Epigenetic and integrative cross-omics analyses of cerebral white matter hyperintensities on MRI

Abstract

Cerebral white matter hyperintensities on MRI are markers of cerebral small vessel disease, a major risk factor for dementia and stroke. Despite the successful identification of multiple genetic variants associated with this highly heritable condition, its genetic architecture remains incompletely understood. More specifically, the role of DNA methylation has received little attention.

We investigated the association between white matter hyperintensity burden and DNA methylation in blood at ∼450 000 cytosine-phosphate-guanine (CpG) sites in 9732 middle-aged to older adults from 14 community-based studies. Single CpG and region-based association analyses were carried out. Functional annotation and integrative cross-omics analyses were performed to identify novel genes underlying the relationship between DNA methylation and white matter hyperintensities.

We identified 12 single CpG and 46 region-based DNA methylation associations with white matter hyperintensity burden. Our top discovery single CpG, cg24202936 (P = 7.6 × 10⁻⁸), was associated with F2 expression in blood (P = 6.4 × 10⁻⁵) and co-localized with FOLH1 expression in brain (posterior probability = 0.75). Our top differentially methylated regions were in PRMT1 and in CCDC144NL-AS1, which were also represented in single CpG associations (cg17417856 and cg06809326, respectively). Through Mendelian randomization analyses cg06809326 was putatively associated with white matter hyperintensity burden (P = 0.03) and expression of CCDC144NL-AS1 possibly mediated this association. Differentially methylated region analysis, joint epigenetic association analysis and multi-omics co-localization analysis consistently identified a role of DNA methylation near SH3PXD2A, a locus previously identified in genome-wide association studies of white matter hyperintensities. Gene set enrichment analyses revealed functions of the identified DNA methylation loci in the blood–brain barrier and in the immune response. Integrative cross-omics analysis identified 19 key regulatory genes in two networks related to extracellular matrix organization, and lipid and lipoprotein metabolism. A drug-repositioning analysis indicated antihyperlipidaemic agents, more specifically peroxisome proliferator-activated receptor-alpha, as possible target drugs for white matter hyperintensities.

Our epigenome-wide association study and integrative cross-omics analyses implicate novel genes influencing white matter hyperintensity burden, which converged on pathways related to the immune response and to a compromised blood–brain barrier possibly due to disrupted cell–cell and cell–extracellular matrix interactions. The results also suggest that antihyperlipidaemic therapy may contribute to lowering risk for white matter hyperintensities possibly through protection against blood–brain barrier disruption.

Yang et al. investigate the association between epigenetic changes and white matter hyperintensities—markers of cerebral small vessel disease—in 9732 adults. Integrative cross-omics analyses implicate blood–brain barrier dysfunction and the immune response in WMH burden, and suggest antihyperlipidemic agents as a potential treatment.

Introduction

Cerebral white matter hyperintensities (WMH) on MRI are indicative of cerebral small vessel disease (cSVD) and are part of the spectrum of brain vascular injury that affects cognitive function, also known as vascular contributions to cognitive impairment and dementia.^1,2 While the pathophysiology of WMH is little understood and probably heterogeneous, it probably has ischaemic and neurodegenerative origins.¹ Historical pathology data suggested chronic ischaemia resulting in demyelination and axonal loss as an underlying mechanism; however, neuroimaging data point to blood–brain barrier (BBB) dysfunction, dysfunctional blood flow linked with impaired cerebrovascular autoregulation, vascular stiffness, periarteriolar inflammation and more recently protein deposition (i.e. amyloid angiopathy).² Genetics play a significant role in WMH with a heritability estimated from 54% to 80%^3–7; however, the genetic variants identified in association studies explain only ∼29% of WMH variance.^8,9 Epigenetic changes such as DNA methylation (DNAm), which regulate gene expression, have emerged as another key component of the genetic architecture of complex traits.¹⁰ Unlike DNA sequence variation, which remains unchanged throughout life, DNAm is plastic and highly sensitive to changes in the environment and ageing.^10,11 To date, its role in cSVD has received little attention. We hypothesized that there may be patterns of DNAm associated with WMH that are common across all populations. We also hypothesized that the interplay between genotype, epigenotype and risk factor exposure underlies cSVD aetiology and used an integrated analytic framework to identify such relationships.

Materials and methods

Overview

This study comprised five analytic parts to implicate novel genes and gene networks in WMH aetiology (Fig. 1). First, we performed an epigenome-wide association analysis to identify DNAm loci, both cytosine-phosphate-guanine (CpG) sites and differentially methylated regions (DMRs), associated with WMH burden. The identified DNAm loci were then annotated for regulatory features, pathways and association with other traits. Second, we investigated the contribution of genetic variation to variation in DNAm at the identified CpGs and used Mendelian randomization (MR) techniques to test for causal association with WMH burden and for the mediating role of expression of nearby genes. Third, we examined the role of DNAm at established WMH genome-wide association study (GWAS) loci. Fourth, we integrated gene expression and expression quantitative trait loci (eQTL) data to prioritize candidate genes associated with the identified CpGs. Last, we performed integrative cross-omics analyses to derive WMH-associated genes networks and their key drivers and to reposition drug targets.

Figure 1.

Overview of the study analytic scheme.

Study subjects

The sample included 9732 middle-aged to older adults of European and African ancestry from 14 community-based studies. Our discovery sample includes 5715 subjects of European ancestry (n = 4610) and of African ancestry (n = 1105) from Atherosclerosis Risk in Communities (ARIC),¹² Biobanking and BioMolecular resources Research Infrastructure,¹³ Cardiovascular Health Study (CHS),¹⁴ Coronary Artery Risk Development in Young Adults (CARDIA),¹⁵ Framingham Heart Study (FHS) offspring study,^16,17 Genetic Epidemiology Network of Arteriopathy (GENOA) study,¹⁸ Lothian Birth Cohort 1936,^19,20 Rotterdam Study^21,22 and Study of Health in Pomerania (SHIP).²³ To replicate our findings, we accessed data on 3398 subjects from the Alzheimer’s Disease Neuroimaging Initiative,^24,25 FHS third generation study,²⁶ the Older Australian Twin Study^27,28 and the Rhineland Study.²⁹ Additionally, we included a secondary replication sample (n = 619) from the BRIDGET Consortium.³⁰ Subjects with history of stroke or dementia were excluded. Details about participating studies and study-specific ethics statements are provided in the Supplementary material. Each study obtained written informed consent from all participants and approval from the appropriate institutional review boards.

WMH burden measurements

Brain MRI was taken in the same or the closest subsequent visit to the visit in which DNAm was measured. In each study, MRI scans were performed and interpreted using standardized procedures without reference to demographic or clinical information. The field strength of the scanners used ranged mostly from 1.5 to 3.0 T. T₁-, T₂- and/or proton-density-weighted scans were obtained for all participants. Most studies used a fully automated segmentation method to quantify WMH burden. MRI procedures and WMH quantification in each study are detailed in the Supplementary material.

DNAm profiling

DNAm levels were measured at ∼450 K CpGs from whole blood samples with the Illumina Infinium Human 450 Methylation BeadChip in most participating cohorts. The GENOA study measured methylation levels at ∼27 K CpGs with the Illumina Infinium HumanMethylation 27 BeadChip, entirely covered by the Human 450 BeadChip. CARDIA, SHIP-TREND, Alzheimer’s Disease Neuroimaging Initiative and Rhineland Study used the Illumina MethylationEPIC BeadChip with a denser coverage of CpGs (∼850 K). Each study independently performed quality control for DNAm data, complying with the agreed minimum quality control guidelines; CpGs with >95% of samples with a detection P < 0.01 and samples with >95% of CpGs with a detection P < 0.01 were selected. DNAm values were then standardized using an intra-array normalization method. The BRIDGET Consortium measured DNAm levels using Hi-seq bisulphate sequencing, and DNAm sites with sample coverage <95% were excluded. Details of DNAm data collection and processing in participating studies are presented in the Supplementary material.

Cohort-level epigenome-wide association analyses

We tested association between DNAm level (untransformed beta values) and WMH burden [ln(WMH+1)] using a linear mixed regression model by ancestry group adjusted for age, sex, study site if applicable, total (intra)cranial volume (cm³), white blood cell proportion (%)³¹ and within-ancestry principal components as fixed effects and technical covariates (i.e. plate, chip-position, row and column) as random effects. In FHS, family structure is also adjusted as a random effect. Multi-ancestry studies with a small number of subjects in each ancestry, namely CHS and CARDIA, performed a pooled-ancestry analysis that also adjusted for ancestry group as fixed effects. Additionally, subgroup analyses by hypertension status were conducted. Hypertension was defined if either systolic or diastolic blood pressure (SBP or DBP) is >140 or 90 mmHg, respectively, or if a subject was taking any antihypertensive medication at the time of MRI measurement. In the BRIDGET study, we tested the association of DNAm with an extreme-SVD phenotype defined as excessive WMH volume with or without brain infarcts accounting for age, sex, country, the sequencing read counts and sample relatedness.³² DNAm measurements and statistical models used in participating studies are described in Supplementary material.

Epigenome-wide meta-analysis and replication analysis

We combined EWAS results based on sample-size-weighted z-score-based fixed-effect method in METAL³³ because WMH was measured on different scales in the various cohorts and because our primary aim was to identify novel DNAm loci for WMH burden rather than estimate effect sizes of methylation probes.³⁴ Hypertensive and normotensive subgroup meta-analyses and ancestry-specific meta-analyses (excluding CHS and CARDIA) were also performed. Study-specific results were corrected for inflation during meta-analysis if inflation was detected [genomic inflation factor (λ) > 1.0]. An association was considered as significant if P was smaller than Bonferroni threshold (∼1.2 × 10⁻⁷). A less stringent threshold was also set as 1.0 × 10⁻⁵ to detect suggestive associations. CpGs on sex chromosomes were not considered because our analytic plan did not account for hemi-methylation on the X chromosome due to chromosome X inactivation in females. Cross-reactive CpGs reported by Chen et al.³⁵ and those showing evidence of heterogeneity (Cochran Q P-value < 0.05) were removed from the results post hoc. In the replication samples, associations for the identified CpGs were tested. CpGs were considered replicated if they were significant at the Bonferroni threshold (0.05 / the number of the CpGs). We plotted epigenetic associations in cis (±50 kb) using R ‘coMET’ package.³⁶

Annotation of regulatory features and traits

We scored genomic positions of the identified CpGs according to RegulomeDB’s³⁷ ranking criteria ranging from one (likely to affect binding and linked to expression of a gene target) to five (minimal binding evidence) and also computed a probability score within a range from zero to one (the most likely to be a regulatory variant). CpGs at the locations with significant regulatory features (rank category one or two, and probability score ≥0.9) are discussed. We also identified enhancers or promoters mapped to CpGs using the database of genome-wide enhancer-to-gene or promoter-to-gene associations computed based on five elements: eQTLs, eRNA co-expression, transcription factor co-expression, capture Hi-C and gene target distance (GeneHancer DB).³⁸ Identified CpGs were also searched in EWAS catalogue³⁹ and EWAS atlas⁴⁰ to identify associated traits reported in previous EWAS. Last, to examine possible correlations among the CpGs, Spearman correlations were calculated in 906 European ancestry and 639 African ancestry subjects from the ARIC study.

DMR analysis

We performed a DMR analysis to identify a group of CpGs that collectively influence WMH burden using two specific methods, Comb-p⁴¹ and DMRCate,⁴² accounting for their spatial correlations. Briefly, Comb-P detects regional enrichment of low Ps at varying distance using the Stouffer–Liptak–Kechris correction for adjacent Ps.⁴¹ DMRcate models Gaussian kernel smoothing within predefined distance (1 kbp in this study) and collapses contiguous significant CpGs (P < 0.05) after multiple testing correction. DMR identified by both Comb-P (Šidák P < 0.05) and DMRCate (FDR < 0.05) was considered significant. To replicate, individual association Ps were pooled at each identified DMRs using DMRCate in the replication samples.

Gene set enrichment analysis of WMH-associated epigenetic loci

Identified CpGs and DMRs were tested for enrichment in gene sets from MSigDB c5 gene ontology database^43,44 and KEGG pathway database,⁴⁵ using ‘gsameth’ and ‘gsaregion’ functions built in R ‘missMethyl’ package.⁴⁶

Shared epigenetics with blood pressure

BP is an influential risk factor for WMH.^47–49 To investigate the shared epigenetics between WMH burden and BP, we performed a pairwise multivariate association test using summary statistics from a previous EWAS of SBP and DBP.⁵⁰ CpGs associated with both traits were tested against the null hypothesis H₀: β_WMH = β_{SBP or DBP} = 0. The test uses Z-scores for each trait and estimates multivariate test statistics accounting for the trait correlation calculated on the basis of the null associations (trait-specific P > 1 × 10⁻⁵). This method is implemented in the ‘metaUSAT’ software.⁵¹ To avoid false positive associations driven entirely by one trait, we included CpGs showing significance (P < 0.001) for both traits. Bonferroni threshold was set at 8.33 × 10⁻³ (=0.05/6) on the basis of the number of associations tested.

Heritability analysis and GWAS of WMH-associated CpGs

Inter-individual variation in DNAm may result from differences in environmental exposures, stochastic variation or genetic influences. To examine the contribution of genetic variation to variation in DNAm at the identified CpGs, we estimated the narrow-sense heritability (h²_meth) in the FHS Offspring Cohort subjects (n = 2377) adjusting for age, sex, blood cell counts, principal components and technical covariates. Body mass index (BMI) and smoking status were additionally corrected in sensitivity analyses.

To further identify genetic variants associated with DNAm levels at the WMH-associated CpGs, we performed GWAS in ARIC European ancestry subjects (n = 984). Genotypes were measured with Affymetrix 6.0 array and imputed from 1000 Genome phase one version three reference using MaCH v.1.0.16. Variants were excluded if minor allele frequency <0.01, sample call rate <95% or imputation quality <0.3. The untransformed methylation beta value was tested for genetic association adjusting for age, sex, top 42 methylation principal components and blood cell components.

Bi-directional MR analysis of the identified CpGs and WMH burden

To determine whether the WMH-associated DNAm level is a causal factor for WMH burden (Forward-MR) or a secondary outcome of WMH burden (Reverse-MR), we performed a bi-directional two-sample MR analysis⁵² for the identified CpGs with at least three instrumental variables (IVs). We identified methylation quantitative trait loci (mQTL) associations in cis (±1 Mb) from the FHS study (n = 4170) that had been validated using ARIC data (n = 963).⁵³ Those mQTLs were clumped at linkage disequilibrium r² < 0.05 for independence. For WMH, the UK Biobank GWAS summary statistics (n = 11 226)⁵⁴ was downloaded from the Cerebrovascular Disease Knowledge Portal (http://www.cerebrovascularportal.org/) on 1 September 2019. Reverse-MR analysis was performed using eight clumped genome-wide significant associations (linkage disequilibrium r² < 0.05). Since the FHS mQTL study shares only significant associations in cis,⁵³ we used the mQTL association statistics from ARIC European ancestry subjects for reverse-MR analysis.

For each CpG in both directions, causal association was tested on the basis of the IVW method in the R package ‘TwoSampleMR’.⁵⁵ To validate the MR result, sensitivity analyses based on weighted median and MR Egger methods, and built-in tests for pleiotropy and heterogeneity were also performed. For existence of pleiotropy (MR Egger intercept test P < 0.05), the Egger regression estimate was assessed instead of the IVW estimate.

Mediating effect of in cis genes between CpGs and WMH burden

To investigate whether expression of nearby genes mediates the relationship between the identified CpG and WMH burden, a two-step MR analysis was performed. We tested the directional relationships: (i) from ‘the exposure (CpGs)’ to ‘the mediator (gene expression)’ (step one); and (ii) from ‘the mediator (gene expression)’ to ‘the outcome (WMH burden)’ (step two) using the identified mQTL IVs, the WMH GWAS associations⁵⁴ and eQTL associations from the GTEx version eight brain eQTL data accessed via eQTL Catalogue (https://www.ebi.ac.uk/eqtl/) on 12 November 2020. Among available GTEx brain tissues, cortex (n = 205), frontal cortex (n = 175), cerebellum (n = 209), cerebellar hemisphere (n = 175) and caudate basal ganglia (n = 194) were selected. MR association based on the IVW method was again tested and sensitivity analysis was also performed. Gene expression with IVW P < 0.05 at both steps was considered as a potential mediator in the association between the identified CpG and WMH burden.

Cis-acting genes associated with the identified CpGs in blood

To functionally annotate the identified CpGs, we tested associations with gene expression in blood in long-range⁵³cis-regions (±5 Mb) in 1966 and 728 European ancestry subjects from FHS and Rotterdam Study, respectively. Expression of the nearest gene/mRNA was regressed on DNAm β score at the CpG adjusting for age, sex, population structure and family structure (FHS only), blood cell counts and technical covariates. Technical covariates and family structure were modelled as random effects. In sensitivity analyses, smoking status and BMI were added to the model. Estimates from two studies were then combined for each gene using the sample-sized based meta-analysis method in METAL.³³

Genes co-localizing with the identified CpGs in brain

To investigate cis-acting genes co-localized with the identified CpGs in the brain, we performed a multiple-trait co-localization (moloc) analysis using brain QTL data. Before this analysis, we examined the inter-individual correlations between DNAm levels in whole blood and in prefrontal cortex at the identified CpGs, using publicly available data.⁵⁶ For CpGs with significant correlation (P < 0.05) between blood and prefrontal cortex, we tested the posterior probabilities for full co-localization (PPFC) that multiple traits (DNAm, gene expression and WMH burden) share causal variants at each locus, given the data. We used coloc priors of 1 × 10⁻⁵. We identified EA-specific GWAS associations⁸ and brain mQTL (n = 543) and eQTL (n = 534) associations accessed via http://mostafavilab.stat.ubc.ca/xqtl/.⁵⁷ If PPFC is >0.7, we considered the gene is significantly co-localized with CpG and WMH burden. Moloc analysis was performed using the R package ‘moloc’.⁵⁸

Epigenetic regulation of known GWAS loci

We next investigated the role of DNAm at established WMH GWAS loci, which may not have been detectable at the genome-wide significance threshold. Among 26 loci reported in the latest WMH GWAS,⁸ we mapped 450 K-array CpGs to 21 loci. EWAS associations at each of these 21 loci were pooled using the Brown’s method (implemented in the package ‘poolr’) adjusting for dependence among CpGs.⁵⁹ For dependency information, we calculated correlation among CpGs in the GWAS loci using ARIC methylation data (906 European ancestry and 639 African ancestry subjects). A GWAS locus with combined P was considered significant if P was smaller than Bonferroni-adjusted threshold (0.05/number of loci tested).

Alternatively, we performed a moloc analysis at the 21 GWAS loci, again using the GWAS and brain QTL data.^8,57 With the priors of 1 × 10⁻⁵, we considered genes with a PPFC >70% as convincingly co-localized with DNAm and WMH burden.

Identification of biological pathways using multi-dimensional data integration

Integrating multi-omics associations for WMH may boost power to identify novel genes influencing WMH burden. We integrated genetic,⁸ transcriptomic⁶⁰ and epigenetic GWAS of WMH using the R package ‘mergeomics’ (version 1.2).⁶¹ To reduce noise in the GWAS data, the top 50% of genetic associations⁸ were included and pruned at r² < 0.5 based on HapMap3 linkage disequilibrium information as recommended.⁶¹ For transcriptomic associations, we used the recent WMH transcriptome-wide association study (TWAS) results.⁶⁰ For epigenetic associations, we used our discovery EWAS. Markers were primarily mapped to the nearest genes. For CpGs, cis-acting genes reported in the MesaEpiGenomics study⁶² were additionally annotated. For each GWAS, EWAS and TWAS, we tested marker-level enrichment with hierarchical permutation size of 20 000 on the basis of biological pathways from pre-defined public databases: KEGG,⁴⁵ REACTOME,⁶³ Biocarta⁶⁴ and the gene ontology knowledgebase.^65,66 Then, we meta-analysed the enriched gene sets from association studies and identified the WMH-associated gene sets (FDR-adjusted P < 0.05).

To describe the regulatory network of the identified gene sets and identify its local hub genes, we performed a weighted key driver analysis (wKDA) using the web-based software Mergeomics version 2.0.⁶⁷ Gene regulatory network was constructed using in-house brain-specific Bayesian network (minimum hub overlap 0.33 and directed edge type)⁶⁸ and visualized via Cytoscape version 3.8.2.⁶⁹

We also conducted an overlap-based drug-repositioning analysis ‘PharmOmics’ based on the identified key driver genes (FDR < 0.05) to predict potential drugs or small molecules targeting WMH.⁷⁰ PharmOmics comprises a curated drug signature database covering 941 drugs, constructed from transcriptomic data across >20 tissues from rat, human and mouse. For our analysis, we selected drug signatures from relevant tissues (in vivo human transcriptome data in cardiovascular and nervous system, and in vitro transcriptome data from murine oligodendroglial precursor cells), and examined the overlap between these drug signature genes and key driver genes from our identified WMH-associated gene sets.

Data availability

The data that support the findings of this study are included in this paper. Full EWAS summary statistics are available in dbGaP at phs000930.v9.p1.

Results

Identification of epigenetic changes associated with WMH burden

Study sample characteristics

In the discovery sample, the mean age ranges from 49.7 years in SHIP to 74.6 in CHS. Sex ratios are balanced in all studies except for GENOA study, which has 72.8% female. ARIC, CARDIA and CHS have both European ancestry and African ancestry subjects, other studies consist of single ancestry subjects (African or European ancestry). In the primary replication sample, subjects from FHS third generation and Rhineland Study (mean age 47.1 and 54.1 years, respectively), which compose 86.2% of the replication study, are younger than most discovery studies and show relatively smaller median WMH burden (0.34 in the FHS third generation study and 0.40 in the Rhineland Study). All subjects in the replication studies are of European ancestry. Demographic characteristics of participating cohorts are shown in Supplementary Table 1.

Novel DNAm loci are associated with WMH burden

In the discovery sample, we identified a novel epigenome-wide significant association between WMH burden and level of DNAm at cg24202936 (Z = 5.38, P = 7.58 × 10⁻⁸) in SEPTIN7P11. Associations at cg24202936 in each study are presented in a forest plot (Supplementary Fig. 1) and regional associations within 50 kb are presented with annotations (Supplementary Fig. 2). At the suggestive significance threshold of 1 × 10⁻⁵, we identified 11 additional loci (Table 1). The associations remained significant (P < 0.05/12 = 4.17 × 10⁻³) after adjusting for BMI, smoking status and SBP and DBP. Quantile–quantile and Miami plots are presented in Supplementary Figs 3 and 4. All subsequent analyses focus on these 12 CpGs, which are referred to as ‘target CpGs’. None of the target CpGs associations were replicated in independent samples and a meta-analysis of the discovery and replication samples showed significant heterogeneity in many of the resulting associations, which was not present in the discovery cohorts (Supplementary Table 2). Target CpGs showed consistent associations with WMH in subgroup analyses by ancestry and hypertension status (Supplementary Tables 3 and 4). Cg06450373 in CDH18 (P = 6.48 × 10⁻⁸) was identified in normotensive subjects (Supplementary Table 5); but not replicated. In a gene set enrichment analysis on discovered CpGs (P < 1.0 × 10⁻⁵), ‘cell–cell junction organization’ was identified as the top pathway [P = 1.32 × 10⁻³, false discovery rate (FDR) = 0.32].

Table 1.

Single-CpG associations with WMH burden in the discovery sample (P < 1 × 10^–5)

CpG	Chr:Position (hg19)	Nearest gene	Reduced model					Full model
			n	Z	P	Q	FDR	n	Z	P
cg24202936	11:50257256	SEPTIN7P11	5359	5.38	7.58 × 10–8	0.03	0.04	4930	5.28	1.30 × 10–7
cg17417856	19:50191637	PRMT1;ADM5	4917	–4.95	7.42 × 10–7	0.15	0.28	4526	–4.40	1.11 × 10–5
cg01506471	7:3990479	SDK1	5359	–4.81	1.52 × 10–6	0.21	0.3	4930	–4.00	6.41 × 10–5
cg14547240	4:15428750	C1QTNF7	5359	–4.71	2.48 × 10–6	0.25	0.3	4930	–4.17	3.10 × 10–5
cg21547371	3:52869521	MUSTN1	5359	–4.65	3.30 × 10–6	0.25	0.3	4930	–4.06	4.95 × 10–5
cg03116124	1:231293208	TRIM67	5129	–4.64	3.54 × 10–6	0.25	0.31	4700	–4.58	4.63 × 10–6
cg06809326	17:20799526	CCDC144NL-AS1	5359	4.57	4.80 × 10–6	0.28	0.34	4930	3.44	5.88 × 10–4
cg13476133	7:44185646	GCK	5359	4.55	5.46 × 10–6	0.28	0.36	4930	4.03	5.65 × 10–5
cg14133539	9:104568	FOXD4	4917	–4.53	5.98 × 10–6	0.28	0.38	4526	–4.45	8.41 × 10–6
cg17577122	22:19511967	CLDN5	5359	4.50	6.88 × 10–6	0.29	0.4	4930	4.79	1.68 × 10–6
cg23586595	4:84034390	PLAC8	5359	4.45	8.45 × 10–6	0.32	0.43	4930	3.93	8.36 × 10–5
cg23054394	3:140784675	SPSB4	5359	–4.42	9.88 × 10–6	0.34	0.45	4930	–4.01	6.07 × 10–5

The reduced model is adjusted for age, sex, study site (if applicable), total (intra)cranial volume (cm³), white blood cell proportion (%), technical covariates and genetic principal components. The full model is additionally adjusted for BMI, smoking status and systolic and diastolic blood pressure measures. Chr = chromosome; EA = European ancestry; FDR = local false discovery rate value; n = number of subjects tested for the CpG; P = P-value; Q = Q-value; SE = standard error; Z = Z-score.

Annotated regulatory functions of target CpGs

We found significant regulatory features from RegulomeDB at the genomic positions of cg24202936 (rank 2b and score 0.93), and cg06809326 (rank 2b and score 0.91) (Supplementary Table 6). Cg24202936 resides near a transcriptional starting site (0.2 kb upstream), and identified as a transcriptional factor binding site computationally annotated with 20 genes (Supplementary Table 6). Previously reported EWAS traits associated with target CpGs are presented in Supplementary Table 7. In particular, cg24202936 was previously reported associated with HIV infection.⁷¹ Cg06450373, cg031161214, cg01506471 and cg14547240 were correlated each other in both ancestries with weak to moderate r (0.23 to 0.55) (Supplementary Fig. 5). In African ancestry, cg23586595 showed weak but significant correlations with cg13476133 (r = 0.32), cg03116124 (r = −0.42) and cg14547240 (r = −0.36). No correlated CpG (|r| > 0.3) was identified for our top CpG, cg24202936, in both ancestries.

WMH-associated DMRs are enriched in immune response-related pathways

We identified 46 DMRs in associations with WMH burden (Supplementary Table 8). Notably, one DMR was in SH3PXD2A, previously identified in GWAS.^8,72–74 Identified DMRs were enriched in several gene ontologies, including signal transducer and activator of transcription family protein binding (FDR = 4.91 × 10⁻³) and defence response to virus (FDR = 5.68 × 10⁻³), which are related to the immune response (Supplementary Table 9). Of the 46 identified DMRs, PRMT1, ABAT, BHMT2, C11orf21, IZUMO1, C5orf66, ENPEP, SLC35F3, FBXO47, SLC45A4, KCTD16, KITLG and UCN3 were replicated (Supplementary Table 8). Of note, ENPEP, SLC35F3 and SLC45A4 were previously reported in BP GWAS.^75–81

Shared epigenetic loci between WMH and BP

At the Bonferroni-corrected threshold (P < 8.33 × 10⁻³), we identified six CpGs associated with both WMH burden and BP (Supplementary Table 10). For WMH-DBP, cg23291754 in MOBKL1A (P = 2.38 × 10⁻⁷) and cg24372586 in GNL1 (P = 7.84 × 10⁻⁷) were identified. For WMH-SBP, cg00711496 in CDC42BPB (P = 1.99 × 10⁻⁷), cg04987734 in C19orf76; PRMT1 (P = 3.09 × 10⁻⁷), cg00934987 in SEPT4 (P = 1.07 × 10⁻⁶) and cg18770635 in KLHDC7B (P = 1.68 × 10⁻⁶) were identified.

Heritability of the WMH-associated CpGs

Significant h²_meth was estimated for cg17417856 (40.4%, P = 1.37 × 10⁻⁸), cg06809326 (26.5%, P = 1.03 × 10⁻⁴), cg23586595 (24.2%, P = 1.47 × 10⁻³), cg17577122 (14.3%, P = 2.80 × 10⁻²) and cg24202936 (15.5%, P = 1.34 × 10⁻²) (Table 2). Additional adjustment for BMI and smoking status did not significantly modify these estimates. In GWAS of the target CpGs in the ARIC European ancestry sample, we observed significant cis-genetic influence on cg06809326, cg13476133 and cg24202936 (Supplementary Fig. 6). This result agrees with a previous publication that included the same dataset.⁵³

Table 2.

Heritability estimates of WMH-associated CpGs

CpG	Nearest gene	Reduced model			Full model
		h 2 met h	SE	P	h 2 met h	SE	P
cg0150647	SDK1	0.02	0.07	0.38	0.01	0.07	0.42
cg03116124	TRIM67	0.01	0.07	0.45	0.01	0.07	0.47
cg06809326	CCDC144NL	0.26	0.07	1.03 × 10–4a	0.27	0.07	9.51 × 10–5a
cg13476133	GCK	0.09	0.07	0.11	0.09	0.07	0.12
cg14133539	FO × D4	0.08	0.07	0.14	0.07	0.07	0.17
cg14547240	C1QTNF7	0.06	0.07	0.20	0.06	0.07	0.18
cg17417856	PRMT1;ADM5	0.40	0.08	1.37 × 10–8a	0.40	0.08	3.06 × 10–8a
cg17577122	CLDN5	0.14	0.08	2.80 × 10–2	0.15	0.08	2.27 × 10–2
cg21547371	MUSTN1	0.00	–	0.50	0.00		0.50
cg23054394	SPSB4	0.00	–	0.50	0.00		0.50
cg23586595	PLAC8	0.24	0.08	1.47 × 10–3a	0.23	0.08	2.51 × 10–3a
cg24202936	LOC441601	0.15	0.07	1.34 × 10–2	0.16	0.07	1.17 × 10–2

h ² _meth = the narrow-sense heritability; SE = standard error. Reduced model is adjusted for age, sex, blood cell counts, principal components of the ancestry and technical covariates. Full model is additionally adjusted for BMI and smoking.

Significant after adjustment for multiple testing burden.

Mendelian randomization analyses between target CpGs and WMH burden

Forward two-sample multiple IV MR analysis was performed for two target CpGs, cg06809326 and cg24202936, which have at least three independent cis-mQTL IVs in Huan et al.⁵³ (Supplementary Table 11). We found a marginally significant causal relationship from cg06809326 to WMH burden (P = 2.91 × 10⁻²). Higher methylation level at the locus is associated with greater WMH burden {odds ratio (OR) [95% confidence interval (CI)] = 1.39 (1.03, 1.87)}. Evidence was lacking for horizontal pleiotropy (P = 0.41) or heterogeneity (P = 0.42) (Supplementary Table 12). In reverse-MR analysis, evidence that WMH causally influence methylation levels at any of the target CpGs was lacking (Supplementary Tables 12 and 13).

Using the same three IVs, we also investigated whether cg6809326 is causally associated with expression of nearby genes (step one). Two cis transcripts were annotated to this CpG in GTEx version eight data.⁸² They both encode a long non-coding RNA designated as CCDC144NL and CCDC144NL-AS1, and we identified one IV for both transcripts. In all five brain tissues, we found evidence of causal association between cg06809326 and both CCDC144NL and CCDC144NL-AS1 (Supplementary Table 14). In step two, a marginal association between CCDC144NL and WMH burden was observed in caudate basal ganglia and cortex (step one P = 1.11 × 10⁻³ and step two P = 3.94 × 10⁻² in caudate basal ganglia; step one P = 1.21 × 10⁻³ and step two P = 4.28 × 10⁻² in cortex).

DNAm at established GWAS loci and WMH burden

We estimated the combined effect of DNAm at each locus from our EWAS results at the 21 established GWAS loci.⁸ Consistent with our DMR results, CpGs at the GWAS locus SH3PXD2A were jointly associated with WMH (P = 8.48 × 10⁻³), but evidence of DNAm effects on WMH at other loci was lacking (Supplementary Table 15). We also conducted a multiple-trait co-localization analysis (moloc)⁵⁸ of brain mQTL and expression QTL (eQTL),⁵⁷ and WMH-associated single nucleotide polymorphisms (SNPs). At 17 out of the 21 GWAS loci, we identified significant co-localization evidence (PPFC > 0.7) (Supplementary Table 16 and Supplementary Fig. 7). At eight loci, the SNPs with the highest PPFC were the sentinel SNPs in the GWAS.

Candidate genes implicated by gene-expression associations with the target CpGs

At the Bonferroni threshold (6.93 × 10⁻⁵ = 0.05/722 cis genes in ±5 Mb of the target CpGs), we identified significant associations between cg23586595 and PLAC8 (P = 2.98 × 10⁻⁷) and between cg24202936 and F2 (P = 6.39 × 10⁻⁵) (Table 3). Adjusting for additional covariates (smoking status and BMI) did not change these associations.

Table 3.

Cis genes (±5 Mb) whose expression is significantly associated with identified CpGs

CpG	Gene	Region (hg19)	n	Z reduced	P reduced	Z full	P full
cg23586595	PLAC8	4:84011211–84138405	2687	−5.13	2.98 × 10–7	−5.11	3.27 × 10–7
cg24202936	F2	11:46740749–46761054	1963	−4.00	6.39 × 10–5	−4.01	6.04 × 10–5

Z-scores and P-values from the reduced and full model are presented. Reduced model is adjusted for age, sex, blood cell counts, principal components of the ancestry and technical covariates. Full model is additionally adjusted for BMI and smoking. PLAC8 = placenta-associated 8; F2 = coagulation factor II.

Cg24202936, cg01506471 and cg06809326 showed significant correlation estimates (|r| > 0.3) between blood and brain (r = 0.33, 0.87 and 0.57, respectively) (Supplementary Table 17) and, thus, were tested for co-localization. We found that mQTLs for cg24202936 and WMH GWAS SNPs colocalize with FOLH1 expression in dorsolateral prefrontal cortex (DLPFC) (PPFC = 0.75) (Supplementary Table 18). Also, suggestive evidence existed for co-localization of cg06809326 mQTLs, CCDC144NL-AS1 eQTLs and WMH SNPs (PPFC = 0.69).

Integrative cross-omics analysis

Integrative cross-omics analysis identifies novel gene regulatory networks

At FDR <0.05, we identified 576 WMH-associated gene sets enriched from the integrated data of GWAS, EWAS and TWAS out of 12 303 gene sets from curated databases.^45,63–66 Top associated gene sets includes ‘regulation of actin cytoskeleton’ (P = 1.14 × 10⁻⁴⁵, 211 genes), ‘telomeres, telomerase, cellular ageing and immortality’ (P = 1.10 × 10⁻³⁵, 18 genes), ‘integrin-mediated cell surface interactions’ (P = 3.17 × 10⁻³⁴, 84 genes), ‘thrombin signalling through proteinase activated receptors’ (P = 1.41 × 10⁻³³, 32 genes) and ‘Nef protein mediated CD4 down-regulation’ (P = 4.70 × 10⁻³², nine genes). All enriched pathways with FDR P < 0.05 are listed in Supplementary Table 19.

We derived two WMH burden-associated gene networks in brain. The first network is comprised of four subnetworks. Five key driver genes (FMOD, COL3A1, SERPING1, SLC13A4 and ISLR) represent a subnetwork of ‘extracellular matrix (ECM) organization, ECM–receptor interaction, focal adhesion and collagen formation’. Additionally, three related subnetworks, ‘smooth muscle contraction’ with key driver TAGLN; ‘G-protein-coupled receptor ligand binding’ with key drivers GAL, ECEL1, ESR1 and NTS; and ‘cytokine signalling in immune system’ with key drivers IFIT1 and RTP4, make up the network (Fig. 2 and Supplementary Table 20). We also identified an independent second network associated with ‘lipid and lipoprotein metabolism’, with key driver gene KNG1. Genes included in each subnetwork are presented in Supplementary Table 21.

Figure 2.

WMH-associated gene networks. WMH-associated genes based on multi-molecular evidence are organized around the 19 key driver genes. (A) WMH-associated network consisting of four subnetworks—extracellular matrix (ECM) organization (FMOD, COL3A1, SEPING1, SLC13A4 and ISLR); smooth muscle contraction (TAGLN); G-protein-coupled receptor ligand binding (GAL, ECEL1, ESR1 and NTS) and cytokine signalling in immune system (IFIT1 and RTP4). (B) WMH-associated network of lipid and lipoprotein metabolism (KNG1). Key drivers and associated gene networks identified in the Mergeomics analysis are coloured in orange. Neighbouring genes are grouped into networks and labelled in random colours.

Overlap-based drug-repositioning analysis of WMH-associated genes

Using drug signatures derived from in vivo cardiovascular and nervous system data, we predicted antihyperlipidaemic drugs, including PPAR-α (peroxisome proliferator-activated receptor-alpha) agonist ‘fenofibrate’, as the top therapeutic target. Using drug signatures derived from murine oligodendroglial precursor cells data, we predicted several small molecules, including a glycogen synthase kinase inhibitor and a phenylalanyl tRNA synthetase inhibitor that may have therapeutic potential for Alzheimer’s disease⁸³ and autoimmune diseases,⁸⁴ respectively (Supplementary Table 22).

Discussion

This first EWAS of WMH burden in 9732 middle-aged to older adults from 14 community-based cohorts identified several novel epigenetic loci. Although we could not independently replicate the association of single CpGs with WMH, probably due to a limited sample size and differences between the discovery and replication sample, functional annotation and bioinformatic analyses provided strong supportive evidence. Moreover, powerful DMR analyses identified 46 DMRs of which 13 were replicated. Integrative analyses of multi-omics information also suggested novel gene networks with key drivers and potential drug targets for WMH.

We identified a novel epigenetic locus, cg24204936, mapping to a pseudogene SEPTIN7P11. Functional integration revealed two candidate genes whose expression may be influenced by variation in DNAm at this locus: F2 in blood and FOLH1 in DLPFC. Prothrombin encoded by F2 plays an essential role in blood clot formation, angiogenesis, tissue repair and vascular integrity. A prothrombotic state or circulating prothrombin has been reported for symptomatic cSVD,^85,86 WMH and stroke.^87,88 However, it remains unclear whether coagulation plays a major role in the aetiology of WMH or is secondary to injury to the cerebral small vessels and white matter.⁸⁹FOLH1 encodes glutamate carboxypeptidase II that catalyses the hydrolysis of N-acetylaspartylglutamate. An elevated level of N-acetylaspartylglutamate in the CSF has been reported in two patients with almost complete absence of myelin in the CNS⁹⁰ and has been proposed as a diagnostic biomarker for rare diseases of the white matter.⁹¹

An epigenetic locus mapping to PRMT1, which encodes a protein arginine N-methylase, was identified in single-CpG and DMR analyses and also as a shared epigenetics locus with BP. The biological link between DNAm at PRMT1 and WMH burden may involve pathways related to endothelial dysfunction, which have previously been implicated in WMH aetiology.⁹²PRMT1, a predominant member of the PRMT family, methylates histone and non-histone proteins to regulate various cellular functions.⁹³PRMT1 is essential for the development of neurons, astrocytes and oligodendrocytes and is critical for myelin formation.⁹⁴ PRMTs also catalyse the formation of ADMA (asymmetric dimethylarginine), which reduces nitric oxide production, promotes endothelial dysfunction in the BBB and triggers the immune response in atherosclerosis.^92,95,96 Higher ADMA levels have been repeatedly associated with cSVD and its monogenic form, cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy.^97–103

Single-CpG association combined with functional genomic analyses and DMR analyses identified a novel epigenetic locus near CCDC144NL;CCDC144NL-AS1 (coiled-coil domain containing 144 family and its antisense RNA1). Cg06809326 is under strong cis-genetic control and brain expression of its nearest gene, CCDC144NL;CCDC144NL-AS1, may mediate the association between DNAm and WMH burden (Supplementary Fig. 8). A TWAS of WMH using blood gene-expression data⁶⁰ did not report a significant association for CCDC144NL;CCDC144NL-AS1 expression, possibly due to its low expression in blood. CCDC144NL-AS1 encodes a long non-coding mRNA transcript that controls expression of target genes by acting as a molecular sponge for various regulatory miRNAs.^104–109In vitro studies have uncovered several of its target genes with potentially relevant function to cSVD. These include matrix metalloproteinases MMP2 and MMP9,¹⁰⁸ F-actin and vimentin,¹¹⁰ and transforming growth factor beta (TGF-β)-activated kinase 1 (TAK1).¹⁰⁶ MMP2 and MMP9 can damage the BBB by triggering recruitment of immune cells¹¹¹ and have been implicated in white matter injury and cSVD.¹¹² F-actin plays an important role in maintaining the shape of endothelial cells and the integrity of the BBB.¹¹³ Disturbed TGF-β signalling has been implicated in the pathogenesis of several monogenic forms of cSVD.^114–117 Deficiency of TAK1 in mouse brain endothelial cells resulted in endothelial cell death, small vessel rarefaction and disruption of the BBB.¹¹⁸

A central role of endothelial dysfunction, possibly resulting in a compromised BBB, in WMH burden¹¹⁹ is further suggested by identified DNAm associations in genes involved in cell junctions. Claudins are integral membrane proteins that comprise tight junctions specifically in brain microvascular endothelial¹²⁰ cells and that regulate BBB permeability.¹²¹ Claudin-5 mapped to cg17577122 is the most enriched tight junction protein in the BBB, and its dysfunction has been implicated in neurodegenerative and neuroinflammatory diseases, and cSVD.^122–126 A recent DMR analysis using DLPFC DNAm levels also identified CLDN5 to be associated with cognitive decline.¹²⁷ In normotensive subjects, we identified a CpG in cadherin 18 (CDH18) that encodes an adherens junction protein, which mediates calcium-dependent cell–cell adhesion. CDH18 is also involved in cell junction organization process and in cell signalling pathways including G-proteins signalling together with F2.¹²⁸

Our DMR analysis, aggregated epigenetic associations using Brown’s method⁵⁹ and moloc analysis using brain QTL data consistently identified an epigenetic association at a known WMH GWAS locus, SH3PXD2A (SH3 and PX-domain-containing protein 2A).^8,72–74 Several genome-wide associations with WMH-related traits have also been reported at SH3PXD2A, including white matter microstructure¹²⁹ and stroke.^130,131SH3PXD2A encodes an adaptor protein (TKS5) involved in the formation of podosomes that act as sites of close contact to as well as degradation of ECM.¹³² Gene set enrichment of identified DNAm loci and integrative cross-omics analyses collectively point to a central role of the ECM in WMH burden. One of the two WMH burden networks identified through the Mergeomics approach centred around key driver genes involved in ECM organization and function and the top associated module was ‘regulation of actin cytoskeleton’. Notably, actin polymerization and disassembly of junctional proteins within microvascular endothelial cells were shown to play a key role in early BBB disruption in a murine model.¹³³

Another network includes genes that function in lipid and lipoprotein metabolism and our overlap-based drug-repositioning analyses suggested antihyperlipidaemic drugs as potential drug targets. A recent MR analysis showed that genetically increased high-density lipoprotein cholesterol level was associated with lower WMH volume and lower risk of small vessel stroke.¹³⁴ Statin therapy for cSVD has also been regarded as promising since individuals with high WMH burden typically carry higher vascular risk factors. Few randomized clinical trials assessing the effect of lipid lowering on WMH progression have been conducted and they have generally provided mixed results.^135–137 While they suggest a possible role of statins, in particular rosuvastatin, in preventing WMH progression, the lack of high-quality data prevents strong evidence-based recommendation at this time.¹³⁸ It has been postulated that statins improve endothelial function and stabilize the BBB in cSVD.^139,140 Studies that investigated membrane proteins including phospholipid flippase (ATP11B) and aquaporin-4 showed that the loss of these proteins cause pathological features of cSVD including endothelial cell dysfunction with reduced tight junctions, nitric oxide, oligodendrocyte progenitor cell maturation block and microglial activation.^126,141

Finally, this study provides further emphasis concerning the long-observed perivascular inflammation as an additional crucial player in cSVD pathology and provides a possible explanation. Interestingly, gene set enrichment analyses identified a possible role of the defence response to viral infection with several DMR-associated genes related to interferon gamma signalling and the innate immune response (DTX3L-PARP9, BNIP3 and IFITM1). Our top associated CpG has been previously reported in an EWAS of chronic HIV infection⁷¹ and our drug-repositioning analysis also identified a HIV antiviral as a possible drug target. Several studies have reported that people with HIV are at higher risk of an increased burden of WMH compared to uninfected controls.^142,143

Several limitations of our study must be acknowledged. First, many of our EWAS discoveries were not independently confirmed. Since a series of functional analyses showed biological relevance, we suspect that the lack of replication may stem from the limited size of the replication sample and from differences between the discovery and replication samples as hinted by the increased heterogeneity in the DNAm association observed in the meta-analysis (Supplementary Table 2). Indeed, variation in WMH burden was smaller in the replication studies than in the discovery studies perhaps due to the younger age of the participants. The younger cohorts, CARDIA (n = 277) with a mean age 53.9 years and SHIP (n = 214) with a mean age 49.7 years, make up only 8.59% of the discovery sample; whereas the Rhineland Study with a mean age 54.1 years and FHS third generation cohort with a mean age of 47.1 years, make up over 86% of the replication sample. Replication of several WMH-associated loci identified through more powerful DMR analyses further underscore an underpowered replication study for single-CpG associations. Additional studies are needed to confirm the findings presented here. Second, we conducted a subgroup analysis stratified by hypertension status, but statistical power in each stratum was limited. A more ideal design to study this and other modifiable risk factors of cSVD will be a longitudinal study or a stratified association study of a larger sample size. Similarly, our study was not sufficiently powered to examine ancestry-specific associations of DNAm with WMH and possible ancestry difference in epigenetic patterns could not be investigated. Third, we did not adjust for additional lifestyle factors or comorbidities to maximize our sample size by minimizing the number of covariates in the models. Our primary goal was to identify novel DNAm loci associated with WMH burden and we cannot exclude the possibility that the identified loci may reflect, in part, variation in those risk factors. Fourth, the currently publicly available brain QTL data are limited to cis-regions of omics markers and, thus, our in silico bioinformatics analyses were restricted only to the CpGs with substantial cis-acting genetic influence. For example, cg17417856 in PRMT1 had a strong heritability estimate (h² = 0.40, P = 1.37 × 10⁻⁸) but was not followed-up because it was under polygenic control. Last, the study was conducted in blood and cell type-specific associations, most notably in brain, may have been missed. To extrapolate the findings in blood to brain, we assessed the correlation with DNAm in brain, and used available brain QTL data. Due to the difficulties of getting both brain DNAm and MRI data from a large population-based sample, an EWAS of WMH burden using brain DNAm may not be easy to achieve. However, findings from this large blood-based study may provide a basis for an epigenetic candidate gene study in the brain.

Funding

We thank all study participants for contributing to this work. This project was largely supported by US National Institutes of Health grant R01NS087541 with additional support from grants U01AG052409 and R01HL105756. A full set of study-specific funding sources is provided in the Supplementary material. R.F.G. is supported by the NINDS Intramural Research Program.

Competing interests

C.S. is a consultant of Novartis pharmaceuticals on a safety study for heart failure. J.W. is supported by the UK Dementia Research Institute, which is funded by the UK Medical Research Council, Alzheimer’s Society and Alzheimer’s Research UK. B.M.P. serves on the Steering Committee of the Yale Open Data Access Project funded by Johnson & Johnson. H.J.G. has received travel grants and speaker’s honoraria from Fresenius Medical Care, Neuraxpharm, Servier and Janssen Cilag as well as research funding from Fresenius Medical Care.

Supplementary material

Supplementary material is available at Brain online.