Integrative co-methylation network analysis identifies novel DNA methylation signatures and their target genes in Alzheimer’s disease

Abstract

INTRODUCTION:

DNA methylation is a key epigenetic marker, and its alternations may be involved in Alzheimer’s disease. CpGs sharing similar biological functions or pathways tend to be co-methylated.

METHODS:

We performed an integrative network-based DNA methylation analysis on two independent cohorts (N=941) using brain DNA methylation profiles and RNA-Seq as well as AD pathology data.

RESULTS:

Weighted co-methylation network analysis identified six modules as significantly associated with neuritic plaque burden. Fifteen hub-CpGs including three novel CpGs were identified and replicated as being significantly associated with AD pathology. Furthermore, we identified and replicated four target genes (ATP6V1G2, VCP, RAD52, LST1) as significantly regulated by DNA methylation at hub-CpGs. In particular, VCP gene expression was also associated with AD pathology in both cohorts.

DISCUSSION:

This integrative network-based multi-omics study provides compelling evidence for a potential role of DNA methylation alternations and their target genes in Alzheimer’s disease.

1. INTRODUCTION

Alzheimer’s disease (AD) is the most common cause of dementia, characterized by the accumulation of amyloid-β plaques and neurofibrillary tangles.[1] As the population ages, AD becomes a widely recognized major public health problem globally.[2] Over time, extensive research efforts have improved our understanding of clinical, radiological, and pathological characteristics of AD. However, the etiology of AD and the process of the disease occurring before the deposition of amyloid plaques remain unclear. Despite longstanding global efforts and numerous failed trials, the FDA just approved aducanumab, the first potentially disease-modifying anti-amyloid treatment.

Besides the well-known familial AD-related genes, such as PS1, PS2, and APP,[3] more than 30 susceptibility genes have been identified in large-scale genome-wide association studies and sequencing studies.[4–6] However, except for APOE ε4 alleles, most genetic variants in AD have relatively small effect sizes, implicating other possible pathogenic mechanisms beyond genotypic variation for explaining the missing heritability. In light of this, several recent studies have investigated the role of epigenetic modifications in AD.[7–9]

Epigenomics is a field of study that examines all epigenetic modifications at a whole genome level. Epigenetic modifications include the processes that alter gene activity without changing the DNA sequence, such as DNA methylation or histone modifications.[10] Major epigenetic features can be interrogated comprehensively by combining cellular, biochemical and molecular techniques with high-throughput sequencing.[11] Researchers can then examine any association between the epigenome and clinical or pathologic phenotypes. For DNA methylation, the technology for epigenome-wide association study (EWAS) is now available.[12]

EWAS is an efficient way to identify differentially methylated positions throughout the epigenome. EWAS using brain tissue-based DNA methylation data has identified several loci associated with AD pathology, and some of the loci have been reported consistently across studies, such as those annotated to ANK1, RHBDF2, and HOXA3.[7–9,13] Furthermore, a recent meta-analysis of EWAS, which used DNA methylation data from six studies, reported CpGs associated with neuropathology.[14] However, EWAS approaches have examined associations of methylation levels of only individual CpGs with a trait of interest. Due to multiple testing correction for many CpGs used in the analysis, biologically meaningful findings with statistical significance at a suggestive level may be missed. In addition, it is possible to assume that CpGs sharing their functions or biological pathways tend to be co-methylated.[15–17] Considering the functional complexity of the genome and the dynamic nature of epigenetic regulation,[18] an analytical approach that takes the correlation between methylation levels of CpGs into account might identify novel CpGs or genes not identified by EWAS. In this regard, a co-methylation network analysis may be beneficial because: (1) it can identify co-methylation network modules that share biological pathways and (2) it can identify novel CpGs or genes by increasing statistical power from a dimension reduction. Currently, however, the application of co-methylation network analysis to epigenetic studies for neurodegenerative diseases is limited,[19,20] and AD-related research that uses epigenome-wide data for co-methylation network analysis is lacking.

Integration with other types of omics data may enhance interpretability of DNA methylome data.[21] There are a few algorithms available regarding the integration of DNA methylation data with gene expression data.[22–25] Of those algorithms, the functional epigenetic modules(FEM) algorithm integrates DNA methylation and gene expression data and identifies gene modules in relation to a phenotype of interest. In this study, however, we tried to identify co-methylation modules using an unsupervised manner and perform downstream analysis with the detected modules.

Hence, we performed a weighted co-methylation network analysis to identify co-methylation modules associated with AD pathology and hub-CpGs in the significant modules. The hub-CpGs significantly associated with AD pathology were further investigated to identify target genes that were regulated by the hub-CpGs by integrating DNA methylation and transcriptomics data. Finally, we performed mediation analysis to investigate whether expression levels of the target genes mediate the association between methylation levels of the hub-CpGs and AD pathology biomarkers. Of note, we identified and replicated CpGs and target genes associated with AD pathology using DNA methylation data from two independent cohorts.

2. METHODS

2.1. Study Participants

Data used in the study were obtained from the Religious Orders Study and the Memory and Aging Project (ROS/MAP; N=740)[26,27] and the Mount Sinai Brain Bank AD Study (MSBB; N=201)[28] as discovery and replication samples, respectively (see Supplementary Material for detail). Clinical and pathological characteristics of participants are shown in Supplementary Table 1.

2.2. Neuropathological Assessment of AD

Postmortem brains in ROS/MAP were processed and examined following a standard procedure as described previously.[9,29] In this study, we used a composite score of neuritic plaque burden and the CERAD neuropathological category[30] as primary phenotype variables in ROS/MAP. In MSBB, we used the CERAD neuropathological category because quantitative neuritic plaque density data were not available (see Supplementary Material for detail).

2.3. DNA Methylation

All DNA methylation and gene expression datasets were downloaded from the AMP-AD knowledge portal (https://www.synapse.org) (syn3168775, syn23650893, syn21347197, syn20801188). In the ROS/MAP cohort, the procedures for DNA extraction, methylation array assay, and quality controls were described elsewhere.[31,32] Briefly, DNA methylation in the prefrontal cortex tissue was interrogated using a bead assay (Infinium Human Methylation450; Illumina). Probes with poor quality based on detection p-values (p>0.01) and cross-reactive and polymorphic probes were removed. K-nearest neighbor algorithm was used for imputation of missing values. The normalization of CpG probes to account for differences between type I and type II probes was performed using the BMIQ algorithm from the Watermelon package.[33] At the sample level, principal component analysis (PCA) was used to evaluate subject-specific data quality. PCA was performed using a random selection of 50,000 autosomal probes for all the samples. Samples that had the first three principal component (PC) values within 3 standard deviations from the means were included. Using these criteria, 748 subjects out of a total of 761 subjects were accepted. Eight subjects had poor bisulfite conversion (BC) efficiency, which is defined as having at least 2 of the 10 BC controls that fail to reach a value of 0.8. After quality control, a total of 420,132 autosomal CpGs for 740 samples were selected for downstream analysis.

In the MSBB cohort, DNA was extracted from postmortem brain tissue in the parahippocampal gyrus using the Qiagen All Prep DNA/RNA Mini Kit according to the manufacturer’s instructions. DNA methylation array assays were carried out using Infinium MethylEPIC BeadChips. Raw data from IDAT files were preprocessed and normalized using the R package minfi, generating β-values to depict methylation levels at each CpG. After removing CpGs with a high missing rate (>0.05) and polymorphisms, missing values were imputed using the K-nearest neighbor algorithm. Finally, 817,331 CpGs for 201 samples were available for replication analysis. For both datasets, we used M-values (logit-transformed β-values) for statistical analysis.

2.4. RNA-Seq Gene Expression

In the ROS/MAP cohort, RNA-Seq expression data were generated from frozen dorsolateral prefrontal cortex tissues. For the MSBB cohort, we used RNA-Seq expression data generated from Brodmann area 36 to match the anatomic location of DNA methylation samples. Detailed procedures for sequencing and data processing for the ROS/MAP[34] and MSBB[35] cohorts have been described in previous studies.

RNA expression levels of protein-coding genes within a range of ±1M base pairs from differentially methylated CpGs were investigated for associations with DNA methylation levels and AD pathology. Expression levels of genes showing significant associations in the ROS/MAP cohort were further investigated for replication in the MSBB cohort.

2.5. Statistical Analysis

2.5.1. Construction of co-methylation network and identification of modules associated with AD pathology

To identify network modules of highly correlated CpGs in an unsupervised manner, we constructed a scale-free, signed co-methylation network using the WGCNA R package. We used the biweight mid-correlation method to calculate the correlation between CpGs with a soft-thresholding power of 12. The minimum number of probes within modules was set to 50. The methylation pattern of CpGs in a module was represented by the module eigen-CpG (ME), which is defined as the 1^st PC of the methylation matrix of the corresponding module. The association between MEs of modules and neocortical neuritic plaque burden was tested using linear models, adjusting for age at death, sex, study (ROS or MAP), experimental batch, BC efficiency, and postmortem interval (PMI). To adjust for multiple testing, we used Bonferroni corrected p-values. We also tested whether the associations between MEs and plaque burden remain after additionally controlling for the neuronal proportion and the 1^st PC. The neuronal proportion for each sample was estimated using the CETS R package,[36] which quantifies neuronal proportions from DNA methylation data using cell epigenotype specific (CETS) marks.

2.5.2. Gene set enrichment analysis

We performed gene set enrichment analyses to identify the biological pathways shared by probes in modules significantly associated with AD pathology using the gometh function in the missMethyl R package. The gometh function tests for gene ontology (GO) terms using a hypergeometric test, considering the number of CpG probes per gene on the Illumina methylation array. To adjust for multiple testing, the false discovery rate (FDR) <0.05 was set as the threshold.

2.5.3. Hub CpG site identification and its association with AD pathology

Instead of using all CpGs in modules significantly associated with neuritic plaques, we identified hub-CpGs highly interconnected in the modules for subsequent analyses. To this end, we selected CpGs as hub-CpGs that were (1) within the top 10%ile intramodular connectivity and (2) had a module membership higher than 0.8. Module membership is defined as the correlation between methylation values of a probe within a module and ME of the module. The selected hub-CpGs were then tested for association with AD pathology. In this step, we used the CERAD neuropathological category instead of neuritic plaque burden because it was available in both ROS/MAP and MSBB. Because the CERAD category is an ordinal variable, we performed ordinal logistic regression. Age, sex, study, batch, BC efficiency, and PMI were used as covariates. The CpGs with FDR corrected p-values <0.05 were investigated for replication analysis using the MSBB dataset. For the MSBB cohort, we included age, sex, and PMI as covariates.

2.5.4. Functional annotation of replicated CpG sites and mediation analysis

To evaluate the functional annotation of the replicated CpGs in AD, we first assessed the association between methylation levels of CpGs and expression levels of genes within ±1M base-pairs from the CpGs using linear regression models. Using FDR correction, the p-values were adjusted for the numbers of genes within ±1M base-pair ranges. For the genes showing significance, we further investigated whether the gene expression levels are associated with AD pathology. Using linear regression models, we tested the association between gene expression levels and neuritic plaque burden. Also using the lavaan R package, we tested whether gene expression levels mediate the association between methylation levels of CpGs and AD pathology for the genes showing significant association with both methylation level and pathologic burden. Age, sex, study, and PMI were used as covariates. For the MSBB cohort, we included age, sex, and PMI as covariates.

2.5.5. Protein-protein interaction network

We constructed a protein-protein interaction (PPI) network using the online visual analytics platform, NetworkAnalyst, to evaluate the connectivity among genes related to our replicated CpGs. We used the STRING interactome database[37] to construct the network. We included genes in which the CpGs were located and the nearest genes for GpGs in the intergenic region. We also included genes showing significant associations with methylation levels of the replicated CpGs. The centrality measures, such as degree or betweenness, as well as GO pathway enrichment for the network, were available in the web platform.

3. RESULTS

3.1. Study participants

A total of 740 participants from the ROS/MAP cohort and 201 participants from the MSBB cohort were used for discovery and replication analyses, respectively (Supplementary Table 1). In terms of clinical diagnoses, 312 (43.3%) ROS/MAP participants and 143 (71.1%) MSBB participants had dementia. Neuropathological evaluations showed that 471 (63.6%) ROS/MAP participants and 112 (55.7%) MSBB participants met the criteria for pathological diagnosis of AD (probable or definite AD).

3.2. Identification of AD-associated co-methylation modules and gene-set enrichment analysis

Co-methylation network analysis identified 67 modules, and the number of CpGs in each module ranged from 53 to 34,797 (excluding the grey module, which comprises unassigned CpGs). Among these modules, six modules were significantly associated with neuritic plaque burden after adjusting for multiple testing using Bonferroni correction (Table 1). These associations remained significant after adjusting for the estimated neuronal proportion and the 1^st PC. For these significant six modules, gene-set enrichment analysis showed that the skyblue3 module was significantly enriched in neuron or synapse related biological pathways (Table 2).

Table 1.

Top 10 associations between MEs of modules and neuritic plaque burden

Module	# of probes	β(SE)	pBonferroni	Adjusted for NeuN+		Adjusted for NeuN+ & PC1
				β(SE)	pBonferroni	β(SE)	pBonferroni
orangered4	795	0.006(0.001)	0.001	0.006(0.001)	0.002	0.006(0.001)	0.002
steelblue	1372	0.006(0.001)	0.002	0.006(0.001)	0.003	0.006(0.001)	0.003
navajowhite2	100	−0.006(0.001)	0.006	−0.005(0.001)	0.013	−0.005(0.001)	0.011
skyblue3	971	−0.006(0.001)	0.008	−0.005(0.001)	0.039	−0.005(0.001)	0.03
floralwhite	261	0.005(0.001)	0.01	0.005(0.001)	0.005	0.006(0.001)	0.004
greenyellow	10810	0.004(0.001)	0.042	0.004(0.001)	0.049	0.003(0.001)	0.007
lightcyan1	557	−0.005(0.001)	0.114	−0.005(0.001)	0.023	−0.005(0.001)	0.013
orange	3413	−0.004(0.001)	0.123	−0.004(0.001)	0.159	−0.004(0.001)	0.169
plum	53	0.004(0.001)	0.124	0.002(0.001)	1	0.002(0.001)	0.279
thistle2	123	−0.004(0.001)	0.191	−0.004(0.001)	0.223	−0.004(0.001)	0.083

ME = Module Eigen-CpG, SE = Standard error, NeuN+ = Neuronal proportion, PC1 = 1^st principal component

Table 2.

Gene-set enrichment analysis of the skyblue3 module

Pathway	Ngenes	Hits	p	p(FDR)
synapse organization	377	26	3.5×10−6	0.006
cytoskeleton-dependent intracellular transport	174	15	6.0×10−6	0.008
nervous system development	2258	82.5	1.5×10−5	0.018
organelle transport along microtubule	82	10	2.0×10−5	0.020
axo-dendritic transport	67	9	2.0×10−5	0.020
regulation of microtubule polymerization or depolymerization	72	9	2.1×10−5	0.020
regulation of cation channel activity	169	14	2.3×10−5	0.020
regulation of microtubule-based process	205	15	2.6×10−5	0.023
transport along microtubule	155	13	2.8×10−5	0.023
neurogenesis	1553	62.5	3.8×10−5	0.028
generation of neurons	1455	59.5	4.8×10−5	0.030
microtubule polymerization or depolymerization	104	10	6.3×10−5	0.038
regulation of transporter activity	264	17	7.4×10−5	0.043
microtubule-based transport	179	13	8.7×10−5	0.047
cell junction organization	627	32	9.1×10−5	0.047
protein localization to cell junction	89	10	9.2×10−5	0.047
regulation of ion transmembrane transporter activity	241	16	9.6×10−5	0.047
microtubule-based process	709	30	9.7×10−5	0.047

FDR = False discovery rate

3.3. Identification of hub-CpG sites in modules and differential methylation analysis

For the six co-methylation modules significantly associated with neuritic plaque burden, we identified hub-CpGs in each module. With 1,358 CpGs satisfying the criteria for hub-CpG, we performed differential methylation analysis using ordinal logistic regression models and identified 140 hub-CpGs as significantly associated with a CERAD neuropathological category after adjusting for multiple testing using FDR correction.

Out of the 140 hub-CpGs identified in the ROS/MAP dataset, 124 CpGs were available in the MSBB dataset. Ordinal logistic regression analysis between methylation values of these CpGs and a CERAD neuropathological category replicated 15 differentially methylated positions (DMPs) with significant associations after FDR correction. Thirteen DMPs were in gene body regions, and two genes (HOXA3 and PRRT1) included more than one DMP (Table 3, Supplementary Table 2). The association between methylation levels of replicated DMPs and pathologic markers are visualized in Fig. 1.

Table 3.

Differentially methylated CpGs and related genes

		Discovery		Replication		Nearest Gene		Functionally associated Genes
									ROS/MAP				MSBB
									Meth-Expr		Expr-Pathology		Meth-Expr		Expr-Pathology
CpG site	Position	β (SE)	pfdr	β (SE)	pfdr	Gene(distance(bp)	Gene-relation	Gene(distance(kbp))†	β	pfdr	β	p	β	p	β	p
cg18668327	2:11262719	0.38(0.13)	0.041	1.33(0.48)	0.048	C2orf50(10460)
cg23859635	2:42795262	1.35(0.26)	2.2E-04	1.54(0.39)	0.003	MTA3(11)	exon,promoter
cg06108383	6:32120899	0.84(0.22)	0.008	1.38(0.36)	0.003	PRRT1(786)	exon,promoter	LST1(565.5)	0.13	0.013	0.055	0.719	0.58	4.2E-06	0.418	0.032
cg17113856	6:32120895	0.97(0.23)	0.003	1.55(0.38)	0.002	PRRT1(790)	exon,promoter	LST1(565.4)	0.14	0.006	0.055	0.719	0.57	5.0E-06	0.418	0.032
cg25251478	6:30853959	0.86(0.27)	0.034	1.35(0.41)	0.011	DDR1(435)	intron	IER3(141.6)	−0.34	0.002	−0.034	0.662	0.32	0.037	0.271	0.133
								LST1(699.9)	0.14	0.037	0.055	0.719	0.47	0.001	0.418	0.032
								ATP6V1G2(660.4)	−0.075	0.037	−0.207	0.443	−0.51	3.0E-05	−1.728	7.1E-10
cg01027532	7:27155974	1.28(0.29)	0.002	1.85(0.50)	0.003	HOXA3(3240)	intron
cg03536885	7:27163820	0.99(0.28)	0.02	1.74(0.54)	0.012	HOXA3(2819)	intron
cg05287480	7:27176127	0.84(0.25)	0.027	1.15(0.33)	0.006	HOXA3(3698)	intron
cg12305431	7:27157855	1.09(0.23)	0.001	1.28(0.37)	0.006	HOXA3(1359)	intron
cg07963670	8:23161764	1.07(0.23)	0.001	1.60(0.42)	0.003	LOXL2(99989)	intron	RHOBTB2(296.4)	0.076	0.028	1.011	1.2E-05	−0.2	0.007	−0.907	0.019
cg08578641	9:34457440	0.80(0.18)	0.002	0.82(0.26)	0.012	FAM219A(1110)	intron	IL11RA(193.3)	−0.081	0.001	−0.703	0.001	0.068	0.071	1.479	0.002
								VCP(600.4)	−0.027	0.026	−0.98	0.042	−0.076	0.007	−2.827	2.1E-05
cg22552736	9:132258136	0.98(0.24)	0.004	1.67(0.45)	0.003	NTMT1(113027)		ENDOG(677.4)	−0.17	0.031	−0.216	0.046	−0.26	0.062	−0.607	0.011
cg17824906	12:1700011	0.78(0.20)	0.009	1.62(0.43)	0.003	WNT5B(16188)	intron	RAD52(600.8)	0.077	0.005	0.061	0.787	0.15	0.022	1.916	3.3E-05
cg24761525	17:46651745	1.24(0.29)	0.003	1.60(0.36)	0.001	HOXB3(104)	exon
cg08230957	19:39087186	0.97(0.21)	0.001	1.53(0.42)	0.003	MAP4K1(547)	intron

Chr = Chromosome, SE = standard error, FDR = False discovery rate, bp = base pairs, kbp = kilo base pairs, Meth=methylation, Expr = Expression

^*The CpG site is on the gene body

^†Genes showing FDR corrected p-value lower than 0.05 were included

Figure 1.

Association between methylation levels of differentially methylated positions and neuritic plaque burden. Functionally annotated CpGs are included. Plots regarding other CpGs are shown in Supplementary Figure 1.

3.4. Functional annotation of the replicated CpGs

3.4.1. Association of gene expression levels with methylation levels of DMPs

Using gene expression data of 546 ROS/MAP participants whose RNA-Seq data were available, we performed association analysis of methylation levels of the replicated DMPs with expression levels of genes within ±1M base pairs from them. Out of 378 genes in the range, expression data of 272 genes were available. The association analysis showed that expression levels of 59 genes were associated with methylation levels of corresponding CpGs at the nominal level of significance (p < 0.05). Of these 59 genes, eight genes (IL11RA, IER3, RAD52, LST1, VCP, RHOBTB2, ENDOG, and ATP6V1G2) remained significant after FDR correction (Table 3). Of these eight genes, four genes (RAD52, LST1, VCP, ATP6V1G2) were replicated in the MSBB cohort. Expression levels of VCP and ATP6V1G2 were negatively associated with the methylation levels of corresponding CpGs, whereas RAD52 and LST1 showed positive associations.

3.4.2. Association of gene expression levels with neuritic plaque burden

For the eight genes significantly associated with methylation levels of corresponding CpGs, we investigated whether expression levels of the genes were also associated with neuritic plaque burden in the ROS/MAP cohort. The association analysis identified four genes (RHOBTB2, IL11RA, VCP, ENDOG) as significantly associated with neuritic plaque burden (Table 3). Of these four genes, two genes (VCP and ENDOG) were replicated in the MSBB cohorts, i.e., expression levels of the two genes were significantly associated with CERAD neuropathological categories (Table 3).

3.4.3. Mediation analysis

For two genes (VCP and ENDOG) that were associated with both methylation levels and neuritic plaque burden, we investigated whether the genes mediated the association between methylation and neuritic plaque burden. The mediation analysis showed that the VCP gene mediated the association between methylation levels of DMPs and neuritic plaque burden in the MSBB cohort (Direct: β(se) 0.208(0.148), p=0.007; Indirect: β(se) 0.058(0.047), p=0.019) and marginally mediated the association between methylation levels of DMPs and neuritic plaque burden in the ROS/MAP cohort (Direct: β(se) 0.202(0.122), p=3.79×10⁻⁵; Indirect β(se) 0.009(0.019), p=0.230) (Fig. 2) despite the same direction of effect. The ENDOG gene mediated the association between methylation levels of DMPs and neuritic burden in the MSBB cohort (Direct: β(se) 0.185(0.313), p=0.047; Indirect: β(se) 0.132(0.217), p=0.040), but not in the ROS/MAP cohort (Direct: β(se) 0.205(0.165), p=2.91×10⁻⁵; Indirect β(se) 0.003(0.026), p=0.743).

Figure 2.

Methylation-gene expression-pathology mediation analysis for VCP and ENDOG gene in (A) the ROS/MAP and (B) MSBB cohorts. Standardized coefficient (p-value) is indicated for each edge in mediation models.

3.4.4. Protein-protein interaction (PPI) network of significant genes

A PPI network was constructed using genes related to the 15 CpGs identified and replicated as significantly associated with neuropathology biomarkers. We included 15 nearest genes and 8 genes significantly associated with methylation levels. The resulting network included 15 genes that we used (Figure 3), of which VCP, RHOBTB2, MTA3, and MAP4K1 showed high degree and betweenness (Supplementary Table 3). Gene-set enrichment analysis showed that the genes in the PPI network were enriched in pathways related to cell cycle such as apoptosis, cell differentiation, and neurogenesis (Supplementary Table 4).

Figure 3.

Protein-protein interaction network. Darker colors indicate higher degree.

4. DISCUSSION

We performed weighted co-methylation network analysis to identify key modules, hub-CpG sites, and possible target genes, which were regulated by hub-CpGs, in Alzheimer’s disease by integrating DNA methylation and RNA-Seq data from two independent cohorts as discovery and replication samples, respectively. Our network analysis identified six modules as significantly associated with neuritic plaque burden. We also identified and replicated 15 hub-CpGs as significantly associated with AD neuropathology biomarkers. In addition, we identified and replicated four target genes that were regulated by hub-CpGs. Our comprehensive and integrative multi-omics analysis identified three novel CpGs (cg18668327, cg01027532, cg03536885) and genes with a significant functional association (LST1, VCP, ATP6V1G2, ENDOG, RAD52) as well as 12 previously reported DMPs. Of note, expression levels of the VCP gene mediated the association between methylation levels of hub-CpGs and AD pathology.

Co-methylation network analysis identified six modules as significantly associated with neocortical neuritic plaque burden. In a similar context where genes with related functions tend to be co-expressed,[38] it is possible to assume that CpG sites that share functions or biological pathways tend to be co-methylated.[15] Indeed, one of the significant six modules was enriched in biological pathways primarily related with neurons and the central nervous system. Among the pathways, neurogenesis and nervous system development might affect brain reserve, which can buffer the effect of dementia-related pathology.[39] Pathways related to microtubule-based transport were also significantly enriched. This may be related to the fact that dystrophic neurites surrounding amyloid cores in neuritic plaques are known as sites of microtubule disruption.[40]

Of 15 DMPs replicated in our analysis, 12 DMPs have been reported in a meta-analysis of epigenome-wide association studies[14] and three DMPs were newly identified. In fact, two of them (cg01027532, cg03536885) were CpGs in the HOXA3 gene region, a gene well-known for showing differential methylation levels. However, cg18668327, which was annotated to C2orf50, was not reported previously. Although the C2orf50 gene is highly expressed in the brain, little research has been done regarding its function or association with the central nervous system. A recent study has reported that C2orf50 is differentially methylated in autistic participants.[41] However, its potential role in neurodegenerative diseases has yet to be investigated in detail.

Because 12 of the DMPs were not functionally investigated in the previous meta-analysis, we examined them using RNA-Seq data to identify functionally associated target genes and their mediation effects. We identified eight target genes for which expression levels were associated with methylation levels of the replicated CpGs. The ATP6V1G2 gene, which is involved in lysosomal transport, has been recently reported to be downregulated in AD brains.[42] ATP6V1G2 encodes subunit G2 of vacuolar ATPase (V-ATPase), and a deficiency can cause neurodegenerative diseases such as AD and PD.[43] This is in line with our finding that methylation levels of cg25251478 were negatively associated with ATP6V1G2 expression levels, and the methylation levels were positively associated with AD pathology.

The VCP gene showed the most consistent results across the cohorts, in terms of methylation-expression and expression-AD pathology relationship. The protein encoded by the VCP gene plays a role in protein degradation, intracellular membrane fusion, DNA repair and replication, and regulation of the cell cycle.[44] Although VCP is known to be related to frontotemporal dementia,[45] a certain type of mutation in the VCP gene (p.Asp395Gly) causes tau aggregates similar to AD neurofibrillary tangles.[46] In light of this finding, VCP has been nominated as a candidate target gene by the Accelerating Medicines Partnership – AD (AMP-AD) Consortium. It is notable that VCP showed the highest centrality measure in the protein-protein interaction network, implying that it might play a significant role in AD pathogenesis.

The expression levels of the LST1 gene were associated with methylation levels of three CpGs on chromosome 6 (cg06108383, cg17113856, cg25251478) in both cohorts. Although the association between the expression level of LST1 and pathology was not significant in the ROS/MAP cohort, the association was significant in MSBB participants. While the definitive function of the LST1 gene remains elusive, there have been some studies showing its association with the immune process. Specifically, LST1 polymorphisms are associated with inflammatory disorders such as systemic lupus erythematosus or rheumatoid arthritis.[47,48] Also, the expression level of LST1 was associated with both presence[49] and the severity of rheumatoid arthritis.[50] With regard to central nervous system diseases, there is a postmortem brain transcriptomics study showing an association between the expression of LST1 and depression.[51] However, studies examining the role of LST1 in AD or other neurodegenerative diseases are lacking. Future investigation of the role of LST1 in AD pathogenesis especially in relation to neuroinflammation may be beneficial.

The ENDOG gene encodes a nuclear encoded endonuclease (endonuclease G) that is localized in the mitochondrion. Our results showed that the expression level of ENDOG was associated with neuritic plaques in both cohorts. A recent study identified polymorphisms of ENDOG as associated with longitudinal change of brain connectivity using the Alzheimer’s Disease Neuroimaging Initiatives dataset.[52] An in vitro study showed that the endonuclease G apoptotic pathway is activated in a certain neurotoxic stimulus.[53] Apoptosis and disruption of brain connectivity can both contribute to the pathogenesis of AD. Further studies elucidating the specific biological contribution of ENDOG in AD pathogenesis are needed.

The expression levels of the RAD52 gene were significantly associated with methylation levels of cg17824906 in both cohorts. Also, in the MSBB cohort, expression levels of RAD52 showed significant mediation effects between methylation levels and AD pathology. The RAD52 protein is known to play a role in homologous recombination, a repair process for double-strand breaks caused by oxidative damage. In an in vitro study, high concentrations of Aβ_1–42 oligomers reduced the RAD52 protein levels in post-mitotic neurons. Interplay between the RAD52 protein-related DNA repair process and AD pathogenesis, which are regulated by epigenetic modification may warrant further investigation.

A few limitations of this study should, however, be noted. First, DNA methylation data in the MSBB cohort was obtained using brain tissue in the parahippocampal gyrus while DNA methylation data in the ROS/MAP cohort was obtained using brain tissue from the prefrontal cortex. Nevertheless, it is notable that our replicated findings are consistent and robust in both cohorts. Second, we used different measures for quantification of AD pathology in two cohorts. In the ROS/MAP cohort, neuritic plaque burden was used as an AD pathology biomarker, and in the MSBB cohort, the CERAD estimates of neuritic plaque pathology were used as an AD pathology biomarker, as neuritic plaque burden was not available. To ensure the robustness of replicated DMPs, we used the CERAD neuropathological category as a pathological marker for association analyses in both cohorts. Third, since the WGCNA algorithm was developed for gene expression analysis, the different nature between methylation and expression data might have affected our results. Of note, however, several recent studies reported significant findings using WGCNA in relation to methylation data.[54,55] Fourth, while we performed cell type-adjusted analysis using a method widely used in previous studies, the method does not account for the heterogeneity of non-neuronal component, which is composed of many different cell types. Last, the results of mediation analysis were not replicated across the cohorts. Further studies in independent larger cohorts are needed to validate the mediation effects.

In conclusion, the integrative approaches of co-methylation network analysis and functional gene expression analysis using multi-omics data (epigenomics and transcriptomics) identified novel CpGs and target genes as significantly associated with pathology biomarkers for AD. The ATP6V1G2, VCP, RAD52, ENDOG, and LST1 genes might be potential targets for further investigation considering their known relevance with neurodegeneration, cell cycle, and inflammation. Genes such as ATP6V1G2 and VCP are known to be related to AD or neurodegenerative dementia, and the possible impact of epigenetic trans-regulation on their roles in AD has not been studied. Our results provide a fundamental basis for further investigation regarding this topic.

KEY RESOURCES TABLE.

Resource Type	Specific Reagent or Resource	Source or Reference	Identifiers	Additional Information
Add additional rows as needed for each resource type	Include species and sex when applicable.	Include name of manufacturer, company, repository, individual, or research lab. Include PMID or DOI for references; use “this paper” if new.	Include catalog numbers, stock numbers, database IDs or accession numbers, and/or RRIDs. RRIDs are highly encouraged; search for RRIDs at https://scicrunch.org/resources.	Include any additional information or notes if necessary.
Deposited Data; Public Database	ROSMAP methylation array data, Homo sapiens	AMP-AD knowledge portal, ROS/MAP cohort	syn3168775
Deposited Data; Public Database	ROSMAP gene expression data (RNA seq), Homo sapiens	AMP-AD knowledge portal, ROS/MAP cohort	syn23650893
Deposited Data; Public Database	MSBB methylation array data, Homo sapiens	AMP-AD knowledge portal, MSBB cohort	syn21347197
Deposited Data; Public Database	MSBB gene expression data (RNA seq), Homo sapiens	AMP-AD knowledge portal, MSBB cohort	syn20801188
Software; Algorithm	WGCNA	Langfelder P and Horvath S, WGCNA: an R package for weighted correlation network analysis. BMC Bioinformatics 2008, 9:559 doi:10.1186/1471-2105-9-559	N/A	Available at Bioconductor
Software; Algorithm	CETS	Guintivano, J., Aryee, M. J. & Kaminsky, Z. A. A cell epigenotype specific model for the correction of brain cellular heterogeneity bias and its application to age, brain region and major depression. Epigenetics 8, 290–302, doi:10.4161/epi.23924	N/A	Direct contact with the developer is needed