Association Between Polygenic Risk Score and the Progression from Mild Cognitive Impairment to Alzheimer’s Disease

Abstract

Background:

Mild cognitive impairment (MCI) is a heterogeneous condition and MCI patients are at increased risk of progression to dementia due to Alzheimer’s disease (AD).

Objective:

In this study, we aim to evaluate the associations between polygenic risk scores (PRSs) and 1) time to AD progression from MCI, 2) changes in longitudinal cognitive impairment, and 3) biomarkers from cerebrospinal fluid and imaging.

Methods:

We constructed PRS by using 40 independent non-APOE SNPs from well-replicated AD GWASs and tested its association with the progression time from MCI to AD by using 767 MCI patients from the ADNI study and 1373 patients from the NACC study. PRSs calculated with other methods were also computed.

Results:

We found that the PRS constructed with SNPs that reached genome-wide significance predicted the progression from MCI to AD (beta = 0.182, SE = 0.061, p = 0.003) after adjusting for the demographic and clinical variables. This association was replicated in the NACC dataset (beta = 0.094, SE = 0.037, p = 0.009). Further analyses revealed that PRS was associated with the increased ADAS-Cog11/ADAS-Cog13/ADASQ4 scores, tau/ptau levels, and cortical amyloid burdens (PiB-PET and AV45-PET), but decreased hippocampus and entorhinal cortex volumes (p < 0.05). Mediation analysis showed that the effect of PRS on the increased risk of AD may be mediated by Aβ₄₂ (beta = 0.056, SE = 0.026, p = 0.036).

Conclusion:

Our findings suggest that PRS can be useful for the prediction of time to AD and other clinical changes after the diagnosis of MCI.

INTRODUCTION

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder with the main neuropathological features of neuronal loss and accumulation of intracellular tau-containing neurofibrillary tangles and extracellular amyloid-β (Aβ) deposits in the brain [1]. It is estimated there are around 5.8 million AD patients in the US and 700,000 people age 65 years and older will die from AD in 2020 [2]. Late-onset AD (LOAD) has onset after the age of 65 years and accounts for 97% of AD in the US. LOAD is sporadic and up to 60-80% of cases are inheritable. As a complex disease, the genetic etiology of AD is still not completely understood. Carrying the E4 allele of apolipoprotein E (APOE4) is a well-known genetic risk factor for AD [3, 4]. Genome-wide association studies (GWASs) have identified about 40 AD-related loci. These loci map by proximity to genes involve in lipid metabolism, inflammation, immune function, endocytosis, and other biological pathways [5, 6]. However, with the exception of the coding SNPs in the APOE gene, SNPs in other regions have minor effects on AD risk and age of onset.

Polygenic risk scores (PRS) is an approach that combines the effects of multiple SNPs based on their effect sizes and can be used for the prediction of risks of multiple diseases including AD [7, 8]. By using the summary data from the GWAS study of the International Genomics of Alzheimer’s Project (IGAP), PRS has been successfully applied to predict life-time AD risk, brain structure changes (e.g., hippocampal cortical thickness, hippocampal volume), cognitive ability changes, and biomarkers from cerebrospinal fluid and plasma [9–15]. Moreover, a Polygenic Hazard Score (PHS) was recently developed with APOE E2 and E4 alleles, and 31 SNPs selecting from the IGAP summary data after applying the step-wise Cox regression and used to quantify individual differences with age-specific genetic risk for AD [16]. Both the PRS and PHS models have comparable performance on the prediction of AD risk [11].

Mild cognitive impairment (MCI) is an intermediate state between the expected cognitive decline of normal aging and AD dementia. About one-third of MCI patients will develop AD over time [17, 18]. The early detection of MCI patients with high risk of AD may provide an opportunity to apply preventive or therapeutic interventions, which may be helpful for preventing or delaying the further conversion to AD [19]. Therefore, it is important to develop models that distinguish clinically heterogeneous MCI with different risks of AD. In this study, we constructed polygenic risk scores (PRS) by using the SNPs from four well-replicated, consortium GWASs and tested its association with the progression from mild cognitive impairment (MCI) to AD as well as cognitive ability and imaging features by using the clinical and biological data from the ADNI study.

METHODS

Participants

In the discovery stage, we used the dataset from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) study (http://adni.loni.ucla.edu). There were 2,253 participants from the ADNI 1, ADNI GO, and ADNI 2 phases with clinical and biological data available (September 2005–January 2020), which included clinical diagnosis (Normal Cognition, MCI, and AD), longitudinal cognitive phenotypes, and neuropathological phenotypes including biomarkers from cerebrospinal fluid (CSF), and the imaging results of PET and MRI. Of them, 916 participants diagnosed with MCI at baseline or during follow-up and 767 samples with GWAS data available remained for further analysis. There were 294 subjects diagnosed with AD during the follow-up period.

An independent sample of MCI patients was selected from the National Alzheimer’s Coordinating Center (NACC) to replicate the findings from ADNI. There were a total of 41,459 participants in the requested NACC dataset, of which 12,068 presented with MCI at baseline or during the following visits. Of these patients, there were 1,373 patients with at least two visits and who had GWAS data available. During the follow up visits, 864 out of the 1,373 MCI patients progressed to AD and 509 patients did not progress to AD but continued to have a diagnosis of MCI at the last visit. The flowcharts of the participant selection and further details of the two datasets have been presented in Supplementary Figure 1 and in the Supplementary Material.

The ADNI and NACC studies were approved by local institutional review boards, and all participants or participant’s guardians provided written informed consent. Additional information about ADNI and NACC studies are available at http://www.adni-info.org and https://www.alz.washington.edu/WEB/study_pop.html, respectively.

Genotyping and imputation

We downloaded the GWAS data from the ADNI database on February 20, 2020, Genotyping was performed with the platforms of Illumina Human610-Quad, Illumina Human OmniExpress, and Illumina Omni 2.5M on 1,674 MCI participants from ADNI 1, GO, and ADNI 2 phases, respectively. We then used SHAPEIT for phasing and performed imputation with minimac4 on the Michigan imputation server (https://imputationserver.sph.umich.edu) with the HRC reference panel (Version r1.1 2016) consisting of 64,940 haplotypes of predominantly European ancestry. For imputation, a set of high-quality SNPs were used: MAF > 0.01; call rate > 95%, Hardy-Weinberg equilibrium test p > 10⁻⁶; allele frequency difference ≤ 0.20 between the sample data and the reference panel. The genotyping data was processed by using PLINK 1.90/2.0.

For NACC, we requested the GWAS data for the 10,256 subjects in NACC AD Centers 1-7, of which genotyping was performed with the platforms of Human660W-Quad_v1_A, HumanOmni Express-12v1_A/H, and humanomniexpressexome-8v1-2_a, respectively (https://www.alz.washington.edu/ADGC/GENOtype.html) [20]. We then conducted quality control and imputation with the same procedures of ADNI.

PRS/PHS construction

There were 40 loci associated with AD risk reported by four well-powered AD GWASs/GWAXs (genome-wide association study by proxy) and 78 leading SNPs were found in those loci, which included two SNPs (rs41289512 and rs12691088) in the genomic region of APOE locus (19q13.32) [5, 6, 9, 21, 22]. All but one SNP rs9271058 were included in the imputation data. We performed clump analysis with PLINK 1.90 for these 77 SNPs based on the ADNI GWAS dataset and the reported p values from the previous GWASs. 41 independent SNPs, including one APOE SNP rs41289512, were selected based on an r-squared threshold greater than 0.1. We did not include rs41289512 in the PRS calculation since this SNP is located in the APOE region and had substantially greater effect size than other SNPs. The other APOE SNP rs12691088 was not included in the PRS calculation as it has moderate linkage disequilibrium (r² = 0.53) with rs41289512. To account for the effect of the APOE gene, we included the APOE E2 and E4 allele counts as covariates in the final regression model instead of including them in PRS calculation. The remaining 40 SNPs were used to construct PRS, weighted by the effect size estimates from the GWAS with the largest sample size for association with LOAD. Genomic data on the PRS SNPs is included in Supplementary Table 1. The PRS was calculated using PRSICE-2 and standardized to z-scores (centering by mean and scaling by standard deviation) [23]. For comparison, we also constructed PRSs by using the results of 4,465 SNPs with p ≤ 1 × 10⁻⁵ in a recent AD GWAS Meta study (including 71,880 cases and 383,378 controls) [21]. SNPs within the 250kb region surrounding APOE (chr19:45159053-45662650; GRCh37/hg19 assembly) were excluded and five different thresholds (i.e., p ≤ 1 × 10⁻⁵, 5 × 10⁻⁶, 1 × 10⁻⁶, 5 × 10⁻⁷, and 1 × 10⁻⁷) were applied for SNPs selection for the PRS construction. The summary results of this GWAS were downloaded from the website of the CTG laboratory (https://ctg.cncr.nl/documents/p1651/AD_sumstats_Jansenetal_2019sept.txt.gz). Desikan et al. [16] reported that PHS combining the effects of APOE E2 and E4 alleles, as well as 31 GWAS SNPs was associated with the age of AD onset. However, for comparison with PRS, we constructed PHS by only using those 31 reported SNPs, without including the APOE E2 and E4 alleles. After linkage-disequilibrium (LD) based clumping, 25 independent SNPs remained (r² < 0.1) in both the ADNI and NACC datasets and their reported effect sizes [log(HR)] in the same publication were used as the weights for PHS construction. Information of the 25 SNPs can be found in Supplementary Table 2 [16]. For method comparisons, we also constructed a Bayesian PRS by using the method of LDpred2 [24, 25]. Briefly, this method estimates posterior mean causal effect sizes for each SNP with a Gibbs sampler by assuming a prior for the genetic architecture and LD information from a reference panel. PRS was then constructed using the posterior effect sizes.

Statistical analysis

The association between demographic and clinical characteristics, PRS, and the progression from MCI to AD was tested using Cox proportional hazards regression with adjustment for age at baseline, sex, years of education, race, top three principal components (PCs) from the genetic data, and the allele copies of APOE E2 and E4. The endpoint event was the occurrence of AD. Survival time (in years) was defined as the duration from baseline for patients with MCI or the date of first diagnosis of MCI to AD occurrence or censoring. The HR from the Cox model analysis indicated the risk of progressing to AD caused by the per 1 SD change in PRS.

We performed log-transformation for the CSF biomarkers (Aβ, tau, and ptau). The volumes of five brain regions [i.e., Ventricles, Hippocampus, Entorhinal, Fusiform gyrus, and Middle temporal gyrus (MidTemp)] were represented as the percentage to the intracerebral volume (ICV). The correlations between PRS and these longitudinal phenotypes were investigated with a linear mixed model by including a random intercept and a random slope of follow-up time with adjustment for age at baseline, sex, years of education, race, top three principal components (PCs), and the allele copies of APOE E2 and E4. The survival and nlme packages in R were used for these analyses. Mediation analysis was performed to quantify the direct and indirect relationships between PRS, Aβ₄₂, PTau, and AD progression by using the method and related R code from one previous publication [26]. Direct and indirect effects were estimated with generalized linear and Cox regression analysis with adjustment for age at baseline, sex, years of education, race, top three PCs from the genetic data, and the number of allele copies of APOE E2 and E4. Bootstrapping was used for significance testing and standard error estimations of these effects. All analyses were conducted with R (version 3.5.1) if not mentioned otherwise.

RESULTS

Characteristics of the study populations

There were 916 participants diagnosed with MCI at baseline or during follow-up. After merging the PRS data and clinical data, 767 samples remained for further analysis and 294 subjects were diagnosed with AD during the follow-up period. It should be noted that there were missing values in the baseline tau/ptau and Aβ levels for 222 subjects. The distributions of demographic and clinical variables were presented in Table 1. In the univariate analysis, we found that age, Aβ₄₂, tau, and ptau levels at baseline, number of APOE E2 and E4 alleles were all significantly associated with the progression to AD among MCI patients (Table 1, p < 0.05).

Table 1.

Characteristic distributions of the demographic and clinical variables in the ADNI dataset

Variables	Total	Event (AD, %)	Beta1	SE1	p 1
Overall	767	294 (38.3)
Follow-up time (y)
Median (Min-max)	3.00 (0.39–13.00)	2.03 (0.46–11.50)
Age (y)			0.028	0.008	0.001
Mean ± SD	73.4 ± 7.40	74.3 ± 7.00
Median (Min-max)	73.6 (55.0–91.4)	74.4 (55.0–88.4)
Sex			−0.013	0.119	0.912
Female	299	114 (38.1)
Male	468	180 (38.5)
Education (y)			−0.001	0.021	0.951
mean ± SD	15.9 ± 2.82	15.9 ± 2.69
Race2			−0.249	0.238	0.295
Non-Hispanic White	700	275 (39.3)
Other races	67	19 (28.4)
Baseline tau3			1.358	0.167	5.07 × 10−16
Mean ± SD [min-max]	5.56 ± 0.43	5.76 ± 0.41
Median (Min-max)	5.55 (4.58–6.72)	5.76 (4.62–6.71)
Baseline ptau3			1.261	0.145	< 2 × 10−16
Mean ± SD [min-max]	3.19 ± 0.49	3.43 ± 0.46
Median (Min-max)	3.18 (2.11–4.52)	3.44 (2.17–4.52)
Baseline Aβ3			−1.523	0.155	< × 10−16
Mean ± SD [min-max]	6.79 ± 0.47	6.55 ± 0.42
Median (Min-max)	6.77 (5.35–7.44)	6.51 (5.59–7.44)
APOE-E4			0.554	0.080	3.35 × 10−12
0	385	108 (28.1)
1	302	143 (47.4)
2	80	43 (53.8)
APOE-E2			−0.492	0.250	0.049
0	708	277 (39.1)
1	59	17 (28.8)

¹Univariate Cox regression.

²There are 222 participants with missing data on tau/ptau/Aβ levels.

³Log-transformed.

There were 1,373 MCI patients with both the genotyping data and survival data in NACC, of which 864 subjects were diagnosed with AD during the follow-up period. The distributions of age, sex, education years, race, APOE E2/3/4 alleles can be found in Supplementary Table 3. We found significant associations between race, number of APOE E2, E3, and E4 alleles and AD survival, which partially replicated the results in the ADNI study.

Survival analysis of PRS/PHS and the progression of MCI to AD

We investigated the association between the PRS/PHS and the progression of MCI to AD by applying Cox proportional hazard models with adjustment for age at baseline, sex, years of education, race, ethnicity, top three significant PCs from the genetic data, and the number of allele copies of APOE E4 and APOE E2. As presented in Table 2, we found that the PRS was associated with the time interval of progression of MCI to AD (beta = 0.182, SE = 0.061, p = 0.003), which indicated that MCI patients with higher PRS scores have increased risk developing AD after MCI. Survival probability was also estimated by the Kaplan-Meier approach. Patients from the ADNI study were divided into three groups based on the PRS quartiles: group 1 (PRS < the 25th percentile), group 2 (< the 75th percentile), and group3 (≥ the 75th percentile). As shown in Fig. 1a, patients in group 2 and 3 had a significantly shorter survival time than those in group 1 [median progression time = NA (as less than 50% patients developing AD at the end of follow-up), 6.04, and 4.03 years for MCI patients in group 1, 2, and 3, respectively; Log-rankp = 0.003]. However, no evidence was found for the correlation between PHS and the progression of MCI to AD (p = 0.221 in Table 2). We also tested the correlation between the Bayesian PRS constructed with the LDpred2 method and MCI progression and found consistent results with the PRS using the Prsice2 method (beta = 0.281, SE = 0.125, p = 0.024) (Supplementary Table 4). We constructed PRSs using SNPs with different thresholds (i.e., p ≤ 1 × 10⁻⁵, 5.0 × 10⁻⁶, 1.0 × 10⁻⁶, 5.0 × 10⁻⁷, and 1.0 × 10⁻⁷) and tested their associations with MCI progression (p = 0.058, 0.044, 0.084, 0.129, and 0.149, respectively). Only PRS constructed with SNPs with p ≤ 5.0 × 10⁻⁶ showed significant association with MCI to AD progression (beta = 0.115, SE = 0.058, p = 0.044) (Supplementary Table 5). However, this association was not replicated in the NACC dataset (Supplementary Table 6; p = 0.224).

Table 2.

Survival analysis of polygenic risk score (PRS)/polygenic hazard score (PHS) and the progression of MCI to AD in ADNI dataset

Variables	PRS			PHS
	Beta1	SE1	p 1	Beta2	SE2	p 2
Age (y)	0.045	0.009	4.59 × 10−7	0.038	0.008	4.46 × 10−6
Male in sex	−0.140	0.126	0.267	−0.146	0.118	0.214
Education (y)	0.019	0.022	0.382	0.015	0.020	0.459
Other races	−0.103	0.240	0.668	−0.186	0.228	0.414
PC3	3.276	2.730	0.230	5.604	2.715	0.039
PC9	−8.854	3.537	0.012	−7.408	2.855	0.009
PC10	−5.879	3.184	0.065	−4.698	2.567	0.067
APOE E4 (allele number)	0.613	0.088	3.82 × 10−12	0.573	0.082	3.10 × 10−12
APOE E2 (allele number)	−0.295	0.253	0.243	−0.170	0.229	0.458
PRS/PHS	0.182	0.061	0.003	0.068	0.056	0.221

¹Multivariate Cox regression model including PRS, age at baseline, sex, years of education, race, significant principal components, the allele numbers of APOE E4 and APOE E2 (#total = 767 and #event = 294).

²Multivariate Cox regression model including PHS, age at baseline, sex, years of education, race, significant principal components, the allele numbers of APOE E4 and APOE E2 (#total = 767 and #event = 294).

Fig. 1.

Kaplan Meier plots of AD progression based on the trichotomized polygenic risk score (PRS). Patients were divided into three groups: group 1 (PRS < the 25th percentile), group 2 (< the 75th percentile), and group 3 (≥ the 75th percentile). A) Patients in the ADNI dataset (Log-rank p = 0.003): n = 767 (with # of deaths = 294), median progression time = 6.04, 4.03 years for MCI patients in group 2, and 3, respectively; B) Patients in the NACC dataset (Log-rank p = 0.021): n = 1373 (with # of deaths = 864), median progression time = 3.26, 3.16, 2.99 years for MCI patients group 1, 2, and 3, respectively.

A previous study reported that Cox proportional hazard and Logistic regression models provided very similar estimates of regression coefficients in studies with short follow up time (5 years or less) [27]. We tested the association of PRS and AD status by using logistic regression in the ADNI dataset and found similar result between PRS and AD risk (beta = 0.195, SE = 0.080, p = 0.015) (Supplementary Table 7).

We performed stratified analysis by the numbers of APOE E2/E3/E4 alleles and found that the PRS using the Prsice2 method was significantly associated with the risk of AD progression in the strata of participants with one copy of APOE E4, null copy of APOE E2, or one copy of APOE E3 (p < 0.05) (Supplementary Table 8). This association was not significant in participants possessing two copies of APOE E4. We then tested the interaction effect of PRS and APOE alleles by including interaction terms in model but did not find statistical significance for any of the interactions (p = 0.564, 0.800, and 0.931 for the interaction terms between PRS and APOE E2, E3, and E4, respectively). No significance was found for the APOE stratified results of PHS association with progression from MCI to AD (Supplementary Table 8).

Correlation between PRS/PHS and longitudinal cognitive impairments

The correlation between the PRS/PHS and longitudinal cognitive scores (MMSE, MOCA, CDR-SB, FAQ, and ADAS-Cog 11/13/Q4) were presented in Table 3 adjusted for age, sex, years of education, race, ethnicity, top 3 significant PCs, and the numbers of APOE E4 and APOE E2 alleles. We found that the PRS was significantly correlated with the worse cognitive performance measured by ADAS-Cog11, ADAS-Cog13, and ADASQ4 (beta = 0.499, 0.800, and 0.373; p = 0.002, 9.0 × 10⁻⁴, and 1.4 × 10⁻⁵, respectively). PRS constructed with GWAS SNPs with p ≤ 5 × 10⁻⁶ also showed significant association with the three cognitive measures (beta = 0.345, 0.650, and 0.346; p = 0.034, 0.007, and 0.0001 for ADAS-Cog11, ADAS-Cog13, and ADASQ4, respectively; Supplementary Table 5).

Table 3.

Associations of polygenic risk score (PRS)/polygenic hazard score (PHS) and longitudinal cognitive impairments for in ADNI dataset

Phenotype	Participants	Observations	PRS			PHS
			Beta1	SE1	p 1	Beta1	SE1	p 1
MMSE	766	3992	−0.091	0.066	0.168	−0.044	0.064	0.490
CDR-SB	763	3956	0.008	0.034	0.826	−0.028	0.034	0.404
MOCA	482	2372	−0.258	0.125	0.039	−0.002	0.120	0.989
FAQ	763	3952	0.085	0.142	0.551	−0.125	0.146	0.393
ADAS-Cog11	766	3987	0.499	0.161	0.002	0.113	0.159	0.477
ADAS-Cog13	765	3964	0.800	0.240	9.0 × 10−4	0.209	0.236	0.377
ADASQ4	767	3993	0.373	0.086	1.4 × 10−5	0.145	0.081	0.077

¹Adjusted for age at baseline, sex, years of education, race, significant principal components, and the allele numbers of APOE E4 and APOE E2.

However, no significant association was found for the other cognitive phenotypes. We also did not find any significant correlation between PHS and the longitudinal cognitive scores.

Correlation between PRS/PHS and longitudinal biomarkers from CSF and imaging

We investigated the correlations between the PRS/PHS and longitudinal biomarkers from CSF as well as the PET- and MRI-based image data. The distributions of the number of observations in these analyses were presented in Supplementary Table 9. As shown in Table 4, there were significant correlations between the PRS and increased levels of tau and ptau in CSF (beta = 0.068 and 0.082; p = 0.002 and 0.001, respectively). We also observed significant correlations between PRS and cortical amyloid burden, which was quantified using ¹¹C-PiB PET and AV45 (¹⁸F florbetapir) PET (beta = 0.175 and 0.040; p = 0.017 and 3.0 × 10⁻⁴, respectively). MCI patients with different PRS scores had different volumes in hippocampus, entorhinal cortex, and middle temporal gyrus ((beta = −0.011, −0.006, and −0.013; p = 1.0 × 10⁻⁴, 7.0 × 10⁻⁴, and 0.054, respectively). We also observed significant correlations between the PHS and FDG and MidTemp (beta = − 0.015; p = 0.015). The PRS constructed with SNPs with p ≤ 5 × 10⁻⁶ in GWAS also presented significant associations with the increases of Aβ (beta = −0.059; p = 0.010), tau (beta = 0.066; p = 0.003), ptau (beta = 0.078; p = 0.002) in CSF, amyloid burden measured with AV45 PET (beta = 0.025; p = 0.020), and the volume of hippocampus (beta = −0.007; p = 0.016) in the ADNI dataset (Supplementary Table 5).

Table 4.

Correlation between polygenic risk score (PRS)/polygenic hazard score (PHS) and longitudinal biomarkers (from CSF, PET, and MRI imaging) in subjects with MCI at baseline or diagnosed during following visits in ADNI dataset

Biomarker	#Participants	#Observations	PRS			PHS
			Beta1	SE1	p 1	Beta1	SE1	p 1
Aβ2	295	712	−0.037	0.023	0.116	−0.018	0.024	0.447
tau2	295	711	0.068	0.022	0.002	0.012	0.023	0.598
ptau2	294	709	0.082	0.025	0.001	0.014	0.026	0.599
PIB	34	81	0.175	0.068	0.017	−0.061	0.069	0.387
FDG	385	1,278	−0.012	0.007	0.072	−0.015	0.006	0.015
AV45	312	817	0.040	0.011	3.0 × 10−4	0.009	0.010	0.374
Ventricles3	700	2,878	0.009	0.040	0.812	−0.033	0.038	0.391
Hippocampus3	652	2,553	−0.011	0.003	1.0 × 10−4	−0.004	0.003	0.142
Entorhinal3	626	2,431	−0.006	0.002	7.0 × 10−4	−0.003	0.002	0.113
Fusiform3	626	2,431	−0.003	0.006	0.608	−0.003	0.006	0.656
MidTemp3	626	2,431	−0.013	0.007	0.054	−0.012	0.007	0.062

¹Adjusted for age at baseline, sex, years of education, race, significant principal components, and the allele copies of APOE E4 and APOE E2.

²Log-transformed.

³These dependent variables were expressed as the percentages to intracranial volume.

Replication in NACC study

We performed an independent replication using the NACC dataset. As shown in Table 5, there is a significant association between the PRS constructed using the Prsice2 method and increased risk of AD progression (beta = 0.094, SE = 0.037, p = 0.009), which is consistent with the finding in the ADNI dataset. We observed similar results using the Bayesian PRS constructed with the LDpred2 method (beta = 0.200, SE = 0.075, p = 0.007) (Supplementary Table 4).

Table 5.

Association between polygenic risk score (PRS) and the longitudinal clinical outcomes in NACC dataset

Outcome	#participants	#events	Beta1	SE1	p 1
Survival (MCI to AD)	1,373 #participants	864 #observations	0.094	0.037	0.009
MMSE	1,227	4,294	−0.138	0.067	0.039
CDR-SB	1,360	5,227	0.033	0.035	0.346
CDR_Global	1,360	5,227	0.004	0.006	0.513
Ventricles2	52	118	−0.005	0.004	0.992
Hippocampus2	52	118	−0.034	0.039	0.393
Entorhinal2	52	118	−0.089	0.097	0.303
Fusiform2	52	118	−0.076	0.055	0.174
MidTemp2	52	118	−0.038	0.027	0.171

¹Adjusted for age at baseline, sex, years of education, race, and the allele copies of APOE E4 and E2.

²These dependent variables were expressed as the percentages to intracranial volume.

The survival probability was also estimated with the Kaplan-Meier approach. As shown in Fig. 1b, patients in groups 2 and 3 had significantly shorter survival times than those in group 1 (median progression time = 3.26, 3.16, 2.99 years for MCI patients in group 1, 2, and 3, respectively; Log-rank p = 0.021).

In the longitudinal analysis, we found a significant correlation between the PRS and the change in of MMSE for MCI patients. However, we could not replicate the associations between the PRS and MRI markers identified in the ADNI study. We performed stratified analysis by age, sex, race, and numbers of APOE E2/E3/E4 alleles and found that the PRS was significantly associated with the risk of AD progression in the strata of older (≥ 55), females, Whites, APOE E4 heterozygous individuals and APOE E2 non-carriers (p < 0.05) (Supplementary Table 10).

Model performance comparison

We evaluated the predictive accuracy of different models by using the Harrell’s C-statistic (Supplementary Table 11). As shown, the APOE model including numbers of APOE E2 and E4 alleles had a significantly increased C-statistic compared with the demographic model that included only demographic variables in the ADNI data (0.635 versus 0.557 for the APOE model and demographic model, respectively; p = 7.37 × 10⁻⁷). After adding PRS calculated by the PRSICE2 method to the APOE model, there is a slight increase in the Harrell’s C-statistic (0.644 versus 0.635 for the PRS1 model and APOE model, respectively; p = 0.124). Similar results were observed for PRS2 model calculated using the Bayesian PRS (Harrell’s C-statistic = 0.642; p = 0.119 compared with the APOE model) and the model including PRS3 constructed with GWAS SNPs with p ≤ 5.0x10⁻⁶ (Harrell’s C-statistic = 0.638; p = 0.384 compared with the APOE model).

Models that included the PHS in contrast with the APOE model did not show any significant improvement in predictive accuracy (0.627 versus 0.627 for the PHS model and APOE model, respectively; p = 0.814). In the NACC dataset, we found that the APOE model had better performance on prediction than the demographic model (Harrell’s C-statistic: 0.575 versus 0.524 for the APOE model and demographic model, respectively; p = 3.52 × 10⁻⁴); however, inclusion of PRS or PHS in the APOE model did not significantly increase the prediction accuracy (Harrell’s C-statistic = 0.585 and 0.615 for PRS and PHS models, respectively; p = 0.418 and 0.579 for prediction comparisons of PRS model versus APOE model, and PHS model versus APOE model, respectively). We also estimated the prediction accuracy of model including the Bayesian PRS and found its Harrell’s C-statistic = 0.574 (p = 0.394 compared with the APOE model). Non-significant results were found for the PRS constructed with GWAS SNPs with p ≤ 5 × 10⁻⁶ (Harrell’s C-statistic = 0.574; p = 0.538 comparing with the APOE model).

Mediation analysis

Reduction of CSF Aβ₄₂ levels and elevation of phosphorylated tau protein are two important pathologic changes in the brains of MCI and AD patients. In this study, we observed that the PRS were associated with the baseline levels of CSF Aβ₄₂ and phosphorylated tau in MCI patients (Supplementary Figure 2). We performed mediation analysis in the ADNI dataset (547 MCI patients with 198 AD event) to test if the effect of the PRS on AD progression are mediated by the two biomarkers. As shown in Supplementary Figure 2, we found that the PRS contributed the increased risk of AD progression through the mechanism mediated by Aβ₄₂ (beta = 0.056, SE = 0.026, p = 0.036). No statistical significance was observed for the direct effect of PRS (beta = 0.067, SE = 0.072, p = 0.355) as well as the indirect effect mediated by phosphorylated tau (beta = 0.036, SE = 0.020, p = 0.075). Our results suggested that about 35.2% [0.056/(0.067 + 0.056 + 0.036) = 0.352] effect of PRS on AD progression was mediated by Aβ₄₂.

DISCUSSION

In this study, we investigated the association between PRSs/PHS and the time of progression to AD in patients with MCI. We found that the PRS constructed with SNPs reported in previously AD GWASs was significantly associated with the progression from MCI to AD in both the ADNI cohort and the NACC study after adjusting for demographic and clinical variables. Further analysis showed that this PRS was associated with increases of ADAS-Cog11/ADAS-Cog13/ADASQ4 scores, tau/ptau levels in CSF, amyloid burden in brain regions, and the decreased volumes of hippocampus and entorhinal cortex. Although several studies have reported the association between PRS and AD risk as well as with AD pathologic-related biomarkers [11, 12, 27–33], there are at least three differences between this study and those previously published. First, in the study design, we used the NACC dataset as an independent replication of our findings on the PRS in the ADNI dataset, while the results of most previous studies were only based on the ADNI dataset. Second, we performed a comprehensive investigation of the correlation between the constructed PRS and longitudinal CSF biomarkers (i.e., Aβ₄₂, tau/ptau levels) and imaging biomarkers (i.e., amyloid burden and the volumes of various structures of the brain). Other studies were limited to investigation of the association of AD PRSs and baseline-levels of Aβ₄₂/tau/ptau, and/or longitudinal changes of volumes. Third, in this study, we performed mediation analysis to explore possible mechanisms underlying the association between the AD PRS and AD progression, which has not been reported in other studies. We found that the effect of the PRS on the time of progression to AD from MCI was partly mediated via Aβ₄₂. Finally, we compared the results of conventional PRS (based on different p value thresholds and LD pruning) with a PRS calculated by a recently published Bayesian method (LDPred2) [24] and with PRSs constructed using SNPs with different p-value selection thresholds. Overall, the study and specifically the mediation analysis places the PRS association results in context with the National Institutes for Health/Alzheimer’s Association (NIH/AA) Research Framework for observational and interventional research in AD [34–37]. The framework provides guidelines to interpret the definition of AD in terms of specific biomarkers related to Aβ deposition, pathologic tau accumulation and neurodegeneration [AT(N)]. The framework considers AD as a continuum where the biomarkers help to stage the severity of the disease and symptoms in context of underlying biology and pathophysiology. The mediation analysis provides estimates of the contribution of the PRS, derived from numerous biological pathways to specific AT(N) biomarker-related processes to the progression of AD. Overall, our findings suggested that the PRS derived from the AD GWASs is useful for the prediction of AD risk and other clinical changes in MCI patients.

It has been established that the progression from MCI to AD can be predicted by the biomarkers from MRI imaging and CSF, APOE genotypes, and different cognitive test batteries [11, 38, 39]. Recent studies also suggested that both PRS and PHS can be used to predict an age-specific risk for developing AD [11, 31, 40]. In this study, we investigated the association between PRSs/PHS and the time from MCI to AD and found that PRSs constructed with SNPs reported in previous AD GWASs can be used for the prediction of progression from MCI to AD in both the discovery and replication studies. Our results are similar with previous studies, which constructed PRS using the non-APOE SNPs from the IGAP GWAS and reported a predictive ability for LOAD progression from MCI of 61–67%, especially in MCI patients with amyloid deposition [28, 33]. By using the effect size estimates from a recent AD GWAS meta-analysis [6], Altmann et. al. constructed PRSs with two p-value cut-offs: p = 0.5 and 1.0 × 10⁻⁵ (genome-wide suggestive) and found the two PRSs had significantly different distributions between cognitively normal individuals and MCI patients, as well as between cognitively normal individuals and AD patients [27]. However, both PRS did not show significant differences between MCI and AD patients. Desikan et al. reported that the PHS constructed from 31 SNPs can significantly predict time to progression from MCI to AD in the ADNI dataset [32]. However, we did not replicate this association in our study. Although both studies used the ADNI dataset, this discrepancy might be due to the difference in SNPs used for PHS construction. Our study applied LD clumping before PHS construction, which resulted in 25 independent SNPs with pairwise r-squared < 0.1, and also excluded the two SNPs in APOE region, while the PHS in the Desikan study was constructed with 31 SNPs including the two APOE SNPs, which had stronger effects compared with SNPs from other loci [16]. Such results suggest that APOE SNPs in the PHS reported in this paper [16] may play the major role in the prediction of AD risk in MCI patients.

In this study, we found that the AD PRS was associated with CSF biomarkers (i.e., CSF Aβ and tau/ptau) and amyloid burden in brain regions. Our results were consistent with two previous studies which reported significant associations between the PRS based on significant SNPs identified in the International Genomics of Alzheimer’s Disease GWAS and baseline levels of CSF Aβ and tau/ptau [27, 29]. The mediation analysis suggested that the association between PRS and AD progression might via a mechanism mediated by Aβ. Since both Aβ plaque and tau deposition precede memory impairment [41], these results suggest that PRS may reflect pathophysiology in early AD stages. However, further functional studies are warranted. It should be noted that the PRSs constructed in another study with two different P-value thresholds showed significant association with the baseline levels of tau/ptau, but not Aβ. Such inconsistencies might be a consequence that PRSs constructed with variants with strong effects are more robust to population structure differences [27, 42].

Both the entorhinal cortex and the hippocampus are essential parts of the medial temporal lobe system that supports declarative memory [43]. It has been shown that early AD associated Braakian atrophy includes the hippocampus, entorhinal and cingulate cortices [44, 45]. In this study, we found that PRS had significant associations with the volume changes of hippocampus and entorhinal cortex. This was consistent with a previous study, which found that the AD PRS comprised of non-APOE SNPs was associated with reduced hippocampal volume in healthy older adults and those with MCI [13].

The present study had several limitations. First, the sample size may not be large enough to detect certain associations, especially for analyses using samples with CSF biomarker data of Aβ and tau. Second, we did not replicate the association between the PRS and imaging markers observed in the ADNI study in the NACC study, which might be due to the small sample size in the NACC dataset for individuals with longitudinal biomarker data. Third, different diagnostic criteria of AD were used across different UDS versions and Alzheimer’s disease centers (ADCs) contributing data to the NACC study, which might cause population heterogeneity for AD diagnoses. Validation of our findings in a larger study is warranted. Finally, AD develops over a period of decades often without evidence of cognitive decline. The interval between MCI and development of dementia represents only one part of the cycle of neurodegenerative events. Other factors, including vascular pathology that is independent of the genetic factors modeled in the PRS may influence the rate of progression of AD development for individuals.

In conclusion, a PRS derived from 40 common AD risk variants with small effect size was found to be associated with the progression time from MCI to AD independent of APOEE4 and E2 risk alleles. The correlation between the PRS and early AD markers (tau/ptau, and atrophy of hippocampus and entorhinal cortex) shows evidence for a genetic modulation of neurodegeneration, and the potential for a combination of PRS and other brain biomarkers to aid in the prediction of MCI patients at risk of AD. Mediation analysis and interpretation of the results in context of the NIH/AA research framework enable use of the PRS in a more precise approach for the design of clinical trials with specific agents that may affect Aβ, pathologic tau, or other biological pathways.