufficient genotype counts for tests; either a three-genotype model with 2 degrees of freedom (DF) or a two-genotype model with 1 DF. A putative rQTL exists when the Aβ42*SNP interaction term is significant when compared with a null model excluding that term. Tests of interaction terms sometimes exhibit type I error inflation due to underestimates of the covariance matrix [15]. Work performed by Bůžková et al. [16] and Voorman et al. [14] showed that sandwich estimators and the parametric bootstrap can be used to obtain valid p values, with the parametric bootstrap as the gold standard. For tests that initially reached the significance threshold, empirical p values were obtained via 200 million parametric bootstraps [5]. A standard 5 × 10⁻⁸ genome-wide significance threshold was used for each rQTL screen.

Context-dependent interactions such as G×E and G×G create single locus patterns such as rQTL. Therefore, each significant rQTL create independent a-priori hypotheses for G×E or G×G interactions. Previous papers on two-stage designs for G×G and G×E studies have shown that each a-priori locus (i.e., loci found in single locus screens) can be treated separately for multiple testing in genome screens for interaction [17‐21]. For each significant rQTL, we performed separate screens for loci that interact (G×G) with each rQTL and their specific traits. We corrected for 5 × 10⁻⁸ divided by the number of independent and significant rQTL from each screen (tau/Aβ42 and ptau/Aβ42).

Traditionally, in human genetics tests of G×G focus only on the additive-by-additive (AA) contrast while interactions between two biallelic loci can have up to four “orthogonal” contrasts that are traditionally parameterized as AA, additive-by-dominance (AD), dominance-by-additive (DA), and dominance-by-dominance (DD). These other types of interactions exist and would be missed with only an AA test. Unfortunately, in natural populations, allele frequencies are not 50/50 which results in unbalanced and sometimes missing cells (two-locus genotypes). The unbalanced cell counts makes the “orthogonal” contrasts no longer orthogonal and sometimes inestimable when there are missing cells [22‐24]. We used a simple interaction test that avoids the need for explicit parameterization of the interaction contrasts and only accounts for interactions with sufficient data for estimation. Our full model incorporates covariates and treats the two-locus genotypes as factors similar to a one-way analysis of covariance (ANCOVA). The reduced model includes the covariates and treats the genotypes from each individual locus as factors [25]. The full model accounts for all single and two-locus interaction effects without explicitly parameterizing them while the reduced model explicitly accounts for just the single locus effects. With sufficient counts for all two-locus genotypes the test has 4 DF. All two-locus genotypes with counts less than five were excluded which ensures that estimation of the interaction effect is appropriate while avoiding the issues of inestimable contrasts or trying to determine explicit contrasts for each particular case. Below is an example of the full model:

where the two-locus genotypes are treated as factors, and below is an example of the reduced model where the genotypes of each SNP are treated as factors:

Parametric bootstrapping [5, 16] was applied to any test that initially met the significance threshold with 100 million replicates.

We attempted to connect the significant rQTL and G×G loci identified in the primary biomarker analyses to AD etiology in three different ways. First, we determined if each rQTL between biomarkers acts as an rQTL between AD risk and either of the two biomarkers. This hypothesis extends from the observation that if a locus affects the relationship between two risk factors, it may also modify the relationship between those risk factors and disease. Therefore, each rQTL was fitted for an analogous rQTL model using logistic regression between Alzheimer’s case/control status and that risk factor using the same CSF biomarker data. Second, we used a large AD case/control dataset to assess whether any of the rQTL or G×G loci are directly associated with AD risk. Finally, we use a different dataset to assess if any of these loci are associated with disease progression via rate of cognitive decline.

We used two datasets for connecting the loci to AD risk through AD case/control data. We used the ADGC as our primary analyses but also report the results from IGAP. While the IGAP study is larger (it includes ADGC), we feel that the ADGC data alone is more homogenous and is a closer match in terms of ethnic composition to our discovery dataset. The ADGC is an NIH-funded collection of GWAS data created for the goal of identifying genetic contributions to late-onset AD. Participants included in this study were from 30 merged datasets [26] combined by Boehme et al. [27] and included 25,666 unrelated individuals carrying either an AD (n = 12,532) or cognitively normal control clinical diagnosis (n = 13,134). Participants were recruited and seen between 1984 and 2012. The genetic data are based on merging the different imputed ADGC datasets. Before merging, SNPs within each dataset with low imputation quality (info < 0.5) were filtered out. Dosage information was used to convert to best guess PLINK allele call format files with an uncertainty cutoff of 0.1. Duplicate samples were identified and removed. Other quality control issues such as physical location discrepancies and strand flipping were identified and appropriately dealt with. A kinship coefficient of 0.0442 was used to screen out third-degree relatives or closer. The resulting number of SNPs in the unrelated dataset was ~ 8.6 million. We only used the SNPs corresponding to our significant rQTL and G×G loci, as simple logistic regression analysis was performed to assess the association between our SNPs and case control status using age, sex, cohort, and two principle components as covariates. Full details on the datasets and the merging process are available at [27].

The IGAP consortium organized AD case/control data from many different studies across the world. For stage 1 of their analysis they used meta-analysis methods for the association between ~ 7 million SNPs and AD case/control data (n = 54,162; 17,008 cases, 37,154 control) from four studies (ADGC, CHARGE, EADI, and GERAD) [28]. We used the results from the stage 1 analyses to determine whether any of the significant rQTL or G×G loci are associated with AD risk. In the preparation of the genetic data, Lambert et al. [28] excluded SNPs with call rates < 95%, imputed using samples of European ancestry in the 1000 Genome Project, and excluded SNPs with MAF < 1%. In each case/control dataset, the association of late-onset Alzheimer’s disease with genotype dosage was analyzed by a logistic regression model including covariates for age, sex, and principal components to adjust for possible population stratification. For the three CHARGE cohorts with incident Alzheimer’s disease data, Cox proportional hazards models were used. For the meta-analysis they undertook a fixed-effects inverse variance-weighted meta-analysis with the standard errors of the beta coefficient scaled by the square roots of study-specific genomic inflation factors estimated before combining the summary statistics across datasets [28].

About half of the samples from the CSF AD biomarker data can be found in both the ADGC and IGAP datasets and only represent ~ 5.5% and ~ 3% of the ADGC and IGAP samples. It is also important to note that the analyses leading to these follow-up tests for association with case/control status in ADGC and IGAP are based on rQTL and G×G analyses using Aβ42, tau, and ptau and not any explicit single-locus test for association with case/control status in the CSF samples.

A total of 1499 individuals with longitudinal cognitive data and genetic data were used to assess the association of the significant rQTL and G×G loci with rate of progression or the rate of cognitive decline. More extensive descriptions of the data and methods are described elsewhere [29]. Participants from this study were enrolled in two different longitudinal studies: the Knight ADRC at Washington University (n = 778) and the ADNI (n = 712). Participants were evaluated in accordance with the Clinical Dementia Rating (CDR) scale, where 0 indicates cognitive normality, 0.5 is very mild dementia, 1 is mild dementia, 2 is moderate dementia, and 3 is severe dementia [30, 31]. The scores in each of the six areas are summed to yield a sum of box (CRD-SB) scores ranging from 0 (no impairment) to 18 (maximal impairment). Only participants that had an AD diagnosis and CDR > 0 at their last visit were included in our analyses. Individuals with dementia caused by neurological diseases other than AD were excluded. For those samples with data available, individuals with Aβ42 values equal or greater than 192 pg/mL (ADNI) [32] or 500 pg/mL (WU) [33] were also excluded. Noninformative longitudinally measured CDR-SB was removed for each participant (noninformative longitudinal data is defined as data where the CDR-SB is either 0 or 18 and remains constant over a period of time). Only individuals with at least two visits and 1.5 years of follow-up were included.

Participants were genotyped with the Illumina 610 or Omniexpress chip (Illumina, San Diego, CA, USA). IMPUT2 v2.3.2 software and the 1000 genome (phase3 NCBI build 37) data were used as reference to impute up to 6 million SNPs. To avoid the possibility of spurious associations of population structure and to confirm ethnicity of each sample, the two principal components scores were used as covariates in the analysis. Only individuals that clustered with the European-American cluster were included in the study.

A linear mixed-model analysis was carried out using R statistical software [14] and the nlme package [34]. A linear mixed-model repeated measure framework was used to account for correlation between repeated measures in the same individual. Change in CDR-SB per year was treated as the independent variable including the following covariants: baseline CDR, baseline age, gender, time (follow-up), level of education, the interaction between baseline CDR and time, and, to avoid the possibility of spurious association due to population substructure, the two first principal components scores were included as covariates, and a random effect for time and individual was included in the model with an AR(1) covariance structure.