Effect of APOE and a polygenic risk score on incident dementia and cognitive decline in a healthy older population

Abstract Few studies have measured the effect of genetic factors on dementia and cognitive decline in healthy older individuals followed prospectively. We studied cumulative incidence of dementia and cognitive decline, stratified by APOE genotypes and polygenic risk score (PRS) tertiles, in 12,978 participants of the ASPirin in Reducing Events in the Elderly (ASPREE) trial. At enrolment, participants had no history of diagnosed dementia, cardiovascular disease, physical disability or cognitive impairment. Dementia (adjudicated trial endpoint) and cognitive decline, defined as a >1.5 standard deviation decline in test score for either global cognition, episodic memory, language/executive function or psychomotor speed, versus baseline scores. Cumulative incidence for all‐cause dementia and cognitive decline was calculated with mortality as a competing event, stratified by APOE genotypes and tertiles of a PRS based on 23 common non‐APOE variants. During a median 4.5 years of follow‐up, 324 participants developed dementia, 503 died. Cumulative incidence of dementia to age 85 years was 7.4% in all participants, 12.6% in APOE ε3/ε4 and 26.6% in ε4/ε4. APOE ε4 heterozygosity/homozygosity was associated with a 2.5/6.3‐fold increased dementia risk and 1.4/1.8‐fold cognitive decline risk, versus ε3/ε3 (p < 0.001 for both). High PRS tertile was associated with a 1.4‐fold dementia risk versus low (CI 1.04–1.76, p = 0.02), but was not associated with cognitive decline (CI 0.96–1.22, p = 0.18). Incidence of dementia among healthy older individuals is low across all genotypes; however, APOE ε4 and high PRS increase relative risk. APOE ε4 is associated with cognitive decline, but PRS is not.


| INTRODUC TI ON
Few studies have measured the effect of Apolipoprotein E (APOE) genotypes and polygenic risk scores (PRS) on incident dementia and cognitive decline in healthy older people. The ASPREE (ASPirin in Reducing Events in the Elderly) cohort offers the opportunity to measure these effects, as recruited participants had no history of cardiovascular disease, dementia or significant physical disability at enrolment. The ASPREE study was a randomised, placebo-controlled trial to determine whether daily low dose aspirin increased survival, free of dementia or persistent physical disability, in 19,114 healthy community-dwelling older people (McNeil et al., 2017). In 2018, ASPREE reported that over an average 4.5 years of follow-up, aspirin did not prolong disability-free survival (McNeil et al., 2018a(McNeil et al., , 2018b(McNeil et al., , 2018c or reduce the risk of dementia or cognitive decline (Ryan et al., 2020).
The APOE gene is the strongest genetic determinant of allcause dementia, especially Alzheimer's disease (AD), with the ε4 allele elevating risk and accelerating age of onset (Q ian et al., 2017;Rasmussen et al., 2018;Saunders et al., 1993;Lee et al., 2018).
Individually, these common genetic variants have low effect sizes, yet when combined into a PRS can enable risk-stratification for dementia indications beyond APOE genotype. There is varying evidence for whether a PRS for dementia can also predict cognitive decline (Chaudhury et al., 2019;Harris et al., 2014;Marden et al., 2016;Verhaaren et al., 2013). Incorporating both APOE genotypes and PRS, alongside conventional risk factors, may enable more accurate risk prediction (Licher et al., 2019;Lourida et al., 2019). This may aid development of therapeutic strategies or prevention, and advance our understanding of the genetic differences between (diagnosed) dementia and cognitive decline.
The predictive performance of PRSs for dementia requires further investigation in well-characterised prospective studies.
Predictive performance can be influenced by factors such as ethnicity, age, study recruitment criteria, clinical diagnostic criteria, neuropsychological assessments used, genotyping platform and genetic variants included (Tan et al., 2017;Chouraki et al., 2016;Cruchaga et al., 2018;Desikan et al., 2017;Escott-Price et al., 2015;Leonenko et al., 2019;Licher et al., 2019;Lourida et al., 2019;Qian et al., 2017;Rasmussen et al., 2018;Saunders et al., 1993;Sleegers et al., 2015;Lee et al., 2018). More studies of cognitively healthy elderly individuals followed prospectively are required to assess variability and predictive accuracy. Here, we report the effects of APOE and PRS on incident dementia and cognitive decline among 12,978 ASPREE participants, where dementia was an exclusionary criterion at entry and adjudicated as a primary trial endpoint.

| Incident Dementia Diagnosis
After standardised cognition and functional measures, participants reporting memory or cognitive problems were assessed by the Victorian Cancer Agency. We thank the trial staff in Australia and the United States, the participants who volunteered for this trial, and the general practitioners and staff of the medical clinics who cared for the participants. We also thank the UK Biobank participants and admin staff. P.L is supported by a National Heart Foundation Future Leader Fellowship (102604).
1.4/1.8-fold cognitive decline risk, versus ε3/ε3 (p < 0.001 for both). High PRS tertile was associated with a 1.4-fold dementia risk versus low (CI 1.04-1.76, p = 0.02), but was not associated with cognitive decline p = 0.18). Incidence of dementia among healthy older individuals is low across all genotypes; however, APOE ε4 and high PRS increase relative risk. APOE ε4 is associated with cognitive decline, but PRS is not.

K E Y W O R D S
Alzheimer's disease, Apolipoprotein E, aspirin in reducing events in the elderly, cognition, cognitive decline, cumulative incidence of dementia, genome-wide association study, polygenic risk score specialists or prescribed dementia medication. Following identification of dementia triggers (3MS<78 or a drop of >10.15 points from the participant's baseline 3MS score, accounting for age and education), additional assessments were conducted, with brain imaging and laboratory analyses collected for adjudication. Each dementia trigger case was reviewed according to the ASPREE protocol for clinical adjudication (McNeil et al., 2018c;Ryan et al., 2020)

| Cognitive decline
The ASPREE cognitive battery included the 3MS for general cognition, the Hopkins Verbal Learning Test-Revised (HVLT-R) delayed recall for episodic memory, the single letter Controlled Oral World Association Test (COWAT) for language and executive function, and the Symbol Digit Modalities Test (SDMT) to measure psychomotor speed. Accredited professionals administered assessments at baseline and year 1, followed biennially during follow-up. As reported previously (Ryan et al., 2020) cognitive decline in participants without a dementia diagnosis was defined as a 1.5 standard deviation decline in 3MS/HVLT-R/SDMT/ COWAT compared with baseline scores, sustained over ≥2 time points.  Figure S2) (Auton et al., 2015;Zheng et al., 2012).

| Genotyping and variant analysis
Imputation was performed using the haplotype reference consortium European panel (Das et al., 2016). Post-imputation quality control removed variants r2 < 0.3. APOE genotype was measured using two directly genotyped variants (rs7412, rs429358) extracted using plink v1.9 (Chang et al., 2015) 2.5 | Polygenic risk score PRS was calculated using 23 common variants (15 genotyped, 8 imputed) associated with AD at genome-wide significance that affect risk independently of APOE Lambert et al., 2013;Ruiz et al., 2014). PRS calculations, using plink v1.9 (Chang et al., 2015), were based on dosage (0,1,2) of SNP effect allele reported from GWAS, multiplied by effect sizes, followed by the sum of products to generate a PRS per participant (Table S1). We used the same 23 SNP PRS and the same PRS calculation methods used in recent analysis of the Rotterdam study (Lee et al., 2018). PRS distribution was divided into low/middle/high tertiles; with mean values of; low −0.56 (range −1.43 to −0.34), middle −0.20 (−0.34 to −0.06) and high 0.16 (−0.06 to 1.86) ( Figure S3). Tertiles were used to ensure equal distribution of samples across PRS groups, and sufficient events numbers occurred in each group for statistical power possible ( Figure S3). In addition, we sought to use the same analysis approach as the Rotterdam Study (Lee et al., 2018), where tertiles were also used.

| Statistical analysis
To determine whether APOE genotype frequencies were under selective pressure due to age and/or trial inclusion/exclusion criteria, we performed Hardy-Weinberg equilibrium (HWE) testing. This compared observed genotype frequencies with those expected in a population under no selective pressure, using chisquared tests. We examined the cumulative incidence of dementia (CID) and cognitive decline, stratified by APOE genotype and PRS tertiles. We used ε3 homozygotes as a reference group for APOEstratified analysis and the low-risk tertile for PRS-stratified analysis. Consistent with other studies Lee et al., 2018) we combined APOE ε3/ε4:ε2/ε4 into a single group, and ε2/ ε2:ε2/ε3 into a single group.
We estimated cumulative incidence of all-cause dementia and cognitive decline during an average of 4.5 years of follow-up, using the Cumulative Incidence Function (CIF) of the etm package (Allignol et al., 2011;Meister & Schaefer, 2008) in R version 3.6.0 (R Core Team, 2013). Data were censored by date of dementia diagnosis, cognitive decline, last contact or death. The age on censored date was used as a time scale in CIF model. Cumulative incidence was calculated up to 95 years, then stratified by APOE genotype and PRS tertiles. Dementia and cognitive decline between PRS tertiles were compared for the whole cohort and further stratified by APOE genotypes. The dementia and cognitive decline models were estimated independently.
We used the Fine and Gray (F&G) method of accounting for competing risk of death, and Cox proportional hazard regression model to calculate dementia hazard ratio of both models, for APOE, PRS and their interaction, adjusted for age at enrolment (continuous, allowing a quadratic function) and sex (Meister & Schaefer, 2008; Lee et al., 2018). We used age on censored date as a time scale in both F&G and Cox models. Hazard ratios for cognitive decline were measured the same way. To test association of APOE genotypes and PRS with cohort characteristics, we used a multivariable regression model with variables; age, sex, follow-up time, education, alcohol use, smoking, diabetes mellitus, hypertension, depression (Center for Epidemiological Studies-Depression-10 scale), family history of dementia (father/mother/sibling), body mass index, blood pressure, cholesterol and triglycerides. Bonferroni multiple test correction at p = 0.002 significance was applied (0.05/17 = 0.002). We followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for reporting of results, see Table S2.

| RE SULTS
Characteristics of the 12,978 genotyped participants are shown in Table 1. Overall, 54.8% were female, 47% had educational attainment <12 years, 3% were current smokers and 25% reported a family history of dementia at enrolment.
Associations of cohort characteristics with APOE genotypes and PRS tertiles are also shown in Table 1. The only associations to survive multiple testing correction were family history of dementia in ε4 heterozygotes/homozygotes, with no cohort characteristics differing between PRS tertiles (Table 1).
We found that APOE genotype frequencies had deviated from HWE (chi-square = 38, p < 0.001) (Table S3) (Tables 1 and 2). Of these, 143 were classified as 'probable AD' and 176 were classified as 'possible AD'. Only 5 cases were classified as 'non-AD related dementia' (Table 1). For cumulative incidence of dementia (CID), we describe results up to age 85 years, representing an approximate centre point between lower and upper age ranges of the ASPREE population at baseline (70 to 96 years). CID in ASPREE was estimated at 7.4% (CI 6.5 to 8.5).
We compared CID between the highest genetic risk group at age 80 (ε4 carriers with high PRS) and the lowest genetic risk group at age 90 (ε2 carriers with low PRS). CID in the highest genetic risk group at age 80 was 6.1% (CI 4.1-9.0) and in the lowest genetic risk at age 90 was 8.8% (CI 4.5-16.7) (Table S6). This corresponded to an approximately 10-year delay in age of onset between these two extreme groups. In sensitivity analysis, we examined interaction between APOE/PRS and found no significant association with incident dementia (p > 0.05).
A total of 1598 (12.6%) participants had cognitive decline (

| DISCUSS ION
In this study, we examined the effect of APOE genotypes and PRS on incident dementia and cognitive decline among 12,978 initially healthy older participants. We found that APOE ε4 and high PRS were associated with increased relative risk of dementia, but overall, cumulative incidence of dementia was low across all genotype groups. PRS effect on dementia risk was modest and delayed compared with APOE ε4, mostly affecting risk after 85 years of age.
APOE ε4 was associated with cognitive decline, but PRS was not, suggesting that APOE genotype has a stronger effect than PRS on both dementia and cognitive decline. We observe that the absence of co-morbidities, atherothrombotic cardiovascular disease and cognitive impairment to age 70 years contributed to the attenuation of incident dementia across all genotypes.  et al., 2018). These differences in CID are substantial, unlikely to be attributable to confounding factors alone between the studies.

TA B L E 1 Characteristics of the ASPREE cohort stratified by APOE genotypes and tertiles of a PRS
Further, in a recent meta-analysis of three population-based cohorts of cognitively normal subjects aged 60-75 years (total N = 11,771), the risk of dementia in APOE ε4/ε4 homozygotes (N = 134) to age 70-75 years was 11.2% (Qian et al., 2017). In ASPREE, however, the risk of dementia to age 75 in ε4/ε4 (N = 200) was only 3.7%. Risk of dementia among ε4/ε4 homozygotes to age 85 years in the Framingham Heart Study (37.6%, N = 67) was also considerably higher than ASPREE (26.6% N = 200) (Qian et al., 2017).
We acknowledge the variation in genetic risk of dementia between ethnic groups (Teruel et al., 2011), yet our study was not designed to assess ethnic differences. We assessed genetic effects in individuals of European ancestry only, and compared results with another similar sized cohort of European ancestry (the Rotterdam Study (Lee et al., 2018)). We did not include individuals with non-European ancestry in the analysis, due to small sample size and the risk of population stratification bias influencing genetic risk estimates.
However, we also observed an attenuated effect of PRS on dementia in ASPREE compared with other studies. We observed only a 2.6% difference in CID between low (7.3%) and high (9.6%) PRS tertiles.
In the Rotterdam study, the observed difference was 9.0% between low (11.6%) and high (20.4%) tertiles to the same age.
In ASPREE, the effect of PRS was more pronounced in APOE ε4 carriers, compared with the reference ε3/ε3 group. However, the PRS effect was attenuated and delayed in age of onset compared with other studies (Peloso et al., 2020;Lee et al., 2018). The PRS effect on dementia risk in ASPREE mostly occurred after the age of 85 years (Figure 2a). We found no significant interaction effect between APOE and PRS in ASPREE, unlike the Rotterdam study (Lee et al., 2018). This again may reflect the attenuation of genetic effects on dementia risk in ASPREE. A recent analysis of the Framingham cohort also reported no significant interaction between APOE and PRS while evaluating dementia risk (Peloso et al., 2020). Further studies with large populations and longer follow-up are required to understand interactions between APOE and PRS in modifying dementia risk. The majority of dementia events observed in our study were classified as either 'probable AD' or 'possible AD', with only 5 cases classified as 'non-AD related'. Therefore, we were unable to undertake sub-group analysis, based on dementia sub-classifications.
Considering the attenuated genetic risk of dementia observed in ASPREE, we query whether other factors further modified risk, beyond the low vascular risk, cognitive screening and absence of cardiovascular disease at baseline. Such factors could include a favourable lifestyle, characterised by healthy diet, regular exercise and high socialisation levels (Licher et al., 2019;Lourida et al., 2019).
Alternatively, the attenuation could be related to the relatively short follow-up period, during which healthy selection effects had not yet dissipated.

F I G U R E 1
Cumulative incidencece of all-cause dementia stratified by APOE genotypes and tertiles of a polygenic risk score (PRS). Cumulative incidence curves for all-cause dementia (a) and cognitive decline (b) were calculated to age 95 years and stratified by APOE genotype, with mortality as a competing event.
Confidence intervals and participants at risk are shown in Table  S4-5 Protective genetic loci not included in the PRS may also have contributed to risk modification, including common variants yet to be identified by GWAS and/or rare high-effect protective variants, including loss-of-function variants in biologically associated genes.
There is growing evidence that protection from dementia risk can be conferred by both common and rare genetic variants, especially in the high-risk APOE ε4/ε4 group (Belloy et al., 2020;Huq et al., 2019).
Further studies are required to examine the effect of protective genetic variants for dementia in ASPREE.
APOE ε4 carrier status was significantly associated with cognitive decline in ASPREE, but PRS was not. This reflects the more modest effect of PRS on cognitive ageing, and/or a divergent genetic aetiology versus APOE genotype (Harris et al., 2014). The association between APOE ε4 and cognitive decline in non-demented individuals has been reported by several studies using comparable cognitive testing (Albrecht et al., 2015;Jager et al., 2012;Reas et al., 2019;Verhaaren et al., 2013;Wisdom et al., 2011). However, few studies have reported a significant effect of PRS on cognitive decline alone (Chaudhury et al., 2019;Harris et al., 2014;Marden et al., 2016;Verhaaren et al., 2013). It appears that PRS derived from GWAS of diagnosed dementia/AD cases are not strong predictors of cognitive decline without dementia during ageing. However, our approach to quantifying cognitive decline may be insensitive or might reflect a different biological process. Alternatively, PRS derived from a GWAS of dementia/AD cases may reflect the functional impairment required for dementia diagnosis, rather than the cognitive aspects.
Strengths of the study include a well-characterised longitudinal cohort with repeated cognitive assessments and dementia adjudication, genetic data for both APOE and PRS variants, longitudinal follow-up to enable survival analysis for dementia and cognition, data available on covariates, adjudicated reports of causes of death to control for competing events and a large number of initially healthy elderly participants.
Limitations of the study include a shorter duration of follow-up compared with other studies (Qian et al., 2017;Lee et al., 2018 (pos-sibly insufficient to overcome a healthy volunteer effect) and limited event numbers in some rarer APOE genotype groups. We also used a broad definition of cognitive decline (CD), characterised by a 1.5 SD inter-individual decline on any cognitive test/domain. This broad measure of CD did not report or quantitate the frequency of domain-specific decline.
In conclusion, our study found that APOE genotypes and PRS effect the relative risk of dementia in a population of healthy older individuals followed prospectively. However, overall CID in the population was low across all genotype groups, reflecting the healthy nature of the population at enrolment. APOE ε4 had a stronger effect than PRS on dementia risk. APOE genotypes affected cognitive decline, whereas PRS did not. Prospective studies of initially healthy older participants with longer follow-up periods are required to further understand the genetic risk of dementia and cognitive decline during ageing, and examine the predictive performance and clinical utility of PRS.

ACK N OWLED G EM ENTS
We thank the ASPREE trial staff in Australia and the United States, the ASPREE participants who volunteered for the trial, and the general practitioners and staff of the medical clinics who cared for the participants. No other conflicts were reported.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available on request from the senior authors of the study (PL and MR).