Vitamin D,Cognition,and Alzheimer’s Disease: Observational and Two-Sample Mendelian Randomization Studies

Observational associations

Using linear regression, we tested the association of log-transformed 25(OH)D with each cognition score (AFT, AFT2, MAT, REYI, REYII, COLTIME) and with the composite cognition score, the PACC score. These scores are described below in Study Outcomes. Since 25(OH)D is confounded by various factors [20] (anthropometric variables, lifestyle habits, and more), we identified covariates available in the CLSA cohort to add to our regression model to partially control for this potential bias. Cognition scores were adjusted for age, sex, smoking status (answered on a questionnaire with possible answers: Yes (I currently smoke), No (I don’t smoke and I never have), Former (I don’t smoke now but I have in the past), high nutritional risk indicator (answered on a questionnaire with possible answers: Not at high nutritional risk, and High nutritional risk), and BMI classification for adults aged 18 and over (the classification includes underweight, normal weight, overweight, obese –Class I, obese –Class II, obese –Class III). Cognition scores were inverse normal transformed. All measures were assessed at baseline and included at most 26,622 individuals. We repeated the analysis stratifying the regression model by sex. To control for multiple testing, we considered Bonferroni significant those associations with p < 7.1×10^-3 (which is equal to 0.05/7, where 7 is the number of cognitive scores tested). All patients gave informed consent in the originalstudy [27].

Instrumental variables for 25(OH)D

Instrumental variables were defined as independent genome-wide significant SNPs (p < 6.6×10^-9) for 25(OH)D levels in 443,734 individuals as per the genome-wide threshold defined by Manousaki et al. (2020) [28]. Identification of independent loci excluded high collinearity (R² > 0.9) SNPs within a 20,000 kilobase window. Details on the conditionally independent approach, imputation quality control, missingness rates, excluded individuals, and software tools employed in the GWAS are provided in detail by Manousaki et al. (2020) [28]. This approach yielded 138 independent SNPs. In order to ensure that we are able to recover all the variants from the outcome GWAS, we restricted our study to SNPs with a MAF > 5%. When a SNP was not present in the results for the GWAS of the PACC, AFT, AFT2, COLTIME, REYI, REYII, and MAT score, we alternatively looked up a proxy SNP in linkage disequilibrium, with an R²≥0.8 using the 1000 G European reference panel [29] (Supplementary Table 2). Therefore, we limited our analysis to 80 SNPs for which we estimated the variance explained in 25(OH)D. We computed the variance explained using the following formula: r² = 2β² f (1 – f), where β and f denote the estimated effect of the SNP on 25(OH)D level and the MAF, respectively. Which we then used to calculate the F-statistic (formula F: r² (N-2)/(1-r²)) where r² is the variance explained and N denotes the sample size.

Next, we repeated our analysis by utilizing instrumental variables chosen in two separate studies and assessed the variants’ pertinence to 25(OH)D levels. These variants were evaluated for their statistical power, biological specificity, and risk of bias caused by horizontal pleiotropy. The first study included 35 variants derived from 307,601 unrelated UK Biobank participants of white European ancestry which had an association which replicated for 25(OH)D in the SUNLIGHT consortium [30] (Supplementary Table 3). The second study included 21 variants derived from 431,489 individuals, comprising 67,992 from the Vitamin D Studies Collaboration (VitDSC), 14,941 from the European Prospective Investigation into Cancer and Nutrition— Cardiovascular Disease (EPIC-CVD), and 348,556 from the UK Biobank [31] (Supplementary Table 4).

Study outcomes

The PACC score is a composite score which tests episodic memory, timed executive function, and global cognition. The composite measure can precisely recognize early signs of cognitive changes prior to mild cognitive impairment (MCI) [13]. Concerning the individual scores, the Rey auditory verbal learning test – trial 1 and 5-min delayed recall (REYI, and REYII respectively) scores represent the number of words correctly recalled within a 90-s span. The first assesses verbal episodic memory, whereas the second assesses delayed recall. The mental alteration test (MAT) score assesses the number of correct consecutive numeric and alphabetical alternations in 30 s. The time to say the ink color of each color name (COLTIME measured in seconds) ascertains the executive and the set switching cognitive domains. Finally, animal fluency (AFT & AFT2) assesses the number animals recited in 60 s. Table 1 summarizes the individual cognitive scores in CLSA.

For each of the cognition scores (PACC, AFT, AFT2, MAT, COLTIME, REYI, REYII) we retrieved the effect size estimates of these 80 variants derived from GWAS performed in the CLSA [21] to perform our MR study. GWAS summary statistics for the cognition scores are available in the GWAS catalogue (See Data Availability section). Since 92.6% of CLSA participants are of European ancestry [27], our analysis focused on individuals with European ancestry individuals (N = 21,065). Population structure was assessed through principal component analysis as detailed elsewhere [27]. We employed REGENIE v3.2.5 [32] to perform association tests, which we adjusted for sex, age, and the first 20 principal components. We stratified our analysis based on age, testing the causal relationship for the entire cohort and for individuals aged 60 and above (N = 13,678). This approach allowed us to compare the effect of age on the cognitive scores.

Mendelian randomization

MR assumptions

Three main assumptions are required in MR analyses. The first MR assumption requires that the genetic instrument must be strongly associated with the exposure. For all sets of instruments, we selected the genome-wide significant variants based on the significance threshold specified in their respective publications (Supplementary Tables 2–4).

The second MR assumption stipulates that genetic instruments should not be correlated with confounders that are a common cause of the exposure and the outcome [34].

The third MR assumption requires the instruments to influence the outcome only via their effects on the exposure (the exclusion restriction assumption), i.e. that horizontal pleiotropy does not influence the results. We, therefore, investigated for the presence of directional horizontal pleiotropy in our study. First, we calculated MR Egger regression slope and intercept, in addition to IVW median and IVW mode. MR-Egger is a weighted linear regression which allows the estimation of an intercept as a measure of the average horizontal pleiotropic effect and produces a slope coefficient that is less prone to bias from directional horizontal pleiotropy. MR-Egger relaxes the exclusion restriction assumption and is valid under the Instrument Strength Independent of Direct Effect (InSIDE) assumption wherein associations of the genetic variants with the 25(OH)D would be independent of direct effects of the genetic variants on the cognitive score. Additionally, we performed a weighted median analysis which weights MR estimates by the effect size/standard error. This approach leverages the fact that causal estimates from non-pleiotropic instrumental variable are more probable to converge toward the median. This approach provides reliable causal estimates when at least 50% of the weight from variants are valid instruments. Additionally, we assessed Cochran’s Q test. The test yields an I² index which gives an estimate of the amount of heterogeneity amongst IVs (I² index≤25%, is considered low heterogeneity, and a I² index > 50% is considered high heterogeneity). Finally, we employed MR-PRESSO global test, which can detect the presence of pleiotropic outlier SNPs. By undertaking this variety of sensitivity analyses, and comparing results using these approaches, each one with different underlying assumptions, we could assess, to some degree, whether our findings were biased by horizontal pleiotropy. We considered results significant when P-values were < 0.05 for all horizontal pleiotropy tests (MR-Egger intercept test, Cochran’s Q test, MR-PRESSO global test).

This study was conducted in accordance with the STROBE-MR guideline [29] (STROBE-MR checklist is provided in Supplementary Material.

Association of late-onset AD polygenic risk score and cognition traits

The effect size estimates of 22 variants associated with the risk of late-onset AD in European ancestry individuals were obtained from the Polygenic Score Catalog (PGS Catalog). Briefly, the PRS was derived using a clumping and thresholding approach from both the UK Biobank and the IGAP cohorts. Further details of the PRS can be obtained from Zhang et al. (2020) [35] and data is available through the accession: PGS000334. We calculated the risk score for everyone in the CLSA and performed a linear regression adjusting for age, sex, and the first 20 principal components for all the cognition traits (PACC, AFT, AFT2, MAT, REYI, REYII, COLTIME). We employed PRSicev2.3.5 [36] for both PRS computation and association.

LD score regression analysis

We next examined the genetic correlation between the PACC score in older individuals and clinically diagnosed AD and AD-by-proxy [37] in European ancestry individuals using summary statistics GWAS. We used linkage-disequilibrium score regression (LDSC v1.0.1) [38] to calculate the genetic correlation between the PACC score in older adults and AD. We performed the analysis using 1000Gp3v5 LD reference data and high-quality variants from HapMap3 as recommended by the LDSC software package.

Observational associations

All of the associations (Table 2) between log-transformed 25(OH)D and the seven cognitive scores (the individual scores and the composite score) adjusted for age, sex, smoking status, nutritional index, and BMI classification were Bonferroni significant (p < 7.1×10^-3). While the direction of effect between 25(OH)D and most of the cognition scores was positive, we observed a negative association between the vitamin D and the performance in the executive and set switching tests. These tests gauge the time taken to name ink colors (COLTIME). This means that higher 25(OH)D levels were associated with improved COLTIME. Thus higher 25(OH)D was consistently associated with improved cognition in observational associations. We repeated the analyses above, but this time stratifying by sex. Our findings revealed that the direction of effect for all the cognition traits remained consistent with the model in which we controlled for sex. The noticeable difference is that the results were more significant in males than females.

Instrumental variables for 25(OH)D

We used 80 common (MAF > 5%) independent SNPs (Supplementary Table 2) explaining 3.8% of the variance in 25(OH)D levels, with an F-statistic of 214.75 as instruments in our MR study. 16 SNPs were absent from the PACC score GWAS and were replaced by proxies in high LD (r² > 0.8). TwoSampleMR (v 0.5.6) identified rs10832289, rs11182550, and rs13336347 as palindromic and were therefore removed from further analyses. An additional 5 SNPs (rs145432346, rs200641845, rs3775150, rs201501563, rs62130059) were removed from further analyses because proxies in the PACC score GWAS could not be found. Of the 35 variants from Sutherland et al. (2022) [30], 5 were replaced by proxies (Supplementary Table 3). Out of the 21 variants from Emerging Risk Factors Collaboration/EPIC-CVD/Vitamin D Studies Collaboration (2021) [31], one (rs187639972) could not be retrieved in the exposure GWAS, and 4 were replaced by proxies (Supplementary Table 4).

GWAS of the PACC score

We performed a genome-wide study of the PACC score in European ancestry individuals from the CLSA. Our analysis was stratified by age, encompassing individuals aged 60 years and above (N = 13,678), as well as the entire cohort (N = 21,065). When we restricted our analysis to individuals aged 60 and older, we found a strong association (p = 3.5×10^– 7) between the PACC score and the well-characterized missense apolipoprotein APOE variant, rs429358 (Supplementary Figure 1). When we did not restrict by age, the association was attenuated (p = 2.5×10^– 4). This locus plays an important role in dementia and AD[39–41].

Mendelian randomization: Cognitive scores as outcomes

For the PACC score, using the set of 80 instruments selected based on MAF and p, MR showed that the genetically predicted increase per 1 SD in natural log of 25(OH)D had no clear estimated effect on the PACC score estimated in the general population (Beta_PACC–All = –0.01, 95% CI –0.06–0.03, p = 0.53). Results were consistent in individuals over 60 (Beta_{PACC–Age >60} = 0.03, 95% CI –0.028–0.08, p = 0.35) (Fig. 2A, Tables 3 and 4). Using the set of instruments from Sutherland et al. (2022) [30], MR also showed no estimated effect on the PACC score (Beta_PACC–All = –0.01, 95% CI –0.05 –0.03, p = 0.78; Beta_{PACC–Age >60} = 0.03, 95% CI –0.02–0.08, p = 0.22) (Fig. 2B). Finally, using the set of instruments from the Emerging Risk Factors Collaboration/EPIC-CVD/Vitamin D Studies Collaboration (2021) [31], MR showed no effect on the PACC score (Beta_PACC–All = –0.01, 95% CI –0.05–0.03, p = 0.71; Beta_{PACC–Age >60} = 0.04, 95% CI 0.0007–0.08, p = 0.05) (Fig. 2C).

OPEN IN VIEWER

Fig. 2

Forest plot of MR studies investigating the genetically predicted increase in the natural log 25(OH)D on PACC, AFT, AFT2, REYI, REYII, COLTIME, and MAT scores. Forest plots of the main study and the sensitivity analyses. Significance threshold as a Bonferroni p < 7.1×10^– 3. MR, Mendelian randomization; IVW(re), random effect inverse weighted variance; Wmedian, weighted median; 25(OH)D, 25-hydroxyvitamin. PACC, AFT, AFT2, REYI, REYII, COLTIME, and MAT are defined in Table 1. A) Forest plot utilizing 80 instruments chosen from the pool of 138 SNPs that are conditionally independent and possess a minor allele frequency (MAF) greater than 5%, while also having a p < 6.6×10^– 9. B) Forest plot utilizing 35 instruments from Sutherland et al. (2022) [30]. C) Forest plot utilizing 21 instruments from the Emerging Risk Factors Collaboration/EPIC-CVD/Vitamin D Studies Collaboration (2021) [31].

OPEN IN VIEWER

Fig. 3

Scatter plots of the genetically predicted increase in natural log 25(OH)D MR analyses. A) PACC score for all individuals. B) PACC score for individuals aged 60 years and above. A genetically predicted increase in the natural log of 25(OH) D does not have a clear effect on a 1 SD of PACC score (for all individuals: beta = –0.02, 95% CI: –0.060– 0.030, p = 0.53; for individuals 60 years and more: beta = –0.03, 95% CI: –0.08–0.03, p = 0.33. Panels A and B presents the MR results which utilize 80 instruments from the pool of 138 SNPs that are conditionally independent and possess a minor allele frequency (MAF) greater than 5%, while also having a p < 6.6×10^– 9. Panels C and D presents the MR results which utilize 35 instruments from Sutherland et al. (2022) [30]. Panels E and F presents the MR results which utilize 21 instruments previously published in Emerging Risk Factors Collaboration/EPIC-CVD/Vitamin D Studies Collaboration (2021) [31]. Each dot represents a genetic instrumental variable. Three lines represent causal estimate (βIV) by the inverse variance weighted (IVW) method (pink), MR– Egger method (blue), and Weighted median method (orange). Error bars represent 95% CIs. MR, Mendelian randomization. For visualization purposes, (B) the IVW line was dashed since it overlaps with the MR-Egger line.

For AFT, AFT2, REYI, REYII, MAT, and COLTIME, the MR results were consistent with those from the PACC score. There was no discernible estimated effect of genetically predicted increase per 1 SD in the natural log of 25(OH)D on any of these scores (Tables 3 and 4).

Sensitivity analysis

We computed MR-PRESSO, Egger’s regression, and Cochran’s Q test to assess the robustness of our results. Our sensitivity analyses shown in Table 5 for the PACC score and the other cognitive scores show that by and large, there is low to moderate heterogeneity since I² values were below 29%. The only moderately elevated heterogeneity value (I² = 28.3%; I² _Pvalue  = 0.014) was for the AFT score when considering individuals in the entire cohort regardless of age. Additionally, the MR-PRESSO global test, which can detect horizontally pleiotropic outlier SNPs, found pleiotropic SNPs for the AFT score in individuals 60 years and older. Assessing the distortion of the causal estimate before and after outlier correction, the MR results changed. We found that MR-Egger intercept test, MR-PRESSO global test and distortion test did not detect horizontal pleiotropy (p-values > 0.05). We therefore did not identify evidence of directional horizontal pleiotropy in the MR-results.

Associations of Alzheimer’s disease PRS and cognition traits

We tested the association between seven cognition scores (PACC, AFT, AFT2, COLTIME, MAT, REYI, and REYII) and the AD PRS in the strata of individuals over 60 years of age (Fig. 4A). Most of the cognition scores were associated with the AD PRS. The direction of effect of the association for each significant trait was negative, except for COLTIME (Supplementary Table 5), as expected, given that this cognitive test provides higher scores for worsened performance. Taken together, this suggests that in individuals over 60, the AD PRS was associated with worsened cognition. Such associations were generally attenuated when taking the entire cohort into account.

OPEN IN VIEWER

Fig. 4

Association of Alzheimer’s disease PRS Cognition scores with cognition scores in CLSA. A) Demonstrates the Alzheimer’s disease PRS association among individuals aged 60 and above (N = 14,234–14,906). B) Illustrates the Alzheimer’s disease PRS association for all individuals (N = 21,840–22,381). The chart displays the associations between polygenic risk scores (PRS) for various cognitive measures including animal fluency score (AFT), animal fluency score 2 (AFT2), Stroop score (COLTIME), mental alteration score, preclinical Alzheimer’s cognitive composite score (PACC), Rey auditory verbal learning score 1 (REYI), and Rey auditory verbal learning score 2 (REYII). The y-axis represents the -log10 p-value indicating the association strength between Alzheimer’s disease PRS and each cognition score. Each bar is annotated with the corresponding p-value representing the association between Alzheimer’s disease PRS and the cognition score. All associations are adjusted for age, sex, and the first 20 principal components.

Genetic correlation of Alzheimer’s disease and the PACC score in older individuals

To quantify the shared genetic etiology of cognition, as measured by the PACC score and AD, we used LD score regression. We found a negative genetic correlation with AD (r _g  = –0.39 [95% CI –0.53–0.23], p = 0.01 [LDSC]). This suggests that the genetic factors improving cognition, as measured by the PACC score tended to be associated with a reduced risk of AD.