Mendelian randomization highlights significant difference and genetic heterogeneity in clinically diagnosed Alzheimer’s disease GWAS and self-report proxy phenotype GWAX

Until now, Mendelian randomization (MR) studies have investigated the causal association of risk factors with Alzheimer’s disease (AD) using large-scale AD genome-wide association studies (GWAS), GWAS by proxy (GWAX), and meta-analyses of GWAS and GWAX (GWAS+GWAX) datasets. However, it currently remains unclear about the consistency of MR estimates across these GWAS, GWAX, and GWAS+GWAX datasets. Here, we first selected 162 independent educational attainment genetic variants as the potential instrumental variables (N = 405,072). We then selected one AD GWAS dataset (N = 63,926), two AD GWAX datasets (N = 314,278 and 408,942), and three GWAS+GWAX datasets (N = 388,324, 455,258, and 472,868). Finally, we conducted a MR analysis to evaluate the impact of educational attainment on AD risk across these datasets. Meanwhile, we tested the genetic heterogeneity of educational attainment genetic variants across these datasets. In AD GWAS dataset, MR analysis showed that each SD increase in years of schooling (about 3.6 years) was significantly associated with 29% reduced AD risk (OR=0.71, 95% CI: 0.60–0.84, and P=1.02E−04). In AD GWAX dataset, MR analysis highlighted that each SD increase in years of schooling significantly increased 84% AD risk (OR=1.84, 95% CI: 1.59–2.13, and P=4.66E−16). Meanwhile, MR analysis suggested the ambiguous findings in AD GWAS+GWAX datasets. Heterogeneity test indicated evidence of genetic heterogeneity in AD GWAS and GWAX datasets. We highlighted significant difference and genetic heterogeneity in clinically diagnosed AD GWAS and self-report proxy phenotype GWAX. Our MR findings are consistent with recent findings in AD genetic variants. Hence, the GWAX and GWAS+GWAX findings and MR findings from GWAX and GWAS+GWAX should be carefully interpreted and warrant further investigation using the AD GWAS dataset.

Importantly, all these AD GWAS, GWAX, and GWAS+GWAX summary datasets are publicly available, which provides strong data support to investigate the causal association between AD and previously reported risk factors using a Mendelian randomization (MR) design [16][17][18]. Meanwhile, there is no strict standard to limit the use of AD GWAS, GWAX, or GWAS+GWAX datasets in MR studies. Hence, some MR studies are based on the AD GWAS datasets from IGAP 2013 or IGAP 2019 [19][20][21][22][23][24][25], of which three MR studies had evaluated the impact of multiple modifiable risk factors on AD [16][17][18]. In 2017, Larsson and colleagues selected 24 potentially modifiable risk factors and found that only increased educational attainment was significantly associated with a reduced risk of AD [16]. In 2020, Wang and colleagues analyzed 45 potentially modifiable risk factors, and eventually highlighted educational attainment and urate levels [17]. In 2021, Andrews and colleagues selected 22 risk factors, and eventually identified educational attainment using polygenic risk scores (PRS) and MR [18]. Meanwhile, others MR studies are based on the AD GWAS+GWAX datasets [26][27][28][29][30][31][32]. However, it currently remains unclear about the consistency of MR estimates from AD GWAS, GWAX, and GWAS+GWAX datasets.
Here, we selected the educational attainment as the risk factor, and evaluated its impact on AD risk using largescale AD GWAS, GWAX, or GWAS+GWAX datasets, as the causal association between educational attainment and AD had been well established in recent MR studies [16][17][18]33].

Study design
MR is based on three principal assumptions. First, the instrumental variables (genetic variants) should be significantly associated with the exposure (educational attainment), generally achieving the genome-wide significant level (P<5.00E-08) [19]. Second, instrumental variables should not be associated with confounders of the exposure (educational attainment) and outcome (AD) [19]. Third, instrumental variables should affect the risk of the outcome (AD) only through exposure (educational attainment) [19]. The second and third assumptions are collectively known as independence from pleiotropy. Figure 1 provides a flow chart about our MR study design.

Educational attainment genetic variants
We selected 162 independent educational attainment genetic variants with the genome-wide significance (P < 5.00E−08) to be the potential instrumental variables, as provided in supplementary Table 1 [34]. Educational attainment is a continuous variable, which is determined by the number of years of schooling completed at or above age 30 [34]. These 162 genetic variants are identified by a large-scale GWAS in 405,072 individuals of European descent including 293,723 individuals in discovery stage (SSGAC) and 111,349 individuals in the independent replication stage (UK Biobank) [34].

AD GWAS dataset
We selected the clinically diagnosed AD GWAS dataset from IGAP 2019 stage 1 including 21,982 AD cases and 41,944 cognitively normal controls of European descent [9]. This GWAS dataset is based on a meta-analysis using participants from four AD consortia including Alzheimer Disease Genetics Consortium, Cohorts for Heart and Aging Research in Genomic Epidemiology Consortium (CHARGE), The European Alzheimer's Disease Initiative (EADI), and Genetic and Environmental Risk in AD/ Defining Genetic, Polygenic and Environmental Risk for Alzheimer's Disease Consortium (GERAD/PERADES) [9]. Table 1 provides the demographic profile about the AD GWAS dataset.

AD GWAX datasets
We selected two GWAX datasets for AD proxy phenotype including GWAX 2018 [13], and GWAX 2021 [15], respectively. Both studies are based on the UK Biobank participants [35]. UK Biobank is a large national and international health resource including the genetic and phenotype data from 502,536 community-dwelling individuals aged between 37 and 73 years recruited in the United Kingdom between 2006 and 2010 [35]. In UK Biobank, a proxy phenotype for AD case-control status was assessed via self-report [13]. Participants were asked to report "Has/did your father or mother ever suffer from Alzheimer's disease/dementia?" [13]. Table 1 provides the demographic profile about these two AD GWAX datasets.

Establishing the Wald estimator
For the same effect allele from each educational attainment genetic variant G j (j = 1, …, 162), we assume that we have successfully extracted their corresponding summary results in educational attainment GWAS dataset including the beta coefficients and their standard errors ( β Xj , se β Xj ), and in the AD GWAS, GWAX and GWAS+GWAX datasets including the beta coefficients and their standard errors ( β Yj , se β Yj ). For a given genetic variant, the causal effect of educational attainment on AD can be consistently estimated as a simple ratio, also called the Wald estimator θ j =β Yĵ β Xj and its Xj [19,36].

MR analysis
Using the Wald estimator from each educational attainment genetic variant, we conducted the MR analysis using four MR methods including inverse-variance weighted (IVW), weighted median, MR-Egger, and MR-PRESSO (Mendelian Randomization Pleiotropy RESidual Sum and Outlier) to combine the variant-specific estimates and get the overall estimate [36][37][38][39]. IVW is the main MR analysis method, which combines the variant-specific Wald estimators by taking the inverse of their approximate variances as the corresponding weights [37]. Weighted median could derive consistent estimates when up to 50% of instruments are not valid [37]. MR-Egger could test the presence of potential pleiotropy and account for this potential pleiotropy using the MR-Egger intercept test [36]. MR-PRESSO could detect and correct for the horizontal pleiotropy via outlier removal (the MR-PRESSO outlier test) [40]. The odds ratio (OR) as well as 95% confidence interval (CI) of AD corresponds to about per 3.6 years increase (about 1 standard deviation (SD)) in EduYears. All the statistical tests were completed using R Packages "MendelianRandomization" [39] and "MR-PRESSO" [40]. The significance threshold is P < 0.05.

Heterogeneity test and paired-samples T test
We performed a heterogeneity test of the Wald estimators using the Cochran's Q statistic and the I 2 statistic [41]. Cochran's Q statistic approximately follows a χ 2 distribution with k-1 degrees of freedom (k is the number of the selected studies) [42].
Q × 100% ranges from 0 to 100%, with 0-25%, 25-50%, 50-75%, and 75-100% corresponding to low, moderate, large and extreme heterogeneity, respectively [42]. Importantly, the Cochran's Q statistic and I 2 assume that the subjects are independent of one another and were selected at random from a larger population. Hence, we only test the heterogeneity in AD GWAS and GWAX including two comparisons GWAS vs. GWAX 2018, and GWAS vs. GWAX 2021. All statistical tests were completed using R Package "meta: General Package for Meta-Analysis. " In line with the heterogeneity test, we further conducted the paired-samples T test to evaluate the average differences of Wald estimators in GWAS vs. GWAX 2018, and GWAS vs. GWAX 2021. Analysis of variance (ANOVA) is widely used to analyze the differences among means from multiple independent (unrelated) groups. However, the selected AD GWAS, GWAX, and GWAS+GWAX datasets are not completely independent with each other. Therefore we could not provide the overall differences among the mean effect sizes across the AD GWAS, GWAX, and GWAS+GWAX datasets. Here, we provide a combined plot using all shared genetic variants to visualize the differences in effect sizes and directions across the six datasets.

MR analysis in AD GWAS dataset
We extracted the GWAS summary statistics of 159 educational attainment genetic variants in the AD GWAS dataset, as provided in supplementary Table 2. Using the MR-Egger intercept test, we did not identify any significant pleiotropic variant. Using MR-PRESSO Global Test, we found two genetic variants rs268134 and rs28420834 to be the pleiotropic variants (Table 2). Hence, we excluded both variants in MR analysis. IVW showed that each SD increase in years of schooling (about 3.6 years) was significantly associated with 29% reduced AD risk (OR=0.71, 95% CI: 0.60-0.84, and P=1.02E−04). Interestingly, evidence from weighted median, MR-Egger, and MR-PRESSO further supported this finding, as provided in Table 3. Meanwhile, all the MR estimates from these four methods are consistent in terms of direction.

MR analysis in AD GWAX datasets
We extracted the GWAS summary statistics of 147 and 159 educational attainment genetic variants in GWAX 2018, and GWAX 2021, respectively, as provided in supplementary  (Table 2). MR analysis indicated no significant causal association between educational attainment and AD, as provided in Table 3. However, two methods showed an increased trend of AD risk with high educational attainment including IVW (OR=1.09, 95% CI: 1.00-1.19, and P=0.051), and MR-PRESSO (OR=1.09, 95% CI: 1.00-1.20, and P=0.053). In GWAX 2021, rs268134 is identified to be a pleiotropic variant and then excluded in MR analysis (Table 2). IVW highlighted that each SD increase in years of schooling could significantly increase 84% AD risk (OR=1.84, 95% CI: 1.59-2.13, and P=4.66E −16). The MR estimates from weighted median, MR-Egger and MR-PRESSO were consistent with the IVW estimate in terms of direction and magnitude (Table 3).

MR analysis in AD GWAS+GWAX datasets
We extracted the GWAS summary statistics of 147, 155, and 159 educational attainment genetic variants in GWAS+GWAX 2018, GWAS+GWAX 2019, and GWAS+GWAX 2021, respectively, as provided in supplementary Table 5-7. Only in GWAS+GWAX 2021, MR-PRESSO indicated rs268134 to be a pleiotropic variant, and then excluded in MR analysis ( Table 2). In GWAS+GWAX 2018, MR analysis showed no significant causal association between educational attainment and AD (Table 3). In GWAS+GWAX 2019, we found that high educational attainment could reduce the risk of AD (Table 3). In GWAS+GWAX 2021, MR analysis further supported the increased risk of AD with high educational attainment using four MR methods, as provided in Table 3. IVW showed that each SD increase in years of schooling could significantly increase 22% AD risk (OR=1.22, 95% CI: 1.08-1.36, and P=1.00E−03). The MR estimates from weighted median, MR-Egger and  MR-PRESSO were consistent with the IVW estimate in terms of direction and magnitude (Table 3).

MR analysis in AD GWAS, GWAX, and GWAS+GWAX datasets using the same genetic variants
These above MR analyses were based on different educational attainment genetic variants as the instrumental variables in AD GWAS, GWAX, and GWAS+GWAX datasets. We further conducted a secondary MR analysis using the same educational attainment genetic variants in each AD GWAS, GWAX, and GWAS+GWAX datasets. Finally, we selected 143 same educational attainment genetic variants available across all datasets by excluding rs268134 and rs28420834, as both are pleiotropic variants. Interestingly, the secondary MR analysis supported these above findings, as provided in Table 4. Figure 2 is a combined plot, which visualizes the differences in effect sizes and directions across the six datasets using the 143 genetic variants.

Discussion
Until now, MR methods had been widely used to determine the causal association between AD and previously reported risk factors using the AD GWAS, GWAX, or GWAS+GWAX summary datasets [16][17][18][19][20][21][22][23][24][25][26][27][28][29][30][31][32]. However, the consistency of MR estimates from AD GWAS, GWAX, or GWAS+GWAX datasets currently remains unclear. Here, we first evaluated the causal association of a well-established risk factor educational attainment with the risk of AD using large-scale GWAS, GWAX, or GWAS+GWAX datasets [16][17][18]33]. MR analysis in the AD GWAS dataset showed that high educational attainment could significantly reduce the risk of AD, which is consistent with recent findings [16][17][18]33]. However, MR analysis indicated no significant causal association between educational attainment and AD in AD GWAX 2018 dataset, and even showed that high educational attainment increased the risk of AD in the AD GWAX 2021 dataset. Meanwhile, MR analysis in AD GWAS+GWAX datasets suggested the ambiguous findings about the causal association between educational attainment and AD. Hence, all these above findings indicated the inconsistency of MR estimates in AD GWAS, GWAX, and GWAS+GWAX datasets. Interestingly, our findings are consistent with recent findings in AD genetic variants. Andrews and colleagues recently summarized and discussed 40 AD susceptibility loci with genome-wide significance, which were identified by at least one of the four studies [10], including GWAS 2013 [6], GWAS+GWAX 2018 [13], GWAS+GWAX 2019 [14], and GWAS 2019 [9]. They found that only 15 were replicated across all the four studies, and 9 were replicated in two or three studies at full genome-wide significance [10]. Table 4 MR analysis of the association between educational attainment and AD using the same educational attainment genetic variants The significance of the association between educational attainment and AD was at P < 0.05 CI confidence interval, IVW, inverse-variance weighted, MR-PRESSO Mendelian randomization pleiotropy residual sum and outlier, GWAS genome-wide association studies, GWAX GWAS by proxy, GWAS+GWAX meta-analyses of GWAS and GWAX We further test the genetic heterogeneity of educational attainment genetic variants across the GWAS, GWAX, and GWAS+GWAX datasets. Interestingly, heterogeneity test indicated evidence of genetic heterogeneity across the GWAS, GWAS, and GWAS+GWAX datasets. Importantly, GWAS vs. GWAX 2021 comparison highlighted the largest number of genetic variants with heterogeneity. Hence, the genetic heterogeneity may have caused the opposite directions about the causal association between educational attainment and AD in clinically diagnosed AD and self-report proxy phenotype.
Our findings are consistent with recent hypothesis that the phenotypic heterogeneity may cause the genetic heterogeneity, and further reduce the statistical power for GWAX and GWAS+GWAX [10,13]. It is known that the AD GWAX is based on the self-report AD proxy phenotype from UK Biobank participants [10,13]. GWAX theoretically could increase the large-scale sample size into traditional AD GWAS, and further improve the statistical power [10,13]. In fact, not all UK Biobank participants could discriminate AD from other dementia subtypes, and exactly reflect the clinically diagnosed AD status, considering the different presentations and genetic architectures [13]. Hence, the incorrect AD classification may reduce the statistical power to detect the true AD risk loci, and further influence the loci uncovered using GWAX and GWAS+GWAX [13]. This hypothesis may explain why only a small fraction of AD susceptibility loci could be replicated across the AD GWAS, GWAX, and GWAS+GWAX datasets. We think that this hypothesis and our findings from heterogeneity test may explain the inconsistency of MR estimates in AD GWAS, GWAX, and GWAS+GWAX datasets. Meanwhile, the biological factors and medical interventions may also have dramatically different effects on different people [43].
Our MR study may have several strengths. First, we selected one large-scale educational attainment GWAS dataset, and six large-scale AD GWAS, GWAX, or GWAS+GWAX datasets, which may provide ample power to detect the causal association between educational attainment and the risk of AD, as reported by recent MR studies [16][17][18]33]. Importantly, all these participants are of European descent, which may further reduce the influence from population stratification. Third, we selected multiple MR methods and tested the pleiotropy. Hence, the MR assumptions did not seem to be violated. Fourth, educational attainment is well-established AD risk factor, as reported by recent MR studies [16][17][18]33]. Hence, evaluating the impact of educational attainment on AD risk may exactly reflect the consistency of MR estimates from AD GWAS, GWAX, or GWAS+GWAX datasets.

Limitations
Our MR study may have some limitations. First, we only selected one AD risk factor educational attainment to evaluate the consistency of MR estimates across AD GWAS, GWAX, or GWAS+GWAX datasets. In fact, Fig. 2 A combined plot visualizing the differences in effect sizes and directions across the six datasets using the 143 genetic variants. GWAS, genome-wide association studies; GWAX, GWAS by proxy; GWAS+GWAX, meta-analyses of GWAS and GWAX several risk factors have been identified to be causally associated with AD risk [23,26,27,[44][45][46][47][48][49][50][51][52]. Hence, our findings should be further verified using other well-established AD risk factors. Second, the educational attainment of GWAS is based on the meta-analysis of SSGAC (293,723 individuals) and UK Biobank (111,349) [34]. Hence, the educational attainment GWAS dataset and GWAX, or GWAS+GWAX may have the overlapped individuals, and may not be independent. Hence, our MR findings from GWAX, and GWAS+GWAX should be further evaluated using independent datasets.

Conclusions
In summary, our MR analysis highlighted the difference and genetic heterogeneity in clinically diagnosed AD and self-report proxy phenotype using large-scale AD GWAS, GWAX, and GWAS+GWAX summary datasets. Hence, the GWAX and GWAS+GWAX findings and MR findings from GWAX and GWAS+GWAX should be carefully interpreted and warrant further investigation using the AD GWAS dataset.