Considering the APOE locus in Alzheimer’s disease polygenic scores in the Health and Retirement Study: a longitudinal panel study

Polygenic scores are a strategy to aggregate the small, additive effects of single nucleotide polymorphisms across the genome. With phenotypes like Alzheimer’s disease, which have a strong and well-established genomic locus (APOE), the cumulative effect of genetic variants outside of this area has not been well established in a population-representative sample. Here we examine the association between polygenic scores for Alzheimer’s disease both with and without the APOE region (chr19: 45,384,477 to 45,432,606, build 37/hg 19) at different P value thresholds and dementia. We also investigate the addition of APOE-ε4 carrier status and its effect on the polygenic score—dementia association in the Health and Retirement Study using generalized linear models accounting for repeated measures by individual and use a binomial distribution, logit link, and unstructured correlation structure. In a large sample of European ancestry participants of the Health and Retirement Study (n = 9872) with an average of 5.2 (standard deviation 1.8) visit spaced two years apart, we found that including the APOE region through weighted variants in a polygenic score was insufficient to capture the large amount of risk attributed to this region. We also found that a polygenic score with a P value threshold of 0.01 had the strongest association with the odds of dementia in this sample (odds ratio = 1.10 95%CI 1.0 to 1.2). We recommend removing the APOE region from polygenic score calculation and treating the APOE locus as an independent covariate when modeling dementia. We also recommend using a moderately conservative P value threshold (e.g. 0.01) when creating polygenic scores for Alzheimer’s disease on dementia. These recommendations may help elucidate relationships between polygenic scores and regions of strong significance for phenotypes similar to Alzheimer’s disease.


Background
The most common form of dementia is Alzheimer's disease (AD), represented in roughly 65% of dementia cases [1]. Alzheimer's disease is thought to arise from a combination of both genetics, environment, and lifestyle factors [2]. The estimated heritability of late onset Alzheimer's disease is around 74% [3]. While large-scale genomewide association studies (GWAS) have identified several Open Access *Correspondence: ebakshis@umich.edu 1 Survey Research Center, Institute for Social Research, University of Michigan, 426 Thompson St., Rm. 3320 ISR-Thompson, Ann Arbor, MI 48104, USA Full list of author information is available at the end of the article genetic loci associated with Alzheimer's disease [4][5][6][7][8][9][10], being a carrier of the Apolipoprotein E (APOE-ε4) allele remains the strongest genetic predictor of late-onset Alzheimer's disease [11]. One copy of APOE-ε4 (inheriting a CC at these two locations from either parent) confers a threefold risk of Alzheimer's disease while two copies (inheriting a CC at these two locations from both parents) a 15-fold increase in risk [12]. The effect of APOE-ε4 is all the more difficult to capture in a single variant GWAS as APOE-ε4 is a haplotype composed of two SNPs: rs7412 and rs429358-which will never be fully be captured in a traditional linear model GWAS framework. However, GWAS have identified many independent SNPs in and near the APOE gene locus. The APOE gene region contains many variants in high linkage disequilibrium within roughly 100 kilobases, including several additional highrisk sites in the translocase of outer mitochondrial membrane 40 (TOMM40) gene.
The largest Alzheimer's disease GWAS meta-analysis to date (N = 94,437) is the from the International Genomics of Alzheimer's Project (IGAP) [6]. This metaanalysis confirmed 20 previously identified Alzheimer's disease risk loci [4] and identified five new genome-wide loci including (IQCK, ACE, ADAM10, ADAMTS1, and WWOX). The IGAP used a three-stage strategy where Stage 1 consisted of genotyped and imputed data on 9,456,058 common and 2,024,574 rare single nucleotide polymorphisms (SNPs) to meta-analyze GWAS from four cohorts (n cases = 21,982; n controls = 41,944). Stage 2 included replication with a custom I-select genotyping chip developed in Lambert et al. 2013 [4] and included 11,632 variants and 18,845 individuals with a metaanalysis of Stage 1 and Stage 2. Finally, Stage 3 replicated 44 variants and meta-analyzed Stages 1 and 2 and 3 for a total of 35,274 cases and 59,163 controls. The associations between millions of genetic loci and Alzheimer's disease are documented in IGAP and available for testing in independent populations.
Many complex diseases may result from the consideration of small individual effects across the genome. Polygenic scores (PGS)-also known as polygenic risk scores/ PRS, though they do not always model a risky phenotype (e.g. "risk" of high education, "risk" of increased height)-are generally derived from the sum of weighted variants across an individual [13,14]. PGS incorporate genome-wide genetic variation into a single, quantitative measure that can be used in modeling as a tool to assess susceptibility. Though conceptually simple, many analytic decisions contribute to different qualities of PGSs including different coefficient of variation (R 2 ), correlations between scores, and areas under the curve. An important consideration for diseases such as Alzheimer's disease-which have a genetic locus like APOE/ TOMM40 conferring much of the genetic risk to the disease-is to determine how the remaining variants in the genome contribute to the disease. Removing a region with many risk variants and deciding which and how many variants to include in a PGS can offer substantively different conclusions. For instance, one study reported a PGS area under the curve of 0.57 for Alzheimer's disease (parental proxy) using 21 SNPs and excluding the APOE region [7], while another study reported using more than 200,000 variants (including APOE) and a PGS area under the curve of 0.84 for Alzheimer's disease [15]. While Alzheimer's disease has a strong genetic locus, there is no benchmark across metrics of PGS construction for the rest of the genome's polygenic contribution.
Population-based studies often assess dementia status, rather that AD, as a trade-off between feasibility for longitudinal, larger samples of more diverse participants versus more specific and intensive clinical assessments. An Alzheimer's disease PGS may be informative for dementia more broadly, and Alzheimer's disease PGS has not been assessed in population-based studies of dementia, other than by proxy in the UK Biobank cohort [5]. Further PGS construction metrics regarding the APOE region and additional SNPs have not been compared. The goals of this manuscript are three-fold. The first aim is to assess the utility of using Alzheimer's disease PGSs in population-based analyses of dementia. The second aim is to evaluate the inclusion of the APOE region in these PGSs with and without a covariate modeling risk directly from the APOE-ε4 allele. The third aim is to test SNP inclusion thresholds in PGS on dementia. We conduct this analysis using the Health and Retirement Study (HRS) in the European ancestries (intentionally plural, as there is no single "European" ancestry) participants.

Health and Retirement Study
The Health and Retirement Study (HRS) is a nationally representative panel study featuring a biennial survey of adults over age 50 and their spouses in the United States [16]. The HRS is sponsored by the National Institute on Aging (NIA U01AG009740) and is conducted by the University of Michigan. The HRS began in 1992 as a means to provide a national resource for data on changing health and economic circumstances associated with ageing at both the individual and population levels. These changes are focused on four broads topics: income and wealth; health, cognition, and use of healthcare services; work and retirement; and family connections [17].
The HRS pre-selected a random one-half of the sample to receive an enhanced face-to-face interview in 2006, which included physical performance tests, anthropometric measurements, blood and saliva samples, and a psychosocial self-administered questionnaire in addition to the HRS core interview. The remaining one-half sample received the same enhanced face-to-face protocol in 2008. The HRS randomly assigned the new 2010 cohort to receive an enhanced face-to-face interview in either 2010 or 2012. Those participants who were not interviewed or did not consent to saliva in 2006 were asked again in 2010.
Salivary DNA was collected using Oragene-250 saliva kits and protocol. DNA extracted from the saliva and was genotyped at the Center for Inherited Disease Research (CIDR) using the Illumina HumanOmni2.5 array (8v1 and 4v1). The Genetics Coordinating Center at the University of Washington, Seattle, WA performed Genotyping Quality Control. SNP annotation aligned to genome build 37/hg 19. The Genetics Coordinating Center calculated genetic principal components (PC) with HapMap controls [18,19]. In addition to selecting independent SNPs with missing call rates < 5% and minor allele frequencies > 5%, the 2q21 (LCT), HLA, 8p23, and 17q21.31 regions were excluded from the initial pool [20]. The final European ancestries sample includes all self-reported non-Hispanic White persons that had PC loadings within ± one standard deviations of the mean for eigenvectors one and two in the PC analysis of all unrelated study subjects. The HRS re-calculated genetic PCs within the group of non-Hispanic Whites of European ancestries to further account for population stratification. These "ancestry-specific principal components" were used in subsequent analyses. Imputation was performed using IMPUTE2 on HRS data phased using SHA-PEIT2. Data were imputed to the 1000 Genomes Project (1000GP) cosmopolitan reference panel phase 3 version 5 (initial release on May 2013, haplotypes released Oct 2014) and are available on the database of genotypes and phenotypes (dbGaP, https ://www.ncbi.nlm.nih.gov/ gap/,phs00 0428.v2.p2).

APOE-ε4
The apolipoprotein E (APOE) gene codes for a protein that binds and transports low-density lipids and is responsible, in part, for removing cholesterol from the bloodstream [21,22]. Variations in this gene affect cholesterol metabolism and may lead to increases in the risk for stroke, heart disease, and may alter the odds of having Alzheimer's disease. Two variants (rs7412 and rs429358) define the APOE genotype resulting in three common isoforms of protein apoE: E2, E3, and E4 encoded by ε2, ε3, and ε4. Of note, the genotyped rs7412 and rs429358 variants failed genotyping quality control in the HRS pipeline and are therefore not included as individual variants in any PGS using genotyped data alone. Using the imputed rs7412 (IMPUTE2 INFO score = 0.99) and rs429358 (IMPUTE2 INFO score = 0.99) variants, we categorized HRS participants as ε2/ε2, ε2/ε3, ε2/ε4, ε3/ε3, ε3/ε4, and ε4/ε4. We analyze two indicator variables for presence of one ε4 allele (1 = yes, 0 = no) or two ε4 allele (1 = yes, 0 = no).

Polygenic score for Alzheimer's disease
We investigate using genome-wide raw genotyped variants and the two imputed APOE variants (rs7412 and rs429358) in the creation of our PGSs. We include SNPs in our PGS analysis at six AD-SNP association P-value thresholds (pT) from the Kunkle et al. [6] summary statistics: pT = (0.001, 0.01, 0.05, 0.1, 0.3, 1.0). For example, a PGS at pT = 0.01 includes only those variants for which the association p-value in the Kunkle et al. [6] GAP meta-analysis on Alzheimer's disease was less than 0.01. We obtained summary statistics from National Institute on Aging Genetics of Alzheimer's Disease Data Storage Site https ://www.niaga ds.org/datas ets/ng000 75. Importantly, our study sample was not included in the Kunkle study of clinical Alzheimer's disease. Thus, the weights are independent of our study sample. We do not include any linkage disequilibrium filtering in our scores (i.e. no clumping or pruning algorithms). That is, we include any directly genotyped variants that pass quality control from the HRS and overlap with those variants from the Alzheimer's disease GWAS summary statistics in our scores, unless otherwise noted. We did not perform clumping or pruning because we are using the set of tag-SNPs from the genotyping chip, which do not contain imputed variants and represent, in essence, an already filtered set of variants.
For our scores with the APOE gene region removed, we removed all variants from the summary statistics on chromosome 19 near APOE (45,384,477 to 45,432,606, build 37/hg 19). This represents the start position of TOMM40 (45,394,477) − 10 kilobases and the stop position of APOC1 (45,422,606) + 10 kilobases. This region was removed in its entirety due to the dense linkage disequilibrium block in European ancestries overlapping these three genes (TOMM40, APOE, APOC1). We compare two PGSs: genotyped PGS including the APOE gene region, genotyped PGS with APOE region removed ( Fig. 1), at six p-values thresholds from the Kunkle et al.

Cognition status
The HRS used a multidimensional measure of cognitive functioning, based on a telephone screening instrument: Telephone Interview for Cognitive Status [23]. Domains assessed using this measure include: memory, mental status, abstract reasoning, fluid reasoning, vocabulary, dementia, and numeracy. In 2009, Langa, Kabeto, and Weir developed an approach for defining dementia and cognitively impaired non-dementia (CIND) in the HRS. A team of dementia experts clinically validated this method using equipercentile equating in the HRS against the Aging, Demographics, and Memory Study (ADAMS). The ADAMS study is a sub-sample of the HRS who received a more extensive neurological battery [24,25]. For self-respondents, the score consists of overall cognitive test performance while the proxy respondents' scores are composed of proxy-rated memory, interviewer-perceived cognition, and impaired activities of daily living. The cut points for this method reflect the prevalence of dementia or cognitive impairment to the expected population prevalence from the ADAMS study. We used repeated measures of the classification of cognitive function data contributed for assessment years 2000−2014 from the HRS imputed cognition researcher contribution data set [26]. For self-respondents, a score from 0 to 6 is categorized as dementia, 7 to 11 is categorized as cognitive impaired not dementia, and 12 to 27 is categorized as normal cognition. For proxy respondents starting in 2000, a score of 6 or higher out of 11 is classified as having dementia, a score of 3 to 5 indicates cognitive impaired not dementia, while 0 to 2 indicates normal cognition [24]. In this analysis, we are only testing the odds of dementia versus normal cognition (1 = dementia, 0 = normal cognition).   Fig. 1 The APOE/TOMM40 genomic locus on chromosome 19. The y-axis corresponds to − log 10 (p-values) of association with Alzheimer's disease in Kunkle et al. [6]. Single nucleotide polymorphisms within the bracketed genomic region were removed from consideration in polygenic score development for scores designed to exclude the APOE region. The variants that make the APOE isoforms are highlighted in purple diamond (rs7412, rs429358) 1942-1947); EBB: early baby boomers (b. 1948-1953); MBB: mid-baby boomers (b. 1954-1959)], and sex (0 = female, 1 = male) are measured at a participant's initial HRS exam. Age (years) and a self-report of doctor diagnosed stroke (0 = none, 1 = stroke, possible stroke/ TIA/mini-strokes, respondent disputes previous waves that indicate condition) are assessed at the same wave as the corresponding cognition visit.

Statistical analysis
Due to the repeated measures in this analysis, we use generalized linear models and the GENMOD procedure in SAS 9.4. We specify repeated measures on the individual, a binomial distribution, and a logit link with an unstructured correlation structure. We include fixed effects covariates in every model: sex, years of education, and five ancestry-specific principal components. The time varying covariates chronological age, year, and stroke history at each interview wave are also included in every model. To evaluate the utility of the PGS in individuals with and without APOE-ε4, we also investigate an interaction effect between APOE-ε4 status and PGS. We use an α of 0.05 as a threshold for significance.

Results
There are 9,872 individuals in the HRS non-Hispanic White, European ancestries analytic sample collected between 2006 and 2010. We removed observations with missing cognition (m = 10,958), observations where the cognitive status was classified as CIND (m = 6905), and observations with a missing stroke status (m = 16). This removed 55 individuals from the analysis. The final analysis included n = 9817 h respondents of European ancestries with a total of m = 51,225 cognitive observations (Fig. 2).
Our analytic sample was 57.8% female with an average age of 63.3 (SD = 10.3) at their first visit during the observation period. On average, participants had 13.2 (2.5) years of education. The average number of cognitive assessments per participant was 5.2 (1.8). A small portion of the sample had a history of stroke at their first visit during the observation period (n = 417, 4.3%), with a higher proportion of males reporting a history of stroke at their first visit (n = 201, 4.9%) than females (n = 216, 3.8%). In unadjusted analyses, the Alzheimer's disease PGS at any pT was not significantly different between males and females and broadly centered at zero with a standard deviation of one. There was no difference in the distribution of number of copies of APOE-ε4 by sex, where the overall proportion with one copy of APOE-ε4 was 24.3% (n = 2382), and two copies of APOE-ε4 was 2.1% (n = 210). Across all observations, accounting for repeated measures within individuals, age and dementia at first visit during the observation period were not significantly different by sex. The proportion of stroke; however, was significantly different by sex (P = 0.0004) with a higher proportion in males than in females (Table 1). APOE-ε4 status is associated (all Ps < 0.05) with PGS (Fig. 3). Additional file 1: AF Table 1 contains a table of correlations between each Alzheimer's disease PGS.
In adjusted, repeated measures regression models, increased age, later year of observation, history of stroke, and lower education were associated with increased odds of dementia in all models (all P < 0.0001; Table 2). In the models using an Alzheimer's disease PGS with the APOE region included, having one copy of APOE-ε4 increased the odds of dementia relative to normal cognition by roughly twofold, while two copies of APOE-ε4 increased the odds of dementia by over fourfold (Table 2), holding all other variables constant. The Alzheimer's disease PGS with the APOE region included was marginally associated with the odds of dementia relative to normal cognition only for the pT cutoff of 0.01 (OR = 1.1 95%CI 1.0 to 1.2), while the Alzheimer's disease PGS at all other P-value thresholds was not significantly associated with the odds of dementia. In the models that included the Alzheimer's disease PGS with the APOE region removed, we observed similar estimates for one and two copies of APOE-ε4 as before, and a similar effect size of ~ 8% increase in the odds of dementia for every one standard deviation increase in the Alzheimer's disease PGSs. Both the Alzheimer's disease PGS (without the APOE region) at pT = 0.001 and 0.01 were significantly associated with the odds of dementia (pT = 0.001 OR = 1.1 95%CI 1.0 to 1.2 and pT = 0.01 OR = 1.1 95%CI 1.0 to 1.2). The Alzheimer's disease PGSs at pT = (0.05, 0.1, 0.3, and 1.0) were not associated with the odds of dementia, relative to normal cognition in this sample.

Sensitivity analysis
Researchers have demonstrated some amount of mortality selection in the oldest individuals in the HRS genetic sample [27], and so we removed the two oldest cohorts (AHEAD and CODA) as a sensitivity analysis (Additional file 1: AF Table 2). We removed a total of 2020 individuals and 8957 observations. The effect of APOE-ε4 remained highly associated with the odds of dementia compared to normal cognition. Though the effect size for the Alzheimer's disease PGSs at each pT and whether or not the APOE region was included were not significantly different from those in Table 2, the P-values associated with these effects were non-significant for all Alzheimer's disease PGS. The slightly attenuated effects are not surprising as the younger cohorts are just now entering ages at which dementia becomes more prevalent. We additionally regressed out the effect of the APOE-ε4 from each PGS and performed the same analyses as in Table 2 (Additional file 1: AF Table 3). The PGSs with APOE-ε4 regressed out correlated with their counterparts at a Pearson's correlation coefficient r > 0.97 (Additional file 1: AF Table 1). As expected, the results from Table 2 and these new analyses in Additional file 1: AF Table 3 are virtually identical. Associations between each PGS and dementia without adjustment for APOE-ε4, and also the association between the three level APOE-ε4 variable and dementia without adjustment for PGS are found in Additional file 1: AF Tables 4 and 5, respectively. Each of these models are adjusted for age, sex, education, year of visit, and stroke history. We found no statistical interaction between APOE-ε4 and a PGS at a

Discussion
In a large, population-based cohort of older, European ancestries Americans, cumulative genetic risk summarized as a PGS is informative of longitudinal dementia odds. We observed that the APOE region requires handling with care in the development of PGS. Specifically, including the APOE region as weighted SNPs in a PGS was insufficient to account for the large risk attributed to    the APOE region. We recommend removing the region in linkage disequilibrium around the APOE locus from the PGS and treating the APOE locus as an independent covariate. In addition, we observed greater performance from PGS developed at a P-value threshold of 0.01 for SNP inclusion, with greater noise from a PGS informed by the full genome in association with this dementia phenotype. Optimized measures of the polygenic nature of dementia allow for more powerful interrogations of genetic and environmental risk for dementia. We observed the APOE-ε4 allele was longitudinally associated with higher risk of dementia, in a dose dependent manner. This observation is consistent with extensive prior research [28,29]. The APOE-ε4 allele is neither necessary, nor sufficient to cause dementia, but the magnitude of increased risk attributed to each copy of the allele is relatively high. The APOE-ε4 allele is in linkage disequilibrium with a ~ 100 kilobase region involving the APOE, APOEC, and TOMM40 genes. Thus, an APOE independent PGS would need to remove the SNPs from the entire APOE region. In excess of the association between the APOE locus and dementia, we observed a small, but significant association between Alzheimer's disease PGS and dementia. We also found that the effect of the PGS on dementia was not significantly different by APOE status. These findings are similar to those observed in clinical populations investigating APOE independent PGS risk of Alzheimer's disease specifically [15,[30][31][32][33]. When building PGS, it is important to have independent study samples between the discovery GWAS and the application PGS. Notably, our study sample was not part of the Kunkle GWAS that generated the weights for the PGS. Our findings show the APOE independent Alzheimer's disease PGS can be successfully implemented in population-based research of a broad dementia phenotype.
Dementia is a disorder with a strong genetic locus of effect (APOE) and substantially weaker effects are scattered throughout the genome. Including the APOE region in PGS without specific measurement of APOE-ε4 is insufficient, and overestimates the polygenic nature of dementia. Similarly, in Amyotrophic Lateral Sclerosis there is a strong main effect locus (C9orf72). Further, there is a significant, albeit modest, proportion of the phenotypic variance explained by a polygenic risk score over and above the C9orf72 region [34]. In contrast, other chronic disease traits, such as obesity, lack a dominant genetic locus and polygenic score development is successful across the entire genome at a higher p-value threshold [35]. Together, these results suggest that in traits with a strong genetic locus, polygenic scores should exclude the primary regions and seek to aggregate the remaining genetic risk as a separate predictor.
We acknowledge several limitations in this research. The first is that our study relied on imputed APOE variant calls. The APOE region is notoriously difficult and labor intensive to measure genotypes [36,37]. Indeed, the two primary APOE SNPs of interest failed quality control metrics on the genotyping array in the HRS. We used the correlation structure of the genome from the 1000GP reference to impute these SNPs with ~ 99% confidence. Second, our study may be subject to mortality selection; however, dementia is primarily a disease of older age and requires survival long enough to manifest symptoms. Mortality selection related to the APOE genotype would only serve to make our observations more conservative. Third, we developed our PGS using weights from a GWAS of primarily European ancestry participants, thereby limiting generalizability to other ancestries. Last, our population-based study assessed a broad phenotype of dementia. There are many types of dementia including Alzheimer's, vascular, and frontotemporal lobe, which have varying genetic architectures, to which we applied a PGS specific for Alzheimer's. As future GWAS become available for dementia subtypes in a clinical population, investigators may be able to classify the utility of PGS in dementia subtypes. Future family-based studies may also consider SNPs strongly associated with early-onset AD, including APP, PSEN1, or PSEN2 mutations.

Conclusion
Dementia has considerable risk attributed to genetic factors. The APOE region is the strongest locus associated with disease and many additional sites confer small risk effects. Incorporating genetic risk from many sites in a polygenic risk score is a useful metric for risk prediction  and etiologic testing in epidemiologic research of complex traits [38]. Our findings demonstrate the APOE region should be removed prior to polygenic risk score development and treated as an independent factor in dementia analyses. More work is needed to assess polygenic scores for Alzheimer's disease for clinical utility and prediction and in diverse ancestries.