Functional Identification of Cancer-Specific Methylation of CDO1, HOXA9, and TAC1 for the Diagnosis of Lung Cancer

Non–small cell lung cancer (NSCLC) is the leading cause of cancer-related mortality worldwide (1, 2). Although improvements in the treatment of advanced stage lung malignancies have been made, including agents targeting specific genetic aberrations, epigenetic therapies, and exploiting the potential of the immune system to assert control over tumor growth, lung cancer remains the main cause of cancer-related deaths (3–5). Cancer-specific molecular changes have utility not only as targets for therapy, but also as biomarkers for the determination of risk of recurrence for early-stage lung cancer. Such prognostic capability may be due to the biologic significance of the alteration or because detection of molecular alterations in lymph nodes may herald a higher stage of disease than is detectable by current pathology standards (6, 7).

There is also much interest in early detection strategies to improve outcomes in lung cancer, which have culminated in the landmark National Lung Screening Trial (NLST). Although the 20% relative reduction in lung cancer mortality in the NLST low-dose computed tomography (CT) screening arm is encouraging, it belies a false positive rate among screening results of 96.4%, which has resulted in some pause among clinicians and payers alike for immediate widespread adoption of the technique (8). Improved techniques or ancillary testing methods to augment both the sensitivity and specificity of screening for lung cancer could augment CT screening.

The most promising nonradiologic ancillary tests involve the detection of cancer-specific events in tissues or fluids carrying tumor cells or tumor DNA, such as lymph node samples, sputum, or plasma. Because cancer-specific DNA methylation events are common and occur early in lung cancer progression, recent studies have used nested methylation-specific PCR (MSP) for detection of promoter methylation in sputum (9, 10). For example, using PAX5α, GATA5, and SULF2 genes derived from studies of genes with known biologic importance in NSCLC demonstrated the ability to predict the outcome of a diagnosis of lung cancer in two high-risk cohorts (11–14). Although these studies demonstrate the feasibility of molecular detection of altered, cancer-specific DNA methylation in sputum, there remains a need for improvement in the panel of markers used. The measure of success expected from a test lies in the frequency of the event (sensitivity) and the absence of the event in normal samples (specificity). In this work, we seek to build upon approaches that define the most highly sensitive and specific markers of cancer, which have often been found to be linked to polycomb-associated sites in embryonic stem cells, toward the deployment of a clinically useful assay (15–17). We hypothesized that the current genes used in strategies to assess presence or absence of lung cancer based on sputum and other bodily tissues and fluids may be augmented by a method combing preclinical and population-based studies to identify the most highly sensitive and specific methylation events in lung cancer.

Here, we report the discovery and characterization of genomic changes in DNA methylation occurring in association with a described biologic program, moving from the study of individual loci to a comprehensive analysis of alterations in NSCLC with the intention of uncovering epigenetic events which may predict a cancer's natural history or be utilized for the molecular detection of disease. This study provides a method for systematic discovery of epigenetic biomarkers which may be used for improving the screening and diagnosis of this deadly disease.

All NSCLC cell lines were purchased from the American Type Culture Collection. H838, H23, H1993, H1568, H2170, and H520 were cultured in RPMI-1640 medium (Mediatech, Inc.); H1869 was cultured in DMEM/F-12 Medium and SK-MES-1 was cultured in Dulbecco's Modified Eagle Medium (DMEM; Mediatech, Inc.). Cell lines H838, H23, H1993, and H1568 were derived from adenocarcinomas and H2170, H520, H1869, and SK-MES-1 were derived from squamous cell carcinomas. Cell lines of squamous carcinoma and adenocarcinoma histology are represented equally so that cancer-specific, rather than histology-specific markers may be elicited by the experimental method. All cell culture media were supplemented with 10% bovine calf serum (BCS) and incubated in humidified air and 5% CO₂ at 37°C. For drug treatments, log phase cells were cultured in growth media containing 10% BCS and 1× penicillin/streptomycin with 5 μmol/L decitabine (Sigma; stock solution: 1 mmol/L in PBS) for 96 hours, replacing fresh media and decitabine every 24 hours. Cell treatment with 300 nmol/L Trichostatin A (TSA; Sigma; stock solution: 1.5 mmol/L dissolved in ethanol) was performed for 18 hours. Control cells underwent mock treatment in parallel with addition of equal volumes of PBS or ethanol without drugs.

RNA was harvested from cells in log phase growth using TRIzol (Invitrogen) and the RNeasy kit with DNase digestion (Qiagen) according to the manufacturer's instructions. RNA was quantified using the NanoDrop ND-100 followed by quality assessment with the 2100 Bioanalyzer (Agilent Technologies). RNA concentrations for each sample was more than 200 ng/μL, with 28S/18S ratios more than 2.2 and RNA integrity scores of 10 (10 scored as the highest). Sample amplification and labeling procedures were carried out using the Low RNA Input Fluorescent Linear Amplification Kit (Agilent Technologies). The labeled cRNA was purified using the RNeasy Mini Kit (Qiagen) and quantified. RNA spike-in controls (Agilent Technologies) were added to RNA samples before amplification. Samples (0.75 μg) labeled with Cy3 or Cy5 were mixed with control targets (Agilent Technologies), assembled on Oligo Microarray, hybridized, and processed according to the Agilent microarray protocol. Scanning was performed with the Agilent G2505B microarray scanner using settings recommended by Agilent Technologies. Microarray data are available in the ArrayExpress database under accession number E-MTAB-1939.

Quality checks for all arrays included visual inspection for artifacts and the distribution of signal and background intensity for red and green channels. All arrays passed quality checks and were used. The statistical platform R and packages from Bioconductor were used for all computation (18, 19). The log ratio of red signal to green signal was calculated after background subtraction and LoEss normalization as implemented in the limma package from Bioconductor (20). Individual arrays were scaled to have the same interquartile range (75th percentile–25th percentile).

RNA was isolated with TRIzol Reagent (Invitrogen) according to the manufacturer's instructions. For real-time PCR (RT-PCR), 1 μg of total RNA was reverse transcribed using SuperScript First-Strand Synthesis System for RT-PCR (Invitrogen). For MSP analysis, DNA was extracted following a standard phenol-chloroform extraction method. Bisulfite modification of genomic DNA was carried out using the EZ DNA methylation Kit (Zymo Research). Primer sequences specific to unmethylated and methylated promoter sequences were designed using MSPPrimer (21). MSP was performed as previously described (22). Ten microliters of all PCR products were loaded directly onto 2% agarose gels containing GelStar Nucleic Acid Gel Stain (Cambrex Corp.) and visualized under UV illumination. Primer sequences and conditions for MSP are available upon request.

Fifty-nine primary lung cancers were obtained from Johns Hopkins Hospital in Baltimore, MD (Cohort A) and 30 from Shinshu University Hospital in Matsumoto, Japan (Cohort B). All tissues were immediately frozen at −80°C after surgical resection. Normal lung cDNA was purchased from DNA Technologies Inc. Six normal lung tissues were obtained from individuals without cancer (five from autopsy and one from lung peripheral to a benign bronchial tumor). Tissue acquisition was conducted under approved guidelines of the Institutional Review Boards from both institutions. Histologic examination was based on World Health Organization classification criteria (23). Clinical staging was done according to Mountain and Dreslers' tumor-node-metastasis classification criteria (24).

We used the DNA methylation data of 409 lung adenocarcinoma samples with 32 matched normal samples as well as 227 lung squamous cell carcinoma samples with 43 matched normal samples from the Cancer Genome Atlas (TCGA) project (25, 26). DNA methylation was measured on the Illumina HumanMethlation 450 K platform (18, 27).

The analysis of DNA methylation data was performed using R/Bioconductor software with the limma package and custom routines for data analysis (18, 19, 28). We selected only those probes for sites situated within CpG-island promoters of genes unmethylated at their promoter sites in all normal TCGA samples (β- value < 0.2). For each probe we estimated a t statistic and P value by fitting a linear model of its differential methylation between tumor and normal samples (29). All probes tested had adjusted P values less than 1 × 10⁻⁴. Figure 1 shows a heatmap of DNA methylation level for each site (in rows) for all tumor and normal samples (in columns). The columns of the heatmap were ordered by unsupervised clustering, whereas rows were ordered top-to-bottom by decreasing value of significance for t statistic for differential methylation. The sites and corresponding statistics for all probes can be found in Supplementary Table S1.

DNA methylation clusters were based on the most variable CpG sites from Fig. 1 and on stage I and II samples. Consensus clustering was applied as implemented in the Bioconductor package ConsensusClusterPlus, with Euclidean distance and partitioning around medoids (pam) was used to derive clusters (30, 31).

P value was computed from the Cox regression (the coxph function of the survival package; refs. 32, 33). Kaplan–Meier curves were made with the help of the survfit function from the same package using TCGA data for stage I and II tumors. The clinical endpoint for analysis was time to death. TCGA samples are not annotated for therapies received; therefore, no control for treatment in analysis is possible but may be assumed to represent the standard of care in the United States. Methylation data were obtained by TCGA from fresh-frozen tumors examined by Infinium HumanMethylation450 as previously described (25). Categorization for groups of comparison for survival outcomes is based on medoid clustering as described in Clustering Analysis.

We selected the most significant CpG site per gene to define binary DNA methylation. For each gene, a sample was labeled DNA hypermethylated if the individual β-value of the gene was more than three times the SD of the mean of all combined β-values of normal samples.

On the basis of a previously designed method to unmask epigenetically silenced cancer-specific, DNA-hypermethylated genes, we treated eight NSCLC cell lines with either the DNA-methylation and DNMT inhibitor, decitabine, or the histone deacetylase (HDAC) class I/II histone deacetylase inhibitor, TSA (34, 35). Gene expression changes determined using Affymetrix microarray for decitabine- or TSA-treated cells were compared with mock-treated cells. This method enables the identification of genes induced specifically by decitabine, an important distinction as decitabine has the capacity to induce gene reexpression of loci silenced predominantly by hypermethylation, whereas TSA alone will fail to induce reexpression (34). The objective of methylation biomarker discovery by decitabine-specific reexpression is to generate a list of genes likely to be silenced by methylation of promoter CpG islands. Decitabine-specific reexpression for a gene is defined as a more than 2.0-fold reexpression on a microarray with decitabine treatment compared with mock-treated cells, less than 1.4-fold reexpression with TSA treatment compared with mock-treated cells, and no basal expression in mock-treated cells as previously described (34, 35). To find genes which would be expected to have higher frequencies of methylation in lung cancer, we refined this list to require the preceding criteria in at least two of eight cell lines. A total of 305 genes were determined to be upregulated by decitabine using these criteria from eight NSCLC cell lines (Supplementary Fig. S1).

The comprehensive analysis of 305 genes in primary tumors to determine their utility would represent a challenging task without additional informatics filters to select the most promising candidates. To refine this list of genes, we applied this functionally derived gene list to primary tumors characterized in the TCGA lung cancer project, and then validated the findings in two, independent single-institution cohorts of primary NSCLC tumors (Table 1). We first tested for tumor specificity among the TCGA tumors, comparing DNA methylation between lung tumors and normal lung tissue. Of the 305 decitabine upregulated genes, 63 genes with a total of 172 annotated CpG island promoter probes on the Infinium 450 K array had a statistically significant ability to differentiate tumor versus normal in TCGA samples as estimated by a linear regression model. In addition, these genes had extremely low methylation (β-values) in TCGA normal samples, thereby defining a group of decitabine -responsive, cancer-specific methylated genes. Data using these probes are represented in a heatmap where rows are ordered from top to bottom by P values based on the ability of an individual methylation array probe to distinguish tumor versus normal. Columns are ordered by unsupervised hierarchical clustering (Fig. 1 and Supplementary Table S1). Maximum estimated P value for each probe was 1 × 10⁻⁴. CDO1, HOXA9, and TAC1 were notable for extremely high rates of DNA methylation in tumors and low methylation in normal samples, and were most effective in distinguishing tumor versus normal based on P value of linear logistic regression model.

Binary methylation values as determined by the single best methylation probe from the promoter CpG islands of CDO1, HOXA9, and TAC1, and were plotted for all NSCLC stages together as well as for stage I alone (Fig. 2 and Supplementary Fig. S2 and Supplementary Table S1). Sensitivity is not limited by histology or tumor stage in the TCGA dataset. In fact, methylation of at least one of these three genes is 98.9% sensitive for tumors stage I–IV and 98.7% sensitive for stage I tumors alone. HOXA9 alone is methylated in 97% of NSCLC TCGA samples. There are limited descriptions of DNA methylation of these genes in human lung cancer in previous studies. Although TAC1 promoter methylation has not been described in lung malignancies, highly prevalent HOX cluster gene methylation, including HOXA9, has been reported in cell lines and a small number of squamous stage I tumors (n = 4) as well as a pool of mixed stage and mixed histology tumors (n = 20; refs. 17, 36). HOXA9 hypermethylation has been described as a potential screening test in combination with SOX1 hypermethylation and DDR1 hypomethylation as assayed by pyrosequencing (37). CDO1 has been reported as a methylated gene in squamous lung tumors (n = 30; ref. 38). CDO1 and TAC1 have been described as high-prevalence cancer-specific methylated genes in breast cancer (35). However, no previous study has described the sensitivity and specificity for a combination of these genes in a large population of NSCLC tumors and validation cohorts.

In addition to their diagnostic utility, we examined the potential prognostic significance of this functionally derived list of cancer-specific methylation. As would be expected from a list of genes with an extremely high prevalence of methylation and no described biologic role in lung cancer, none of the 63 genes examined individually was associated with survival outcome in TCGA (data not shown). To examine whether methylation of these genes taken as a group reflects biologic differences in tumors, we clustered all TCGA lung cancer samples using medoid clustering, a method for defining optimal numbers of groups within a dataset. When taken together, the 63 cancer-specific hypermethylated genes form three groups, adenocarcinoma-predominant, squamous-predominant, and a mixed group. These clusters demonstrate a marginal association with survival in the TCGA tumors (P = 0.04; Supplementary Fig. S3). From our previously published markers of outcome in early-stage, resected lung cancer, our strongest associations with outcome came from questions pertaining to cancer-specific methylation confirmed in lymph nodes, thus a diagnostic or staging paradigm. As the TCGA contains only samples of primary tumors and no associated lymph nodes, there is no ability to assess concordance of methylation between tumor and lymph node. When examining tumor-only questions from our previous work, we find general agreement with the moderate prognostic capacity of methylation of four genes when examined in tumor only, highlighting the need to refine a highly sensitive and specific diagnostic markers for the molecular staging of NSCLC (ref. 6; Supplementary Fig. S4).

Previous studies have suggested that genes with polycomb marks in chromatin surrounding the transcription start sites are predisposed to aberrant DNA methylation silencing in cancer (15, 39, 40). In embryonic stem cells, polycomb association occurs in the context of bivalent chromatin marks containing both active histone 3 lysine 4 trimethylation (H3K4me3) and repressive histone 3 lysine 27 trimethylation (H3K27me3) marks. Of the 63 cancer-specific hypermethylated genes, 45 (71.4%) are considered bivalent genes silenced by polycomb-repressive complex in progenitor cell states, a rate much higher than the presence of these marks among all genes (21% using estimated 4,413 bivalent genes among an estimated 21,000 total human genes, P < 0.0001; refs. 15, 38). CDO1, HOXA9, and TAC1, are all polycomb associated in embryonic stem cells (Fig. 1 and Supplementary Table S1).

To confirm the high prevalence of DNA methylation for these genes in other primary lung tumors, we then validated the sensitivity of these three genes in two independent cohorts of NSCLC tumor samples using MSP (Table 1; Fig. 3). Primers for CDO1, HOXA9, and TAC1 were designed and tested on tumor samples from cohorts in the United States and Japan. As was observed for these genes on the Infinium platform within TCGA data, there was no methylation in seven normal lung samples when examined using MSP. In contrast with normal lung, among the American cohort A and Japanese cohort B, respectively, 94.9% and 83.3% of the tumor samples were methylated for at least one of these three genes. Because this three-gene panel has near-zero methylation β-values by Infinium and MSP in normal tissues and is found to have stage-independent hypermethylation in cancer, these genes fulfill critical characteristics for designing a threshold for methylation in clinical assays and for identifying the earliest stages of NSCLC (Fig. 3).

Using an experimental model to derive a list of candidate cancer-specific, hypermethylated, polycomb-associated genes in lung cancer, we validated a three-gene test in a large publicly available database and two independent cohorts to describe a highly sensitive, highly specific diagnostic test for NSCLC. In the present study, we use a functional approach to identify three genes, CDO1, HOXA9, and, TAC1, in which we describe cancer-specific DNA methylation without regard for the biologic implication of that cancer-specific methylation. When examining diagnostic sensitivity, we find a remarkable concordance between TCGA samples, derived entirely from American hospitals, and our American validation cohort with sensitivities of 98.9% and 94.9%, respectively. Diagnostic sensitivity in the Japanese cohort is similar but lower at 83.3%. Although some variation may be due to sampling, we can also reasonably hypothesize that this reflects other established differences in the NSCLC populations of American and Japan and highlights the need to tailor a test precisely to target populations. Although an 83% sensitivity of detection far exceeds any mutational detection approach currently available, it may be possible to provide an even better three-gene test if these genes were chosen from among highly methylated genes determined from analysis of lung cancers in Japanese populations.

In addition, we have explored whether these cancer-specific alterations may have prognostic value. As might be expected, these genes without an established role in the pathogenesis of lung cancer and/or an extremely high prevalence of methylation prove to be of no prognostic value when examined individually. Indeed, in our previously published study of four genes, there was limited prognostic value when knowledge of methylation status is known for the tumor only. In addition, our previous study suggested that the presence of cancer-specific methylation in histologically negative lymph nodes, particularly mediastinal (N2) nodes, was most prognostic of recurrence and lung cancer associated (6).

An interesting characteristic of the genes elicited by this functional screen for novel cancer-specific biomarkers is a high degree of overlap with polycomb-associated genes. H3K4me3 and H3K27me3 define a bivalent chromatin state that denotes a low-transcriptional, poised state for a group of genes in progenitor and stem cells highly enriched for developmental processes (41). These genes, largely active during development of differentiated tissues, are downregulated by the polycomb-repressive complex when a chromatin bivalent state exists and are largely devoid of DNA methylation. These loci are particularly vulnerable to DNA methylation during the process of carcinogenesis (15). Although the mechanism that underlies epigenetic silencing transitioning from the polycomb-repressive complex to DNA methylation would suggest little or no alteration in gene expression in some cases, assaying these methylation changes remains useful as highly sensitive and specific hallmarks of tumor tissue and are therefore excellent candidates as diagnostic biomarkers. In addition, because different stem and progenitor populations show variation in distribution of chromatin-bivalency, the methylation marks at polycomb-associated DNA may signal subtle differences in the cell of origin.

For the molecular detection of disease in lymph nodes for staging and for approaches for early detection involving sputum, plasma, or fine needle aspirates, molecular alterations present in the vast majority of tumors will be the most sensitive and efficient means of detection. Through the characterization of hypermethylated loci reported here, we have developed a highly sensitive, highly specific test for identifying cases of NSCLC which may serve these purposes. A three-gene methylation assay with sensitivity in tumors approaching 100% may allow for the detection or diagnosis of disease in tissues remote from the primary tumor without specific knowledge of methylation of those genes in the tumor itself. The present study demonstrates the performance of a three-gene test in primary tumor samples for which inadequate diagnostic methods currently exist. With improvements in detection of DNA methylation in blood and sputum, the sensitivity of detection in additional types of biospecimens, including plasma and sputum samples, can now be tested (42).

L.V. Neste is an inventor and has ownership interest in a patent for the detection and prognosis of lung cancer. S.B. Baylin has received speakers bureau honoraria from MDxHealth and is a consultant/advisory board member for Bionumberick Pharm. J.G. Herman has commercial research support from and is a consultant/advisory board member for MDxHealth. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J. Wrangle, E.O. Machida, A. Hulbert, L.V. Neste, J.G. Herman

Development of methodology: J. Wrangle, E.O. Machida, A. Hulbert, N. Franco, L.V. Neste, K.E. Schuebel, N. Ahuja

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. Wrangle, E.O. Machida, A. Hulbert, N. Franco, W. Zhang, M. Tessema, J. Licchesi

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J. Wrangle, E.O. Machida, L. Danilova, A. Hulbert, S. Glockner, L.V. Neste, C.M. Hooker, J.G. Herman

Writing, review, and/or revision of the manuscript: J. Wrangle, E.O. Machida, L. Danilova, M. Tessema, L.V. Neste, N. Ahuja, J. Amano, S.A. Belinsky, S.B. Baylin, J.G. Herman, M.V. Brock

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): E.O. Machida, N. Franco, L.V. Neste, H. Easwaran, N. Ahuja, J. Amano

Study supervision: N. Ahuja, J.G. Herman

The authors thank NIH Lung SPORE CA058184, DOD W81XWH-12-1-0323, and MDxHealth for support and also Leslie Cope and Mary Lou Wrangle.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.