Leveraging DNA methylation quantitative-trait loci to characterize the relationship between metabolic traits and Alzheimer’s disease

Background: The relationship between DNA methylation, common metabolic risk and Alzheimer’s disease (AD) is not well understood. Methods: Summary statistics integrating DNA methylation quantitative trait loci (mQTLs) and several genome-wide association study data were used. Network with bidirectional mendelian randomization (MR) analysis was performed to examine the causal association among metabolic traits, DNA methylation and AD. Results: Our study showed that cis-mQTLs determined DNA methylation to higher total cholesterol (TC) was associated with higher AD risk (β [95% CI] =0.007 [0.002-0.013], P=0.005). The ndings were robust in sensitivity analyses with different instrumental variables. We found no evidence to support causal associations of cis-mQTLs determined obesity and T2D with AD, and vice versa. Conclusion: Overall, our study showed that the cis-mQTLs determined DNA methylation to higher TC was associated with higher AD risk, whereas the relation of cis-mQTLs determined AD and metabolic dysregulation was unlikely to be causal. studies showing no or controversial association between dyslipidemia and AD risk In our study, cis-mQTLs determined DNA methylation to higher TG is associated with a higher AD risk (i.e., SNP → DNA methylation → higher TC → AD). Several studies reported that lipid-lowering medications-statins were of a protective effect against the development of AD 37]. In addition, high cholesterol in late life was associated with decreased AD risk [35, 36], which may be explained by the timing of the cholesterol measurements in relationship to age and the clinical onset of AD. Taken together, inherited lifetime exposure to higher TC was associated with higher AD risk, however, the patients with higher TC and took lipid-lowering medications might reduce the risk of AD. Further analysis of AD outcomes in total cholesterol intervention trials is warranted. In addition, our ndings suggest the imperative need for further investigation of the possibility that lipid-lowering medications might be of an independent effect on the prognosis of AD risk without dyslipidemia.

First, the causal effects of cis-mQTL determined DNA methylation on metabolic risk factors and AD are estimated (I). Next, the causal effects of cis-mQTLs determined metabolic traits on AD are analyzed (II). Finally, the causal effects of cis-mQTLs determined AD on metabolic traits are estimated (III).

Statistical analysis
The rst two-sample MR analyses was aimed to evaluate the potential causal relationship between DNA methylation at 41,760 CpG sites (The threshold of statistical signi cance was P < 1 × 10 − 10 , as shown in Figure S1) and 9 complex traits. We only investigated CpG sites using cis-mQTL with the lowest P-value in order to improve the speci city of the IV for each CpG. Therefore, MR effect estimates were calculated using the Wald ratio. Then these cis-mQTLs corresponding to the casual DNA methylation for each mediator were as IVs removing the overlapping cis-mQTLs between AD and metabolic traits for further MR analysis.
The MR analysis required that the cis-mQTL SNPs in the instruments were not of horizontal pleiotropy (a genetic variant affects the outcome via a different biological pathway from the mediator under investigation) and in linkage equilibrium (LD). IVs for each mediator were composed of independent cis-mQTLs pruned at LD r 2 < 0.1. The LD proxies were de ned using 1000 genomes European samples [25]. We conducted sensitivity analyses based on the cis-mQTL SNPs removing the overlapping cis-mQTLs among metabolic traits and based on the prior cis-mQTL SNPs pruned at LD r 2 < 0.1.
For forward MR, we used a conventional inverse-variance weighted (IVW) MR analysis, in which the SNPs-AD estimate was regressed on the SNPs-metabolic traits estimate with the intercept term set to zero, weighted by the inverse-variance of SNPs-AD estimate, and vice versa for reverse MR. The heterogeneity between SNPs was estimated by Cochran Q statistic. Random-effects IVW model was used if heterogeneity existed, otherwise xed-effects IVW model was performed. In addition, the MR-Egger method was used to assess the robustness of estimates to potential violations of the standard IV assumptions attributing from the directional pleiotropy.

Results
As shown in Table S2-10, we identi ed 3597 causal CpGs for BMI, 2420 for WHR, 2307 for FG, 2157 for FI, 24 for HDL, 3200 for LDL, 4100 for TC, 3008 for TG, and 2806 for AD, were identi ed in this MR analysis. These identi ed cis-mQTLs determined DNA methylation could be potential IVs for metabolic traits and AD.
The causal CpGs that overlapped between the metabolic traits and AD is presented in Fig. 2. Of note, there were 291 causal CpGs that overlapped between BMI and AD, 143 (49.14%) of which changed in the same direction of BMI and AD ( Figure S2 and Table S11). There were 179 causal CpGs that overlapped between WHR and AD, 80 (44.69%) of which changed in the same direction of WHR and AD ( Figure S3 and Table S12). There were 152 causal CpGs that overlapped between FG and AD, 60 (40.79%) of which changed in the same direction of FG and AD ( Figure S4 and Table S13). There were 121 causal CpGs that overlapped between FI and AD, 58 (47.93%) of which changed in the same direction of FI and AD ( Figure S5 and Table S14). There was only one causal CpG that overlapped between HDL and AD, which performed the same direction of HDL and AD (Figure S6and Table S15). There were 264 causal CpGs that overlapped between LDL and AD, 151 (57.20%) of which changed in the same direction of LDL and AD (Figure S7and Table S16). There were 95 causal CpGs that overlapped between TC and AD, 45 (47.37%) of which changed in the same direction of TC and AD ( Figure S8 and Table S17). There were 234 causal CpGs that overlapped between TG and AD, 108 (46.15%) of which changed in the same direction of TG and AD ( Figure S9 and Table S18). There were 1803 causal CpGs that were speci c to AD, accounting for 64.26% of the total.
We further performed a bidirectional MR to evaluate the causal associations between metabolic traits and AD. For forward MR analysis, we used the causal cis-mQTL SNPs associated with each metabolic trait as IVs removing the overlapping cis-mQTL between AD and metabolic traits for further MR analysis. We  Table 1, the cis-mQTLs determined DNA methylation to higher TC was associated with higher AD risk (β [95%CI] = 0.007 [0.002-0.013], P = 0.005). In addition, MR-Egger regression showed no evidence of directional pleiotropy for the associations of TC with AD (β [intercept] = 0.000, P = 0.417). The ndings were robust in sensitivity analyses with different IVs. The cis-mQTLs determined DNA methylation to higher BMI (β [95%CI] = 0.010 [0.002-0.016], P = 0.010) and HDL (β [95%CI] = 0.100 [0.022-0.177], P = 0.011) were associated with higher AD risk in sensitivity analyses, however, the ndings were not robust with different IVs. In addition, MR-Egger regression showed evidence of directional pleiotropy for the associations of HDL with AD (β [intercept] =-0.001, P = 0.017). Therefore, the relationship of HDL and BMI with AD seemed to be less clear as compared with TC in our study. In addition, no evidence supported the causal associations of cis-mQTLs determined FG and FI with AD. IVs1 is based on cis-mQTLs corresponding to the casual DNA methylation for each metabolic trait removing the overlapping cis-mQTL between AD and metabolic traits.
IVs2 is based on cis-mQTLs corresponding to the casual DNA methylation for each metabolic trait removing the overlapping cis-mQTL among metabolic traits.
Beta is the estimated effect size. IVs1 is based on cis-mQTLs corresponding to the casual DNA methylation for each metabolic trait removing the overlapping cis-mQTL between AD and metabolic traits.
IVs2 is based on cis-mQTLs corresponding to the casual DNA methylation for each metabolic trait removing the overlapping cis-mQTL among metabolic traits.
Beta is the estimated effect size. Taking these putative associations forward, we evaluated the potential for reverse causal relationships by performing MR of AD against metabolic traits. For reverse MR analysis, we used the casual cis-mQTL SNPs associated with AD as IVs removing the overlapping cis-mQTLs between AD and metabolic traits.
Since we were using different databases, we couldn't extract all candidate cis-mQTL SNPs as IVs. Therefore, we considered AD to BMI (N SNPs = 1807), AD to WHR (N SNPs = 2040), AD to FG (N SNPs = 1930), AD to FI (N SNPs = 1940), AD to HDL-cholesterol (N SNPs = 1980), AD to LDL-cholesterol (N SNPs = 1778), AD to TC (N SNPs = 1953), and AD to TG (N SNPs = 1784). As shown in  IVs1 is based on cis-mQTLs corresponding to the casual DNA methylation for AD removing the overlapping cis-mQTL between AD and metabolic traits.
IVs2 is based on cis-mQTLs corresponding to the casual DNA methylation for AD removing the overlapping cis-mQTL among metabolic traits.
Beta is the estimated effect size. IVs1 is based on cis-mQTLs corresponding to the casual DNA methylation for AD removing the overlapping cis-mQTL between AD and metabolic traits.
IVs2 is based on cis-mQTLs corresponding to the casual DNA methylation for AD removing the overlapping cis-mQTL among metabolic traits.
Beta is the estimated effect size.

Discussion
To our knowledge, this is the rst study to investigate the causal relationship between DNA methylation, metabolic traits, and AD using a network with bidirectional MR design integrating cis-mQTLs and summary GWAS data. Our study showed cis-mQTLs determined DNA methylation to higher TC was associated with higher AD risk, whereas the relation of the cis-mQTL determined AD and metabolic traits were unlikely to be causal.
The causal relationship between metabolic indicators and AD has been widely explored, and the ndings showed that metabolic dysregulation related genetic risk factors did not predict AD risk [17,20,21,[26][27][28]. There are two possible explanations for these negative results. First, the ndings might be faced with the bias of weak IVs due to the complexity of metabolic dysregulation and the small fraction of metabolic dysregulation variance accounting for the genetic variants [20,21,28,29]. Completely ruling out an alternative direct causal pathway is a challenge for all MR analyses [30,31], particularly for complex traits determined by both multiple genetic variants and complexly environmental exposures. Therefore, a promising approach was urgently needed to reduce the weak instrument bias and provide new clues to the common molecular mechanisms and biological processes of complex traits including AD and metabolic dysregulation. DNA methylation of CpGs is strongly associated with genetic loci, might explaining additional phenotypic variation in diseases besides genetic variants. Thus DNA methylation was considered as IVs (or intermediate phenotype) in a wide range of MR studies to infer the causal association between DNA methylation and complex diseases [11,12,14,15], bridging the GWAS gap regarding SNPs to diseases. Therefore, we explored the casual DNA methylation for metabolic traits and AD with the cis-mQTLs determined DNA methylation as IVs. Incorporating the cis-mQTLs information into GWAS analyses, demonstrated a high potential to increase the power of GWAS in identifying loci associated with metabolic traits and AD, and improve the explanation of traits variance. The large number of SNPs was included as IVs for metabolic traits and AD; therefore, the sensitivity analysis was performed to rule out the issue of pleiotropy and linkage disequilibrium. The ndings were robust in sensitivity analyses with different IVs. Secondly, these null ndings suggest that the associations between metabolic factors and AD could attribute from the reverse causation bias [20]. Our bidirectional MR might reduce the bias, and the reverse MR analyses showed no signi cant association of cis-mQTLs determined AD with metabolic traits. The consistent causal direction indicated that the association of AD with the metabolic dysregulation were unlikely to be causal.
The genetic linkage and association studies have identi ed some AD susceptibility genes, a number of which are related to cholesterol metabolism or transport [3,17,32]. Lipid metabolism play an important pathway involved in the development of AD [32,33]. However, contradictory evidence comes from epidemiological studies showing no or controversial association between dyslipidemia and AD risk [34][35][36]. In our study, cis-mQTLs determined DNA methylation to higher TG is associated with a higher AD risk (i.e., SNP→DNA methylation→higher TC→AD). Several studies reported that lipid-lowering medications-statins were of a protective effect against the development of AD [36,37]. In addition, high cholesterol in late life was associated with decreased AD risk [35,36], which may be explained by the timing of the cholesterol measurements in relationship to age and the clinical onset of AD. Taken together, inherited lifetime exposure to higher TC was associated with higher AD risk, however, the patients with higher TC and took lipid-lowering medications might reduce the risk of AD. Further analysis of AD outcomes in total cholesterol intervention trials is warranted. In addition, our ndings suggest the imperative need for further investigation of the possibility that lipid-lowering medications might be of an independent effect on the prognosis of AD risk without dyslipidemia.
In the present study, we found no evidence to support causal associations of cis-mQTLs determined T2D and obesity with AD, and vice versa. The result was in line with the previously reported MR result [20,28,29]. For metabolic dysregulation, T2D was the only risk factor with convincing evidence for an association with AD [5,6,38]. The increasing evidence showed that obesity in midlife was associated with the risk of AD [5,39,40]. A wide range of studies suggest the genetic and pathological links between AD and T2D/obesity [17,40,41]. DNA methylation plays an important regulatory role in the pathogenesis of T2D, obesity and AD [11,14]. In addition, our study identi ed that AD and T2D/obesity share several common genetic and epigenetic architectures. These ndings indicated that the casual DNA methylation overlapped between AD and T2D/obesity might attribute to pleiotropy, shedding light on molecular mechanisms of DNA methylation underlying these comorbidities.
Almost half of the overlapped epigenetic architectures changing in AD and metabolic dysregulation were in the reverse direction. It is important to recognize that the relationship between AD and metabolic dysregulation is far from simple. These ndings open the path toward a more detailed investigation of etiologic processes linking metabolic dysregulation and AD. The understanding of the epigenetic mechanisms of these diseases are likely only the tip of the iceberg. The potential downstream effects of DNA methylation (known as cis-gene expression-quantitative trait methylation analysis) on disease etiology still needs to be further veri ed. To the best of our knowledge, this is the rst study to explore possible biological mechanisms in the causal pathway from DNA methylation to AD. There is considerable merit for using cis-mQTLs along with metabolic traits related cis-mQTLs to reveal much broader and more complex networks underlying genetic variant-AD associations. However, most causal DNA methylation of CpG-sites directly affect AD and are independent of metabolic dysregulation (SNP→DNA methylation→AD), which also suggests that AD is a highly heritable disease and DNA methylation is the critical the heritable epigenetic marks of the genome linked to AD.
Given the multifactorial etiology of AD, multidomain interventions that target several risk factors and mechanisms simultaneously might be necessary for an optimal preventive effect on AD. The previous study indicated that a third of AD cases might be attributable to modi able factors such as diabetes mellitus, mid-life obesity and hypertension, physical activity, depression, smoking and low educational attainment [5,6,42]. Aside from metabolic traits, other modi able factors also play important role in AD. Future researches are supposed to explore the association with DNA methylation, other modi able factors and AD.

Limitation
The main strength of our study is the large sample size accrued from the GWAS summary statistics, enabling us to examine the causal relationship among DNA methylation, metabolic risk factors, and AD. However, some limitations should be considered when interpreting the ndings. Since all participants in summary data are of European ancestry, population strati cation is less likely to be a problem. However, the results of this study are of less necessarily to be validated in other ethnic groups. We only investigated CpG sites using cis-mQTL with the lowest P-value in order to improve the speci city of the IV for each CpG, which also could induce the bias of winner curse. Although the causal network is generally high-dimensional and unknown, the direction of potential causal relationships between layers of data can often be deduced from external biological knowledge. The putative causal associations reported in the network with bidirectional MR study are not de nitive due to the cross-sectional observational nature of our study. Individual patient-level genotype and phenotype information are unavailable, which limit our exploration of other exposures. Our approach has the attractive advantage of enabling the interrogation of the potential epigenetic-complex trait interplay on a much wider scale by foregoing the requirement that ''omic'' data, environmental exposures and phenotypes are measured in the same sample for future study.

Conclusion
Taken together, we found that cis-mQTLs determined DNA methylation to higher TC was associated with higher AD risk. In contrast, the relation cis-mQTLs determined AD and metabolic dysregulation was unlikely to be causal. Further analysis of AD outcomes in total cholesterol intervention trials is warranted. In addition, our ndings provided evidence that endeavors leveraging cis-mQTLs data could help to further characterize the complex networks of relationships between DNA methylation and complex traits. The continued research for additional risk factors for characterizing the relationship between DNA methylation and AD are still warranted.

Declarations
Each point represents the association of a cis-mQTL SNP with CpG. In the plot, we only kept the cis-mQTL SNP having the smallest P-value for a CpG in the cases of multiple SNPs associated with the same CpG. The horizontal coordinate represents the location of a CpG within the chromosome. Points represent -log10 p values (y axis) for CpGs according to their location (x axis). Effects that survived the multiple-testing threshold in our analysis (P < 1×10 -10 ) are above the red horizontal line. Figure S2. Venn diagram for the casual DNA methylation of BMI and AD.
The plot was performed by R package "VennDiagram". The number of DNA methylation of CpG-sites overlapping between BMI and AD is shown. The consistency means that the causal CpGs overlapping between BMI and AD changes in the same direction.
AD indicates Alzheimer disease; BMI: body mass index. Figure S3. Venn diagram for the casual DNA methylation of WHR and AD.
The plot was performed by R package "VennDiagram". The number of DNA methylation of CpG-sites overlapping between WHR and AD is shown. The consistency means that the causal CpGs overlapping between WHR and AD changes in the same direction.
AD indicates Alzheimer disease; WHR, waist-hip ratio.    The plot was performed by R package "VennDiagram". The number of DNA methylation of CpG-sites overlapping between LDL and AD is shown. The consistency means that the causal CpGs overlapping between LDL and AD changes in the same direction.
AD indicates Alzheimer disease; LDL, low density lipoprotein cholesterol. Figure S8. Venn diagram for the casual DNA methylation of TC and AD.
The plot was performed by R package "VennDiagram". The number of DNA methylation of CpG-sites overlapping between TC and AD is shown. The consistency means that the causal CpGs overlapping between TC and AD changes in the same direction.
AD indicates Alzheimer disease; TC, total cholesterol. Figure S9. Venn diagram for the casual DNA methylation of TG and AD.
The plot was performed by R package "VennDiagram". The number of DNA methylation of CpG-sites overlapping between TG and AD is shown. The consistency means that the causal CpGs overlapping between TG and AD changes in the same direction.
AD indicates Alzheimer disease; TG, triglycerides. Figure 1 Network with bidirectional Mendelian randomization analysis framework. The solid lines depict the true potential causal diagram. The dashed lines represent the parameters that need to be estimated. AD, Alzheimer disease; mQTL, DNA methylation quantitative trait loci; MR, mendelian randomization Upset plot for the casual DNA methylation overlapping between metabolic traits and AD. The plot was performed by R package "UpSetR". The number of DNA methylation of CpG-sites overlapping between metabolic traits and AD is shown. AD indicates Alzheimer disease; BMI: body mass index; FI, fasting insulin; FG, fasting glucose; HDL, high density lipoprotein cholesterol; LDL, low density lipoprotein cholesterol; TC, total cholesterol; TG, triglycerides; WHR, waist-hip ratio.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.