Leukocyte-adjusted epigenome-wide association studies of blood from solid tumor patients

Abstract

Epigenome-wide studies of DNA methylation using blood-derived DNA from cancer patients are complicated by the heterogeneity of cell types within blood and the associated cell lineage specification of DNA methylation signatures. Here, we applied a novel set of analytic approaches to assess the association between cancer case-status and DNA methylation adjusted for leukocyte variation using blood specimens from three case-control cancer studies (bladder: 223 cases, 205 controls; head and neck: 92 cases, 92 controls; and ovarian: 131 cases, 274 controls). Using previously published data on leukocyte-specific CpG loci and a recently described approach to deconvolute subject-specific blood composition, we performed an epigenome-wide analysis to examine the association between blood-based DNA methylation patterns and each of the three aforementioned solid tumor types adjusted for cellular heterogeneity in blood. After adjusting for leukocyte profile in our epigenome-wide analysis, the omnibus association between case-status and methylation was significant for all three studies (bladder cancer: P = 0.047; HNSCC: P = 0.013; ovarian cancer: P = 0.0002). Subsequent analyses revealed that CpG sites associated with cancer were enriched for transcription factor binding motifs involved with cancer-associated pathways. These results support the existence of cancer-associated DNA methylation profiles in the blood of solid tumor patients that are independent of alterations in normal leukocyte distributions. Adoption of the methods developed here will make it feasible to rigorously assess the influence of variability of normal leukocyte profiles when investigating cancer related changes in blood-based epigenome-wide association studies.

Introduction

Genome-wide association studies (GWAS) have impacted our comprehension of genetic susceptibilities to a broad variety of complex diseases. However, DNA sequence variations identified through GWAS are estimated to account for less than 30% of phenotypic variability in humans,¹ and explain only a small portion of familial cancer risk.² Epigenetic mechanisms likely account for some of this variability and this possibility has been a catalyst for the development of epigenome-wide association studies (EWAS).³ In particular, the involvement of epigenetics in carcinogenesis is widely appreciated.⁴ There is intense interest in DNA methylation as cancer biomarkers, owing to the chemical stability of these marks, their involvement in crucial genomic functions, as well as the potential for modification by clinically relevant environmental exposures.

Despite the superficial similarities between GWAS and EWAS, there are fundamental differences between genetic and epigenetic variation that must be considered.³ When using blood derived DNA, different cell lineages show striking differences regarding epigenetic marks,^5-11 which is in sharp contrast to the uniform allelic variation in SNPs among different blood cell types. This means that global epigenetic patterns are complex functions of normal blood cell development.^12,13 However, there is also evidence that exposures to endogenous and exogenous agents can alter DNA methylation^14-16 and may differentially affect tissues depending on the degree of contact with the tissue, cellular proliferation rate, and other characteristics impacting susceptibility to epigenetic insult. Since personal characteristics (e.g., age, gender, and race), behaviors (e.g.. smoking and alcohol consumption), exogenous exposures, and pathologic states can modulate the circulating immune profile,^17-23 there is a high potential for confounding by inter-subject immune variations. Cell lineage specification of DNA methylation has not been formally and rigorously interrogated until now in the EWAS literature. Given that EWAS is inherently more confounded by the mixtures of cells in samples compared with GWAS, new analytical strategies must be developed to simultaneously assess cell lineage epigenetic patterns when searching for alterations that are disease and/or exposure related.

In order to evaluate the influence of leukocyte distribution on DNA methylation it is necessary to access the relevant epigenetic subtypes within blood. There are three general approaches in evaluating proportions of circulating leukocytes²⁴: manual hematological cytology, automated cell counting with differential, and antibody-based flow cytometry (FACS). The first two approaches are widely used clinically but do not delineate epigenetically distinct lymphocyte subsets, such as T cells, B cells or natural killer (NK) cells, or T cell subsets, including CD4+, CD8+, and regulatory T cells (Treg). FACS has the potential to delineate these and other subsets but generally requires fresh blood samples and is too costly and laborious for routine comprehensive blood profiling.^25-28 Moreover, as all of these approaches require intact cells, they cannot be applied to studies in which intact cells are not accessible or only extracted DNA is available. In response to these limitations, and because so few population data sets for EWAS include detailed cell type information, we have developed a relevant immune cell epigenetic reference library and statistical approach⁶ that can be readily applied to peripheral blood methylomic data (which is particularly well suited for this purpose due to its relative stability and strong role in differentiation) for evaluating leukocyte subtypes. This approach, which has been independently validated,²⁹ renders the aforementioned hematological methods unnecessary for direct ascertainment of subtype data from populations under study, and allows for application to peripheral blood DNA (whereas the latter approaches all require intact cells). We judge that these issues urgently need attention as demonstrated by numerous recent reports of cancer-related methylation profiles from blood in patients with various solid tumor types, including breast,³⁰ ovarian,^31,32 pancreatic,³³ bladder,³⁴ head and neck,³⁵ colorectal,³⁶ and lung cancers.³⁷ A very recent paper showed genetic risk for the immune disorder rheumatoid arthritis was mediated by DNA methylation, but this association was only observed after adjusting for normal leukocyte distributions.³⁸ As observed cancer-associated methylation in these studies may be substantially driven by cancer-associated shifts in the proportions of circulating leukocytes,^6,8 it remains unclear if additional variation in DNA methylation can be detected in blood that is independent of interpersonal leukocytic variability.

Here we present a novel approach to blood-based EWAS in three independent case-control studies of solid tumor patients designed to assess cancer-associated methylation states that arise independently of shifts in relative leukocyte proportions in the blood. Our data suggest that using lineage-specific DNA methylation information to account for leukocyte composition can overcome confounding by cell mixtures, and promises to greatly improve blood-based epigenome-wide association studies.

Results

EWAS were performed on three separate previously described population-based case-control studies of bladder cancer (223 cases, 205 controls),^39,40 head and neck squamous cell carcinoma (HNSCC; 92 cases, 92 controls),^35,41 and a publicly available data set for ovarian cancer (131 cases and 274 controls).³² A description of each of the three study populations is provided in Table 1.

Table 1. Description of each of the three study populations by case-control status.

	Bladder Cancer (New Hampshire)				HNSCC (greater Boston)				Ovarian Cancer (United Kingdom)
	Cases (n = 223)	Controls (n = 205)	P value		Cases (n = 92)	Controls (n = 92)	P value		Cases (n = 131)	Controls (n = 274)	P value
Age, median years (range)	66.0 (25–74)	64.0 (28–74)	0.05a		58.0 (31–84)	59.0 (32–86)	0.54a		67.0 (49–91)	64.0 (52–78)	0.08a
Gender
Female	52 (23.3%)	65 (31.7%)	0.06b		28 (30.4%)	28 (30.4%)	>0.99b		266 (100%)	274 (100%)	>0.99b
Male	171 (76.7%)	140 (68.3%)			64 (69.6%)	64 (69.6%)			0 (0.0%)	0 (0.0%)
Race
White	223 (100%)	205 (100%)	>0.99b		84 (91.3%)	85 (92.4%)	>0.99b		—	—	—
Non-White	0 (0.0%)	0 (0.0%)			8 (8.7%)	7 (7.6%)			—	—
Cigarette smoking
Never-smoker	40 (17.9%)	58 (28.3%)	<0.001b		17 (18.5%)	32 (34.8%)	0.04b		—	—	—
Former smoker	111 (49.8%)	112 (54.6%)			59 (64.1%)	47 (51.1%)			—	—
Current smoker	72 (32.3%)	35 (17.1%)			16 (17.4%)	13 (14.1%)
HPV serology (E6, E7 or L1)
Negative	—	—	—		66 (71.7%)	83 (90.2%)	0.002b		—	—	—
Positive	—	—			26 (28.3%)	9 (9.8%)			—	—

Abbreviations: HNSCC = head and neck squamous cell carcinoma; ^aWilcoxon rank-sum test for a difference between cases and controls; ^bFisher's exact test for a difference between cases and controls.

Cancer-associated differential methylation

For each study, differential methylation between cases and controls was examined among all autosomal loci (26 486 CpG sites) and among CpGs that did not significantly vary between leukocyte types (leukocyte-specific differentially methylated positions [DMPs]). After exclusion of 9849 putative leukocyte DMPs and adjustment for false-discovery rate,⁴² we observed 3,987 differentially methylated loci (Q ≤ 0.05) for bladder cancer cases relative to controls (Table S1), representing 24% of the 16 637 non-leukocyte-specific loci. Of the 3987 differentially methylated loci, 2105 loci were observed to be hypomethylated among bladder cancer cases, with a median methylation β-value that was 0.017 lower (roughly corresponding to a 2% decrease in methylation), and 1882 were hypermethylated, with a median β-value that was 0.032 higher among these cases compared with controls. A much lower proportion of non-leukocyte-specific loci were found to be differentially methylated in peripheral blood between the HNSCC cases and controls: after adjusting for multiple comparisons, there were only 4 significantly differentially methylated CpGs (Q ≤ 0.05), all of which were hypomethylated in cases, with a median difference in β-value of 0.015, constituting less than 0.1% of total non-leukocyte specific CpGs. Importantly, the sample-size for the HNSCC study (n = 184) was much smaller than for the bladder cancer study (n = 428) impacting our power to detect significant differences. Likewise, a smaller number of differentially methylated loci were detected between ovarian cancer cases and controls, with 17 differentially methylated loci (Q ≤ 0.05). Of these, 9 were hypomethylated in cases, with a median difference in methylation β-value that was 0.009 lower, and 17 were hypermethylated with a median difference that was 0.003 higher.

By contrast, the frequency of differentially methylated loci changed when differential methylation was assessed for all 26,486 autosomal CpG loci (Table S2). This was particularly evident for the HNSCC and ovarian cancer studies.

Leukocyte-adjusted epigenome-wide association studies

Since potentially important information may be lost through simple exclusion of putative DMPs, we applied an approach to blood-based EWAS that takes into account all autosomal CpG loci, while adjusting for leukocyte composition using subject-specific estimates inferred through epigenetic deconvolution of blood (Fig. 1).⁶ The omnibus cancer-specific associations with DNA methylation in blood (Fig. 2A) were significant for all three studies, both prior to adjusting for leukocyte composition (β) and after adjusting for leukocyte composition (α). However, the volcano plots contrasting effect size and significance level for individual loci for each respective study (Figs. 2B–D) clearly illustrate the difference in results that were obtained before and after adjusting for leukocyte composition (if there were no difference, the points would perfectly overlap). This was not unexpected, given that a significant cancer-associated difference in circulating immune profile (Г) was observed in all three studies. Cancer-associated differences in the proportions of the specific leukocyte types are presented in Figure 3. Further, we conducted sensitivity analyses to rule out the possibility that treatment or racial/ethnic differences were driving the observed case-associated blood methylation for bladder cancer and HNSCC, respectively. We excluded all non-White subjects in the HNSCC study (7 cases, 8 controls; all bladder cancer subjects were White, and race was not available for the publicly available UK ovarian cancer cohort) and all bladder cancer cases that received treatment other than surgery (55 cases; all ovarian and HNSCC blood draws occurred prior to treatment), which yielded similar results (Figs. S1 and S2).

Figure 1. Schematic of epigenome-wide association study (EWAS) approaches using blood. (A) Typical EWAS approach using blood, where the marginal effect of an exposure on blood DNA methylation is examined without accounting for the effect leukocyte distribution on the methylation profile. (B) Leukocyte-adjusted EWAS approach to examine the direct effect of an exposure or phenotype (e.g., case-status) on DNA methylation in blood.

Figure 2. Epigenome-wide analysis of cancer-associated DNA methylation and shifts in leukocyte composition in peripheral blood for bladder, head and neck (HNSCC) and ovarian cancers. (A) Omnibus P values for cancer-associated DNA methylation relative to control subjects, unadjusted (β) and adjusted (α) for leukocyte composition, and cancer-associated shifts in relative leukocyte proportions in the blood (Γ). Leukocyte-adjusted (α; depicted as black diamonds) and –unadjusted (β; blue open circles) associations for individual CpG loci for each respective solid tumor type are presented in the volcano plots (B–D), which depict –log(P values) vs. regression coefficients for case-status adjusted for age. The models for bladder cancer and HNSCC were additionally adjusted for sex and smoking habit. The horizontal red line on each plot represents the –log value where P = 0.05. Note: β = Association between case-status and methylation without adjustment for leukocyte composition; α = Association between case-status and methylation adjusted for leukocyte composition; Γ = Association between case-status and leukocyte composition

Figure 3. (A) Relative percent differences for each leukocyte subtype are for cases compared with controls in studies of bladder cancer, HNSCC, and ovarian cancer. (B) Average leukocyte proportions in the control subjects from each respective case-control study. Relative leukocyte proportions were determined by applying the deconvolution methods of Houseman et al. (2012) to the peripheral blood methylation profile of each study subject. Note: *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001; ****P ≤ 0.0001; Black error bars represent the standard error for each estimate. Bladder cancer and HNSCC estimates are adjusted for age, gender, and smoking status (current/former/never); ovarian cancer estimates are adjusted for age.

The top ten CpG loci (based on ranked absolute α coefficient for each respective leukocyte-adjusted EWAS) are presented in Table 2, along with CpG island status, distance from transcription start site and information on the associated gene. There were no individual CpGs common to all three cancers among their respective top 50 loci (Table S3), although bladder and ovarian cancer each had one of their top hits located within the promoter region of DPPA3 (albeit at separate loci); this gene encodes developmental pluripotency-associated protein 3, which is an enzyme associated with embryonic development and stem cell differentiation.⁴³

Table 2. Top 10 cancer-associated CpG loci in blood for bladder, head and neck (HNSCC) and ovarian cancer after adjustment for leukocyte composition, based on ranked test statistic.

Locus ID	Regression Coefficient (α)		Chromosome	Located in a CpG Island	Distance from TSSa (bases)	Associated Gene	Transcription Factor	Associated Gene Descriptionb	CpG-me Correlationc with Gene Expression in Tumor (P value)
Bladder Cancer
cg07510052	+0.020		5	Yes	41	HIGD2A	No	Proposed subunit of cytochrome c oxidase	n/a
cg06572974	-0.057		15	No	1200	CHD2	Yes	Alters expression through chromatin structure modifications. Alternatively spliced transcript variants encoding distinct isoforms have been found	ρ = 0.18 (0.02)
cg00429618	+0.014		9	Yes	440	ZNF322B	No	Zinc Finger Protein 322 Pseudogene	ρ = 0.04 (0.62)
cg25307081	+0.031		16	Yes	140	BRD7	Yes	Component of the SWI/SNF chromatin-remodeling complex. Alternative splicing results in multiple transcript variants	n/a
cg07081888	+0.040		19	Yes	470	NALP4	No	Involved in the activation of proinflammatory caspases	n/a
cg08950375	+0.031		2	No	580	COL6A3	No	Encodes the α-3 chain of type VI collagen	ρ = -0.03 (0.68)
cg09425215	-0.027		15	Yes	428	CHD2	Yes	See above	n/a
cg20938689	+0.031		5	Yes	131	COX7C	No	Catalyzes the electron transfer from reduced cytochrome C to oxygen	n/a
cg00461841	-0.038		16	No	760	ATF7IP2	No	Recruiter that couples transcriptional factors to general transcriptional machinery apparatus, modulating transcription regulation and chromatin formation	ρ = 0.18 (0.02)
cg20150565	-0.026		20	Yes	743	ZNFX1	Yes	NFX1-Type Zinc Finger-Containing Protein 1	ρ = 0.06 (0.47)
HNSCC
cg03835292	-0.034		22	Yes	607	LGALS2	No	Binds β-galactoside. Function is presently unknown.	ρ = 0.08 (0.37)
cg05473677	-0.034		3	Yes	330	OSTalpha	No	heterodimerizes with OSTbeta for bile acid export from enterocytes into blood	ρ = -0.05 (0.60)
cg08241785	0.022		5	No	41	F2RL2	No	Plays an essential role in hemostasis and thrombosis	ρ = -0.02 (0.84)
cg00725635	0.022		1	No	82	B3GALT2	No	Membrane-bound glycoprotein with diverse enzymatic functions	ρ = -0.03 (0.76)
cg17628717	-0.020		7	Yes	154	HECW1	No	Mediates ubiquitylation and subsequent degradation of DVL1	ρ = -0.08 (0.38)
cg11435943	-0.022		18	No	706	SERPINB12	No	Inhibitor of trypsin and plasmin. Reported to be downregulated in oral squamous cell carcinoma (Shiiba M et al. [2010]).	ρ = 0.12 (0.17)
cg02449608	-0.014		19	No	696	C19orf18	No	Uncharacterized	ρ = 0.04 (0.64)
cg25943702	-0.026		22	No	99	BRD1	No	Component of the MOZ/MORF complex which has a histone H3 acetyltransferase activity	ρ = -0.03 (0.75)
cg22346765	-0.019		6	No	594	UNC5CL	No	Inhibits NF-kappa-B-dependent transcription	ρ = -0.02 (0.82)
cg03266453	0.020		2	Yes	296	EN1	Yes	Homeobox protein involved in developmental control	ρ = 0.12 (0.19)
Ovarian Cancer
cg01980222	-0.032		6	No	3	TREM2	No	Forms a receptor signaling complex with TYROBP and triggers activation of the immune responses in macrophages and dendritic cells. Alternative splicing results in multiple transcript variants encoding different isoforms	ρ = 0.06 (0.31)
cg24211388	-0.041		6	No	196	AIF1	No	Involved in vascular inflammation and phagocytosis and thought to be involved in negative regulation vascular smooth muscle cell growth. Overexpressed in recurrent ovarian tumors (Jinawath et al. [2010]). Three transcript variants encoding different isoforms have been found	ρ = 0.07 (0.29)
cg00974864	-0.039		1	No	152	FCGR3B	No	Low affinity receptor for the Fc region of gamma immunoglobulins (IgG). Several transcript variants encoding different isoforms have been found	ρ = -0.09 (0.13)
cg06872381	-0.046		12	Yes	13	DPPA3	No	Involved in epigenetic chromatin reprogramming in the zygote	ρ = 0.15 (0.01)
cg06543018	+0.020		3	Yes	332	RBP1	No	Involved in the transport of retinol from the liver storage site to peripheral tissue. Decreased expression has been associated with malignant transformation of ovarian epithelium (Roberts D et al. [2002]). Multiple transcript variants encoding different isoforms have been found	ρ = -0.14 (0.02)
cg26205131	+0.020		5	Yes	667	SLC6A3	No	Dopamine transporter with reported SNP associations with body mass index (Azzato EM et al. [2009])	ρ = -0.23 (0.0001)
cg12910797	-0.055		17	Yes	88	HOXB3	Yes	Homeobox protein involved in developmental control	ρ = -0.23 (0.0002)
cg19910382	-0.050		2	No	17	FABP1	No	Involved in fatty acid uptake, transport, and metabolism	ρ = 0.10 (0.11)
cg00645579	-0.037		11	Yes	1213	IRF7	Yes	Key regulator of type I interferon (IFN)-dependent immune responses. Has been identified as a BRCA1 transcriptional target (Buckley NE et al. [2007]). Multiple transcript variants have been identified	ρ = -0.17 (0.005)
cg09964921	-0.034		21	No	1194	KCNE1	No	Involved in potassium ion channel formation. Alternative transcript variants have been identified	ρ = 0.10 (0.12)

^a Absolute distance in bases from transcription start site; ^bDescription adapted from www.GeneCards.org; ^cMethylation-expression Spearman correlation coefficients (ρ) for tumor were calculated using Infinium 450K HumanMethylation and mRNA-Seq v2 data publicly available through The Cancer Genome Atlas (http://cancergenome.nih.gov/). Paired methylation-expression data was available for 165 bladder cancers, 133 HNSCC, and 258 ovarian cancers.

Motivated by the hypothesis that altered DNA methylation observed in cases may be specifically targeting genes with common sequence features, we evaluated the DNA sequences within 1kb of the CpG dinucleotides associated with altered DNA methylation (after leukocyte adjustment) for the presence of transcription factor binding sites (TFBS). There were 29 TFBS that were overrepresented in the bladder cancer study, 27 for the HNSCC study, and 33 for the ovarian cancer study. Seventeen TFBS were overrepresented in at least two of the studies, while one of the overrepresented TFBS was common to all three cancers (EVI1). The entirety of observed overrepresented TFBS proximal to cancer-associated loci is presented in Figure 4.

Figure 4. Overrepresentation of transcription factor binding sites (TFBS) within 1kb of cancer-associated CpG loci (after leukocyte-adjustment) in the blood of patients with bladder, head and neck (HNSCC), and ovarian cancer. Each circle depicts overrepresented TFBS motifs for each respective cancer; those listed in overlapping regions indicate overrepresented TFBS common to two or more of the cancer types.

Discussion

We have presented an approach for conducting blood-based EWAS in solid tumor studies that, along with the approaches recently described by Liu et al.³⁸ and Elliott et al.⁴⁴ for non-cancer associated disease and smoking exposure, respectively, offers a practical solution for dealing with cellular heterogeneity of blood samples in EWAS, advancing the methodology in this field. In applying these methods, we have provided support for the existence of cancer-specific methylation patterns in the blood of bladder cancer, HNSCC, and ovarian cancer patients that were independent of differences in leukocyte profile. Our observation of cancer-specific DNA methylation profiles in blood is in line with other reports of such findings in the literature.^{8,30-34,36,37} Importantly, our results suggest that differential methylation patterns in the blood of patients with solid tumors may arise as a result of biological mechanisms beyond what can be attributed to differential methylation between the major normal leukocyte cell types. It should be noted that it is presently unknown to what degree the observed differences in blood reflect the methylation patterns of the tumor. Further, while these findings are provocative, the cross-sectional measurement of blood methylation at diagnosis does not allow us to draw any definitive conclusions on the temporality of the observed epigenetic marks, although a recent prospective study identified differentially methylated loci in the blood of breast cancer patients that preceded diagnosis.⁴⁵ As such, the source of the observed cancer-specific methylation differences remains unclear at the present.

While it is accepted that circulating tumor DNA is detectable in the blood stream of a subset of patients with solid tumors, including bladder cancer,^46-51 HNSCC,^52-55 and ovarian cancer,^56-59 it is improbable that the observed methylation differences in peripheral blood between cancer cases and controls are driven by this phenomenon, particularly in the absence of enrichment for free DNA in the plasma, as circulating tumor DNA, when present, occurs at an extremely low concentration in blood⁶⁰ and, thus, would have a negligible impact on the methylation signature of peripheral blood.

The analytic applications presented here required high-density epigenome-wide methylation data from sorted non-pathologic human leukocyte samples for the identification of putative leukocyte lineage-specific DMPs, along with corresponding epigenomic data for the peripheral blood samples from the case-control studies for conduct of the EWAS. Even though our analyses and epigenetic library were derived from the Infinium 27K BeadArray, approximately 94% of the loci from the 27K platform are featured on the Infinium 450K BeadArray platform, and any of the loci derived from the array could be extended to sequencing data based on their annotated genomic coordinates. Our findings were bolstered by the population-based nature of the case-control studies, strengthening the generalizability, and by the availability of well-characterized epidemiologic data, particularly for the bladder cancer and HNSCC study populations. These attributes made it possible to control for possible confounding by exposure and/or personal characteristics related to the cancer of interest.

Among the potential limitations of the study was the cross-sectional procurement of blood samples upon study enrollment. While the case-control study design allows us to identify cancer-associated DNA methylation in blood, the lack of temporality precludes us from knowing whether the observed differences in methylation preceded the development of the disease or if they occurred as a direct or indirect result of the cancer itself. Prospective studies are needed to establish temporality and determine the time frame for which epigenetic alterations occur in blood during the carcinogenic process of solid tumors. Additionally, the moderate sample size may have limited our statistical power to detect associations (particularly for the HNSCC study). Further, while the focus of this manuscript was the application of a practical EWAS approach to account for cellular heterogeneity, independent validation of the findings reported herein for each of the three solid tumor types should be performed in future studies. Further, although we have identified differentially methylated loci across major leukocyte types (i.e., CD4+ T cells, CD8+ T cells, NK cells, B cells, monocytes, and granulocytes), we cannot rule out additional epigenetic distinctions, either from unmeasured subpopulations or activated versions of measured cell types (including possible cancer-specific immune subtypes or compartments) as a source of residual confounding. While several of the top loci are situated within cancer-associated genes, many others lie within genes with known immune function, raising the prospect that they are unaccounted DMRs in unmeasured leukocyte subtypes. This may be best exemplified by the presence of a top 10 loci for ovarian cancer that is associated with TREM2, a gene with known involvement in immune activation of macrophages and dendritic cells.⁶¹ It is also important to note that cancer-specific changes that occur solely in a minor subtype may not be detected using this method. However, these would also not likely be detected in unadjusted assessments in the absence of isolation and purification of the single subtype for all study subjects. Finally, the vast majority of participants for the three studies presented herein were Caucasian, which may affect the generalizability of our findings to other races/ethnicities. Future studies should seek to incorporate additional racial/ethnic subgroups to determine if these findings apply across different populations, as well as extending similar research to other solid tumor types, particularly those with reported alterations in blood methylation.

Our findings underscore the importance of adjusting for interpersonal leukocyte variability in blood-based EWAS. The methods described herein present new opportunities for researchers to account for such variations for the conduct of these studies, particularly as large international consortiums are underway to establish epigenetic roadmaps of human cells,⁶² from which similar information on DMPs can be derived. This is particularly important, as blood is commonly collected and archived as part of many epidemiologic studies, and thus the ability to accurately analyze disease associated DNA methylation modifications may open new doors for biomarker discovery or individual risk assessment. Elucidation of the mechanisms driving cancer-associated methylation changes in peripheral blood and their impact on cancer risk will enhance our understanding of the epigenetic contribution to the carcinogenic process and will better inform biomedical scientists and clinicians as to how to use and interpret blood-based methylation data.

Methods

Study populations

Epigenome-wide methylation analyses were performed on three separate population-based case-control cancer studies. The first included 223 incident bladder cancer cases, identified from the NH State Cancer Registry from July 1, 1994 through June 30, 1998 as part of a previously described study,^39,40 with available blood samples and 205 healthy population-based control subjects identified through public records, also with available blood samples and no prior history of cancer (Gene Expression Omnibus [GEO] accession number: GSE50409). Upon enrollment, consenting subjects underwent personal interviews furnishing sociodemographic, health behavior and exposure information. Blood was collected for cases on average within 1-y of diagnosis; however only 24.7% of cases (n = 55) received any treatment beyond surgery (i.e., radiation, BCG, or chemotherapy). The second study consisted of 92 incident head and neck squamous cell carcinoma (HNSCC) cases and 92 cancer-free control subjects with available blood samples (GEO: GSE30229), randomly selected from a previously described case-control study of head and neck cancer.^35,41 Briefly, the study was comprised of incident head and neck cancer patients from the greater Boston area and population-based controls from the same region with no prior history of cancer, frequency-matched on age (+/− 3 y), gender, and town of residence. Subjects completed a self-administered questionnaire that provided data on sociodemographic, health behavior and exposure history. All case blood draws were performed prior to initiation of radiation or chemotherapy. Serologic HPV16 testing for E6, E7, and L1 viral proteins was performed on all subjects using sandwich ELISA assays for antibody detection, as previously described.^63,64 The third study was obtained from a publicly available Infinium HumanMethylation27 blood methylation data set (GEO: GSE19711) and included 266 postmenopausal women diagnosed with primary epithelial ovarian cancer (131 cases collected pre-treatment and 135 post-treatment) and 274 postmenopausal cancer-free controls collected as part of the United Kingdom Ovarian Cancer Population Study (UKOPS).³² Of the ovarian cancer cases, only the 131 pre-treatment cases were used in our analyses; all cases with samples collected post-treatment were excluded.

All cases and controls enrolled in the study provided written informed consent as approved by the Institutional Review Boards of the participating institutions.

DNA methylation array

DNA was extracted from peripheral blood samples from the bladder cancer and HNSCC case-control studies using the DNeasy Blood and Tissue Kit (Qiagen) and sodium bisulfite-converted using the EZ DNA methylation kit (Zymo Research) according to manufacturer’s protocols. Methylation was measured using the Infinium HumanMethylation27 BeadArray (Illumina), which interrogates 27 578 CpG loci across 14 495 genes, at the University of California San Francisco Institute for Human Genomics Core Facility. Outliers were detected using array control probes supplied by Illumina to diagnose problems such as poor bisulfite conversion, batch or beadchip effect or color-specific problems. Specifically, Mahalanobis distances between arrays were determined based on fitted mean vector and variance-covariance matrix, and arrays with large distances inconsistent with multivariate normality were discarded. The methylation status for each individual CpG locus was calculated as the ratio of fluorescent signals, ranging from 0 to 1, using the average probe intensity for the methylated (M) and unmethylated (U) alleles.

Statistical analysis

Identification of putative leukocyte DMPs

Putative leukocyte-specific DMPs were determined using a series of linear mixed effects models fit to each of the 26,486 autosomal CpG loci from methylation data for previously obtained samples of sorted leukocytes⁶ (GEO: GSE39981), with a random intercept for beadchip to adjust for possible plate effect. Briefly, sorted non-pathological human peripheral blood leukocytes (n = 47) were previously obtained from different anonymous individuals through AllCells, LLC. Leukocytes were isolated by magnetic activated cell sorting (Miltenyi Biotec Inc.) and purity was confirmed by fluorescence activated cell sorting. The major cell types obtained included NK cells (n = 9), B cells (n = 5), CD4+ T cells (n = 8), CD8+ T cells (n = 2), “other” pan-T cells (n = 6), monocytes (n = 5), and granulocytes (n = 12). This yielded an F-statistic for differences across leukocyte types for each CpG locus.

Differential methylation

Differential methylation between cases and controls was examined among all autosomal loci (26 486 CpG sites) and for CpGs that did not significantly vary between leukocyte cell types (leukocyte-specific DMPs). To maximize the likelihood that all leukocyte-specific DMPs were excluded from the latter analysis, all CpG loci with an unadjusted P ≤ 0.05 were considered as putative DMPs (n = 9849), leaving 16 637 autosomal CpG loci that did not vary between leukocyte cell types (henceforth referred to as non-DMP CpGs) for the subsequent differential methylation analyses of the bladder cancer, HNSCC, and ovarian cancer case-control studies. Differentially methylated CpGs were identified by subtracting the median methylation β-value of controls from that of cases (Δβ) for each locus. To account for multiple testing, false discovery rate (FDR) estimation and Q-values were computed using the methods of Benjamini and Hochberg.⁴² CpGs were considered to be hypermethylated in cases if median β was greater in cases than controls (Δβ > 0) and hypomethylated if lower (Δβ < 0). Differential methylation was considered to be significant where Q ≤ 0.05.

Leukocyte-adjusted epigenome-wide association studies

The subject-specific leukocyte distribution (ω_i, where i = 1,…,N_p), a d₀ × 1vector such that ${\sum }_{k=1}^{d0}\omega ik=1$, was estimated using constrained projections based on peripheral blood methylation profiles, as previously described by Houseman et al.⁶ Here N_p, p ∈ {1, 2, 3}, indicates the number of samples in the three cancer case-control data sets considered, and d₀ indicates the number of unique cell-types (e.g., CD4+ T-lymphocytes, CD8+ T-lymphocytes, B cells, etc.). Several regression parameters were estimated: β_j, representing the association of case-status and methylation β value at CpG j, unadjusted for ω; α_j, representing the corresponding association adjusted for ω; and Г, the association of case-status and ω. Г, β and α were each adjusted for age, sex and smoking for bladder cancer and HNSCC, and age for ovarian cancer. Statistical inference was achieved by permutation, where null distributions were obtained by permuting case-status with respect to methylation values and other covariates, using an omnibus test adjusted for multiple-comparisons, constructed by comparing the observed average F-statistic across all CpGs to the corresponding quantity obtained from the permutation distribution. Volcano plots were generated for each adjusted and unadjusted EWAS by plotting the –log10 of the P value (y-axis) from the regression model against the case-coefficient from the respective case-coefficient. Volcano plots for the leukocyte-adjusted (represented by black diamonds) and -unadjusted (represented by blue open circles) models were overlaid for each respective case-control study to contrast the results.

To explore the biological relevance of our findings, we performed a gene set enrichment-based analysis to examine overrepresentation of any of 258 putative transcription factor binding site motifs (TFBS) located within 1kb of cancer-associated methylation of loci in blood CpG loci. TFBS data was obtained from the tfbsConsSites track of the UCSC Genomes Browser site (TFBS Z-score > 2). Overrepresented TFBS were determined using a Kolmogorov-Smirnov test, stratified by CpG island status of the target CpG. Stratification was implemented using a weighted chi-square statistic, with weights based on the number of CpGs within each of the two CpG island strata, individual stratum-specific chi-squares obtained by inverting the individual Kolomogorov-Smirnov P-value, and null distribution determined by simulation (10⁷ samples). Significance was considered where the summary P ≤ 0.05.

All statistical analyses were conducted using the R statistical package (v. 2.11.1).

10.4161/epi.28575