Effect of Pathway-specific Polygenic Risk Scores for Alzheimer’s Disease (AD) on Rate of Change in Cognitive Function and AD-related Biomarkers among Asymptomatic Individuals

Abstract

Background:

Genetic scores for late-onset Alzheimer’s disease (LOAD) have been associated with preclinical cognitive decline and biomarker variations. Compared with an overall polygenic risk score (PRS), a pathway-specific PRS (p-PRS) may be more appropriate in predicting a specific biomarker or cognitive component underlying LOAD pathology earlier in the lifespan.

Objective:

In this study, we leveraged longitudinal data from the Wisconsin Registry for Alzheimer’s Prevention and explored changing patterns in cognition and biomarkers at various age points along six biological pathways.

Methods:

PRS and p-PRSs with and without APOE were constructed separately based on the significant SNPs associated with LOAD in a recent genome-wide association study meta-analysis and compared to APOE alone. We used a linear mixed-effects model to assess the association between PRS/p-PRSs and cognitive trajectories among 1,175 individuals. We also applied the model to the outcomes of cerebrospinal fluid biomarkers in a subset. Replication analyses were performed in an independent sample.

Results:

We found p-PRSs and the overall PRS can predict preclinical changes in cognition and biomarkers. The effects of PRS/p-PRSs on rate of change in cognition, amyloid-β, and tau outcomes are dependent on age and appear earlier in the lifespan when APOE is included in these risk scores compared to when APOE is excluded.

Conclusion:

In addition to APOE, the p-PRSs can predict age-dependent changes in amyloid-β, tau, and cognition. Once validated, they could be used to identify individuals with an elevated genetic risk of accumulating amyloid-β and tau, long before the onset of clinical symptoms.

Introduction

Late-onset Alzheimer’s disease (LOAD) is an age-related neurodegenerative disease that is clinically manifested by a progressive deterioration of cognitive function, memory, and social ability. Abnormal accumulation of proteins such as amyloid-β (Aβ) and tau are two hallmarks that play important roles in LOAD pathology long before the clinical symptoms of neurodegeneration are evident [1]. Under the amyloid hypothesis, an imbalance between Aβ clearance and Aβ production is considered the underlying cause for the initiation of LOAD through the formation of extracellular senile plaques in the brain [2]. Previous studies have provided evidence that neurobiological pathways, such as amyloid-β protein precursor (AβPP) processing, altered cholesterol metabolism, endocytosis, and tau pathology, are closely linked to Aβ production and clearance [3–9]. Tau, on the other hand, is hypothesized to trigger the progression of LOAD by forming insoluble filaments and accumulating intracellular neurofibrillary tangles of hyperphosphorylated tau in the brain. These accumulations block axonal transport and finally harm the synaptic communications between neurons. In addition to the four pathways affecting Aβ, neurobiological pathways of LOAD that are related to tau accumulation among LOAD patients include immune response and axonal development [3–7,9–12].

Genetics play a major role in LOAD. LOAD is highly polygenic, and the heritability estimates from twin studies range from 58% to 79% [13]. The apolipoprotein E (APOE) gene is the strongest known genetic risk factor for LOAD, with the APOE ε4 allele conferring increased risk and the APOE ε2 allele conferring a protective effect relative to the APOE ε3 allele [14]. A meta-analysis of genome-wide association studies (GWAS), which included more than 94,000 individuals with European genetic ancestry, confirmed 20 previously reported risk loci and discovered five novel, susceptibility single-nucleotide polymorphisms (SNPs) [15]. However, except for APOE, most of the discovered genetic variants only exhibit tiny effects on the risk of LOAD, and therefore the prediction from any single genetic variant is limited. Polygenic risk scores (PRSs), on the other hand, sum the effects of multiple independent SNPs and convert the overall genetic burden to a single score. This score has been found to serve as a good predictor of disease risk [16–18]. Although an overall PRS that combines genetic variants across the genome is more commonly used and may be more powerful in the prediction of the overall cognitive status or LOAD risk, a pathway-specific polygenic risk score (p-PRS) that sums individual SNPs under a specific neurobiological pathway may be more appropriate in predicting a specific biomarker or cognitive component (such as the amyloid-β 42/40 ratio, phosphorylated tau, or executive function) underlying LOAD etiology [19,20].

To date, a constellation of studies has been published to examine the prediction performance of p-PRS on LOAD disease risk, cognitive decline, and biomarker variation among people with or without LOAD; however, the study findings are mixed. Previous research from our group examined the prediction performance of p-PRSs under three pathways on cognition, Aβ burden in the brain as measured with Pittsburgh compound B Positron Emission Tomography (PiB-PET), and cerebrospinal fluid (CSF) Aβ and tau using a prospective cohort of 1,200 asymptomatic individuals [19]. We found that p-PRSs under all three pathways were not predictive of the global or domain-specific cognitive scores, whereas p-PRSs in the Aβ and cholesterol pathways were good predictors of variations of PiB amyloid accumulation and CSF Aβ and tau. However, the predictive performance was significantly reduced with the exclusion of the APOE variants. Another study investigated the effect of p-PRSs under seven pathways on cortical thickness using a longitudinal population cohort of 544 individuals [21]. Significant associations with cortical thickness were discovered in the AβPP metabolism, cholesterol metabolism, and endocytosis pathways when APOE was included; however, only the AβPP metabolism pathway remained significant after adjustment for the APOE variants. A recent study estimated the risk of LOAD among 1,779 Dutch individuals using p-PRSs in five major pathways involved in LOAD [22]. They found that all p-PRSs except for angiogenesis were significantly associated with increased risk of LOAD, regardless of adjustment for the APOE variants. Several reasons may exist for the discrepant results among the existing LOAD-related p-PRS analyses, but it is likely because different LOAD outcomes are used and the methods for pathway-gene-variant mapping did not draw from a comprehensive body of literature. In addition, age is the strongest factor associated with variation in the endophenotypes and cognitive decline, but it was not considered as more than a covariate in the existing literature when assessing the predictive performance of p-PRSs on cognition and LOAD-related biomarkers. A recent study leveraging a 25-year longitudinal cohort of non-demented individuals showed that the overall LOAD genetic risk on cognitive decline is age-related during the life course [23].

In the present study, we updated findings from Darst et al. (2017) [19] with five additional years of follow up data from an ongoing longitudinal cohort of cognitively healthy adults enriched for a parental history of AD from the Wisconsin Registry for Alzheimer’s Prevention (WRAP) to explore the potential of p-PRSs in the prediction of cognitive decline and changes in LOAD-related biomarkers over time. Specifically, after a comprehensive review of the existing literature on the LOAD disease pathways and functions of the genes identified by the recent case-control GWAS meta-analysis, we constructed weighted p-PRSs for AβPP metabolism, cholesterol metabolism, endocytosis, tau pathology, immune response, and axonal development. For each p-PRS, we tested its association with a global cognitive composite score (Preclinical Alzheimer Cognitive Composite – 3 (PACC-3)), domain-specific cognitive composite scores (Immediate Learning, Delayed Recall, and Executive Function), and biomarkers of Aβ accumulation (CSF Aβ42 and CSF Aβ42/40 ratio), neurodegeneration (CSF total tau [T-tau]), and tau pathology (CSF phosphorylated tau [P-tau]), while taking heterogeneity in genetic risk by age into account. To check the robustness of the results, we further performed a replication analysis using an independent sample of cognitively healthy individuals from the Wisconsin Alzheimer’s Disease Research Center (ADRC).

Methods

Study participants

Data leveraged in this study originated from WRAP, an ongoing longitudinal prospective cohort study of middle-aged adults who were cognitively healthy at enrollment and spoke English (N > 1,500). WRAP is enriched for participants with a parental history of clinical AD, increasing the proportion of individuals who will experience AD pathology and cognitive decline during the course of the study. The details of the study design have been described elsewhere [24]. The WRAP study began recruiting participants in 2001 with an initial follow-up after four years and subsequent follow-up every two years. In general, the participants were between 40 and 65 years old at baseline. Siblings of WRAP participants were allowed to enroll. Participants were given an extensive battery of neuropsychological tests at each visit. The maximum number of WRAP visits available at the time of analysis, using freeze20May, was seven. In the present study, the sample was limited to self-reported non-Hispanic white (NHW) participants to match the race and ethnicity of the participants in the GWAS meta-analysis from which the weights for the PRS were drawn. We excluded data from the baseline WRAP visit because the cognitive outcome examined in this study cannot be computed using the neuropsychological tests administered in the first WRAP visit. Data from the seventh visit of WRAP were excluded because data collection in the seventh visit is ongoing and data from this visit were only available for less than 50 participants. Compared to the previous p-PRS study on the LOAD-related outcome using WRAP [19], the present study includes additional data from approximately two more WRAP visits per participant. This study was conducted with the approval of the University of Wisconsin Institutional Review Board, and all subjects provided signed informed consent before participation.

Neuropsychometric assessments

Participants were given a battery of neuropsychological tests for the assessment of cognitive function at each WRAP visit. In the present study, we measured the overall cognitive performance using the PACC-3 score based on work by Donohue and colleagues [25]. Specifically, this WRAP composite score is computed by standardizing and averaging performance from three tests that assess the memory and executive function of participants: Rey Auditory Verbal Learning (RAVLT; Trials 1–5), Wechsler Memory Scale-Revised (WMS-R) Logical Memory II total score (LMII), and Wechsler Adult Intelligence Scale-Revised (WAIS-R) Digit Symbol Coding total items completed in 90 seconds [26]. In addition to the overall cognitive performance, we also examined domain-specific cognitive performance for immediate learning, delayed recall, and executive function [27]. The immediate learning domain-specific composite score was derived from RAVLT total trials 1–5, WMS-R Logical Memory I total score (LMI) , and Brief Visuospatial Memory Test-Revised (BVMT-R) immediate recall score. A delayed recall domain-specific composite score is constructed based on the sum of the RAVLT long-delay free recall score, WMSR logical memory delayed recall score, and BVMT-R delayed recall score. The executive function domain-specific composite score is obtained based on Trail-Making Test part B (TMT-B) total time to completion, Stroop Neuropsychological Screening Test color-word interference total items completed in 120 seconds, and WAIS-R Digit Symbol Coding. All three domain-specific cognitive composite scores are calculated by averaging standardized test scores as previously described [27]. The z-score for TMT-B was multiplied by −1 before the inclusion into the composite so that higher z-scores indicate better performance for all tests.

CSF collections, quantification, and analysis

A subset of WRAP participants consented to a lumbar puncture to obtain CSF. Details and methods for the WRAP CSF processing have been described elsewhere [28]. In brief, 22 mL of CSF were collected through gentle extraction and combined into a 30 mL polypropylene tube. All CSF samples were processed in one batch at the Clinical Neurochemistry Laboratory at the Sahlgrenska Academy of the University of Gothenburg in Sweden using the Roche NeuroToolKit robust prototype assays (Roche Diagnostics International Ltd, Rotkreuz, Switzerland) under strict quality control procedures as previously described [29]. CSF measurements examined in the present study include Aβ42, Aβ42/40 ratio, T-tau, and P-tau. Previous studies have indicated that CSF Aβ42 levels are negatively associated with amyloid burden; however, higher levels of CSF T-tau and Ptau are signals of increased tau pathology [30]. We also examined the CSF Aβ42/40 ratio because it has greater predictive and diagnostic power in early diagnosis of LOAD compared to the individual biomarker CSF Aβ42 alone [31].

DNA collection, genotyping, and quality control

Details about genomic data collection have been described elsewhere [19,32]. Briefly, we used the PUREGENE DNA Isolation Kit to extract DNA from whole blood samples, and then we used UV spectrophotometry to quantify DNA concentrations. Of the 23 SNPs included in the analysis, 21 were genotyped in 1,448 individuals using competitive allele-specific PCR-based KASP™ genotyping assays (LGC Genomics, Beverly, MA). Duplicate quality control (QC) samples had 99.9% genotype concordance, and all discordant genotypes were set to missing. The QC was carried out using PLINK v1.07 [33]. Individuals with high missingness of alleles (>10%) were removed. A total of 1,415 individuals remained in the sample after QC procedures. All 21 SNPs had call rates >95% and were in Hardy-Weinberg equilibrium (HWE).

Two SNPs (rs12459419 from CD33 and rs593742 from ADAM10) that were not genotyped by the KASP™ assays were extracted from genome-wide genotyping performed using the Illumina Infinium Expanded Multi-Ethnic Genotyping Array (MEGA^EX) at the University of Wisconsin Biotechnology Center. Standard quality control procedures for genome-wide data were performed and have been described previously [32]. Genotypic data from individuals of European genetic ancestry were then imputed using the Michigan Imputation Server and the Haplotype Reference Consortium (HRC) reference panel. SNPs rs12459419 and rs593742 were imputed with a high quality (R² > 0.8). A total of 1,198 individuals with data for all 23 SNPs remained after QC.

Mapping variants to pathways

Information needed to construct a pathway-specific PRS includes knowledge about the relationship between genetic variants and genes and the relationship between genes and pathways [22]. The current literature mainly addresses the challenge of assigning the variant-gene relationship by assuming a 1:1 variant-gene relationship with the closest gene for variant mapping. For the gene-pathway mapping, most previous studies of p-PRS in LOAD leveraged a traditional approach by referring to the gene-pathway relationships identified in a single bioinformatic database, single literature, or a union set of the gene-pathway relationships identified from the combination of bioinformatic database and limited literature [19,21,34]. However, the genetic functions of a specific gene have yet to be consistently defined across the literature, and the functional annotation of genes differs across various databases [22]. These uncertainties make the gene-pathway assignment under the traditional approach less accurate, incomparable, and might not comprehensively reflect the underlying biological mechanism it intends to represent. A recent study proposed a novel approach to constructing p-PRSs by including a weight representing the degree of involvement of a given gene in the preselected pathways in calculating p-PRSs to allow for uncertainty in the gene-pathway assignment [22]. However, the idea of assigning a weight to represent the probability that a gene belongs to a particular pathway based on the number of times each gene was associated with each pathway according to the sources referred to (bioinformatic databases and literature) still creates uncertainties in the accuracy and validity of the resulting pPRS (underweight or overweight of a particular variant). For example, if a gene fully contributes to a pathway biologically but was not widely explored by the sources referred to, it may be underestimated (e.g., ABCA7 received a weight of 0.11 in the immune response pathway, but most literature indicates ABCA7 is an important gene for the immune response pathway). The validity of their method highly depends on the number of sources used and the quality of information from these sources.

To address the limitations in the traditional pathway-gene-variant mapping method in p-PRS studies that did not consider uncertainties in gene-pathway mapping and relieve concerns about the validity of the novel approach in the p-PRS construction proposed recently, we combined the merits of these two approaches and designed a new but conservative strategy to map genetic variants to various LOAD pathways using the following steps (Supplementary methods).

Step 1 (selection of pathways), we comprehensively browsed pathways explored in the past LOAD-related p-PRSs studies published in peer-reviewed journals between 2017 and early 2020 to determine pathways that had been widely explored [19,21,22]. From the previous p-PRS studies, we included the six pathways that were included in at least one of the three papers: AβPP metabolism, cholesterol metabolism, endocytosis, tau pathology, immune response, and axonal development.

Step 2 (variant-gene mapping), we selected the closest gene to all variants that were genome-wide significant, as identified by the most recent and largest International Genomics of Alzheimer’s Project (IGAP) GWAS meta-analysis on diagnosed AD [15]. In addition, we included genomewide significant variants from three closest genes (MEF2C, NME8, and CD33) identified in previous GWAS meta-analyses [35–37]. These were widely mentioned in previous AD review papers and were marginally significant in Kunkle et al. (2019). In the variant-gene assignment, we assumed a 1:1 variant-gene relationship and included the closest gene for variant mapping.

Step 3 (gene-pathway mapping), we extensively browsed recent review papers (which summarized gene-pathway relationships based on literature, bioinformatic database, etc.) on LOAD pathology published between 2017 and early 2020, with the number of citations set to higher than 5 [3–9]. Then we counted the number of times the genes identified in step 2 presented in any of the specific pathways in the papers we reviewed. A specific gene was finally counted toward one of the pathways identified in step 1 only if more than 50% of the reviewed literature showed evidence that this gene belongs to that particular LOAD pathway.

We finally included 22 genes in the main analysis under six pathways: AβPP metabolism (CLU, SORL1, ABCA7, PICALM, ADAM10, APOE), cholesterol metabolism (CLU, SORL1, ABCA7, APOE), endocytosis (SORL1, ABCA7, PICALM, BIN1, CD2AP, PTK2B, FERMT2, SLC24A4, APOE), tau pathology (BIN1, FERMT2, CASS4, APOE), immune response (CLU, ABCA7, CR1, INPP5D,HLA-DRB1, TREM2, EPHA1, MS4A6A, CD33, MEF2C), and axonal development (EPHA1, FERMT2, CASS4, SPI1, NME8) (Supplemental figure 1). Our method considers the uncertainties in the gene-pathway mapping by restricting the gene-pathway assignment to those where more than 50% of the peer-reviewed literature shows evidence that the gene belongs to a particular LOAD pathway. Therefore, we have more confidence that the gene-pathway assignment is biologically meaningful. We also address the uncertainties in the weight because we only consider the gene-pathway relationship for which we have reasonable confidence of an “actually existing” biological relationship on an absolute scale. We also discard those gene-pathway relationships not commonly discussed and explored by the recent literature. Even though our method is a bit conservative (we may discard some false-negative gene-pathway relationships), we believe it can more accurately construct a p-PRS that reflects the underlying biological mechanism it intends to represent (more true-positive gene-pathway relationships).

Polygenic risk score and pathway-specific polygenic risk scores

The APOE risk score has many advantages over more traditional APOE coding and was calculated according to the odds ratios (ORs) of ε2, ε3, and ε4 genotypes based on rs7412 and rs429358 in the meta-analysis of APOE genotype frequencies from AlzGene [38–40]. Specifically, we constructed an APOE score using the ε2/ε2 genotype as the reference (ε2/ε2, OR=1): OR(ε2/ε3) = 1.38, OR(ε3/ε3) = 2, OR(ε2/ε4) = 4.45, OR(ε3/ε4) = 6.78, OR(ε4/ε4) = 25.84 [19]. Then, we log-transformed and added the score to the corresponding PRS/p-PRS. For genes other than APOE, the most significant variant from each of the 21 genes identified by the IGAP GWAS meta-analysis was used in the construction of PRS and p-PRS. PRS and p-PRS were calculated using the formula $PR{S}_i=\frac{{\displaystyle {\sum }_{n=l}^k\ln \left(O{R}_n\right)*C_n}}{M}$, where i represents the ith individual whose PRS is calculated by summing all SNPs n in the pathway from the first SNP l to last SNP k; OR is the odds ratio of the risk allele for SNP n from the IGAP GWAS meta-analysis; C is the number of risk alleles for SNP n for individual i; and M represents the number of non-missing SNPs under each predetermined pathway observed in individual i. In addition to the p-PRS, an overall PRS by including all 22 genes was constructed to examine the overall genetic effect by summing SNPs in all pathways of LOAD being investigated on the outcome of interest. A higher PRS/p-PRS indicates a higher genetic risk for LOAD. Since the effect size of APOE alone is known to be large, we excluded APOE for the pathways that theoretically should include APOE to examine the p-PRS on the outcome beyond APOE alone. We also tested the association between the APOE score and the outcome of interest to quantify the effect of APOE alone. To facilitate comparison across various pathways, all PRS, p-PRSs, and APOE scores were standardized with a mean of 0 and a standard deviation of 1.

Statistical analysis

We developed a set of linear mixed-effects models to examine the genetic association with cognitive outcomes and LOAD-related biomarkers while accounting for within-family (sibship) and within-individual correlations and missing data. All regression analyses related to linear mixed effects models were performed using the MIXED procedure implemented in Statistical Analysis Software (SAS) 9.4. Following the previous literature, we included random intercepts for family and study subjects [19,23]. WRAP investigators have reported the nonlinear effect of age on cognitive decline, and we therefore included a linear age, quadratic age, and cubic age in the model with cognitive outcomes to achieve better model fit [24]. For the biomarker analysis, we first used spaghetti plots to check the individual trajectory in the change of biomarkers by age and then determined the appropriate functional form of age based on the visualization of individual trajectories. To better model the dynamic relationship between aging, genetic risk, and LOAD-related outcomes, we further included an interaction term between PRS/p-PRS and all age terms to control for the potential age-related genetic risk on all outcomes of interest. In addition to the PRS/p-PRSs, age, and interaction between age and PRS/p-PRSs mentioned above, additional covariates include sex, education, practice effect (only adjusted in cognitive-related outcomes and quantified by the number of tests completed prior to the current test), and the first five genetic principal components [41]. We assessed the variance explained by each PRS/p-PRS and each interaction term between age and PRS/p-PRS using the incremental likelihood ratio $r^2\left(r_{LR}^2\right)$ [42,43]. The significance of the interactions between each PRS/p-PRS and all age terms is evaluated by the likelihood ratio test (LRT). Upon discovering significant interactions, we further probe the nature of the interactions by investigating the conditional effect of each PRS/p-PRS at different age values and test them versus zero for significance using the simple slope approach. All simple slope estimates were calculated using the PLM procedure implemented in SAS 9.4.

Replication analysis

We replicated all analyses performed in the WRAP sample within the Wisconsin ADRC, which began enrolling participants in 2009. Because the Wisconsin ADRC administered a different battery of neuropsychological tests compared to WRAP, we could only replicate our analyses for the overall cognitive performance (PACC-3) and CSF-related outcomes. We replicated our findings using two samples from the Wisconsin ADRC. The first sample is the IMPACT cohort, for which the enrollment criteria (age range of 45–65 at baseline, cognitively unimpaired, and enriched for a parental history of AD) are the most similar and comparable to the WRAP sample. To broaden the age range represented, we supplemented the IMPACT sample with the Wisconsin ADRC healthy older controls (HOCs), which includes people who are older than 65 at enrollment and do not meet the National Institute on Aging and Alzheimer’s Association criteria for mild cognitive impairment (MCI) or the National Institute of Neurological and Communicative Disorders and Stroke and Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) criteria for probable AD. We called this combined sample the All Healthy Controls (AHC) sample. All replication analyses in the Wisconsin ADRC were restricted to people who self-identified as NHW.

The Wisconsin ADRC administered a different battery of neuropsychological tests than WRAP, which resulted in a substantial missingness in the score of Logical Memory II Delayed Recall and Wechsler Adult Intelligence Scale-Revised, Digit Symbol. To make the best use of the current information, we consulted neuropsychologists in the Wisconsin ADRC and created a PACC-3-TMT score to replace PACC-3 in the replication analysis. Specifically, we converted the Craft Story score to an estimated Logical Memory score based on a published crosswalk table and followed the previous practice of substituting the Digit Symbol score with the total time to completion in the TMT-B test [26,44]. Since the published crosswalk table is only available for converting the Craft Story score to the Logical Memory score for the first five visits, we restricted our replication analysis using only data from the first five Wisconsin ADRC visits. The final PACC-3-TMT score was computed by standardizing and averaging the results from RAVLT, the estimated Logical Memory score, and the reversed coded TMT-B test results. The substantial missingness in the score of the Logical Memory score and Digit symbol makes it difficult to assess the correlation between PACC-3 and PACC-3-TMT in the Wisconsin ADRC; therefore, we used the same method to construct a PACC-3-TMT score in the WRAP and assessed the correlation between PACC-3 and PACC-3-TMT in the WRAP sample.

The Wisconsin ADRC employed the same methods of collection, processing, and quantification for the CSF data as those in WRAP [19]. Details about genomic data collection and QC have been described elsewhere [19,45]. Briefly, the top significant SNPs from 21 genes except for ADAM10 were genotyped by LGC Genomics (Beverly, MA) using the same competitive allele-specific PCR-based KASP™ genotyping assays as in WRAP. All SNPs had call rates >95% and were in Hardy-Weinberg equilibrium (HWE). Two individuals with high missingness of alleles (>10%) were excluded from subsequent analyses. Only the AβPP metabolism pathway-specific PRS and overall PRS were affected by the exclusion of the ADAM10 gene, but we expect the impact will be small due to the small effect size (β=−0.065) of the top significant SNP from ADAM10. The methods of constructing the PRS, p-PRS, and APOE scores are the same as those in WRAP analysis.

We utilized a linear mixed-effects model to examine the genetic association with the overall cognitive performance and LOAD-related biomarkers by accounting for within-individual correlations and allowing for missing data. All analyses were performed using SAS 9.4. All other statistical methods in the replication analyses are the same as those described in the WRAP analysis, except for the exclusion of genetic principal components as covariates because genome-wide data and, thus, genetic principal components for the full sample are not available in the Wisconsin ADRC genomic dataset.

Results

Descriptive statistics for samples and participants

Table 1 presents the demographic features for WRAP participants included in this study. Briefly, a total of 1,175 individuals with available genetic, cognitive, and demographic data remained in the sample for up to five visits (~8 years) after the implementation of the inclusion criteria. A subset of 197 WRAP participants had CSF data for up to five visits. Demographic characteristics are comparable between the cognitive and CSF samples. The sample with CSF is slightly older at baseline than the full WRAP sample because WRAP CSF collection began later during the WRAP study. WRAP participants are generally highly educated, female, and enrolled at middle age, and a majority have a parental history of AD. The APOE score is not available for five participants in the full sample and one participant in the CSF sample because of missing allele information for either rs7412 and/or rs429358. Following the previous literature, we decided to keep these individuals in the analysis because data are available on other genetic variants that we are interested in, and the magnitude of missingness is small.

Table 1.

Participant characteristics for WRAP at baseline

Variable	Full sample (N=1,175)	Biomarker sample (N=197)
Age	58.57 (6.52)	61.98 (6.64)
Education (years)	15.81 (2.24)	16.16 (2.14)
Gender (male)	353 (30%)	69 (35%)
Family history of AD	858 (73%)	142 (72%)
Max visits
1	41 (4%)	70 (36%)
2	85 (7%)	42 (21%)
3	182 (15%)	60 (30%)
4	369 (31%)	23 (12%)
5	498 (42%)	2 (1%)
APOE genotypes
ε2/ε2	4 (0.3%)	0 (0)
ε2/ε3	96 (8%)	21 (11%)
ε3/ε3	619 (53%)	107 (54%)
ε2/ε4	39 (3%)	6 (3%)
ε3/ε4	369 (31%)	56 (28%)
ε4/ε4	43 (4%)	6 (3%)
CSF Aβ42 (pg/mL)	N/A	899.31 (391.24)
CSF Aβ42/40	N/A	0.06 (0.02)
CSF T-tau (pg/mL)	N/A	208.82 (69.89)
CSF P-tau (pg/mL)	N/A	18.34 (6.71)

Mean (SD); n (%); Aβ: amyloid-β; P-tau: phosphorylated Tau; T-tau: total Tau.

“Max visits” refers to number of cognitive assessments of outcomes included in these analyses (left column) or lumber puncture draws per participant (right column).

Cognitive outcomes

Statistically significant interactions (p < 0.05) between polynomial age and both the APOE risk score (Figure 1a) and all PRSs that included APOE (Supplementary figure 2a) were observed for all cognitive outcomes (regression output for first model in supplementary excel table; full regression outputs are available upon request). However, when APOE was excluded from the PRS (Figure 1a), significant p-PRSs*polynomial age interactions were only observed under the endocytosis pathway for the immediate learning composite score; under AβPP metabolism, endocytosis, and tau pathology pathways for the delayed recall composite score; under endocytosis and tau pathology pathways, and the overall PRS for the executive function composite score; and under the endocytosis pathway and the overall PRS for the PACC-3 composite score.

Figure 1. Likelihood ratio test results of the interactions between genetic predictor (APOE and PRS/p-PRSs excluding APOE) and polynomial age.

Figure 1 presents the -log₁₀(P) from the likelihood ratio tests for the interaction between polynomial age and genetic risk for all outcomes in WRAP and Wisconsin ADRC. Since the effect of PRS with APOE is primarily driven by APOE, for simplicity, we only report the LRT test statistics when we use APOE and PRS without APOE as the predictor. The likelihood ratio test statistic is calculated as the ratio between the log-likelihood of the nested model (model without interaction terms) to the full model (model with polynomial age*genetic risk terms). All association analyses are performed using the linear mixed effect model and adjusted for within-individual and within-family correlation. In addition to PRS/p-PRSs, age (linear, quadratic, and cubic), and their interactions, additional covariates include gender, education years, practice effect, and the first five genetic principal components of ancestry. The red dashed horizontal reference line represents p-value = 0.05.

Figure 2 presents the simple slope estimates of annual cognitive change by a one standard deviation change in genetic risk (APOE and PRS/p-PRSs excluding APOE) and by age in five-year increments between 55 and 80 years old for WRAP participants. Only the p-PRS for endocytosis, cholesterol metabolism, and AβPP metabolism had simple slope estimates that were significantly different from zero beginning at age 75 or 80 for at least one cognitive composite score. The simple slope estimates of APOE-related annual cognitive change were not statistically different from zero prior to age 65 for every cognitive outcome. However, the simple slope estimates accelerate in growth and become statistically significant once WRAP participants reached the age of 65. Simple slope estimates of annual cognitive change for PRS/p-PRS while including APOE are shown in Supplementary figure 3.

Figure 2. Model predicted simple slope of genetic predictor (APOE and PRS/p-PRSs excluding APOE) on cognition at different age with 95% confience interval in WRAP.

Figure 2 presents the model predicted simple slope (based on the estimates) of PRS/p-PRSs on immediate learning (2a), delayed recall (2b), executive function (2c), and PACC3 (2d) at various age points in WRAP with 95% confidence intervals. Since the effect of PRS with APOE is primarily driven by APOE, for simplicity, we only report the simple slope estimates for APOE and PRS without APOE. Each row within the figure depicts the simple slope estimates for one genetic predictor (a pathway, the overall PRS, or APOE. All association analyses are performed using the linear mixed effect model and adjusted for within-individual and within-family correlation. In addition to PRS/p-PRSs, age (linear, quadratic, and cubic), and their interactions, additional covariates include gender, education years, practice effect, and the first five genetic principal components of ancestry. Simple slope estimates were calculated based on parameters from the linear mixed effect model. The red dashed vertical line indicates the age threshold when statistically significant simple slope estimates of the genetic predictor were first observed in the lifespan.

We used a reduced set of WRAP participants who have complete data in all PRS/p-PRSs to compare the performance of each PRS/p-PRS in explaining the amount of variation in the overall and domain-specific cognitive composite scores, as measured by$r_{LR}^2$ and presented in Supplementary table 1. Consistent with Darst (2017), the largest $r_{LR}^2$ for a single PRS/p-PRS is about 0.2% when APOE is included and 0.1% when APOE is excluded, which indicates that almost none of the model variance was explained by any of the PRS/p-PRS. When the interaction between age and PRS/p-PRS was included, an additional 1% of the model variation was explained for all pathways when APOE was included, but the additional gain in explained model variation substantially decreased after APOE was excluded.

CSF biomarker outcomes

Like the cognitive outcomes, statistically significant interactions between polynomial age and both the APOE risk score (Figure 1b) and all PRSs that included APOE (Supplementary figure 2b) were observed for all Ab outcomes (full regression outputs are available upon request). However, when APOE was excluded from the PRS (Figure 1b), significant interactions were observed between age and overall PRS, as well as p-PRSs under AβPP metabolism, cholesterol metabolism, and immune response pathways for Ab42; between age and p-PRSs under AβPP metabolism and cholesterol metabolism pathways for Ab42/40. However, for P-tau and T-tau, there was no statistically significant interaction between age and APOE risk score. When APOE was excluded from the PRS, significant interactions were observed between age and overall PRS, as well as pPRSs under AβPP metabolism, cholesterol metabolism, endocytosis, and immune response pathways for P-tau and between age and overall PRS, as well as p-PRSs under AβPP metabolism, endocytosis, and the immune response pathway for T-tau.

Figure 3 presents the simple slope estimates for annual biomarker change by a one standard deviation change in genetic risk (APOE and PRS/p-PRSs excluding APOE) and by age in five-year increments between 55 and 80 years old for WRAP participants. The simple slope estimates of APOE-related annual biomarker change became statistically significant once WRAP participants reached the age of 55 for the Aβ42/40 ratio, 60 for Aβ42, 70 for P-tau, and 75 for T-tau. Only the p-PRS for AβPP metabolism, cholesterol metabolism, and immune response, and the overall PRS had simple slope estimates that were significantly different from zero beginning at age 65 for the Aβ42/40 ratio and at age 70 or 75 for Aβ42. For tau, only the p-PRS for AβPP metabolism, cholesterol metabolism, endocytosis, and immune response, and the overall PRS have simple slope estimates that were significantly different from zero, generally beginning at age 65 or 70 for P-tau and at age 65 to 75 for T-tau. Simple slope estimates of annual biomarker change for PRS/p-PRS while including APOE are shown in Supplemenatary figure 4.

Figure 3. Model predicted simple slope of genetic predictor (APOE and PRS/p-PRSs excluding APOE) on biomarker at different age with 95% confience interval in WRAP.

Figure 3 presents the model predicted simple slope (based on the estimates) of PRS/p-PRSs on amyloid-β 42 (3a), amyloid-β 42/40 ratio (3b), phosphorylated tau (3c), and total tau (3d) at various age points in WRAP with 95% confidence interval. Since the effect of PRS with APOE is primarily driven by APOE, for simplicity, we only report the simple slope estimates for APOE and PRS without APOE. Within each figure, each panel depicts the simple slope estimates of APOE and PRS/p-PRSs excluding APOE score for a specific pathway. All association analyses are performed using the linear mixed effect model and adjusted for within-individual and within-family correlation. Spaghetti plots determine the functional form of age for all biomarker analyses. In addition to PRS/p-PRSs, age, and their interactions, additional covariates include gender, education years, and the first five genetic principal components of ancestry. Simple slope estimates were calculated based on parameters from the linear mixed effect model. The red dashed vertical line indicates the age threshold when statistically significant simple slope estimates of the genetic predictor were first observed in the lifespan.

We used a reduced set of WRAP participants who have complete data in all PRS/p-PRSs to compare the performance of each PRS/p-PRS in explaining the amount of variation in the LOAD-related biomarkers, as measured by $r_{LR}^2$ and presented in supplementary table 2. When APOE is included, single PRS/p-PRS can explain on average a 3~4% variance in Aβ42, 7–8% variance in Aβ42/40, and 1% variance in T-tau and P-tau. Adding an interaction between PRS/p-PRS and age resulted in an additional 3~4%, 1~3%, 1–2% gain in the variance explained for Aβ42, Aβ42/40, T-tau and P-tau, respectively. For amyloid-β outcomes, the gain in variance explained by addition of the PRS/p-PRS and the additional gain in model variation explained by the interaction between age and PRS/p-PRSs was substantially decreased after APOE was excluded. However, the removal of APOE from PRS/p-PRS does not substantially affect the additional variance explained by the PRS and interaction term for the tau outcomes.

Replication analysis

We used the AHC combined sample from the Wisconsin ADRC to replicate our findings in WRAP. Table 2 details the Wisconsin ADRC participant characteristics for the cognition analysis. The mean baseline age for the AHC cohort is about 60. The mean education is just over 16 years. About 35% of participants in the AHC cohort are males, and 70% have a family history of AD. The basic characteristics are similar between the WRAP and AHC cohorts. Similar characteristics were observed in the biomarker samples.

Table 2.

Participant characteristics for the Wisconsin ADRC

	Wisconsin ADRC AHC
Variable	PACC3-TMT (N=427)	Biomarker Biomarker (N=259)
Baseline age	59.95 (8.26)	60.53 (8.08)
Education (years)	16.35 (2.46)	16.20 (2.39)
Gender (male)	149 (35%)	80 (31%)
Family history of AD	299 (70 %)	192 (74 %)
Max visits
1	31 (7%)	203 (78%)
2	56 (13%)	34 (13%)
3	60 (14%)	6 (2%)
4	34 (8%)	11 (4%)
5	246 (58%)	3 (1%)
6		1 (0.4%)
7		1 (0.4%)
APOE genotype
e2/e2	1 (0.2%)	1 (0.4%)
e2/e3	49 (11%)	30 (12%)
e3/e3	219 (51%)	132 (51%)
e2/e4	14 (3%)	7 (3%)
e3/e4	124 (29%)	76 (29%)
e4/e4	20 (5%)	13 (5%)
CSF Aβ42 (pg/mL)	N/A	957.40 (381.79)
CSF Aβ42/40	N/A	0.07 (0.01)
CSF T-tau (pg/mL)	N/A	194.47 (73.58)
CSF P-tau (pg/mL)	N/A	17.00 (6.91)

Mean (SD); n (%); AHC: all healthy controls; Aβ: amyloid-β; P-tau: phosphorylated Tau; T-tau: total Tau.

For the cognition analysis, the correlation between PACC-3 and PACC-3-TMT is about 0.93 in the WRAP sample. Surprisingly, in the Wisconsin ADRC, there was no significant interaction between polynomial age and the APOE risk score for PACC-3-TMT (Figure 1c). Significant interactions were observed between polynomial age and p-PRS under the AβPP metabolism and cholesterol metabolism pathways (Figure 1c), with simple slope estimates significantly different from zero by age 70 for both of these p-PRSs (Supplementary figure 5). Likelihood ratio test results of the interactions between polynomial age and PRS/p-PRS while including APOE are shown in Supplementary figure 2c. Model predicted simple slopes of PRS/p-PRSs on PACC-3-TMT at different ages while including APOE are shown in Supplementary figure 6.

A reduced set of Wisconsin ADRC participants who have complete data in all PRS/p-PRSs were used to compare the performance of each PRS/p-PRS (supplementary table 3). Similar to the WRAP findings, the largest $r_{LR}^2$ for a single PRS/p-PRS is about 0.2% in the AHC sample. The largest $r_{LR}^2$ for PRS/p-PRSs with and without APOE increased by 0.4% and 0.7% with the addition of the interaction between PRS/p-PRSs and age, respectively.

Significant interactions between age and genetic risk were also observed in the biomarker analysis (Figure 1d; Supplementary figure 2d). Specifically, significant interactions between age and the APOE risk score were observed for the Aβ42/40 ratio and P-tau (Figure 1d). Significant interactions were also observed between age and p-PRS under the cholesterol metabolism pathway for Aβ42/40 and between age and overall PRS, as well as p-PRSs under endocytosis and tau pathology pathways for both P-tau and T-tau. No significant interactions were observed for Aβ42, although suggestive evidence of interactions were observed between age and p-PRSs under AβPP metabolism (p=0.07) and cholesterol metabolism (p=0.06) pathways.

Supplementary figure 7 presents the simple slope estimates for annual biomarker change by a one standard deviation change in genetic risk (APOE and PRS/p-PRSs excluding APOE) and by age in five-year increments between 55 and 85 years old for Wisconsin ADRC participants. The simple slope estimates of APOE-related annual biomarker change were statistically significant by age 55 for Aβ42 and the Aβ42/40 ratio, 60 for P-tau, and 65 for T-tau, and the absolute effect size increases with age. Only the p-PRS for AβPP metabolism and cholesterol metabolism had simple slope estimates that were significantly different from zero beginning at age 60 to 65 for Aβ42 and the Aβ42/40 ratio. For tau, only the p-PRS for endocytosis, tau pathology, and axonal development, and the overall PRS have simple slope estimates that were significantly different from zero, beginning at age 70 to 80 for P-tau and T-tau. Simple slope estimates of annual biomarker change for PRS/p-PRS while including APOE are shown in Supplemenatary figure 8.

A reduced set of Wisconsin ADRC participants with complete data in all PRS/p-PRSs was used to compare the performance of each PRS/p-PRS (Supplementary table 3). When APOE is included, the variance explained by a single PRS is about 3% for Aβ42, 7% for Aβ42/40, 1% for P-tau, and less than 1% for T-tau. The additional interaction between age and p-PRSs contributes to the added variance explained in Aβ42, Aβ42/40, T-tau, and P-tau by about 1%, 2%, 2%, and 1.5%, respectively. When APOE was excluded, p-PRSs and age–PRS interaction under all pathways contributed substantially less variance, as shown using Aβ42 and Aβ42/40 ratios, which is consistent with the WRAP findings. For tau-related outcomes, when APOE is excluded, the performance of PRS/p-PRSs and the interaction between PRS/p-PRSs and age under most pathways (except for endocytosis) deteriorated.

Discussion

In the present study, we updated findings from Darst et al. and investigated the potential of pathway-specific PRSs in predicting rate of change in cognitive function and biomarkers of amyloid-β deposition, tau pathology, and neurodegeneration among asymptomatic individuals in WRAP [19]. With five additional years of data collection, GWAS summary statistics with a larger sample size, our comprehensive variant-pathway mapping method, and the inclusion of an ageinteraction effect, we found p-PRSs and the overall PRS can predict preclinical changes in cognition and biomarkers beyond the effect of APOE and the effects of APOE and the PRS/p-PRSs on rate of change in cognition, amyloid-β, and tau outcomes vary by age.

Although genetics play a large role in the development and expression of LOAD, the complex relationships in the etiology between age, APOE, and non-APOE PRS/p-PRSs are not generally considered [46]. Our study shows the genetic risk from APOE, PRS, and p-PRSs under certain pathways on rate of change in cognition, amyloid-β, and tau are age related in both WRAP and the Wisconsin ADRC. Including an interaction term between genetic risk and age in the model could also lead to the model explaining additional variance. The APOE effect trajectory follows the approximate timeline of LOAD pathology, with an effect of the APOE risk score on amyloid pathology becoming detectable earlier, around by age 55, followed by an effect on tau pathology around age 60 to 65, and an effect on cognition around age 65. A similar trend was found for pPRSs excluding APOE, for example, AβPP metabolism and cholesterol metabolism, but the timeline was generally shifted five to ten years later, with significant effects of the p-PRS on AD biomarkers beginning around age 60 to 65 and on cognition beginning around age 70 to 80. Our findings are consistent with a recent study that leveraged Alzheimer’s Disease Neuroimaging Initiative (ADNI and UK Biobank (UKBB) samples and concluded both APOE and PRS predicted AD risk and presented age-related effects, but the effects of APOE were stronger in younger groups (age <80) [47]. Zimmerman et al. examined the age-related genetic effect of APOE and PRS in UKBB. Similar to our findings, they reported that an AD PRS modified the association between age and cognition, that APOE ε4 allele carriers experienced earlier cognitive decline than noncarriers did, and that models using the PRS that excluded APOE ε4 had attenuated and later modification of age associations compared to when APOE was included in the PRS [48]. These patterns explain the reason for observing significant associations between PRS/p-PRSs and amyloid-β outcomes in Darst et al., but not for tau and cognition outcomes, since the effects of PRS/p-PRSs on amyloid-β accumulation can be detected at a younger age than the effects on tau and cognition, and the sample analyzed in Darst et al. was too young to detect polygenic effects on tau and cognition.

Results are mixed for current p-PRS studies on LOAD risk, cognitive decline, and biomarker variation. This is partially due to the discrepancy in the sample characteristics across different studies and outcomes being investigated, but also due to the methodology of attributing specific genetic variants to their corresponding biological pathway. Most existing studies on p-PRSs have been based entirely on limited or even single-review papers and bioinformatic databases to map a specific genetic variant to a pathway. However, the genetic functions of a specific variant have not been consistently defined across the literature, and the functional annotation of genes might differ across various databases [22]. Our study utilized a conservative but comprehensive variant-pathway mapping method via only matching variants and corresponding pathways for which we have good confidence from the literature of an “actually existing” biological relationship.

Pathway-specific PRSs often predicted earlier changes in AD-related outcomes than the overall PRS could, especially for biomarker outcomes. Our results demonstrate, when APOE is excluded, p-PRSs for AβPP and cholesterol pathway can predict changes in Aβ42 about 15 years earlier than the overall PRS, and this finding was replicated in the Wisconsin ADRC. Even though p-PRSs for AβPP and cholesterol metabolism pathways also show potential for predicting earlier changes in Aβ42/40 ratio compared to the overall PRS, only the finding for the cholesterol metabolism pathway was replicated in the Wisconsin ADRC. Our results also show p-PRSs under certain pathways can predict adverse change in tau and cognition outcomes earlier than the overall PRS in WRAP can, but these findings were not fully replicated in the Wisconsin ADRC, which may warrant further investigation as longitudinal data focusing on the preclinical stage of AD with a larger sample size become available.

One finding – that the tau pathology PRS is not predictive of tau outcomes after the exclusion of APOE in WRAP, but is predictive of both P-tau and T-tau in the Wisconsin ADRC - may require further investigation once we have a larger sample size and longer follow-up time. There are two possible explanations for the discrepancies in the effect of the tau pathology PRS between WRAP and the Wisconsin ADRC. First, the genes known to be related to the tau pathology pathway may not be well established and may not fully reflect the biological pathway of tau since only three SNPs in addition to APOE were included in the tau pathology pathway and all these SNPs overlapped with the SNPs included in the other disease-related pathways (Supplemental figure 1). Second, the AHC sample that was extracted from the Wisconsin ADRC includes a sample of older healthy controls (enrollment age >= 65) and tau-levels are higher in the older age groups [49].

The present study has limitations. First, although we tend to match genetic variants and biological pathways for which we have good confidence of a biological variant-pathway relationship, our variant-pathway mapping method is conservative and may not comprehensively reflect the genetic role in a specific disease pathway. Our results’ accuracy is subject to the knowledge of biological function of genes and pathways at the time of performing this study. It would be crucial to modify the variant-pathway mapping as additional knowledge becomes available. Second, we only considered the significant variants as identified from the most recent IGAP case-control GWAS meta-analysis as the weight to construct pathway-specific PRSs; however, a larger panel of SNPs from the recent GWAS by proxy (GWAX) and a combined study of GWAS and GWAX may provide additional insights into the variant-pathway mapping [50,51]. This was not considered in the current study. Third, results from the AHC sample extracted from the Wisconsin ADRC are not absolutely comparable with the WRAP findings because the AHC sample was constructed based on two separate cohorts with different characteristics (e.g., age) Additional replication analyses may be carried out once the data for an older IMPACT cohort (more like WRAP) become available. Fourth, our conservative strategy restricts the source of information for gene-pathway mapping to recently published LOAD-pathology-related papers with moderate to high citations but doesn’t directly consider information sources from available bioinformatic databases (e.g., Gene-ontology). Future p-PRS-related work should explore the possibility of an unbiased way for variant-gene-pathway mapping by integrating information from the literature and various bioinformatics databases and, at the same time, take “uncertainties” into consideration. Fifth, the sample size was relatively small, especially for the biomarkers, which may have prevented us from detecting a true association with a PRS/p-PRS that has a small effect. A longitudinal cohort with a larger sample size is needed to validate the findings reported by the current study. Finally, we did not correct for multiple testing in the current study, but instead chose to include a replication cohort to protect against false positive findings.

In conclusion, in addition to APOE, the pathway-specific PRSs can predict age dependent changes in amyloid-β, tau, and cognition. Once validated, they could be used to identify individuals with an elevated genetic risk of accumulating amyloid-β and tau, long before the onset of clinical symptoms. This information could be useful for selection of high risk participants for clinical trials and, as effective therapeutic targets further develop, p-PRSs could be used to determine an individual’s risk for accumulating amyloid and the predicted age of onset so that resources could be used effectively in screening individuals for amyloid accumulation with more expensive and invasive, but accurate tests. This idea is being explored in other diseases, such as breast cancer [52].

Data Availability

The data supporting the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.