Background

Oral and pharyngeal cancer are major public health concerns in the USA, where there were an estimated 42,440 new cases of oral and pharyngeal cancer diagnoses (it is the eighth most common form of cancer in men) and 8390 deaths in 2014 [1]. This problem is even more pronounced on the global scale, with 442,760 incident cases and 241,458 deaths worldwide in 2012 [2]; rates are particularly high in parts of Western Europe, Southeast Asia, and Oceania. The relatively high mortality is, in part, due to the fact that the majority of patients initially present at an advanced stage [3], which is associated with a much poorer prognosis [4]. Additionally, oral and pharyngeal cancer carries a very high morbidity, often with disfigurement and impairment of basic functions, such as talking, swallowing, eating, and breathing [3], that is exacerbated by more advanced disease and the associated disease treatments. Taken together, these considerations underscore the critical importance of early detection in reducing the adverse impact of this disease.

DNA methylation is a very common epigenetic event associated with the genesis of oral and pharyngeal carcinoma, often preceding the onset of frank malignancy [5]. DNA methylation occurs primarily in the context of CpG dinucleotides [6], which are disproportionately concentrated in enriched regions referred to as CpG islands. CpG islands are commonly situated in the 5′ promoter region of genes where their methylation is generally associated with transcriptional repression. However, methylation of CpG islands situated in inter- and intra-genic enhancer regions can also impact the timing or spatial patterns of gene expression [7]; there is mounting evidence that methylation of CpG islands located in the gene body can lead to increased transcriptional activation [8, 9]. Furthermore, regional methylation can impact the expression of non-coding RNA [7], the sequences of which are commonly situated in intronic or intergenic regions. Methylation of CpG islands can arise aberrantly during disease development and progression [6] but can also occur as part of normal biological processes, such as X-inactivation, imprinting [5], or tissue differentiation [10–14].

Currently, no proven screening techniques are in widespread use for oral and pharyngeal cancer aside from visual inspection and palpation, which are provided by dentists and clinicians on an opportunistic basis, lack sensitivity (particularly for pharyngeal tumors), and vary according to the skill of the clinician performing the exam. Oral rinse can be utilized as a non-invasive ascertainment technique for detection of DNA methylation in these cancers [15–27] and therefore has potential in biomarker-based screening applications, particularly among high-risk groups or for post-treatment surveillance. While the existing literature has primarily focused on a limited set of candidate promoter regions, epigenome-wide strategies offer a more comprehensive approach for discovery. Systematic evaluation of methylation over predefined aggregate regions, such as CpG islands, can help to mitigate issues relating to false discovery rate and technical noise that can complicate epigenome-wide assessment of large numbers of individual loci [28]. Hence, the goal of this study was to begin to address these needs through epigenome-wide interrogation via the Infinium HumanMethylation450 BeadArray for identification and validation of a novel sentinel CpG island methylation profile in non-invasive oral rinse samples that may be useful in predicting oral and pharyngeal carcinoma.

Results

The study population included 154 cases with incident initial primary oral or pharyngeal squamous cell carcinoma from the greater Boston area and 72 cancer-free controls, from whom oral rinse samples were obtained. DNA methylation was interrogated in the oral rinse samples using the Infinium HumanMethylation450 BeadArray (Illumina, San Diego, CA), which contains probes for more than 450,000 CpG loci across 99 % of annotated human genes. A general schematic of our analytic workflow is presented in Fig. 1. Study participants were randomly partitioned into either a training or testing set at a 2:1 ratio, which resulted in 157 subjects in the training set (101 cases, 46 controls) and 76 in the testing set (51 cases, 25 controls); 3 of the original 226 samples (2 cases, 1 control) failed initial quality control measures and were excluded from the analyses. A description of the study population for the training and testing sets by case-control status is presented in Table 1.

Fig. 1
figure 1

Schematic of the SS-RPMM algorithm for identification of a CpG island methylation profile predictive of oral and pharyngeal carcinoma case status

Full size image
Table 1 Characteristics of study subjects in the training and testing sets according to case-control status for oral and pharyngeal carcinoma
Full size table

There were a total of 32,465 autosomal CpG islands represented by at least three CpG probes on the HumanMethylation450 BeadArray. After fitting a series of individual linear mixed-effects models for average methylation across each of these CpG islands and ranking them according to absolute t-statistic using only the training set data, we used a semi-supervised recursively partitioned mixture modeling (SS-RPMM) algorithm [29], through which we determined that a methylation classifier based on 22 CpG islands formed the optimal number of top CpG loci for discriminating between cases and controls, resulting in seven distinct methylation classes. A description of each of these 22 CpG islands, along with their bioinformatic attributes, is presented in Table 2.

Table 2 Description of the 22 CpG islands used to establish the methylation classifier in oral rinse samples
Full size table

We then validated this class structure and tested its predictive power by applying the latent structure of the methylation classes established in the training set to the testing set. A heatmap of the methylation profiles of each of the 22 CpG islands for the subjects assigned to the testing set by methylation class is presented in Fig. 2a. Two methylation classes, rRL and rRRR (denoted according to left and right branches on the dendogram from the clustering procedure), particularly stand out as being “case-heavy” (Fig. 2b), which collectively include 33 cases and only one control. When considered together, the association of these two classes with case status relative to all other classes is remarkably strong, with a crude odds ratio (OR) = 43.8 (95 % CI 8.1, 816.7) and adjusted OR = 76.9 (95 % CI 11.8, 1818.2). The sensitivity and specificity of classes rRL and rRRR for correctly predicting case status is 64.7 and 96.0 %, respectively.

Fig. 2
figure 2

a Heatmap showing average methylation of the 22 CpG islands comprising the classifier for each of the 76 testing set subjects, clustered according to methylation class. The numbers in the x-axis correspond to the rank-ordered M number displayed for each respective CpG island in Table 2. b The distribution of cases (n = 51) and controls (n = 25) from the testing set across the seven methylation classes

Full size image

To further assess the performance of the classifier in the testing set, we constructed receiver operating characteristic (ROC) curves for the logistic regression models and calculated the corresponding area under the curve (AUC) (Fig. 3). When considering the association between case status and methylation class alone (i.e., no other independent covariates in the model), the AUC was 0.84 (95 % CI 0.75–0.93). After additional adjustment for age, sex, smoking pack-years, and alcohol consumption, the AUC increased to 0.92 (95 % CI 0.86, 0.98).

Fig. 3
figure 3

Receiver operator characteristic (ROC) curves with corresponding area under the curve (AUC) and 95 % confidence intervals (95 % CI) for logit regression for the association of oral and pharyngeal carcinoma case status in the testing set with a methylation class only; b age, sex, smoking pack-years, and alcohol consumption only; and c methylation class adjusted for age, sex, smoking pack-years, and alcohol consumption

Full size image

In an effort to explore possible drivers of these findings, we assessed the sociodemographic and clinical characteristics of oral and pharyngeal carcinoma cases for each “case” class (rRL, n = 20; rRRR, n = 13) relative to cases in all other classes (n = 18) (Table 3). No significant differences were observed across classes. However, while non-significant, class rRRR had a higher fraction of cases with seropositivity for HPV16 e6/e7 antibodies (p = 0.11), which were performed on study subjects as a biomarker of HPV16-transformed invasive tumors [30]. Additionally, the case classes contained a somewhat higher, albeit non-significant, proportion of advanced stage cases, although it is notable that both of these cases also contain a sizable fraction of localized cancers and that the majority of cases have relatively smaller (T1–T2) tumors.

Table 3 Characteristics of testing set cases for each of the two case-associated methylation classes (rRL and rRRR) relative to those in the other five classes
Full size table

To evaluate the biological significance of the 22 CpG islands that form the oral rinse methylation classifier, we downloaded Infinium HumanMethylation450 BeadArray data from The Cancer Genome Atlas (TCGA; http://cancergenome.nih.gov/) for all oral and pharyngeal carcinoma cases with paired adjacent normal tissue (34 pairs). With respect to the CpG islands forming the classifier, 20 of 22 were found to be significantly differentially methylated in tumor tissue relative to the adjacent normal tissue, based on the Wilcoxon signed-rank test and adjusting for false discovery rate (FDR) using the methods of Benjamini and Hochberg [31]. The results of this analysis are presented in the right-hand columns of Table 2. Of note, the two CpG islands that were not significantly differentially methylated exhibited broad variability from the minimum to maximum differentials, which could potentially add information on a subset of samples to the classifier, particularly when considered in conjunction with other CpG islands.

Discussion

We have identified a CpG island methylation classifier that can be used with oral rinse samples for predicting incident oral and pharyngeal carcinoma with a high degree of accuracy. Several other studies have examined the potential utility of DNA methylation in oral rinse samples for predicting head and neck cancers [15–25], but our findings represent the strongest predictive panel reported to date that was validated in an independent study set, with an impressive adjusted AUC of 0.92 (several studies reporting high sensitivity and/or specificity established methylation cut-points using the same set of samples that predictive accuracy was tested, leaving them susceptible to issues from over-fitting). By applying a rigorous two-stage analysis of Infinium HumanMethylation450 BeadArray data with an agnostic genome-wide assessment that encompassed all annotated CpG islands, including those outside of the gene promoter context, our study provides contrast to the majority of existing studies, which with few exceptions employ a candidate-gene approach centered on promoter methylation. As such, this study has given rise to novel targets, the majority of which, to our knowledge, have not been previously reported.

Adding further strength to our findings, 20 of the 22 CpG islands were observed to be differentially methylated in tumor tissue relative to adjacent normal tissue. One of the two CpG islands that was not differentially methylated in the TCGA tumors overlaps the promoter region of DKK1, which has been reported to be hypermethylated in head and neck carcinoma [32, 33], and lower expression has been associated with increased risk of lymph node metastasis and poorer outcome [34, 35], although other studies report conflicting findings [36, 37]. The other CpG island that was not differentially methylated in the TCGA tumors is associated with the zinc-finger protein ZCCHC14, which is an intriguing locus, as SNPs in that gene have been associated with nicotine dependence [38]. The CpG island is located in the 3′UTR of ZCCHC14 and overlaps putative microRNA-binding sites for miR-542-3p and miR-615-3p [39].

The strengths of this study include the relatively large number of cases and controls with oral rinse samples compared to the majority of studies in the current body of literature; the use of average CpG island methylation in conjunction with the broad coverage from Infinium HumanMethylation450 BeadArray helps to mitigate technical noise that is often an impediment with single locus analysis; and the inclusion of oropharyngeal and hypopharyngeal cases along with oral cavity cases broadens the potential applicability of this panel. Additionally, our agnostic, data-driven approach with the inclusion of all CpG islands rather than select, candidate promoter methylation can also be viewed as a strength. This does not diminish the importance of the candidate genes used in prior studies, but we have now added novel CpG island loci for investigation in future studies. One potential weakness of this study is its retrospective design, although, conversely, the case-control design has the advantage of providing us with a much larger number of oral and pharyngeal carcinoma cases than we would be able to obtain using a population-based prospective study. Future studies will be aimed at prospective validation the 22 CpG island methylation panel as a screening tool in a high-risk population and as a potential tool for use in post-treatment follow-up surveillance for head and neck cancer patients. Additionally, due to coverage limitations of the Infinium HumanMethylation450 BeadArray, it is plausible that we may have missed some CpG islands that could potentially play an important epigenetic role in oral and pharyngeal carcinoma by restricting our analysis to CpG islands containing at least three CpG probes on the array. However, this platform offers excellent coverage of CpG-dense regions, allowing us to analyze >32,000 distinct CpG islands (as defined by Hidden Markov Model), and remains among the best available options for epigenome-wide analysis of a large number of samples [40–42].

Conclusions

Although further expanded testing is warranted in a prospective setting, this panel may have utility for early detection of disease, particularly in targeted, high-risk populations. Importantly, methylation panels used in conjunction with non-invasive oral rinse samples, such as that described herein, may ultimately prove valuable as an aid for post-treatment follow-up surveillance; again, further prospective testing of this methylation classifier is warranted to determine its applicability to such applications. Continued discovery and development of clinically relevant biomarkers that can help with early detection of incident and/or recurrent head and neck cancer will ultimately have a positive impact on public health by reducing morbidity and mortality associated with this devastating disease, both in the USA and worldwide.

Methods

Study population

The study population included 154 cases with incident initial primary squamous cell carcinoma arising in the oral cavity (ICD-9: 141.1–141.5, 141.8, 141.9, 143–145.2, 145.5–145.9, 149.8, 149.9), oropharynx (ICD-9: 141.0, 141.6, 145.3, 145.4, 146, 149.0, 149.1), or hypopharynx (ICD-9: 148) diagnosed between October 2006 and June 2011 at major teaching hospitals located in Boston, MA (Brigham and Women’s Hospital, Beth Israel Deaconess Medical Center, Boston Medical Center, Dana-Farber Cancer Institute, Massachusetts Eye and Ear Infirmary, Massachusetts General Hospital, and New England Medical Center) as part of a population-based study of head and neck cancer in the greater Boston area (Collaborative Study of Head and Neck Diseases (CoHANDS)) that has been previously described [43, 44]. For inclusion in the study, cases were required to reside in the greater Boston area or any of 162 contiguous cities and towns within an approximately 1-h drive from Boston at the time of diagnosis. Cases with a prior history of malignancy other than non-melanoma skin cancer were excluded from the analyses. Cancer-free control subjects (n = 72) were randomly selected from 567 controls that were recruited into CoHANDS using a population-based design [45] during the same time frame as the cases. All patients included in the analyses provided written informed consent prior to enrollment in the study, as approved by the institutional review boards of Brown University and the participating institutions listed above.

Sample collection, DNA extraction, and bisulfite modification

Upon enrollment into CoHANDS (and prior to initiation of treatment for cases), subjects were asked to vigorously swish with approximately 30 ml of commercial alcohol-free mouthwash (Act™) for 30 s. Samples were then centrifuged into cell pellets and stored at −80 °C in cryovials until DNA extraction. DNA was extracted using the QIAamp Blood Kit (Qiagen, Valencia, CA) using the spin protocol for DNA purification from blood or body fluids. Extracted DNA was bisulfite modified using the EZ-96 DNA Methylation-Direct Kit (Zymo Research, Irvine, CA) according to Illumina’s recommendations for the Infinium HumanMethylation450 BeadArray.

Infinium HumanMethylation450 BeadArray

The Infinium HumanMethylation450 BeadArray assay was performed in three batches at the University of California San Francisco (UCSF) Institute for Human Genomics Core Facility (first and second batches) and University of Cincinnati (UC) Genomics, Epigenomics and Sequencing Core (third batch). Approximately 500 ng of bisulfite-modified genomic DNA was provided to the respective facility for initial processing of the BeadArrays, with samples randomized to BeadChip positions to mitigate any impact of potential batch or chip effects. Raw image files were preprocessed using the RnBeads pipeline in R [46]. All array data points are represented by fluorescent signals from both methylated (Cy5) and unmethylated (Cy3) alleles, and average methylation level (β) is derived from the ~18 replicate methylation measurements, β = (max(Cy5, 0))/(|Cy3| + |Cy5| + 100). Beta (β) = 1 indicates complete methylation; β = 0 represents no methylation. Outliers were assessed using quality control plots generated through the RnBeads pipeline designed to diagnose problems such as poor bisulfite conversion or signal intensity issues. Functional normalization was performed using minfi.funnorm [47] following background correction with the normal-exponential using out-of-band probes (NOOB) method [48]. Any probes with a detection p value >0.01 or that contained a single-nucleotide polymorphism (SNP) in the probe sequence were filtered out of the dataset prior to analysis. To account for any residual batch or chip effects, methylation data were adjusted using the ComBat method [49] via the SVA package in Bioconductor. The dataset supporting the results of this article are available the Gene Expression Omnibus (GSE7097: http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE70977).

Statistical analysis

Average methylation was calculated for each autosomal CpG island (determined by the Hidden Markov Model approach, which provides a more accurate approach for identification of CpG islands than expected CG content-based filtering strategies [50]) that spanned at least three CpG loci on the HumanMethylation450 BeadArray (after filtering). It was our intent that assessment of CpG island methylation (as opposed to the individual locus approach) would provide more stable estimates that are less readily influenced by outliers due to technical variation.

A semi-supervised recursively partitioned mixture modeling (SS-RPMM) algorithm [29] was applied to identify a novel set of CpG islands for which methylation in oral rinse samples were predictive of oral and pharyngeal carcinoma case status. This method is based both on the semi-supervised procedure proposed by Bair and Tibshirani [51, 52] and recursively partitioned mixture models (RPMM) developed by Houseman et al. [53]. To avoid over-fitting the data and provide for validation of the model, subjects were randomly partitioned into either a training set (for the initial analysis) or a testing set (for subsequent validation) at a 2:1 ratio (frontloaded to increase the precision of the classifier identified in the training set), stratified by case-control status to ensure an equal distribution between sets. A series of linear mixed-effects models were then fit to logit-transformed average methylation (M) values to identify CpG islands most associated with case status, and were adjusted for age, sex, smoking pack-years, and alcohol consumption (typical number of alcoholic beverages per week) with a random-effect term for batch/processing site (UCSF or UC). CpG islands were ranked based on the absolute value of the t-statistic for case status. The top M loci were selected using a nested cross-validation procedure to train a classifier for case/control status by fitting a RPMM to the training data to cluster subjects using the M selected loci. To predict class membership in the testing set, the latent class structure from the RPMM fit to the training data was applied using an empirical Bayes procedure. Unconditional logistic regression was used to calculate the magnitude of the association between methylation class and oral and pharyngeal carcinoma, controlling for potential confounding covariates (age, sex, smoking pack-years, and alcohol consumption). Receiver operating characteristic (ROC) curves and corresponding area under the curve (AUC) were generated to assess the performance of the DNA methylation classifier.