A prognostic predictor panel with DNA methylation biomarkers for early-stage lung adenocarcinoma in Asian and Caucasian populations

The incidence of lung adenocarcinoma (LUAD) is increasing worldwide with different prognosis even in early-stage patients. We aimed to identify a prognostic panel with multiple DNA methylation biomarkers to predict survival in early-stage LUAD patients of different racial groups. The methylation array, pyrosequencing methylation assay, Cox regression and Kaplan-Meier analyses were conducted to build the risk score equations of selected probes in a training cohort of 69 Asian LUAD patients. The risk score model was verified in another cohort of 299 Caucasian LUAD patients in The Cancer Genome Atlas (TCGA) database. We performed a Cox regression analysis, in which the regression coefficients were obtained for eight probes corresponding to eight genes (AGTRL1, ALDH1A3, BDKRB1, CTSE, EFNA2, NFAM1, SEMA4A and TMEM129). The risk score was derived from sum of each methylated probes multiplied by its corresponding coefficient. Patients with the risk score greater than the median value showed poorer overall survival compared with other patients (p = 0.007). Such a risk score significantly predicted patients showing poor survival in TCGA cohort (p = 0.036). A multivariate analysis was further performed to demonstrate that the eight-probe panel association with poor outcome in early-stage LUAD patients remained significant even after adjusting for different clinical variables including staging parameters (hazard ratio, 2.03; p = 0.039). We established a proof-of-concept prognostic panel consisting of eight-probe signature to predict survival of early-stage LUAD patients of Asian and Caucasian populations.


Background
Lung cancer is the leading cause of cancer-related deaths with an increasing incidence of lung adenocarcinoma (LUAD) subtype worldwide [1]. Prognosis may vary in patients with the same stage tumor because cancer is characterized by genetic, epigenetic, and phenotypic changes that result in a tremendous variability in clinical behavior [2,3]. Therefore, the development of additional molecular markers for survival prediction of LUAD is required.
Therefore, in the current study we focus on the development of survival predictors in early-stage LUAD patients by performing genome-wide methylation analysis and pyrosequencing quantitative methylation assay to select eight DNA methylation probes in a training cohort of 69 patients recruited in Taipei Veterans General Hospital (TVGH). We also included certain clinical parameters that are known to affect prognosis [2,3,[17][18][19] along with the selected eight-probe panel to the Cox regression analysis. The relevance of our finding has been validated in a cohort of 299 patients as part of The Cancer Genome Atlas (TCGA) project.

Patients and tissue samples
A total of 69 surgically resected LUAD patients in early stage (stages I and II) were recruited from Taipei Veterans General Hospital (TVGH), after obtaining appropriate institutional review board permission (#98-03-18A) and informed consent from the patients. These LUAD patients with checked clinical data and sufficient amount of DNA available for successful genome-wide methylation and pyrosequencing quantitative methylation assays were defined as a training cohort. A validation cohort of 299 LUAD patients with clinical follow-up data and methylation microarray data available from TCGA were collected. The mean follow-up period for training cohort was 82 months (range 9-157 months) and for validation cohort was 37 months (range 12-242 months). The end of the follow-up in TVGH was defined as January 2016 and TCGA as April 2015. Patients with clinicopathological characteristics are shown in Table 1.

Genomic DNA extraction and sodium bisulfite conversion
Genomic DNA from primary tumor tissue samples of 69 patients from TVGH were extracted using proteinase K digestion and phenol-chloroform extraction. A total of 1 μg genomic DNA was used for bisulfite conversion using the EpiTect Bisulfite kit (Qiagen, Duesseldorf, Germany) according to the manufacturer's protocols.

The genome-wide methylation analysis platform
The Illumina Infinium HumanMethylation27 BeadChip (27,578 CpG dinucleotides for 14,495 genes) was adapted for DNA methylation detection according to manufacturer's manual. DNA methylation levels were reported as β-values by calculating the ratio of intensities between locus-specific methylated and unmethylated bead-bound probes. The β-value is a continuous variable, ranging from 0 (unmethylated) to 1 (fully methylated). The methylation array data can be viewed online under GEO accession number GSE83845.

Pyrosequencing assay
To quantify cytosine methylation in individual CpG sites of candidate methylation probes identified by methylation array, bisulfite-converted DNA was analyzed using a pyrosequencing system (PyroMark Q24, Qiagen, Hilden, Germany). Specific pyrosequencing primer and PCR primer were designed for "target" CpG sites in the probes to be analyzed. Pyrosequencing was carried out in accordance with the manufacturer's protocol (Qiagen). The target CpG sites were evaluated by converting the resulting pyrograms to numerical values for peak heights. Primer sequences are listed in Additional file 1: Table S1, and the genomic map of the detected CpG sites are shown in Additional file 1: Figure S1.

Data processing and statistical analysis
Receiver operating characteristic (ROC) curve analysis was performed to determine the accuracy of the established CpG panel [area under the curve (AUC), sensitivity, and specificity]. The univariate and multivariate Cox regression analyses were conducted to explore the relationship between patient survival and several explanatory variables for defining the hazard ratio (HR) and confidence intervals (CI) of cancer death risk of variables using the Statistical Package for the Social Sciences version 17.0 (SPSS Inc., Headquarters Chicago, IL, USA). Overall survival curves were calculated according to the Kaplan-Meier method. p < 0.05 was considered statistically significant.

Marker discovery in genome-scale DNA methylation dataset
In the marker selection phase of this study, we collected surgically dissected tumors of 69 early-stage LUAD patients from TVGH to form a training cohort for genome-wide methylation array analysis using Illumina Infinium HumanMethylation27 BeadChip. Procedures were performed as described below and shown in Fig. 1. First, we obtained 9384 qualified probes after removing potentially problematic probes, probes containing SNP, repeat sequencing, probes not in CpG island, probes in the X-chromosome and the non-differential probes with β-value greater than 0.9 or less than 0.1 in all samples after methylation analysis. Second, we selected 2815 informative probes with large variance of β tumor (top-30 % ranked) among the patients. Third, for each probe, a supervised principal components (Superpc) analysis [20] was applied to compare the survival distributions between patients with methylation levels above and below the mean level of all tumor tissues analyzed; 100 probes were chosen based on these tests. Fourth, since pyrosequencing assay is a highly sensitive method for detection of DNA methylation [12,17,21], we performed pyrosequencing methylation analyses. Of these probes, pyrosequencing was successfully designed and performed for 34 probes. A high concordance in quantifying the CpG methylation level was observed between DNA methylation array and pyrosequencing assay (Fig. 2). Fifth, eight specific candidate probes corresponding to eight genes showed significant correlation with survival by univariate Cox regression, including Angiotensin II receptor-like 1 (AGTRL1), Aldehyde dehydrogenase 1 family member A3 (ALDH1A3), Bradykinin receptor B1 (BDKRB1), Cathepsin E (CTSE), Ephrin A2 (EFNA2), NFAT activating protein with ITAM motif 1 (NFAM1), Semaphorin 4A (SEMA4A), and Transmembrane protein 129 (TMEM129) ( Table 2). Sixth, we applied receiver operating characteristic (ROC) curve analysis to determine the diagnostic sensitivity and specificity of eight-probe panel in the validation cohort of  Sensitivity and specificity of the eight selected probes by ROC analysis We examined the sensitivity and specificity of the eight selected probes by ROC curve analysis in the training cohort of 69 early-stage LUAD patients from TVGH.
The area under the curve (AUC) of eight-probe together was 0.802 (Fig. 3a), indicating that the eight-probe signature showed good sensitivity and specificity in the ROC analysis. To assess the accuracy of the prognostic predictor panel, ROC curve analysis was performed on another randomly selected eight probes from the top 34 candidate probes. The AUC was 0.602 (Fig. 3b), suggesting a stronger prediction power of the specifically selected eight probes than the randomly selected probes. Thus, we defined this eight-probe signature as the prognostic predictor panel of early-stage LUAD.
The risk score calculation and survival prediction of the eight-probe panel by Kaplan-Meier method In the clinical validation phase, we first built the risk score for the eight selected methylation probes using the multivariate Cox regression analysis in the TVGH training cohort of 69 early-stage LUAD patients. These DNA methylation probe covariates were weighted by the regression coefficients to calculate the coefficient and hazard ratio for each patient. The risk score for each patient was derived from sum of methylation value of each probe multiplied by the corresponding coefficient, as following equation: risk score = AGTRL1 methylation value × (-0.015) + ALDH1A3 methylation value × (-0.023) + BDKRB1 methylation value × (-0.034) + CTSE methylation value × (0.022) + EFNA2 methylation value × (0.010) + NFAM1 methylation value × (-0.017) + SEMA4A methylation value × (-0.012) + TMEM129 methylation value × (-0.006). Example of risk score calculation for two patients is shown in Additional file 1: Figure S2. Furthermore, we used the risk score calculation ranging from -1.03 to -4.95 to classify patients into two groups by median value of -2.63 in the TVGH training cohort of 69 early-stage LUAD patients (upper panel, Fig. 4a). The Kaplan-Meier overall survival analysis was performed to show the relative survival in each of the two groups identified by the risk score calculation (middle panel, Fig. 4a). Patients with high risk score indeed had a short median survival time (MST) of 58.9 months compared with other patients. The difference in the MST and 95 % confidence interval (CI) between the two groups was highly significant (lower panel, Fig. 4a). Therefore, the median risk score (as -2.63) was chosen as the cutoff value for survival prediction in the TVGH cohort.
We further applied our risk score model to determine whether our finding could be validated in another cohort of 299 early-stage LUAD patients whose follow-up data were available in TCGA project, and methylation level was also determined by the Infinium Methylation array. The risk score calculated with the median value (as 0.47) classified the 299 TCGA patients into two groups (upper panel, Fig. 4b). Such a calculation predicted a subset of patient with a high risk score showing poorer survival with MST of 50.9 months (middle panel, Fig. 4b) with statistical significance (lower panel, Fig. 4b). These results indicated that the prognostic predictor panel consisting of the selected eight-probe showed a strong prediction value in the TCGA validation cohort.

Univariate and multivariate Cox regression analysis of the eight-probe panel
To determine whether the eight-probe panel is an independent variable associated with poor survival of earlystage LUAD patients, we performed the univariate and multivariate Cox regression model in both TVGH and TCGA cohorts. The univariate Cox regression analysis revealed that patients with risk score > median of the eight-probe panel, stage IIA, stage IIB, or lymph node metastases had poor outcome (p = 0.009, HR = 2.37, 95 % CI = 1.24-4.53 for risk score > median of the eight-probe panel; Table 3). Notably, multivariate Cox regression analysis showed that the eight-probe panel correlated with a relative risk of death of 2.03 (p = 0.039), even after adjusting for the tumor staging and metastasis status ( Table 3), suggesting that the eightprobe panel was an independent risk factor of poor outcome.
To further define the prognostic effects of the eightprobe panel in early-stage LUAD patients, univariate and multivariate Cox regression analyses were performed in the TCGA validation cohort of 299 early-stage LUAD patients. Univariate Cox regression analysis revealed that patients with the risk score > median of the eight-probe panel had poor outcome, with a relative risk of death of 1.66 (p = 0.038) ( Table 3). However, the eight-probe panel showed a borderline significance by the multivariate analysis in the TCGA cohort.

Discussion
The incidence of LUAD is increasing worldwide [1]. Patients with the same stage of lung cancer may have different prognosis [22]. Development of prognostic markers is especially important in the patients with early-stage lung cancer, in whom clinical oncologists need selection factors to decide whether adjuvant therapy is necessary. In the present study, we develop a prognostic predictor panel for early-stage LUAD patients. This panel consists of eight DNA methylation probes corresponding to eight specific genes, including AGTRL1, ALDH1A3, BDKRB1, CTSE, EFNA2, NFAM1, SEMA4A, and TMEM129. The risk score calculated using the eight-probe panel served as an independent prognosis biomarker by Cox regression model and the multivariate analysis in our recruited patients. Therefore, the risk scores calculated from this eight-probe panel are valuable biomarkers for prognostic evaluation for early-stage LUAD patients to be tested in other cohorts. Recently, Heller et al. identified a total of 12 genes that were differentially methylated in tumors compared with surrounding tissues in stage I, II or III Caucasian non-small cell lung cancer patients. Among the 12 genes, only the methylation patterns of HOXA2 and HOXA10 were independent prognostic factors in lung squamous cell carcinoma patients [15]. In addition, Esteller and the associates used methylation array to establish methylation profiles of stage I Caucasian non-small cell lung cancer and identified that methylation of two or more genes in HIST1H4F, PCDHGB6, NPBWR1, ALX1, and HOXA9 correlated with an increased risk of cancer recurrence [16]. Interestingly, HOXA9 promoter methylation was associated with high risk in stage I LUAD patients of two independently cohorts by another study [23]. To date, all studies that have been executed in an attempt to find markers for clinical use do not include patients from different racial groups. In our study, the prognostic predictor panel comprising eight DNA methylation biomarkers was an independent risk factor of poor outcome in Asian LUAD patients. We further applied our risk score model to determine whether our finding could be validated in another cohort of 299 early-stage LUAD patients whose follow-up data were available in TCGA project. The new coefficient and hazard ratio were defined according to the methylation value of the eight probes given in TCGA database of these patients. The Kaplan-Meier overall survival analysis showed that TCGA patients with risk score greater than median value had a shorter MST compared with other patients (Fig. 4b). However, the result of multivariate Cox regression was only close to significance in the Caucasian LUAD patients ( Table 3). One of the limitations of the current TCGA study is that we are unable to acquire the data on treatment or surgery performed on the TCGA patients (Table 1). We believe that these results could be improved after including data from more patients when they are available in TCGA dataset or by validating in other cohorts of Caucasian LUAD patients.
The identification of the eight probes that can predict the clinical outcome in patients may reveal causes of the cancer development and tumorigenesis. For example, Angiotensin II receptor-like 1 (AGTRL1) and Bradykinin receptor B1 (BDKRB1) are G-proteincoupled receptors (GPCRs). GPCRs, which represent by far the largest family of cell-surface molecules involved in signal transduction, have recently emerged as crucial players in tumor growth and metastasis [24]. AGTRL1 is Apelin receptor. Apelin is an angiogenic factor secreted by tumor cells in order to promote the formation of new vessels necessary for tumor growth [25]. In addition, crosstalk between BDKRB1 and EGFR has been shown to maintain tumor growth in the breast cancer [26]. Aldehyde dehydrogenase 1 family, member A3 (ALDH1A3) is the retinoic acid biosynthesis enzyme, and plays a major role in the detoxification of aldehydes generated by alcohol metabolism and lipid peroxidation. Promoter hypermethylation of ALDH1A3 has been reported to be a prognostic marker for lung cancer, gastric cancer, and invasive bladder cancer [27][28][29][30]. Cathepsin E (CTSE) prevents tumor growth and metastasis by catalyzing the proteolytic release of soluble trail from tumor cell surface [31]. Ephrin A2 (EFNA2), which belongs to ephrins family, regulates cell adhesion, motility, survival, proliferation, and differentiation. Semaphorins 4A (SEMA4A) suppresses endothelial cell migration and proliferation in vitro and angiogenesis in vivo mediated by vascular endothelial growth factor [32]. Further characterization of the probes validated in our panel could help to dissect the mechanism of LUAD tumorigenesis and progression.
The advantages of our prognostic predictor panel are as follows. First, the methylation level of the eight probes could be analyzed by DNA methylation array or pyrosequencing in patients. Second, the stepwise multivariate Cox regression analysis, in which the coefficients were obtained for the selected eight probes, could generate the risk score equations specifically for the cohort of patients to be tested. Third, any newly recruited patients could be assigned into risk groups once the risk score equations are determined. Therefore, the prognostic predictor panel could calculate the risk score not only in the Asian but also in the Caucasian LUAD patients. However, some technical limitations such as sample collection and preprocessing as well as experimental procedures of DNA methylation array or pyrosequencing assay need to be controlled to avoid batch effects. In addition, clinical variables such as adjuvant therapy and surgical methods may affect outcome prediction. Large-scale, multicenter and prospective studies are necessary to validate our risk score model in early-stage LUAD patients.

Conclusions
Our study provides a proof-of-concept prognostic prediction panel consisting of eight methylated probes that are closely associated with survival in the earlystage LUAD patients. This prediction panel could be useful in stratifying patients according to the Coxmodel and risk score before further treatment for early-stage LUAD patients who in dire need of intensive care.