Causal Associations Between Modifiable Risk Factors and the Alzheimer’s Phenome

Abstract

Objective:

The purpose of this study was to infer causal relationships between 22 previously reported risk factors for Alzheimer’s disease (AD) and the “AD phenome”: AD, AD age of onset (AAOS), hippocampal volume, cortical surface area and thickness, cerebrospinal fluid (CSF) levels of amyloid-β (Aβ₄₂), tau, and ptau₁₈₁, and the neuropathological burden of neuritic plaques, neurofibrillary tangles (NFTs), and vascular brain injury (VBI).

Methods:

Polygenic risk scores (PRS) for the 22 risk factors were computed in 26,431 AD cases/controls and the association with AD was evaluated using logistic regression. Two-sample Mendelian randomization (MR) was used to infer the causal effect of risk factors on the AD phenome.

Results:

PRS for increased education and diastolic blood pressure were associated with reduced risk for AD. MR indicated that only education was causally associated with reduced risk of AD, delayed AAOS, and increased cortical surface area and thickness. Total- and LDL-cholesterol levels were causally associated with increased neuritic plaque burden, although the effects were driven by single nucleotide polymorphisms (SNPs) within the APOE locus. Diastolic blood pressure and pulse pressure are causally associated with increased risk of VBI. Furthermore, total cholesterol was associated with decreased hippocampal volume; smoking initiation with decreased cortical thickness; type 2 diabetes with an earlier AAOS; and sleep duration with increased cortical thickness.

Interpretation:

Our comprehensive examination of the genetic evidence for the causal relationships between previously reported risk factors in AD using PRS and MR supports a causal role for education, blood pressure, cholesterol levels, smoking, and diabetes with the AD phenome.

Late-onset Alzheimer’s disease (AD) is a debilitating neurological condition characterized by progressive neurodegeneration and deterioration of cognitive function leading to dementia.¹ The primary neuropathological hallmarks of AD are the aggregation of extracellular amyloid-β (Aβ) peptides into amyloid plaques and of intracellular hyperphosphorylated tau into neurofibrillary tau tangles (NFTs) often accompanied by cerebrovascular injury or other pathologies.¹

In the absence of any disease-modifying therapies, the number of people living with dementia in the United States is expected to exceed 13.8 million by 2050.¹ Observational studies have identified potentially modifiable risk factors that could be targeted in intervention studies to reduce the risk of dementia or delay its onset, thereby significantly reducing the population prevalence of AD and related dementias.² From these studies, it has been estimated that 40% of AD cases may be attributable to preventable causes, such as low educational attainment, hearing loss, traumatic brain injury, hypertension, alcohol consumption, obesity, smoking, depression, physical inactivity, social isolation, air pollution, and diabetes.^3,4 However, the quality of evidence for interventions targeting these risk factors to reduce the risk of dementia or cognitive decline is mixed.^2,5 Lifestyle interventions that target modifiable risk factors are entirely dependent on accurate causal relationships being established between modifiable risk factors and AD. In observational studies, a correlation between a risk factor and AD cannot be reliably interpreted as evidence of a causal relationship due to potential confounding or reverse causation. Therefore, unless those modifiable factors specifically exacerbate disease progression, disease reduction strategies targeting them will not be successful.

Methods of causal inference that exploit genetic information, such as polygenic risk scores (PRS) and Mendelian randomization (MR), can overcome some of the limitations of observational studies. PRS are a measure of an individual’s genetic propensity to a trait and can be used in cross-trait analyses to test whether genetic liability for one trait is associated with disease risk for a second.⁶ Although this does not imply that the trait causally modifies disease risk, because there are several alternative explanations, such a PRS-disease association would be expected if the trait were causal of disease, and thus PRS can be used to prioritize putative causal risk factors.⁶ MR uses genetic variants as proxies for environmental exposures to infer a causal relationship between an intermediate exposure and a disease outcome. MR is akin to conducting a “genetic randomized control trial,” with the risk factors (genotypes) randomly allocated (from parents to offspring), independent of confounding factors that influence the risk factors and disease and unaffected by reverse causation.⁷ Although MR can be used to infer causal relationships between traits, it typically has lower statistical power than tests of PRS-disease associations.⁶ The results from PRS and MR analyses can be integrated and compared with those from traditional epidemiological studies and randomized control trials (RCTs). By integrating the results from multiple approaches that have different sources of bias, it is possible to rigorously assess the evidence regarding causality of interventions targeting modifiable risk factors to reduce the risk of dementia.⁸

In this study, we used PRS and MR to infer causal relationships among potentially modifiable risk factors for dementia (alcohol consumption, smoking, hearing loss, diabetes mellitus, obesity, dyslipidemia, blood pressure depression, sleep, social isolation, physical activity, and diet) reported in the Dementia prevention, intervention, and care: 2020 report of the Lancet Commission⁴ and the World Health Organization’s Guidelines for Risk Reduction of Cognitive Decline and Dementia⁵ and the AD phenome (AD status, AD age of onset survival [AAOS], cerebrospinal fluid [CSF] levels of Aβ₄₂, tau and hyperphosphorylated tau [ptau₁₈₁], hippocampal volume, cortical surface area and thickness, and the neuropathological burden of neuritic plaques, NFTs, and vascular brain injury [VBI]). Based on these analyses, we identified a subset of modifiable risk factors that represent the most promising targets for public health initiatives to reduce AD burden in the population.

Methods

Genomewide Association Summary Statistics

We obtained genomewide association study (GWAS) summary statistics (GWAS-SS) for each exposure and outcome of interest (Table 1). Modifiable risk factors included alcohol consumption,⁹ the alcohol use disorder identification test (AUDIT),¹⁰ moderate-vigorous physical activity (MVPA),¹¹ lipid traits,¹² systolic blood pressure (SBP), diastolic blood pressure (DBP), pulse pressure (PP),¹³ type 2 diabetes (T2D),¹⁴ body mass index (BMI),¹⁵ meat-related diet and a fish- and plant-related diet,¹⁶ depression,¹⁷ insomnia symptoms,¹⁸ sleep duration,¹⁹ social isolation,²⁰ smoking initiation,⁹ cigarettes per day,⁹ educational attainment,²¹ and hearing difficulty.²² These summary statistics were used to generate PRS and as the exposures in the MR analyses.

TABLE 1.

GWAS Datasets Utilized in this Study

Study	Trait	Cohort/consortium	N	Age	Females, %
Exposures
Liu et al 2019	Alcohol consumption	GSCAN; 23andMe	941,280	—	—
	Smoking initiation	GSCAN; 23andMe	1,232,091	—	—
	Cigarettes per day	GSCAN; 23andMe	337,334	—	—
Sanchez-Roige et al 2019	Alcohol use disorder test	UKBB; 23andMe	141,932	—	—
Wells et al 2019	Hearing difficulty	UKBB	250,389	—	—
Xue et al 2018	Type 2 diabetes	DIAGRAM; UKBB; GERA	659,316	—	—
Yengo et al 2018	BMI	UKBB; GIANT	690,495	—	—
Willer et al 2013	Total cholesterol	GLC	188,577	54.94	56.58
	LDL cholesterol
	HDL cholesterol
	Triglycerides
Evangelou et al 2018	DBP	UKBB; ICBP	757,601	—	—
	SBP
	PP
Howard et al 2019	Depression	UKBB; PGC; deCODE; iPSYCH; GeneScotland; GERA; 23andMe	807,553	—	—
Jansen et al 2018	Insomnia symptoms	UKBB; 23andMe	1,331,010	—	—
Dashti et al 2019	Sleep duration	UKBB	446,118	57.3	54.1
Day et al 2018	Social isolation	UKBB	452,302	—	—
Lee et al 2018	Educational attainment	UKBB; SSGAC; 23andMe	1,131,881	63.8	54.7
Klimentidis et al 2018	Moderate-vigorous physical activity	UKBB	377,234	—	—
Niarchou et al 2020	Meat-related diet	UKBB	335,576	—	54%
	Fish and plant-related diet	UKBB	335,576	—	54%
Outcomes
Lambert et al 2013	Late onset AD	IGAP	54,162	71	58.4
Kunkle et al 2019	Late onset AD	IGAP	63,926	72.6	58.5
Huang et al 2017	AAOS	IGAP	40,255	77.5	60.35
Deming et al 2017	CSF Aβ42	Knight-ADRC	3,146	71.8	49.57
	CSF Ptau181
	CSF Tau
Hibar et al 2015	Hippocampal volume	ENIGMA	13,688	39.9	51.8
Hibar et al 2017	Hippocampal volume	ENIGMA; CHARGE	26,814	54.3	55.3
Grasby et al 2020	Cortical surface area	ENIGMA	33,709	45.9	51.9
	Cortical thickness
Beecham et al 2014	NPs	ADGC	4,914	74.7	65.4
	NFTs
	VBI

AAOS = Alzheimer’s disease age of onset; AB = amyloid-β; AD = Alzheimer’s disease; BMI = body mass index; CSF = cerebrospinal fluid; DBP = diastolic blood pressure; HDL = high-density lipoprotein; LDL = low-density lipoprotein; NFTs = neurofibrillary tangles; NPs = neuritic plaques; PP = pulse pressure; Ptau₁₈₁ = hyperphosphorylated tau; SBP = systolic blood pressure; VBI = vascular brain injury.

GWAS-SS for the AD phenome consisted of late-onset AD status,²³ AAOS,²⁴ CSF levels of Aβ₄₂, ptau₁₈₁ and total tau (Tau),²⁵ hippocampal volume,²⁶ cortical surface area and thickness,²⁷ neuropathological burden of neuritic plaques, NFT burden and VBI.²⁸ Although VBI is not a classical neuropathological hallmark of AD, concurrent cerebrovascular injury is a common neuropathological finding in AD.²⁹ As such, VBI was included in the AD phenome to disentangle whether a particular risk factor may influence the clinical symptomatology of AD via vascular or amyloid/tau pathways. Due to data use restrictions associated with evaluating alcohol intake and education phenotypes in the most recent GWAS of AD and hippocampal volume, we used an earlier GWAS for AD³⁰ and hippocampal volume³¹ for estimating the causal effect of alcohol intake and educational attainment on these phenotypes. These summary statistics were used as outcomes in the MR analyses.

GWAS-SS that were mapped to earlier human genome builds were lifted over to Human Genome Build 19.³² GWAS-SS were standardized using a pipeline,³³ that (1) aligns effect alleles to the alternate allele on the forward strand of the human genome reference build and normalizes indels, (2) annotates variants with marker names using chromosome:position:ref:alt, 1000 Genomes rsIDs (phase III), and database-single-nucleotide polymorphism (dbSNP) rsIDs (b151) (3) where allele frequencies are missing, annotates allele frequencies using non-Finnish Europeans from gnomAD (version 2.1), and (4) convert summary statistics to VCF and TSV files.

Alzheimer’s Disease Genetics Consortium

Individual-level genetic and phenotypic data used to compute and test the association of PRS were obtained from the Alzheimer’s Disease Genetics Consortium (ADGC), a large multicenter project composed of 34 separate cohorts with the goal of performing genomewide analyses of AD. The recruitment and genotyping of ADGC samples has been described in detail elsewhere.^23,34 Briefly, genotype data in each cohort underwent stringent quality control (QC) checks, with variants excluded if the call rate < 0.95, not in Hardy–Weinberg equilibrium (p < 1 × 10^–6), and samples excluded if the call rate was < 0.95, discordant sex was reported based on X chromosome heterozygosity, cryptic relatedness, and non-European ancestry. Related individuals were determined within and across cohorts by identity-by-descent (IBD) using KING, ³⁵ with individuals excluded based on a proportion of IBD < 0.1875, corresponding to less than halfway between second-and third-degree relatives. Ancestry was determined empirically by projecting samples onto principal components from known ancestral populations in the 1000 Genomes Project, with samples determined to be European population outliers if they were ± 6 SD away from the European population mean on the first 10 principal components using PC-Air³⁶ and PLINK.³⁷ Single nucleotide polymorphisms (SNPs) that were not directly assayed were imputed on the Michigan Imputation Server individually for each of the cohorts or subcohorts using all ethnicities of the Haplotype Reference Consortium (HRC) 1.1 reference panel.³⁸ Eagle was used for phasing and Minimac3 was used for imputation. Following imputation, poorly imputed (r² < 0.8) or rare (minor allele frequency [MAF] < 0.01) variants were removed and the cohorts merged for joint analysis. Following this merger, variants with low call rate due to differential imputation (< 95%) were removed, and then samples with low call rates (< 95%) were removed. Within-ancestry principal components were created using PLINK to correct for residual population stratification within the European population subset. After sample QC, 26,431 participants were available (Table 2). Written informed consent was obtained from study participants or, for those with substantial cognitive impairment, from a caregiver, legal guardian, or other proxy, and the study protocols for all populations were reviewed and approved by the appropriate institutional review boards (IRBs).

TABLE 2.

Demographic Characteristics of ADGC

Variable	Cases (n = 13,312)	Controls (n = 13,119)
Female	7,699 (57.8%)	7,785 (59.3%)
APOE e4+	7,690 (57.8%)	3,085 (23.5%)
Age, yr	73.4 (8.3)	76.6 (8.3)

ADGC = Alzheimer’s Disease Genetics Consortium.

Polygenic Risk Scores

PRSice-2 was used to construct the PRS for each of the exposures of interest in ADGC.³⁹ PRSice generates PRS as the sum of all alleles associated with the exposure of interest exceeding a given p value threshold (p_t), weighted by their effect size estimated in an independent GWAS on the trait. SNPs were clumped to obtain variants in linkage equilibrium with an r² > 0.001 within a 10 MB window and PRS were constructed across a range of p_t (p_t = 5e-8, 1e-6, 1e-5, 1e-4, 1e-3, 0.01, 0.05, 0.1, 0.2, 0.3, 0.4, and 0.5). We then performed a principal component analysis (PCA) on the resulting PRS and used the first PRS-PCA in subsequent association tests.⁴⁰ The PCA reweights the variants included in the PRS to achieve maximum variation across all the p_t.⁴⁰ This PRS-PCA approach avoids the optimization step in standard pruning and thresholding models to determine the optimal p_t, which can inflate type 1 error and result in overfitting.⁴⁰ As a sensitivity analysis, we further excluded variants located ±250 kb of the APOE e4 defining SNP, rs429358. The association between each exposure PRS and AD was evaluated using logistic regression adjusting for age, sex, APOE ε4 dose, and 10 principal components. The Benjamini & Hochberg false discovery rate was used to account for the multiple testing across the 22 different exposures.

Mendelian Randomization Analysis

We inferred causal relationships between each modifiable risk factor (exposures) and the AD phenome (outcomes) using 2-sample MR, in which the associations for the exposure- and outcome-genetic instrumental variables are obtained from GWAS-SS generated from different, non-overlapping samples.

Genetic Instruments

For each exposure, we constructed 2 different sets of instrumental variables (IVs), corresponding to independent 1) genomewide significant SNPs (p < 5 × 10⁻⁸) and 2) SNPs of at least borderline significance (p < 5 × 10⁻⁶). Increasing the number of SNPs used as IVs increases the phenotypic variance explained and, thus, has the potential to increase statistical power. However, if the additional variants included violate the core MR assumptions then they may instead reduce power, biasing the results toward the null by introducing weak instrument bias. To obtain independent SNPs, linkage disequilibrium (LD) clumping was performed by excluding SNPs that have an r² > 0.001 with another variant with a smaller p value association within a 10 MB window using PLINK.³⁷ For genetic variants that were not present in the outcome GWAS, PLINK was used to identify proxy SNPs that were in LD (r² > 0.8; European reference population). Finally, the exposure and outcome GWAS datasets were harmonized so that the effect size for the exposure and outcome corresponded to the same effect alleles. Genetic variants that were palindromic with ambiguous allele frequencies (AF > 0.42), or that had incompatible alleles, were removed. Instruments that were genomewide significant for the outcome were removed. As a sensitivity analysis, we further excluded variants located ± 250 kb from the APOE ε4 defining SNP, rs429358. The proportion of variance in the phenotype explained by each instrument and F-statistic were calculated, as previously described.^41,42

Statistical Analysis

For each genetic variant, we calculated an instrumental variable ratio estimate by dividing the SNP-exposure by SNP-outcome and the resulting coefficients were combined in a fixed-effects meta-analysis using an inverse-variance weighted (IVW) approach to give an overall estimate of causal effect.⁷ The IVW method assumes that all SNPs included in the causal estimate are valid instruments - that is, that they do not violate any of the underlying MR assumptions, in particular horizontal pleiotropy, whereby genetic variants have direct effects on multiple phenotypes, could lead to false inference of causal associations.⁷ In order to account for potential violations of the assumptions underlying the IVW analysis, we conducted sensitivity analyses using alternative MR methods known to be more robust to horizontal pleiotropy in particular, but at the cost of reduced statistical power. The alternative approaches included (1) Weighted Median Estimator (WME), which tests the median effect of all of the IV variants, allowing 50% of variants to exhibit horizontal pleiotropy⁷; (2) Weighted Mode Based Estimator (WMBE), which clusters variants into groups based on the similarity of causal effects and reports the final causal effect based on the cluster with the largest number of variants⁷; and (3) MR-Egger regression, which allows all variants to be subject to direct effects that bias the estimate in the same direction.⁷

The MR-Egger regression intercept was used to verify the absence of pleiotropic effects of the SNPs on the outcome.⁷ To further confirm the absence of distortions in the causal effects due to heterogeneity or horizontal pleiotropy, we used the MR Pleiotropy Residual Sum and Outlier (MR-PRESSO) test to detect and correct for horizontal pleiotropic outliers.⁴³ Where heterogeneity was detected (the MR-PRESSO Global Test) and significant outliers were detected (MR-PRESSO Outlier Test), the outliers were removed.

We report the IVW results for the set of IV variants (at p < 1 × 10⁻⁸ or 5 × 10⁻⁶) with the smallest p value, outliers were removed if detected. Where there was evidence of horizontal pleiotropy or heterogeneity (MR-PRESSO Global Test p < 0.05 or an MR-Egger Intercept p < 0.05), we report the IVW results for which the sensitivity analyses were also significant and the effect direction was concordant with the IVW results. To account for multiple testing, we report q-values, a false discovery rate-based measure of significance.⁴⁴ Power analyses were conducted using the non-centrality parameter-based approach using the observed IVW coefficient.⁴⁵

As a further sensitivity analysis to evaluate potential reverse causation, we also conducted MR in the reverse direction with the AD phenome as the exposure and the modifiable risk factors as the outcome using the same analysis pipeline as detailed above. Variants located ± 250 kb from the APOE ε4 defining SNP, rs429358, were excluded. For the neuropathology and CSF endophenotypes, analyses were limited to selecting nominally significant SNPs (p < 5 × 10⁻⁶) due to lack of genome-wide significant SNPs available to include as instruments, and additionally, are likely to be underpowered.

All statistical analyses were conducted using R version 3.6.3. MR analysis was performed using the “TwoSampleMR” version 0.4.22 package.⁷ A Snakemake workflow was constructed that automates the PRS and MR analysis pipelines and allows for multiple exposure – outcomes datasets to be run in parallel.⁴⁶

The SNPs used as IVs, their harmonized effects, and outliers are presented in Supplementary Tables S1 and S2. The causal estimates for each p value threshold, MR method, and pre- and post-outlier removal and bidirectional results are presented in Supplementary Tables S3 to S5.

Results

Polygenic Risk Score Analysis

We evaluated the association of 22 PRS for potentially modifiable risk factors with AD in ADGC (Table 3). After correction for multiple testing, a 1 SD higher PRS for educational attainment was associated with decreased risk of AD (odds ratio [OR] = 0.93, confidence interval [CI] = 0.91, 0.96). Additionally, a 1 SD increased DBP was nominally (false discovery rate [FDR] < 0.1) associated with reduced risk of AD (OR = 0.96, CI = 0.94, 0.99). These results remained unchanged after excluding variants located in the APOE locus.

TABLE 3.

Association of Polygenic Risk Scores for Potentially Modifiable Risk Factors on Alzheimer’s Disease

	Including APOE			Excluding APOE
Exposure	b (se)	p	FDR	b (se)	p	FDR
Educational attainment	−0.07 (0.014)	8.90E-07	2E-05	−0.07 (0.014)	8.90E-07	2E-05
DBP	−0.038 (0.014)	0.007	0.076	−0.038 (0.014)	0.007	0.076
Total cholesterol	0.034 (0.014)	0.017	0.127	0.014 (0.014)	0.336	0.825
Meat diet	−0.028 (0.014)	0.043	0.239	−0.026 (0.014)	0.063	0.459
LDLs	0.024 (0.015)	0.093	0.409	−0.001 (0.014)	0.965	0.968
Social isolation	0.022 (0.014)	0.115	0.423	0.022 (0.014)	0.115	0.635
SBP	−0.019 (0.014)	0.184	0.579	−0.019 (0.014)	0.184	0.81
Moderate-to-vigorous PA	0.017 (0.014)	0.233	0.64	0.017 (0.014)	0.233	0.825
Cigarettes per day	−0.014 (0.014)	0.305	0.746	−0.014 (0.014)	0.31	0.825
Hearing difficulties	−0.013 (0.014)	0.351	0.771	−0.013 (0.014)	0.351	0.825
AUDIT	−0.012 (0.014)	0.393	0.786	−0.012 (0.014)	0.393	0.825
Alcohol consumption	0.011 (0.014)	0.451	0.807	0.011 (0.014)	0.45	0.825
PP	−0.01 (0.014)	0.491	0.807	−0.012 (0.014)	0.419	0.825
Smoking Initiation	−0.009 (0.014)	0.545	0.807	−0.009 (0.014)	0.516	0.872
BMI	−0.008 (0.014)	0.55	0.807	−0.008 (0.014)	0.573	0.901
Depressive symptoms	−0.005 (0.014)	0.743	0.992	−0.005 (0.014)	0.743	0.961
Type 2 diabetes	0.004 (0.014)	0.807	0.992	0.004 (0.014)	0.807	0.968
Sleep duration	0.002 (0.014)	0.871	0.992	0.002 (0.014)	0.871	0.968
Fish and plant diet	0.002 (0.014)	0.913	0.992	0.001 (0.014)	0.968	0.968
Insomnia symptoms	−0.001 (0.014)	0.942	0.992	−0.001 (0.014)	0.939	0.968
Triglycerides	−0.001 (0.014)	0.967	0.992	−0.006 (0.014)	0.652	0.956
HDL	0 (0.014)	0.992	0.992	0.005 (0.014)	0.728	0.961

AUDIT = alcohol use disorder identification test; BMI = body mass index; DBP = diastolic blood pressure; HDL = high-density lipoprotein; LDL = low-density lipoprotein; PA = physical activity; PP = pulse pressure; Ptau₁₈₁ = hyperphosphorylated tau; SBP = systolic blood pressure.

Mendelian Randomization Analysis

We used MR to infer causal relationships among 22 potentially modifiable risk factors and 11 AD outcomes, across 2 sets of IV variants corresponding to 2 different p value thresholds. We observed 12 exposure-outcome pairs that were significant at an FDR < 0.05 and that either showed no evidence of heterogeneity or horizontal pleiotropy, or in the presence of heterogeneity or horizontal pleiotropy, the additional MR sensitivity analyses were significant (Fig; Table 4). The descriptive statistics for the number of SNPs, PVE, F-statistics, and power for each exposure are presented in Supplementary Table S6, with estimates for the individual exposure-outcome pairs are presented in Supplementary Tables S3 to S5.

FIGURE:

Putative causal associations between modifiable risk factors and the AD phenome. Shown are the best IVW results for each causal association, with colors representing the standardized effect sizes - for LOAD, NP, NFT, and AAOS red indicates increased risk / earlier onset and blue reduced risk / delayed onset, for CSF levels and Hippocampal volume, red indicates increased levels / volume and blue reduced levels / volume. “.” FDR < 0.1; * FDR < 0.05; ** FDR < 0.01; *** FDR < 0.001. Causal estimates bracketed in red or orange indicate significant causal effects that showed no evidence for horizontal pleiotropy or where sensitivity analyses were also significant. AAOS = Alzheimer’s disease age of onset; AD = Alzheimer’s disease; AUDIT = alcohol use disorder identification test; BMI = body mass index; CSF = cerebrospinal fluid; FDR = false discovery rate; IVW = inverse-variance weighted; NFT = neurofibrillary tangle; NP = neuritic plaque; PA = physical activity. [Color figure can be viewed at www.annalsofneurology.org]

TABLE 4.

Causal Association of Potentially Modifiable Risk Factors on Alzheimer’s Disease and Alzheimer’s Endophenotypes

			IVW		MR-Egger	WMBE	WME	MR-PRESSO Global	MR-Egger Intercept
Exposure	pt	SNPs	b (se)	q-value	b (se)	b (se)	b (se)	p	p
LOAD
Educational attainment	5E-08	478	−0.44 (0.071)	7.73E-08	−0.55 (0.27)*	−0.44 (0.11)***	−0.43 (0.38)	0.04156	0.68
AAOS
Educational attainment	5E-06	716	−0.28 (0.058)	5.67E-05	−0.19 (0.21)	−0.31 (0.09)***	−0.28 (0.25)	0.003	0.62
Type 2 diabetes	5E-06	218	0.072 (0.017)	9.93E-04	0.035 (0.044)	0.045 (0.033)	0.067 (0.034)*	5E-04	0.34
NPs
LDLs	5E-08	74	0.7 (0.19)	0.006	0.67 (0.32)*	0.51 (0.33)	0.21 (0.41)	0.2495	0.91
Total cholesterol	5E-06	122	0.69 (0.18)	0.004	0.7 (0.31)*	0.8 (0.31)**	0.71 (0.44)	0.8558	0.95
NFTs
Total cholesterol	5E-06	123	0.33 (0.11)	0.042	0.28 (0.23)	0.15 (0.19)	0.32 (0.2)	0.0051	0.76
VBI
DBP	5E-06	608	0.073 (0.017)	0.001	0.1 (0.041)*	0.068 (0.028)*	0.022 (0.056)	0.58812	0.47
PP	5E-08	384	0.058 (0.017)	0.013	0.14 (0.045)**	0.055 (0.028).	0.033 (0.068)	0.17312	0.057
Hippocampal volume
Total cholesterol	5E-06	125	−0.065 (0.02)	0.019	−0.034 (0.035)	−0.084 (0.034)*	−0.061 (0.032).	0.0974	0.26
Cortical surface area
AUDIT	5E-06	51	5,400 (1,400)	0.003	−4,800 (9,900)	3,400 (2,300)	460 (4,000)	0.0088	0.3
Educational attainment	5E-06	707	4,600 (440)	1.17E-22	2,300 (1,800)	3,000 (730)***	370 (3,300)	<4e-05	0.18
Insomnia symptoms	5E-06	375	−4,100 (1,300)	0.028	−4,900 (8,200)	−2,700 (2,000)	700 (5,800)	0.0016	0.92
Type 2 diabetes	5E-06	217	−460 (140)	0.016	−410 (400)	−360 (310)	−340 (310)	<1e-04	0.89
Cortical thickness
BMI	5E-08	506	−0.0092 (0.0024)	0.003	−0.0062 (0.0075)	−0.0077 (0.0041).	−0.019 (0.011).	<4e-05	0.67
Educational attainment	5E-06	710	0.01 (0.0031)	0.016	0.012 (0.012)	0.011 (0.0049)*	0.013 (0.014)	<4e-05	0.87
Sleep duration	5E-06	191	0.016 (0.0045)	0.01	0.012 (0.02)	0.018 (0.007)*	0.036 (0.016)*	0.0034	0.83
Smoking initiation	5E-06	554	−0.022 (0.0065)	0.012	−0.08 (0.029)**	−0.019 (0.0099).	0.0083 (0.028)	<4e-05	0.04

AAOS = Alzheimer’s disease age of onset; AUDIT = alcohol use disorder identification test; BMI = body mass index; DBP = diastolic blood pressure; IVW = inverse-variance weighted; LDL = low-density lipoprotein; MR = Mendelian randomization; MR-PRESSO = Mendelian randomization pleiotropy residual sum and outlier; NFTs = neurofibrillary tangles; NPs = neuritic plaques; PP = pulse pressure; SNPs = single nucleotide polymorphisms; VBI = vascular brain injury; WMBE = Weighted Mode Based Estimator; WME = Weighted Median Estimator.

^*= p < 0.05;

^**= p < 0.01;

^***= p < 0.001.

Genetically predicted higher educational attainment was associated with significantly (1) lower risk of Alzheimer’s disease (OR = 0.64, CI = 0.56, 0.74), (2) delayed AAOS (hazard ratio [HR] = 0.76, CI = 0.67, 0.85), (3) increased cortical surface area (β mm² = 4,600, CI = 3,737.6, 5462.4), and (4) increased cortical thickness (β mm = 0.01, CI = 0, 0.02). Genetically predicted higher diastolic DBP (OR = 1.08, CI = 1.04, 1.11) and PP (OR = 1.06, CI = 1.02, 1.1) were associated with significantly increased risk of VBI. Genetically predicted longer sleep duration was associated with significantly increased cortical thickness after outlier removal (β mm = 0.02, CI = 0.01, 0.02). Genetically predicted smoking status was associated with significantly reduced cortical thickness (β mm = −0.02, CI = −0.03 to −0.01]). Genetically predicted type 2 diabetes was associated with significantly earlier AAOS (HR = 1.07, CI = 1.04, 1.11). Genetically predicted increased low-density lipoproteins (OR = 2.01, CI = 1.39, 2.92) and total cholesterol levels (OR = 1.99, CI = 1.4, 2.84) were associated with significantly increased risk of neuritic plaques. Additionally, increased total cholesterol levels were associated with reduced hippocampal volume (β = −0.06, CI = −0.1 to −0.03]). However, after excluding variants in the APOE locus, these associations between total cholesterol and low-density lipoproteins with neuritic plaques and total cholesterol and hippocampal volume were nonsignificant.

For the risk factors that were causally associated with the AD phenome, bidirectional MR analysis indicated that increased cortical surface area and cortical thickness were significantly associated with increased educational attainment and a higher risk of smoking, respectively.

A further 5 risk factors, including AUDIT, diabetes, BMI, total cholesterol, and insomnia, were causally associated with the AD phenome in the IVW analysis (see Table 4; Fig), however, there was evidence of heterogeneity and the sensitivity analyses were nonsignificant suggesting that the observed associations were not robust to violations of MR underlying assumptions.

Discussion

Using genetic variants as proxies for modifiable risk factors, we applied PRS and MR analyses to investigate the association of putative modifiable risk factors with the AD phenome. PRS for higher educational attainment and DBP were observed to be associated with reduced risk for AD. In the MR analysis, only higher educational attainment was causally associated with a reduced risk of AD. Additionally, education was causally associated with a delayed AAOS, increased cortical surface area, and increased cortical thickness. There was no evidence that education was causally associated with AD neuropathology or CSF biomarkers, supporting the hypothesis that education mitigates dementia risk via cognitive reserve rather than affecting AD pathogenesis.⁴⁷ Bidirectional MR analysis also indicated a bidirectional effect of cortical surface area on education.

The lack of causal associations between modifiable risk factors and AD may reflect heterogeneity in the underlying pathogenesis that can lead to clinical phenotypes analogous to AD. An endophenotype is usually less genetically complex than the disorder it underlies due to the endophenotype being influenced by fewer genetic risk factors than the disease as a whole and reflecting a single pathophysiological pathway of the overall clinical disorder. As endophenotypes can be measured in both cases and controls, there is greater power to detect an association due to the effect allele influencing the endophenotype even in asymptomatic carriers. As such, we expanded our MR analysis to infer causal relationships between modifiable risk factors and AD endophenotypes to evaluate how potential risk factors may influence the underlying pathophysiological pathways of AD. We observed (1) higher total-cholesterol and LDL-cholesterol levels to be causally associated with increased risk of neuritic plaque burden, although causal effects were driven by variants located within the APOE locus; (2) higher DBP and PP causally associated with increased risk of VBI; (3) higher total cholesterol was causally associated with decreased hippocampal volume; (4) smoking status was causally associated with reduced cortical thickness; and (5) longer sleep duration was causally associated with increased cortical thickness.

Observational studies have indicated that lifestyle interventions targeting modifiable risk factors can either prevent or delay the age of onset of dementia. In particular, low educational attainment, hearing loss, traumatic brain injury, hypertension, alcohol consumption, obesity, smoking, depression, physical inactivity, social isolation, and diabetes.^3,4 However, with the exception of educational attainment, our analyses did not provide strong evidence of a causal association with these risk factors and AD or AAOS. The lack of a causal association between these risk factors and AD could be due to insufficient power in our analyses, but, alternatively, may be a result of confounding or reverse causation in observational studies. For instance, increased physical activity is generally associated with a reduced risk of dementia,³ however, a recent meta-analysis found that the protective association with dementia was observed when physical activity was measured < 10 years before dementia diagnosis, but when measured > 10 years before dementia onset no association with dementia was observed – consistent with reverse causation driving the observed protective association.⁴⁸ Additionally, although these risk factors may not be associated with AD pathogenesis, they may be associated with the pathogenesis of other dementia subtypes. For instance, the observed association between blood pressure and VBI suggests that while reducing blood pressure in late life may have limited utility in the prevention of AD, it may reduce the risk of vascular dementia by reducing the risk of VBI and therefore affect the risk for all-cause dementia, but not specifically affect the risk of AD.

The association of modifiable risk factor PRS with clinically diagnosed AD has not been extensively studied, although several studies have conducted phenome-wide scans to evaluate the association of AD PRS with a wide range of diseases and other traits. Using data from the UK Biobank (n = 334,398), Richardson and colleagues found that an AD PRS composed of 124 SNPs and inclusive of APOE (p_t ≤ 5e-05) was associated with 72 of 551 traits (FDR < 0.05).⁶ In particular, a higher AD PRS was associated with lower DBP and BMI, reduced risk of self-reported diabetes, shorter sleep duration, increased risk of self-reported high cholesterol, and increased amount of moderate-physical activity.⁶ Similarly, a second study by Korologou-Linden and colleagues evaluated the association of an AD PRS composed of 18 SNPs, inclusive of APOE (p_t ≤ 5e-08) across 15,403 traits in the UK Biobank (n = 334,968).⁴⁹ A higher AD PRS was associated with 165 traits and, in particular, with lower DBP, lower BMI, increased total cholesterol, levels, reduced risk of self-reported diabetes, increased oily fish consumption, increased sleeplessness or insomnia, reduced sleep duration, increased amount of moderate-physical activity, and increased risk of self-reported depression.⁴⁹ In a follow-up MR analysis of these traits, only moderate-physical activity was observed to be causally associated with an increased risk of AD.⁴⁹

Two earlier studies used MR to evaluate the association of potentially modifiable risk with AD cases-control status.^50,51 First, Østergaard and colleagues evaluated the association of 13 risk factors with AD and observed that higher SBP, high-density lipoprotein (HDL) cholesterol, and smoking quantity were associated with a reduced risk of AD, whereas higher total cholesterol and low-density lipoprotein (LDL) cholesterol were associated with increased risk.⁵¹ No significant associations were observed for BMI, diabetes, insulin resistance, triglycerides, smoking initiation, or education, and after variants in the APOE locus were excluded from the analysis, the cholesterol levels were no longer significantly associated with AD risk.⁵¹ Second, Larsson and colleagues evaluated the association of 22 risk factors with AD, finding that years of education, intelligence, and 25-hydroxyvitamin D were associated with a reduced risk of AD, whereas coffee consumption was associated with increased risk.⁵⁰ No significant associations were observed among alcohol consumption, serum folate, serum vitamin B₁₂, homocysteine, cardiometabolic factors, or C reactive protein with AD.⁵⁰

The results of this study should be interpreted in conjunction with knowledge of its limitations and those of MR in general. First, we are unable to validate the predictive ability of the exposure PRS as individual level data on the exposure traits were unavailable in the ADGC cohorts and we thus assume that each PRS is a good measure of genetic liability for that trait. Second, although we cannot exclude that our findings may be affected by weak instrument bias, the F-statistics for all of the analyses were greater than 10, indicating that the instrument strength was sufficient for MR analysis.⁴² However, in 2-sample MR, weak instrument bias is in the direction of the null, thus, we cannot exclude low power as an explanation for the null results.⁵² In particular, the samples size for the CSF and neuropathological outcomes are small and as such the MR analyses are potentially underpowered. Third, we cannot completely rule out violations of the independence and the exclusion restriction assumption, particularly in regard to pleiotropy.⁵³ Nevertheless, we used several methods to infer robust causal estimates, including outlier removal using MR-PRESSO and WMBE, WME, and MR-Egger sensitivity analyses. Finally, it is assumed that both samples used to generate the GWAS summary statistics used in the MR model come from comparable populations. In evaluating the demographics of the studies used in this analysis, the exposures have an average age of 56.1 to 63.8 years, whereas outcomes, with the exception of hippocampal volume and cortical sufrace area and thickness, have an average age of 71 to 74.7 years. As such, some of the results reported here may be subject to survivor bias.⁵⁴ Nevertheless, the bias introduced by survival effects is large for exposures that strongly affect survival. However, when selection effects are weak or moderate, selection bias does not adversely affect causal estimates.⁵⁴

Despite these limitations, this study has significant strengths. We assessed the causal effect of multiple modifiable factors strongly hypothesized as affecting AD risk. In addition, we used the largest GWAS for AD and the exposure traits available at the time of analysis, allowing us to include the largest possible number of instruments for the exposures, resulting in increased statistical power. Finally, rather than limiting our analyses to AD case/control status, we expanded our MR analysis to include AD endophenotypes.

In conclusion, this study used large exposure and outcome GWAS to conduct PRS and MR analyses to infer causal relationships between potentially modifiable risk factors and the AD phenome. The PRS analysis indicated that a higher genetic predisposition education and DBP influenced AD risk. In the follow-up MR analysis, only genetically predicted higher education was observed to have a causal association with reduced AD risk. Expanding our analysis to additional AD endophenotypes, we observed that higher genetically predicted cholesterol levels and blood pressure were associated with increased risk of neuritic plaque burden and VBI, respectively, suggesting that these risk factors influence the development of neurodegenerative disease pathology. The results from this analysis should be considered within a triangulation of evidence framework by integrating the findings from different study types, such as observational studies and RCTs, all of which have different sources of bias to identify and prioritize modifiable risk factors that are robust.