An Integrated Prognostic Classifier for Stage I Lung Adenocarcinoma based on mRNA, microRNA and DNA Methylation Biomarkers

Abstract

Introduction

Up to 30% Stage I lung cancer patients suffer recurrence within 5 years of curative surgery. We sought to improve existing protein-coding gene and microRNA expression prognostic classifiers by incorporating epigenetic biomarkers.

Methods

Genome-wide screening of DNA methylation and pyrosequencing analysis of HOXA9 promoter methylation were performed in two independently collected cohorts of Stage I lung adenocarcinoma. The prognostic value of HOXA9 promoter methylation alone and in combination with mRNA and miRNA biomarkers was assessed by Cox regression and Kaplan-Meier survival analysis in both cohorts.

Results

Promoters of genes marked by Polycomb in Embryonic Stem Cells were methylated de novo in tumors and identified patients with poor prognosis. The HOXA9 locus was methylated de novo in Stage I tumors (P < 0.0005). High HOXA9 promoter methylation was associated with worse cancer-specific survival (Hazard Ratio [HR], 2.6; P = 0.02) and recurrence-free survival (HR, 3.0; P = 0.01), and identified high-risk patients in stratified analysis of Stage IA and IB. Four protein-coding gene (XPO1, BRCA1, HIF1α, DLC1), miR-21 expression and HOXA9 promoter methylation were each independently associated with outcome (HR, 2.8; P = 0.002; HR, 2.3; P = 0.01; and HR, 2.4; P = 0.005, respectively), and, when combined, identified high-risk, therapy naïve, Stage I patients (HR, 10.2; P = 3x10⁻⁵). All associations were confirmed in two independently collected cohorts.

Conclusion

A prognostic classifier comprising three types of genomic and epigenomic data may help guide the postoperative management of Stage I lung cancer patients at high risk of recurrence.

INTRODUCTION

Lung cancer remains the leading cause of cancer-associated deaths worldwide.¹ The 5-year survival rate for all stages is below 17%, owing to the fact that most patients are diagnosed with locally advanced or metastatic disease, with few therapeutic options.² However, with the advent of Low-Dose spiral Computed Tomography (LDCT) screening, it is expected that the number of lung cancers diagnosed at an early stage will rise sharply. In the recent National Lung Screening Trial (NLST), up to 60% of the cancers diagnosed after positive LDCT screening were Stage I.³ The recommended treatment for Stage I Non-small-cell lung cancer (NSCLC) patients is surgery, which may be followed by chemotherapy in patients with pathologically high-risk, margin-negative Stage IB tumors.⁴ Still, up to 30% surgically-treated Stage I patients will die within 5 years of diagnosis.⁵ Biomarkers that molecularly categorize Stage I patients after tumor resection and identify high-risk patients who may benefit from adjuvant chemotherapy, as well as low-risk patients who could be spared, would lead to improved clinical management.⁶

Large scale analysis of the lung adenocarcinoma (ADC) genome, transcriptome and methylome has revealed integrated subtypes characterized by idiosyncratic combinations of molecular alterations that underscore the heterogeneity of this disease.⁷ As a result, any one molecular biomarker may correctly classify tumors as high-risk based on a particular underlying biology, and misclassify others driven by a different set of genomic or epigenomic changes. Thus, a multi-omic prognostic classifier derived from independent analyses of different types of molecular platforms may provide a more robust biomarker of risk. We have previously developed prognostic classifiers for Stage I lung ADC based on coding and non-coding gene expression and their combination.^8–10 Here, we interrogated the lung cancer epigenome to find prognostic DNA methylation biomarkers and subsequently evaluated their combination with biomarkers based on mRNA and microRNA (miRNA) expression.

Epigenetic abnormalities are frequent in cancers and contribute to cancer initiation, progression and response to treatment.¹¹ In NSCLC, hypermethylation at CpG-dense sequences in gene promoters (CpG islands) is associated with cigarette smoking¹², histological subtype^{13, 14}, progression^{15, 16}, clinically-relevant molecular subtypes^{17, 18}, and patient prognosis^{19, 20}. Here, from genome-wide screening of differential DNA methylation in adjacent tumor and non-tumor tissues from three cohorts, HOXA9 promoter methylation emerged as a candidate prognostic biomarker. We further evaluated HOXA9 promoter methylation by pyrosequencing in 217 primary tumors from two cohorts and its prognostic value alone and in combination with mRNA and miRNA biomarkers. Cox regression and Kaplan-Meier survival analysis were conducted in each cohort separately as well as combined. Our study follows the recommendations and best practices set forth for tumor marker prognostic studies.^{21, 22} Alone, or in combination with mRNA and miRNA biomarkers, HOXA9 promoter methylation could provide useful prognostic information on patients with Stage I lung ADC.

PATIENTS AND METHODS

Study design and patient demographics

We retrospectively analyzed tissue samples from patients with early stage lung ADC accrued by 3 studies at the Metropolitan Baltimore area of the United States (NCI cohort, n=35 Stage I/II for microarray and n=87 Stage I for pyrosequencing validation), the Haukeland University Hospital (Bergen, Norway; Norway cohort, n =17, Stage I/II), and the National Cancer Center Hospital (Tokyo, Japan; Japan cohort, n=38 for microarray and n=113 for pyrosequencing, Stage I). Patients in the NCI cohort were recruited between 1999 and 2012. Survival time for this cohort, which is a case series within the NCI-MD case control study OH98CN027, was determined by a combination of searching the National Death Index (www.cdc.gov/nchs/ndi.htm) and, subsequently, the Social Security Death Index for those still alive according to the National Death Index. Cases with cause of death other than cancer were considered censored for survival analysis. The Norway cohort was recruited between 1988 and 2003. The survival endpoint for this cohort was also cancer-specific death. The Japan cohort was recruited from National Cancer Center Hospital between 1998 and 2008. The clinical endpoint examined in this cohort was relapse. Each of the studies received approval from the corresponding institutional review board, and informed consent was obtained from all patients. These three patient cohorts have been described elsewhere.⁸

At the time of surgery, a portion of tumor specimen and non-involved adjacent lung tissue were flash-frozen and stored at −80°C to be used in molecular studies. Clinical and pathological information was obtained from medical records and pathology reports. Tumor staging was originally based on AJCC 6^th edition, and later updated to AJCC 7^th edition, except for seven cases, for which the tumor size could not be determined from the original biopsy report (Table 1).

Table 1.

Demographic and Clinical Characteristics of study cohorts.

	Microarray cohorts		Pyrosequencing cohorts
	NCI-MD (n=35)	Japan (n=38)	NCI-MD (n=87)	Norway (n=17)	Japan (n=113)
Age-years
mean (SD)	65.0 (11.6)	59.5 (7.3)	64.9 (10.6)	63.5 (10.4)	59.7 (6.4)
Range	47–88	39–76	32–88	42–82	39–76
Sex (%)
Female	19 (54.3)	21 (55.3)	42 (48.3)	7 (41.2)	60 (53.1)
Male	16 (45.7)	17 (44.7)	45 (51.7)	10 (58.8)	53 (46.9)
Race
Caucassian	26 (74.3)	0 (0.0)	63 (72.4)	17 (100.0)	0 (0.0)
African American	9 (25.7)	0 (0.0)	24 (27.6)	0 (0.0)	0 (0.0)
Asian	0 (0.0)	38 (100.0)	0 (0.0)	0 (0.0)	113 (100.0)
Smoking History (%)
Never smoker	4 (11.4)	21 (55.3)	6 (6.9)	0 (0.0)	62 (54.9)
< 20 Packyears	5 (14.3)	5 (13.1)	7 (8.0)	7 (41.2)	17 (15.0)
≥ 20 Packyears	16 (45.7)	12 (31.6)	65 (74.7)	9 (55.9)	34 (30.1)
Smoker NOSb	9 (25.7)	0 (0.0)	3 (3.4)	0 (0.0)	0 (0.0)
Unknown	1 (2.9)	0 (0.0)	6 (6.9)	1 (5.9)	0 (0.0)
Smoking Pack-years
mean (SD)	36.6 (27.3)	39.2 (27.0)	55.0 (31.3)	31.6 (27.0)	34.9 (23.5)
Tumor size (cm)
mean (SD)	2.8 (1.3)	2.3 (0.7)	2.7 (1.1)	3.5 (1.4)	2.5 (0.8)
Range	1.2 – 6.5	1.2 – 4.2	0.9 – 6.8	2.0 – 5.0	0.9 – 4.6
Unknown	1	0	4	10	0
AJCC TNM 7th (%)
IA	19 (54.3)	30 (78.9)	55 (63.2)	6 (35.3)	81 (71.7)
IB	11 (31.4)	8 (21.1)	32 (36.8)	5 (29.4)	32 (28.3)
IIA	3 (8.6)	0 (0.0)	0 (0.0)	0 (0.0)	0 (0.0)
IIB	1 (2.9)	0 (0.0)	0 (0.0)	0 (0.0)	0 (0.0)
I or IIa	1 (2.9)	0 (0.0)	0 (0.0)	6 (35.3)	0 (0.0)
Adjuvant Therapy (%)
None	26 (74.3)	38 (100.0)	68 (78.2)	17 (100.0)	113 (100.0)
Chemotherapy	5 (14.3)	0 (0.0)	8 (9.2)	0 (0.0)	0 (0.0)
Chemotherapy plus Radiotherapy	1 (2.9)	0 (0.0)	3 (3.4)	0 (0.0)	0 (0.0)
Radiotherapy	1 (2.9)	0 (0.0)	3 (3.4)	0 (0.0)	0 (0.0)
Unknown	2 (5.7)	0 (0.0)	5 (5.7)	0 (0.0)	0 (0.0)

^acannot determine, tumor size not available

^bNOS, not otherwise specified

DNA methylation array hybridization and analysis

Total genomic DNA was extracted using DNeasy Blood and Tissue Kit (QIAGEN, Valencia, CA). DNA quality and yield were determined using NanoDrop^™ Spectrophotometer (Thermo Fisher Scientific, Wilmington, DE). One microgram DNA was denatured and bisulfite converted using the EZ DNA Methylation Gold, kit (Zymo Research Corp, Irvine, CA). Bisulfite-converted DNA was hybridized to Infinium Human Methylation27 Beadchips using the Illumina HD methylation assay kit (Illumina Inc., San Diego CA). Sample pairs were randomized on Beadchips and Batch, and pairs of Tumor/Non-tumor Adjacent Tissues were always run in the same batch in order to reduce confounding between tumor phenotype and array processing. See Supplementary Materials and Methods, Supplementary Digital Content 1, for details on methylation data preprocessing.

Level 3 Methylation (individual probe/tumor β-values) and clinical data for The Cancer Genome Atlas Lung Adenocarcinoma Study (TCGA_LUAD) were downloaded from the Broad Institute distribution site (http://gdac.broadinstitute.org/) and analyzed by paired t-test. Methylation β-values and clinical/demographic data from 117 Stage I ADC patients from a previously published study²⁰ were downloaded from GEO accession GSE39279.

Functional enrichment analysis

Gene lists were uploaded to DAVID (http://david.abcc.ncifcrf.gov/) for Gene Ontology analysis.²³ Gene Set Enrichment Analysis (GSEA) implemented within Partek and χ² tests were used to determine enrichment of gene lists with respect to annotated functions or gene sets. Lists of genes differentially methylated between T and NT or differentially expressed between High and Low methylation clusters were uploaded to Ingenuity Pathway (IPA Ingenuity Systems, www.ingenuity.com) for a Core Analysis with default parameters. The output consisted of Canonical Pathways, Upstream Regulators, Diseases, and Biological Functions associated with those genes. For this analysis, differentially methylated genes were uploaded along with their associated M-value fold-changes, and differentially expressed genes along with expression fold-changes.

Analysis of DNA methylation by Pyrosequencing

Bisulfite conversion was performed on 500ng genomic DNA using EpiTect Fast DNA Bisulfite Kit (QIAGEN). Pyrosequencing was performed as per manufacturers’ instructions on a PyroMark – Q96 MD pyrosequencing instrument using the following Qiagen PyroMark CpG Assays: Hs_HOXA9_07_PM (CpG island associated with HOXA9) and Hs_HOXA9_10_PM (CpG island associated with miR-196b). Manufacturer’s recommendations were followed for PCR conditions: 15 min PCR activation step at 95°C, 45 cycles of 30 s Denaturation at 94°C, 30 s Annealing at 56°C and 30 s Extension at 72°C, followed by a 10 min Final Extension at 72°C using 10–20ng of bisulfite-converted DNA.

Profiling mRNA and miRNA expression

Total cellular RNA was extracted using miRNA Kit (QIAGEN), according to the manufacturer’s instructions. RNA quality was determined using Agilent Bioanalyzer. RNA (150ng) was labeled and amplified using the Ambion Illumina TotalPrep-96 RNA Amplification Kit (Life Technologies, Grand Island, NY) to generate biotinylated cRNA, and 750ng cRNA was hybridized to Illumina HumanRef-8 v3 Expression Beadchip arrays (Illumina Inc.). Samples were randomized over two plates as described above. Profiling of miRNA expression was performed on 100 ng total RNA using the Nanostring nCounter Human miRNA Expression Assay Kit version 1.6 (Nanostring, Seattle, WA) following processing protocol recommended by the manufacturer. Samples were randomized over several batches. See Supplementary Materials and Methods, Supplementary Digital Content 1, for details on preprocessing of mRNA and miRNA expression data. miR-21 expression values normalized on the basis of the average expression of the five most highly expressed miRs that do not include miR-21(miR-720, miR-26a, miR-126, miR-16, and miR-29a) were derived from Nanostring nCounter as in our earlier study.⁹

Measurement of 4-protein-coding gene classifier by qRT-PCR

Total RNA extracted from frozen tissues was subjected to qRT-PCR for evaluation of the prognostic 4-gene signature previously described.⁹ TaqMan assays (Life Technologies) included BRCA1 (ID Hs00173233_m1), HIF1A (ID Hs00936371_m1), DLC1 (ID Hs00183436_m1), and XPO1 (ID Hs00418963_m1). 18S (ID Hs03003631_m1) was used as a normalization control. The prognostic score was calculated as (0.104 x BRCA1) + (0.133 x HIF1A) + (−0.246 x DLC1) + (0.378 x XPO1), were BRCA1, HIF1A, DLC1 and XPO1, are the normalized linear expression values of each of the four genes in each sample.

Survival analysis

Survival time was calculated from date of surgery to date of either last known follow up or date of death due to lung cancer (NCI, Norway cohorts) or lung cancer recurrence (Japan). Proportional hazards assumptions were verified by visual inspection of log-log plots and using a non-zero slope test of the Schoenfeld residuals.²⁴ HR and 95% CI were estimated using a univariable Cox proportional hazards regression model and a model adjusted for covariates smoking (categorical), sex (categorical), and age (continuous), as well as stage of disease (IB vs IA), race (categorical), and cohort membership (categorical) when appropriate. Smoking status was dichotomized into ever/never smokers, and ≥ 20 (heavy smokers) or < 20 (moderate smokers) packyears of smoking, which was defined as the number of cigarettes smoked per day/20 multiplied by years of smoking. Patients were dichotomized on the basis of the mean percent methylation in pyrosequencing analysis (≥ 40%/< 40% mean methylation) or on the basis of β-values (≥ 0.4/< 0.4 β-value) to evaluate the association between HOXA9 methylation and survival. Statistical tests were two-sided, and were considered significant at P < 0.05, unless otherwise indicated. Kaplan–Meier plots and log-rank test were evaluated using Graphpad Prism v6.0 (Graphpad Software Inc.). Cox regression was carried out using Stata 13.1 (Stata-Corp LP). A combined score was generated that categorized patients according to the cumulative number of high values for HOXA9 promoter methylation, miR-21 and 4-gene signature, zero, one, two or three. The association between this combined score and survival was assessed for significance by Log-rank Ptrend and Cox regression.

Data Availability

The methylation, mRNA and miRNA microarray data discussed in this publication have been deposited in NCBI’s Gene Expression Omnibus²⁵ and are accessible through GEO Series accession numbers GSE63384, GSE63459 and GSE63805, respectively.

RESULTS

Characteristics of patients in the microarray and pyrosequencing cohorts

Demographic and clinical characteristics of the microarray and pyrosequencing cohorts are shown in Table 1. All tumors in the NCI microarray cohort (n=35) comprise Stage I ADC by the original American Joint Committee on Cancer (AJCC) 6^th edition staging, however, restaging to AJCC 7^th edition resulted in the re-classification of four cases to Stage II. All other cases in microarray and pyrosequencing cohorts were Stage I by AJCC 7^th edition, except for seven cases (1 from NCI microarray cohort and 6 from Norway pyrosequencing cohort) that were Stage I by AJCC 6^th but could not be evaluated by AJCC 7^th because the tumor size could not be determined. Adjuvant therapy is not recommended for Stage I NSCLC⁴, thus the majority of patients included in our analysis did not receive any. Most patients in the NCI cohort and about half of the patients in the Norway cohort are heavy smokers (defined as ≥ 20 Packyears), in contrast to 70% of patients in the Japan cohort who are moderate smokers (defined as < 20 Packyears) to non-smokers. The NCI and Norway cohorts showed similar 5-year survival rates, gender and age at diagnosis. Thus, they were combined in all statistical analyses, in order to increase power.

Polycomb-marked gene promoters are methylated de novo in early stage lung ADC

We first analyzed the overall changes in DNA methylation using Illumina Human Methylation27 Beadchips in paired tumor (T) and non-tumor (NT) adjacent tissues from the NCI microarray cohort (n=35). We identified 1,309 individual CpG sites showing differential methylation in tumors, with 964 CpG sites (associated with 772 genes and 4 miRNAs) hypermethylated and 345 CpG sites (associated with 319 genes) hypomethylated at adjusted FDR < 0.05, and fold-change > 2 (Supplementary Table S1, Supplementary Digital Content 1). As expected, hypermethylation occurred preferentially at CpG islands (Supplementary Fig. S1, Supplementary Digital Content 1).

Functional characterization of hypermethylated genes using Gene Ontology and INTERPRO identified the homeobox domain as the most significantly enriched category (FDR = 3.46 x 10⁻³²). In addition, molecular functions and biological processes involving homeobox, such as transcription factor activity, development and morphogenesis, were among the top Ontology Terms (Supplementary Table S2, Supplementary Digital Content 1). Among genes hypomethylated in tumors, categories populated by cytokines, including immune response, defense response, and cell-cell signaling, were significantly enriched, albeit to a lesser degree (FDR = 3.76 x 10⁻⁰⁵, FDR = 1.30 x 10⁻⁰⁴, FDR = 0.0032, respectively; Supplementary Table S3, Supplementary Digital Content 1). Ingenuity Pathway Analysis identified embryonic development and inflammatory disease as the top biological functions among hypermethylated (P = 1.21 x 10⁻⁴⁰) and hypomethylated (P = 5.51 x 10⁻¹³) genes, respectively. Interestingly, genes hypermethylated in tumors defined Polycomb Group (PcG) proteins EZH2 and RNF2 as activated transcriptional repressors (z-score = 2.096, P = 6.48 x 10⁻¹⁵ and z-score = 2.236, P = 4.46 x 10⁻¹⁰, respectively; Supplementary Table S4, Supplementary Digital Content 1). Genes marked by lysine 27-trimethylated histone H3 (H3K27me3) and targets of Polycomb Repressor Complex 2 (PRC2) and its subunits, SUZ12 and EED in Embryonic Stem Cells (ESC)²⁶ were over-represented among hypermethylated genes (GSEA Enrichment P = 1.00 x 10⁻⁸³ to P = 4.19 x 10⁻⁶²; Supplementary Table S5, Supplementary Digital Content 1). Because PRC2 target genes have large CpG islands that are represented by multiple probes in the microarray, we repeated this analysis by collapsing probes to genes, with similar results (χ² test P < 0.0001). GSEA was replicated using genome-wide DNA methylation data from 38 Stage I lung ADC and paired non-tumor adjacent tissues from Japanese patients (Table 1) as well as DNA methylation data on Stage I and matched non-tumor samples measured on the more comprehensive Human Methylation450 Beadchips by the TCGA_LUAD⁷. In these datasets, we confirmed an enrichment of genes marked by H3K27me3 and targets of PRC2, SUZ12 and EED in ESC among those methylated in Stage I lung ADC (χ² test P < 0.0001 and GSEA Enrichment P = 2.65 x 10⁻¹⁰⁴ to P = 1.6 x 10⁻⁷¹; Supplementary Table S6, Supplementary Digital Content 1).

ESC-like chromatin changes are associated with worse clinical outcome

Recent studies suggest that tumors acquire embryonic stem cell-like gene expression signatures ²⁶. It has been proposed that an epigenetic switch, whereby genes repressed through H3K27me3 and polycomb marking in ESCs are stably silenced in tumors through DNA hypermethylation, underscores such reprogramming of gene expression.²⁷ Expression of EZH2 Polycomb Repressor has also been associated with survival in NSCLC.^28–30 Thus, we hypothesized that the ESC-like chromatin changes that we observed in lung ADC could be associated with prognosis. Using the NCI microarray cohort, we focused on 55 probe sets (corresponding to 47 genes) that were hypermethylated in tumors and also marked by H3K27me3 in ESC (Supplementary Table S7, Supplementary Digital Content 1). Hierarchical clustering based on these 55 probe sets separated the tumors into low and high methylation groups (Fig. 1A). A higher frequency of patients with shorter cancer-specific survival was observed in the high methylation group (cluster 2, Fisher’s exact test P = 0.04), but no significant differences in the distribution of tumors according to stage (I vs II or IA vs IB), race, sex, or smoking were found (Supplementary Table S8, Supplementary Digital Content 1). Kaplan-Meier survival analysis confirmed that patients in the high methylation cluster had shorter cancer-specific survival (Log-rank test P = 0.05; Fig. 1B). This analysis was replicated in the Japanese microarray cohort, in which hierarchical clustering also resulted in classification of patients into a high methylation cluster with shorter relapse-free survival (Log-rank test P = 0.009; Fig. 1, C and D). Clusters defined by methylation have functional relevance, as genes up-regulated in high methylation samples included EZH2 (Supplementary Table S9, Supplementary Digital Content 1), were associated with a lung cancer poor survival signature³¹ (Supplementary Table S10, Supplementary Digital Content 1), and as a network reflected activation of E2F, MYC and CCND1, and inactivation of TP53 (Supplementary Table S11, Supplementary Digital Content 1). In addition, tumor suppressive miRNAs, miR-218, and miR-99a, were down-regulated, and oncogenic miR-21 was up-regulated in high methylation tumors (Supplementary Table S12, Supplementary Digital Content 1; lung cancer expression pattern from mir2disease [http://www.mir2disease.org/]). Thus, ESC-like chromatin changes are associated with more aggressive disease and poor prognosis in early stage lung ADC.

Fig. 1.

Methylation of genes marked by H3K27me3 in Embryonic Stem Cells (ESC) is associated with clinical outcome. Hierarchical clustering based on hypermethylated CpGs of genes marked by H3K27me3 in ESC in the NCI (A) and Japan (C) microarray cohorts. Each column represents an individual patient and each row an individual CpG probe. Patients in the high methylation clusters had shorter cancer-specific (B) or relapse-free (D) survival, respectively, in Kaplan-Meier survival analysis.

HOXA9 promoter methylation stratifies lung cancer outcome in two independent patient cohorts

Among the 55 probe sets identified above in the NCI cohort, 3 were associated with HOXA9 (Supplementary Table S7, Supplementary Digital Content 1). Two of them also showed the highest statistical significance for differential methylation in T vs NT analysis. Pyrosequencing analysis of DNA methylation in a subset of 20 paired T/NT tissues confirmed de novo methylation at the HOXA9 locus in Stage I lung ADC (paired t-test P = 0.0004; Supplementary Fig. S2A, Supplementary Digital Content 1). In the Illumina platform, a miRNA adjacent to HOXA9, miR-196b, is annotated to the HOXA9 probe sets. Pyrosequencing analysis showed that a smaller CpG island that overlaps miR-196b was also methylated in tumors (paired t-test P = 0.0001; Supplementary Fig. S2B, Supplementary Digital Content 1). Mean methylation was below 20% in non-tumor tissues.

HOXA9 promoter methylation was analyzed by pyrosequencing in a validation cohort comprising 104 Stage I ADC from NCI (n=87) and Norway (n=17). We observed a highly significant correlation between pyrosequencing and microarray values for individual tumors at both loci (R² = 0.7, P < 0.0001) and found that samples in the poor prognosis cluster had a mean methylation of 40%, while samples in the better prognosis cluster had a mean methylation of 17%. High HOXA9 methylation (above 40%) was observed in 23/104 (22%) Stage I ADC and associated with shorter cancer-specific survival in the Kaplan-Meier (Log-rank test P = 0.03; Fig. 2A) and Cox regression analyses (hazard ratio [HR], 2.6; P = 0.02), independent of stage (IB vs IA), and smoking history (Table 2). Because a majority of Stage I patients do not receive adjuvant therapy, we focused the analysis only on chemotherapy naïve patients, and similarly found HOXA9 methylation to be associated with shorter cancer-specific survival (HR, 3.8; P = 0.01, Table 2). Multivariable analysis adjusted for adjuvant therapy yielded a similar result (HR, 3.7; P = 0.006, Supplementary Table S13, Supplementary Digital Content 1). Methylation at the miR-196b CpG island was not associated with survival and was not pursued further. In order to be clinically useful, an outcome predictor should be replicated in independently collected patient cohorts. Therefore, we sought to evaluate the association of HOXA9 methylation and lung cancer prognosis in a cohort of 113 Stage I lung ADC patients from Japan (Table 1). We confirmed that the HOXA9 locus was methylated in tumors (paired t-test P < 0.0001; Supplementary Fig. S3, Supplementary Digital Content 1) and that methylation of HOXA9 was associated with relapse in the Kaplan-Meier (Log-rank test P = 0.004; Fig. 2A) and Cox regression (HR, 3.0; P = 0.01) analyses, independent of stage (IB vs IA) and smoking history (Table 2).

Fig. 2.

Kaplan-Meier analysis of HOXA9 promoter methylation in Stage I lung ADC. HOXA9 methylation values were dichotomized based on ≥ 40%/< 40% mean methylation in pyrosequencing analysis. The associations were evaluated in two independent cohorts, including the combined NCI/Norway cohort (left panels) and the Japan cohort (right panels). Shown are the analysis of Stage I (A), and the subgroup analysis of Stage IA (B) and Stage IB (C). P values were calculated by log-rank test.

Table 2.

Univariable and Multivariable Cox Regression of HOXA9 promoter methylation in two cohorts.

HOXA9 promoter methylationa	Univariable			Multivariableb
	N	HR (95% CI)	P	HR (95% CI)	P
NCI/Norway cohort
Stage I	99	2.30 (1.01–4.79)	0.027	2.60 (1.16–5.80)	0.020
Stage I (therapy näive)	80	2.58 (1.10–6.05)	0.029	3.77 (1.37–10.4)	0.010
Stage IA	59	3.27 (1.18–9.03)	0.022	3.55 (1.08–11.6)	0.036
Stage IB	34	1.93 (0.58–6.42)	0.285	1.79 (0.46–1.04)	0.402
Japan cohort (therapy näive)
Stage I	113	3.21 (1.38–7.44)	0.007	3.02 (1.28–7.16)	0.012
Stage IA	81	2.32 (0.74–7.32)	0.150	2.28 (0.67–7.79)	0.189
Stage IB	32	3.62 (1.0–13.2)	0.051	4.37 (1.15–16.6)	0.030
Combined cohort (therapy näive)
Stage I	193	2.55 (1.44–4.54)	0.001	3.06 (1.62–5.82)	0.001
Stage IA	131	2.16 (0.99–4.73)	0.055	3.44 (1.35–8.77)	0.009
Stage IB	56	3.48 (1.32–9.18)	0.012	3.41 (1.27–9.24)	0.015

^aHOXA9 methylation values were dichotomized based on ≥ 40%/< 40% mean methylation in pyrosequencing analysis.

^bAdjusted for stage, smoking history, sex, and age, as well as race and cohort membership when appropriate. Upon restaging to AJCC 7th edition, there were 6 cases in the Norway cohort for which it could not be distinguished whether they were TNM stage IB or II. These are included in univariable analyses and excluded in multivariable analyses.

N: The number of available data for a particular variable in the univariable analysis.

NOTE: Bold, significant values < 0.05.

Stratified analysis of Stage IA and IB patients identified high-risk patients in both cohorts, albeit with loss of power (Fig. 2B, 2C, Table 2). The clinical endpoint reported in the NCI/Norway cohort is cancer-specific death, whereas it is relapse-free survival in the Japan cohort. Because post-recurrence survival of surgically treated patients is reported to be less than a year even for Stage I diagnosis³², we decided to combine the two cohorts for the stratified analysis of stage, focusing only on therapy naïve patients. In the combined cohort, HOXA9 methylation was associated with outcome in Stage I (HR, 3.1; P = 0.001), Stage IA (HR, 3.4; P = 0.009) and Stage IB (HR, 3.4; P = 0.015) patients independent of smoking, race, and cohort membership (Table 2).

A DNA methylation signature that identifies poor prognosis Stage I NSCLC was recently published.²⁰ This signature included HOXA9 promoter methylation as one of five markers associated with relapse-free survival in NSCLC from the United States and Europe, but did not differentiate between adeno and squamous cell carcinoma histologies. CpG sites associated with the other four genes included in this signature (HIST1H4F, PCDHGB6, NPBWR1, and ALX1), although methylated in tumors, did not fit our selection criteria for follow-up by pyrosequencing. We sought to validate the prognostic value of HOXA9 methylation alone within the subset of Stage I ADC patients from that study using publically available data.²⁰ In this third, independently collected cohort, we found that HOXA9 methylation was significantly associated with a greater risk for shorter relapse-free survival in the Kaplan-Meier (Log-rank test P = 0.0002; Supplementary Fig. S4, Supplementary Digital Content 1) and Cox regression (HR, 4.4; P = 0.001) analyses, independent of stage (IB vs IA) and smoking. Thus, HOXA9 methylation can identify high-risk Stage I lung ADC patients from diverse ethnic backgrounds and smoking habits.

A combination of RNA and DNA biomarkers identifies patients with poor prognosis in two independent cohorts of Stage I lung ADC

We previously developed RNA-based prognostic biomarkers that stratify Stage I lung ADC patients.^8–10 A combined high miR-21 and 4-protein-coding gene classifier (based on expression of XPO1, BRCA1, HIF1α, and DLC1) identified a subset of Stage I lung ADC patients with poor prognosis in the same NCI/Norway and Japan cohorts utilized here.⁹ High miR-21, high 4-gene score and HOXA9 methylation were each independently associated with outcome in Stage I lung ADC patients in NCI/Norway and Japan cohorts (Supplementary Table S14, Supplementary Digital Content 1). As they are statistically independent, we hypothesized that a classifier incorporating all three “omic” biomarkers might further refine patient stratification. For this analysis, we categorized patients according to the number of combined high values of HOXA9 promoter methylation, miR-21 and 4-gene signature. An increasing combined score identified high-risk patients in both cohorts (Table 3; Fig. 3; Supplementary Fig. S5, Supplementary Digital Content 1). Importantly, a high score conferred a greater risk for cancer-specific death among therapy naïve patients from the NCI/Norway cohort (HR, 43; P = 0.001, Table 3). Similarly, an increasing combined score conferred a greater risk for shorter relapse-free survival in Stage I patients from Japan (HR, 6.2; P = 0.005, Table 3). The association remained significant when evaluating therapy naïve patients in the combined Stage I cohort (HR, 10.2; P = 3x10⁻⁵, Table 3), as well as in stratified analysis of Stage IA (Fig. 3B) and Stage IB (Fig. 3C), even after adjusting for covariates (Supplementary Table S15, Supplementary Digital Content 1).

Table 3.

Univariable and Multivariable Cox Regression of the combined 4-gene classifier, miR-21 expression and HOXA9 DNA methylation in two cohorts.

Combined Biomarker	Univariable			Multivariablea
	N	HR (95% CI)	P	HR (95% CI)	P
NCI/Norway cohort
0 (all low risk)	24	Reference		Reference
1 (one high risk)	30	2.66 (0.70–10.0)	0.149	2.11 (0.53–8.49)	0.291
2 (two high risk)	30	4.51 (1.29–15.7)	0.018	3.96 (1.07–14.7)	0.040
3 (all high risk)	7	12.0 (2.82–51.0)	0.001	18.5 (3.72–91.9)	4.E-04
		Trend P = 5.E-04		Trend P = 7.E-04
NCI/Norway cohort (therapy näive)
0 (all low risk)	18	Reference		Reference
1 (one high risk)	25	2.61 (0.53–12.9)	0.241	2.13 (0.37–12.1)	0.395
2 (two high risk)	25	4.35 (0.95–19.9)	0.058	5.23 (0.96–28.6)	0.056
3 (all high risk)	4	13.2 (2.17–80.4)	0.005	43.1 (4.52–411)	0.001
		Trend P = 0.005		Trend P = 0.002
Japan cohort (therapy näive)
0 (all low risk)	26	Reference		Reference
1 (one high risk)	36	0.21 (0.02–2.06)	0.183	0.25 (0.03–2.48)	0.239
2 (two high risk)	28	2.64 (0.70–9.94)	0.152	2.91 (0.75–11.3)	0.123
3 (all high risk)	23	6.51 (1.84–22.9)	0.004	6.22 (1.74–22.2)	0.005
		Trend P = 3.E-05		Trend P = 0.0001
Combined cohort (therapy näive)
0 (all low risk)	44	Reference		Reference
1 (one high risk)	61	1.01 (0.32–3.17)	0.992	0.75 (0.23–2.47)	0.633
2 (two high risk)	53	3.37 (1.25–9.07)	0.016	3.14 (1.11–8.80)	0.030
3 (all high risk)	27	7.28 (2.66–20.0)	1.E-04	10.2 (3.43–30.3)	3.E-05
		Trend P = 9.E-07		Trend P = 4.E-07

^aAdjusted for stage, smoking history, age and sex, as well as race and cohort membership when appropriate. Upon restaging to AJCC 7th edition, there were 6 cases in the Norway cohort for which it could not be distinguished whether they were TNM stage IB or II. These are included in univariable analyses and excluded in multivariable analyses.

N: The number of available data for a particular variable in the univariable analysis.

NOTE: Bold, significant values < 0.05.

Fig. 3.

Kaplan-Meier analysis of a combined prognostic biomarker in Stage I lung ADC. Patients in the combined NCI/Norway cohort (left panels) and the Japan cohort (right panels) were categorized according to the combined number of high values for HOXA9 methylation, miR-21 and 4-protein-coding gene signature. An increasing combined score conferred greater risk for poor outcome in Stage I (A), and within subgroup analysis of Stage IA (B) and Stage IB (C). P values calculated by log-rank test.

DISCUSSION

The integrative analysis of “omic”, clinical and epidemiological data for single patients is at the core of a precision medicine strategy.³³ We previously identified prognostic biomarkers of early stage NSCLC based on mRNA^{9, 10} and miRNA⁸ expression, inflammatory cytokines^{34, 35}, and urinary metabolites.³⁶ Here, we report the identification and validation of HOXA9 promoter methylation as a prognostic biomarker and its combination with mRNA and miRNA-based biomarkers into a simple score that is a robust classifier of risk in patients with Stage I lung ADC. Our study addresses a clinically relevant objective: reliably identifying high-risk stage IA patients who could benefit from adjuvant chemotherapy, and Stage IB patients at low risk of recurrence who could be spared its toxic effects.

Our finding has implications for the postsurgical management of patients with Stage I lung ADC. Large clinical trials have failed to provide conclusive evidence of benefit for adjuvant chemotherapy in Stage IB NSCLC.^37–40 In part, that has been due to the lack of reliable biomarkers of risk that can guide therapeutic decisions. By current guidelines, adjuvant chemotherapy is recommended for patients with high-risk Stage IB tumors.⁴ A tumor diameter > 4 cm is one of the high-risk factors to be considered when determining treatment with adjuvant chemotherapy.⁴ In our patient cohorts, among those patients with Stage IB disease defined as high-risk based on high methylation, miR-21 and 4-protein-coding gene scores, only 2/13 had tumors > 4 cm in diameter and would have been eligible for adjuvant chemotherapy under current guidelines. However, 9/11 in the Japan cohort experienced relapse and 2/2 in the NCI/Norway cohort died from lung cancer within 5 years after surgery. Therefore, a molecular classifier may help with risk stratification of Stage IB patients.

The importance of replication and rigorous validation of prognostic biomarkers in independent cohorts cannot be overstated, in light of problems with some prognostic signatures that preclude their clinical application.²² The prognostic value of miR-21 expression and the 4-gene classifier has been verified in several cohorts of Stage I lung ADC.^8–10 HOXA9 was recently identified as one of five methylation markers associated with relapse-free survival in Stage I NSCLC.²⁰ Through an independent analysis of genome-wide methylation, we have identified HOXA9 as a prognostic biomarker for Stage I ADC in two ethnically and geographically different cohorts. We further showed its statistical independence from mRNA and miRNA-based biomarkers and validated the ability of the combination to categorize Stage I patients. The potential utility of a prognostic classifier based on HOXA9 methylation, 4-gene classifier and miR-21 should be tested prospectively as a stratification factor for validation in future adjuvant chemotherapy trial.⁶

In cancer cells, marking by Polycomb proteins and H3K27me3 mediate de novo DNA methylation.⁴¹ Methylation of ESC polycomb targets⁴² and a stem cell-like gene expression signature²⁶ provide a mechanism for the acquisition of stem cell-like features that are observed in many tumors. This process is presumably the result of an epigenetic switch in which ESC developmental genes are stably silenced through DNA methylation in cancer cells.²⁷ The presence of an ESC signature of gene expression is associated with poor differentiation and worse overall survival in lung ADC.⁴³ Our results indicate that an embryonic stem cell-like epigenetic state characterized by high methylation of HOXA9 and other developmental genes is associated with prognosis in early stage lung ADC. HOXA9 methylation could be a surrogate that identifies lung ADC with stem cell-like features or even subpopulations of cancer stem cells within lung ADC that are responsible for recurrence and resistance to therapy. This hypothesis should be addressed in future functional studies.

The development of molecularly targeted therapies against oncogenes that are somatically activated or translocated in tumors has revolutionized the treatment paradigm of lung cancer.⁴⁴ Still, over 50% of lung ADC show no clinically actionable DNA alterations.⁷ Our study supports the contribution of epigenetic alterations to the molecular taxonomy of lung ADC. ADC belonging to the high methylation cluster defined herein have many features in common with CpG island methylator phenotype-High (CIMP-H) lung ADC described previously^{7, 18}, including pervasive DNA methylation of developmental genes, poor prognosis¹⁸ and MYC overexpression.⁷ Unlike DNA aberrations, epigenetic alterations are reversible, making them attractive therapeutic targets.¹¹ We found that EZH2 is overexpressed in ADC with high methylation, consistent with its previously described association with poor lung cancer prognosis^28–30. Pharmacological inhibition of EZH2 elicited growth repression in lung cancer cells^{45, 46} and remarkable tumor regression in xenografts of lymphoma with EZH2-activating mutations.⁴⁷ We hypothesize that CIMP-H methylation profile or the presence of HOXA9 promoter methylation could be a marker for sensitivity to EZH2 inhibition in the absence of EZH2 mutations.

In summary, we have developed and validated a prognostic classifier that comprises three types of genomic data, to enhance clinical management of Stage I lung ADC. Our exploration of the lung cancer methylome in relation to gene and miRNA expression contributes to the molecular taxonomy of lung cancer and may have therapeutic implications. Thus, our approach exemplifies the power of precision medicine to harness diverse molecular data to better categorize disease and inform treatment.

Supplementary Material

Supplemental Digital Content

Fig. S1. Hierarchical clustering of based on CpG sites differentially-methylated in Stage I ADC compared to non-tumor adjacent tissues.

Fig. S2. Confirmatory pyrosequencing analysis of DNA methylation at the HOXA9 locus in Stage I ADC from a subset of the NCI microarray cohort.

Fig. S3. Methylation Beta-values for HOXA9 probe cg26521404 in Stage I ADC samples from Japan.

Fig. S4. Kaplan-Meier analysis of HOXA9 promoter methylation in a published cohort of Stage I lung ADC (J Clin Oncol 2013;31(32):4140-7).

Fig. S5. Kaplan-Meier analysis of a combined prognostic biomarker in Stage I lung ADC.

Table S1. CpG sites differentially-methylated in Stage I ADC compared to non-tumor adjacent tissues.

Table S2. Functional characterization of genes hypermethylated in tumors using Gene Ontology and INTERPRO.

Table S3. Functional characterization of genes hypomethylated in tumors using Gene Ontology and INTERPRO.

Table S4. Summary of Ingenuity Pathway Analysis of genes differentially methylated in tumors.

Table S5. Gene Set Enrichment Analysis of genes differentially methylated in tumors from the NCI microarray cohort.

Table S6. Gene Set Enrichment Analysis of genes differentially methylated in tumors from the Japan microarray cohort.

Table S7. Hypermethylated probe sets corresponding to genes marked by H3K27me3 in ESC.

Table S8. Association between methylation cluster and clinical-demographic variables in the NCI microarray cohort.

Table S9. Gene expression differences between high and low methylation clusters in the NCI microarray cohort.

Table S10. Gene Set Enrichment Analysis of genes differentially expressed between high and low methylation clusters in the NCI microarray cohort.

Table S12. miRNA expression differences between high and low methylation clusters in the NCI microarray cohort.

Table S13. Univariable and Multivariable Cox Regression of HOXA9 promoter methylation in two cohorts.

Table S14. Univariable and Multivariable Cox Regression of 4-protein-coding gene classifier, miR-21 expression and HOXA9 promoter methylation in two cohorts and their overall combination.

Table S15. Univariable and Multivariable Cox Regression of High combined 4-protein-coding gene classifier, miR-21 expression and HOXA9 methylation in the combined NCI/Norway and Japan cohorts.