Genome-wide 450k Illumina DNA methylation datasets were extracted from the Gene Expression Omnibus (GEO) database using the GEOquery R package (Davis and Meltzer, 2007). Raw DNA methylation values were intra-array normalized using the beta mixture quantile dilation (BMIQ) method (Teschendorff et al., 2013) and normalized beta methylation values were used for all further analyses.

Table 1 summarizes the genome-wide DNA methylation datasets of AD brain and whole blood samples, and plaques and whole blood samples of atherosclerotic patients used in our study. Also, one dataset containing samples of intracranial aneurysm arteries was included. Genome-wide methylation levels were measured using the 450k Illumina arrays in every dataset. The dataset_ID is used to refer to each dataset in the main text.

Table 2 summarizes the genome-wide DNA methylation datasets of whole blood chronic disease samples. Genome-wide methylation levels were measured using the 450k Illumina arrays in every dataset. The dataset_ID is used to refer to each dataset in the main text.

Genome-wide DNA methylation analysis was performed using the limma moderated t-test (Ritchie et al., 2015) for AD whole blood (GSE59685) and atherosclerosis whole blood (GSE107143) datasets. Genome-wide similarity was determined by correlating the resulting t-statistics using the Pearson’s correlation test.

The DMPs in whole blood samples of atherosclerosis (GSE107143) that we detected previously (Istas et al., 2017) were used to compare with the different AD datasets (whole blood and brain tissues). Athero-DMPs were selected based on an FDR < 0.15 and a delta beta > 0.05 using the limma moderated t-test (Ritchie et al., 2015), resulting in 712 athero-DMPs. Two tailed t-tests were performed to determine the significance level in the AD datasets for each of the 712 athero-DMPs. For each AD dataset, the percentage of overlapping genes (POG) was calculated by dividing the number of athero-DMPs with an unadjusted p-value < 0.05 with the 712 athero-DMPs. The consistency of the overlapping DMPs was calculated by dividing the overlapping DMPs with a same direction of methylation change as in the atherosclerosis whole blood dataset (i.e., hypermethylated in AD dataset and hypermethylated in atherosclerosis dataset) with the total number of overlapping DMPs.

The different DNA methylation datasets were also compared by correlating the t-statistics of each athero-DMP across the datasets using the Pearson’s correlation test.

To detect co-methylation consensus modules between atherosclerosis and AD in whole blood, the weighted correlation network analysis (WGCNA) R package was used (Langfelder and Horvath, 2008). First, the most variable probes were selected based on an median absolute deviation (MAD) threshold of 0.03 in at least one dataset. In this way 97,375 probes remained for further analysis. The blockwiseConsensusModules function in the WGCNA R package was subsequently used to construct weighted co-methylation networks and detect consensus modules across the two datasets (Zhang and Horvath, 2005). We used the soft-threshold power of 7, a minimum module size of 30 probes, a maximum block size of 20,000, and a dendrogram cut height of 0.25 for module merging as input parameters. The consensus module eigengenes (i.e., first principal component of the module) were associated with disease (either atherosclerosis or AD), and the modules with a significant association (p-value < 0.05) in both datasets were used for further analysis. The module membership of each probe in the modules was calculated by correlating the module eigengenes with the DNA methylation beta values. A module membership close to 1 or −1 indicates high connection with the module. The gene significance values of each probe in the modules were calculated using the t-statistics of the association between the beta-values and the disease groups. The Pearson’s correlation was used to correlate these significance values across different datasets. Probes in the significant modules were mapped to different genomic regions, including gene elements [transcription start site (TSS), gene bodies, untranslated regions (UTRs), intergenic regions] and CpG island (CGI) elements (CGI shelves, shores, and islands) using the Illumina manifest annotation file and Gm12878 ENCODE chromatin segmentation states obtained from the UCSC genome browser. The enrichment of module probes in one of the genomic regions was calculated using the Fisher’s exact test. Probes in the significant modules were mapped to genes using the Illumina manifest annotation file. Pathway enrichment was performed using the Ingenuity Pathway Analysis (IPA) software and the Fisher’s exact test as statistical test.

Module preservation across different AD and CVD datasets were performed using the module Preservation function in the WGCNA R package (Langfelder et al., 2011). One hundred permutations were performed to calculate the preservation z-scores for each dataset. Z-scores higher than 10 indicate strong preservation, between 2 and 10 weak to moderate preservation and below 2 no preservation.

Cell type fractions were calculated using the method described by Houseman et al. (2012). The EpiDISH R package was used to perform the calculations (Teschendorff et al., 2017). For the whole blood datasets, we used the centDHSbloodDMC.m whole blood reference dataset containing 333 CpG probes of the seven major blood cell types [B-cells, natural killer (NK)-cells, CD4+ T-cells, CD8+ T-cells, granulocytes, and monocytes]. To estimate the cell counts in the atherosclerosis vascular tissues, we created a new reference methylome. For the smooth muscle cells, fibroblasts, and endothelial cells, we retrieved 450k Illumina methylation data of AoSMC, ProgFib, and HUVEC from the ENCODE project (GSE40699), respectively. Raw methylation values were intra-array normalized using the BMIQ method (Teschendorff et al., 2013) and normalized beta-values were subsequently used for further analysis. Immune cell (IC) reference methylomes were obtained from the study of Reinius et al. (2012). Next, differences in methylation across the different cell types were calculated using limma linear models (Ritchie et al., 2015) comparing each cell type with the rest of the samples: IC vs. rest of samples, AoSMC vs. rest of samples, ProfFib vs. rest of samples, and HUVEC vs. rest of samples. For each cell type the top 100 significant CpG probes with the largest methylation difference were selected and combined to obtain 357 unique CpG probes. The beta values of the ICs were averaged to obtain the final reference methylome (Supplementary Table 1 ). This reference methylome was subsequently used to estimate cell counts and IC infiltration in the vascular tissues using the reference based method described by Houseman et al. (2012) and implemented in the EpiDISH R package (Teschendorff et al., 2017). Information about IC infiltration of The Cancer Genome Atlas (TCGA) cancers were obtained from a recent study examining immunogenomic profiles of different cancers (Thorsson et al., 2018). TCGA level-3 450k Illumina methylation data were retrieved using the TCGAbiolinks R package ( Table 3 ) (Colaprico et al., 2016). Beta-values were used for subsequent analyses.

To compare methylation profiles between atherosclerosis (GSE107143) and AD (GSE59685) in whole blood, we first compared the genome-wide significance of each CpG probe in both datasets. Using the limma moderated t-test, we performed differentially methylation analysis on both whole blood datasets. We found a weak positive correlation between the t-statistics (Pearson correlation: 0.108, P < 2.2e⁻¹⁶) in both datasets, indicating that at the genome-wide level the similarity between the methylation profiles in atherosclerosis and AD is limited ( Figure 1A ). Next, we checked more specifically, whether the top significant CpG-probes found in our atherosclerosis dataset (Istas et al., 2017) were also differentially methylated in the AD dataset. We first selected the most significant CpG-probes by setting the threshold for differentially methylation at FDR < 0.15 and delta beta > 0.05. In this way 712 CpG-probes were selected which we called athero-DMPs. T-tests were performed to determine the significance level in the AD dataset for each of the 712 athero-DMPs. We found several probes which were also found to be differentially methylated in AD (p-value < 0.05) (Figure 1B). Of particular interest, the directionality of the methylation change was very similar in both datasets.

Using different publicly available DNA methylation datasets ( Table 1 ), we performed the same analyses in multiple AD brain tissues. For each AD dataset, we selected the 712 athero-DMPs and correlated the t-statistics resulted from the comparison in methylation between AD patients and healthy controls. Again, for some of the brain tissue we could find a similar methylation profile compared to atherosclerosis blood samples ( Figures 1C, D and Supplementary Figure 1 ). Especially in frontal lobe, frontal cortex, and superior temporal gyrus, the hypo- and hypermethylated atherosclerosis DMPs corresponded with hypo- and hypermethylation in AD, respectively ( Supplementary Figure 1 ). The POG in each tissue was rather limited, with values ranging between 4 and 32% ( Table 4). However, the direction of methylation change for most of the overlapping CpG sites were highly consistent with the atherosclerosis whole blood dataset (consistency %) (Table 4). For example, 23.2% of athero-DMPs were also found significant in AD whole blood samples of which 98.2% of the DMPs were directional consistent with the athero-DMPs. The percentage of consistency was the lowest for the AD cerebellum samples (49.4 and 61.8%) and a dataset with AD superior temporal gyrus samples (45.1%). Except for cerebellum tissues, there was a positive correlation between the t-statistics of the athero-DMPs in the atherosclerosis whole blood dataset and the other AD brain and whole blood datasets, with correlation coefficients ranging from 0.14 to 0.69 (Figure 1C). In contrast, in cerebellum samples no strong correlation could be found (correlation coefficient ∼ 0). Furthermore, cerebellum samples did not correlate with the other brain AD tissues (Figure 1D).

We next used WGCNA to identify a consensus co-methylation module of atherosclerosis and AD in whole blood. One hundred seventy-five consensus modules could be found. We used the module eigengenes to associate the different modules with atherosclerosis and AD disease state. We found 25 and 16 modules significantly associated with atherosclerosis and AD, respectively. Three consensus modules were found to be both significant (p-value < 0.05) in atherosclerosis and AD. Module ME91 was positively associated with atherosclerosis and negatively with AD, module ME54 was positively associated in both datasets, and module ME21 was negatively associated in both datasets.

Next, we calculated the gene significance and module membership of the probes in the consensus modules. We defined the gene significance as the t-statistic of the association between the CpG-probe methylation value and disease state, and the module membership as the correlation coefficient between the module eigengene and the CpG-probe methylation beta value. The closer the module membership is to 1 or −1 the more important the probe is in the module. In general, a module membership close to 1 or −1 is highly connective and therefore represents a hub in the network. There was a strong correlation between the gene significance in the two datasets for module ME21 (Pearson’s correlation: 0.845) (Figure 2A). The same was true for the module membership (Pearson’s correlation: 0.989) ( Figure 2B). As expected the gene significance and module membership was also highly correlated ( Supplementary Figure 2). Because a less strong correlation could be found with module ME54 (data not shown), we decided to focus only on module ME21 as the consensus module ( Supplementary Table 2 ).

We next mapped the CpG probes in the consensus module to different genomic regions relative to gene elements (TSS, gene bodies, etc.), CGIs, and chromatin segmentation states. Interestingly, we found an enrichment in 5’UTR regions, CpG-poor regions outside CGIs, active promoters, strong and weak enhancers, transcriptional transition, and elongation states ( Figure 2C ). In addition, the consensus module CpG probes were strongly depleted in CGIs, repressed chromatin states, and heterochromatin.

The consensus module CpG probes were subsequently mapped to genes. IPA pathway analysis showed a strong enrichment in T cell regulatory and immune pathways, including T- and B cell receptor signaling, Th1 and Th2 pathway, IL-10 and IL-8 signaling, and NF-κB signaling ( Figure 2D ). In addition, genes were enriched in immunological and inflammatory diseases, and functions related to cellular development, growth, proliferation, and movement ( Figure 2E ).

We further analyzed whether module 21 was preserved in other AD methylation datasets of different brain tissues. The preservation z-scores for all AD brain tissues, except for cerebellum, were between 2 and 10, suggesting weak to moderate preservation ( Figure 3A ). In cerebellum, there was no indication of module preservation (z-score < 2). We next calculated for each AD dataset the gene significance values (t-statistics) of the CpG probes in the consensus module, and performed pairwise correlation across the different AD datasets. Except for cerebellum, all the other AD datasets showed a positive correlation with the gene significance values of the whole blood datasets, and relative to each other ( Figure 3B ).

We next wondered whether the same pattern could also be found in other methylation datasets related to CVD and atherosclerosis. We extracted 450k Illumina data from carotid plaques, plaques after cerebrovascular event and arteries with intracranial aneurysm (Table 1). Here the preservation was much stronger, with z-scores higher than 10 in the atherosclerotic plaques after a cerebrovascular event and in intracranial aneurysm, while for the carotid plaque datasets we found moderate module preservation (Figure 3C). Again, we could find strong positive correlations between the gene significance values of the different datasets (Figure 3D). Of note, the highest correlation with the whole blood dataset could be found with the atherosclerotic plaque dataset with a cerebrovascular event. In contrast, there was no evidence of correlation between the carotid plaque dataset and the dataset with a cerebrovascular event.

The enriched pathways in T cell activation and function indicate that part of the methylation changes may be due to differences in cell type heterogeneity in the samples analyzed. We therefore estimated cell type composition in the atherosclerosis and AD whole blood datasets. B-cells and CD4+ T-cell levels were both reduced while granulocyte levels were increased in atherosclerosis and AD samples as compared to healthy blood samples (Figure 4A). In addition, many CpG probes in the consensus module seem variable across the different blood IC types (Supplementary Figure 3 ). In this respect, the positive correlations seen with the AD brain and atherosclerotic plaque tissues could reflect the infiltration of IC in the brain and arterial wall respectively. To verify this hypothesis, we estimated IC fraction in the vascular tissues using a new reference methylome created from methylation profiles of aortic smooth muscle cells (AoSMC), fibroblasts (ProgFib), human umbilical vein endothelials cells (HUVEC), and IC (see Methods section for details). As expected, an increase in IC was observed in plaque tissue compared to healthy aorta tissue (Figure 4B ), and monocytes were the main infiltrated blood cell type observed in the plaques ( Supplementary Figure 4 ). More surprisingly, the AoSMC fraction was relatively lowered in plaque material. HUVEC and ProgFib fractions didn’t show substantial differences. Methylation values from the reference methylomes of the 500 most significantly DMPs in aorta plaques revealed that the hypermethylated profile in atherosclerotic plaques (Zaina et al., 2014) was mainly due to an overall hypermethylation in IC compared to the other cell types. Similarly, the small fraction of hypo-DMPs could also be attributed to hypomethylated CpG sites in IC ( Supplementary Figure 5 ). Furthermore, a strong correlation was found between the consensus module eigengene and the estimated IC fraction in aorta and carotid atherosclerotic plaque tissues, supporting our hypothesis ( Figure 4C ).

To further prove that our methylation profile measures an immune component, we made use of IC infiltration information of TCGA cancers obtained from a recent study (Thorsson et al., 2018). As expected, in almost all cancers there was a negative correlation between the module eigengenes and leukocyte fraction, stromal fraction, and lymphocyte infiltration signature score ( Supplementary Figure 6 ). Thus, tumors with methylation profiles resembling the methylation consensus module demonstrated more IC infiltration. This is completely in line with our observations and supports our conclusions.

Since inflammation is a common hallmark of many chronic aging diseases, we further checked whether the consensus immuno-methylation module identified was also prevalent in other chronic inflammation and age-dependent diseases. We therefore reanalyzed Illumina 450K DNA methylation profiles of whole blood samples of Parkinson’s disease (PD), schizophrenia, obesity, osteoporosis, and multiple sclerosis (MS) (Table 2), and correlated the gene significance values of the CpG probes in the consensus module across all the diseases ( Figure 5A). A strong positive correlation could be observed with PD (r: 0.91). Also obesity (r: 0.43) and osteoporosis (r: 0.21) showed a moderate positive correlation. On the other hand, schizophrenia (r: −0.39) and MS (r: −0.51) demonstrated a negative correlation.

The high correlation found in PD could again be attributed to a shift in CD4+ T cell and granulocyte blood levels. However, no such changes were detected in obesity and osteoporosis (Figure 5B). In osteoporosis, NK cell levels were slightly higher as compared to healthy samples. In obesity and schizophrenia, IC types didn’t change dramatically. MS patients had opposite cell type distribution in comparison to atherosclerosis, AD and PD patients with higher CD4+ T- and B-cell levels and lower granulocyte levels.