DNA methylation profiling analysis identifies a DNA methylation signature for predicting prognosis and recurrence of lung adenocarcinoma

Abstract

The aim of the present study was to identify candidate prognostic DNA methylation biomarkers for lung adenocarcinoma (LUAD), since the modern precise medicine for the treatment of LUAD requires more biomarkers and novel therapeutic targets of interest. DNA methylation profiling data of LUAD were downloaded from The Cancer Gene Atlas portal. Differentially methylated genes (DMGs) were screened to differentiate between samples designated as good and bad prognosis. LUAD-associated methylation modules were obtained with the weighted correlation network analysis (WGCNA) package, followed by function enrichment analysis. Optimal prognostic DMGs were selected using the LASSO estimation-based Cox-PH approach and were used to construct a prognostic risk scoring system. The training set was dichotomized by risk score, into high- and low-risk groups. The differences in overall survival (OS) time or recurrence-free survival (RFS) time between the two groups were evaluated using a Kaplan-Meier curve. A total of 742 DMG samples were screened for good and bad prognosis. WGCNA identified three LUAD-associated modules, which were primarily associated with cytoskeleton organization, transcription and apoptosis. A nine-gene prognostic methylation signature was determined, which included C20orf56, BTG2, C13orf16, DNASE1L1, ZDHHC3, FHDC1, ARF6, ITGB3 and ICAM4. A risk score-based methylation signature classified the patients in the training set into high- and low-risk groups with significantly different OS or RFS times. The prognostic value of the methylation signature was successfully verified in a validation set. In conclusion, the present study identified a nine-gene methylation signature for the prediction of survival and recurrence in patients with LUAD and improved the understanding of the alterations in DNA methylation in LUAD.

Introduction

Lung adenocarcinoma (LUAD) is the most predominant subtype of non-small cell lung cancer (NSCLC), with increased incidence over the past decades worldwide (1,2). LUAD is usually observed in the peripheral region of the lungs, with a poor overall five-year survival rate of 15% worldwide in 2008 (3). Due to resistance to radiation therapy, LUAD is often treated surgically (4). Nonetheless, approximately a third of patients relapse within five years of surgical removal (5).

DNA methylation, a primary epigenetic modification in the mammalian genome, often occurs at CpG islands, leading to altered tumor suppressor gene transcription (6). Aberrant DNA methylation plays a key role in the progression and metastasis of LUAD, reflecting important biological features in the etiology (7). Zhu et al (8) identified a group of genes with differentially methylated loci in LUAD. Han et al (9) found that methylated PTPRF (protein tyrosine phosphatase, receptor type F), HOXD3 (homeobox D 3), HOXD13 and CACNA1A (calcium voltage-gated channel subunit alpha1 A) may be potential markers of LUAD, based on DNA methylation profiling analysis. Furthermore, Sandoval et al (10) suggested a prognostic signature based on five hypermethylated genes for early stage NSCLC. Additionally, Kuo et al (11) established a proof-of-concept prognostic signature of eight methylated probes for survival prediction in Asian and Caucasian populations with early-stage LUAD. Despite these remarkable findings, there is a lack of a prognostic DNA methylation signature for LUAD.

In the present study, genome-wide methylation analysis was carried out on the methylation data of 425 patients with LUAD, with corresponding clinicopathological features from The Cancer Genome Atlas (TCGA). LUAD-associated co-methylation modules were mined with the weighted correlation network analysis (WGCNA) package. Furthermore, a group of differentially methylated genes (DMGs) predictive of survival were identified for LUAD by performing differential DNA methylation, correlation, univariate Cox regression and L1 penalized (LASSO) Cox proportional hazards (PH) regression analyses. These findings may potentially contribute to a deeper insight into the epigenetic landscape of LUAD and improve the prognostic evaluation of patients.

Materials and methods

Data resources

DNA methylation data and the corresponding survival information of 425 LUAD tissue samples were downloaded from TCGA portal (https://gdc-portal.nci.nih.gov/) on May 26th, 2018, based on the Illumina Infinium Human Methylation 450 BeadChip platform, and were used as the training set in the present study. The GSE39279 dataset was downloaded from the National Center for Biotechnology Information Gene Expression Omnibus (NCBI GEO) database (http://www.ncbi.nlm.nih.gov/geo/), based on the Illumina Infinium Human Methylation 450 BeadChip platform, consisting of the gene methylation data of 443 NSCLC samples. Among these samples, 155 LUAD samples with available survival information were selected as the validation set. The clinicopathological features of the training set and the validation set are shown in Table I.

Table I.

Clinicopathological characteristics of patients in the training set and the validation set.

Clinicopathological characteristics	Training set (n=425)	Validation set (n=155)
Age, years (mean ± SD)	65.12±10.04	65.11±10.85
Sex, male/female	198/227	76/79
Death, dead/alive/-	120/305	–
OS time, months (mean ± SD)	22.13±28.39	–
RFS time, months (mean ± SD)	18.802±26.34	54.68±45.62
Recurrence, yes/no	87/228	68/87

SD, standard deviation; -, information unavailable; OS, overall survival; RFS, recurrence-free survival.

Differential DNA methylation analysis

A bad prognosis was defined as patients who died or who had a survival time <12 months, whereas a good prognosis indicated living patients who survived >24 months. According to the annotation profiles provided by the platform, only the DNA methylation loci in CpGs were reserved. The comparison of samples associated with good and bad prognosis from the TCGA set, identified differentially methylated CpGs (DM CpGs), using the limma package (12) of R language (version 3.34.7; http://bioconductor.org/packages/release/bioc/html/limma.html). A strict cut-off value was set at false discovery rate (FDR)<0.05 and |Log₂fold change (FC)|>0.1. The genes mapped by the identified DM CpGs were defined as the DMGs.

Co-methylation analysis

In order to analyze the inter-correlation among the identified DM CpGs, co-methylation network analysis was carried out using the WGCNA R package (version 1.63; http://cran.r-project.org/web/packages/WGCNA/index.html), as previously described (13). Briefly, a thresholding power function (β) of 5 was chosen to fit a scale-free network. Topological overlap matrix (TOM) was then calculated to measure the correlations between the methylation levels of two genes. The resulting hierarchical clustering dendrogram was obtained, followed by selection of the modules with a minimum module size of 100 and a minimum cut height of 0.95, using the Dynamic Tree Cut algorithm. These identified DM CpGs were then mapped to the modules obtained by WGCNA analysis. The enrichment of target DM CpGs in each module was assessed by hypergeometric-based test (14), using the following formula: f (k, N, M, n) = C (k, M) × C (n-k, N-M)/C (n, N).

The modules with P<0.05 and fold enrichment >1 were further selected as LUAD-associated modules and subjected to Gene ontology (GO) enrichment analysis using the DAVID 6.8 software (15,16). This revealed the biological functions of the DMGs clustered in these modules.

Correlation of DNA methylation level with gene expression level

The genome-wide expression data of the LUAD samples in the TCGA set was obtained. The correlation between the overall methylation level and the overall gene expression level of the DMGs, included in the LUAD-associated modules, was analyzed with the cor.test function of R language by calculating the Pearson's correlation coefficient (PCC) (17). Subsequently, the PCC of the overall methylation level of each individual gene, with its overall gene expression level was also computed. The genes with negative PCC and P<0.05 were used for further analysis.

Construction of a prognostic risk scoring model based on the training set

Univariate Cox regression analysis was performed to identify the prognosis-associated DMGs from the aforementioned genes with negative PCC, using the survival package of R language (18) (http://bioconductor.org/packages/survivalr/), with log-rank P<0.05 as the cutoff.

A LASSO estimation-based Cox-PH model (19) was used to select the optimal panel of genes predictive of prognosis from these prognosis-related DMGs by the penalized package (version 0.9–50) (20) of R language. Combining the Cox-PH coefficients of the optimal genes with their methylation levels, a risk scoring model was constructed for predicting survival as follows: Risk score = ∑ coef_gene × methylation_gene, where coef_gene represents the Cox-PH coefficient of a gene and methylation_gene represents the methylation level of a gene.

A risk score was assigned to each patient in TCGA dataset and the median risk score for all patients was calculated to divide patients into a low-risk group (risk score below the median value) and a high-risk group (risk score above the median value). To estimate the overall survival (OS) time and recurrence-free survival (RFS) time of the patients in the two risk groups, the Kaplan-Meier estimate (21) was used together with the Wilcoxon log rank test. The areas under the receiver operating characteristic (AUROC) curves were used to evaluate the prognostic ability of the risk scoring model and tested in the validation set.

Results

DM CpGs between samples with good and bad prognosis

With regard to the DNA methylation data of the TCGA set, 15654 methylated loci in CpGs were retained, according to the Illumina 450 K methylation platform. This set was comprised of 41 samples with bad prognosis and 72 samples with good prognosis. As shown in Fig. 1A, a total of 742 DMGs (3.79×10⁻⁷ <P-nominal <6.84×10⁻⁴, 2.77×10⁻⁵ <FDR <0.05) were identified between the samples with good and bad prognosis. Kernel density plotting of log₂FC showed that of the 742 DMGs, 289 were hypermethylated, whereas 453 were hypomethylated in the samples with good prognosis (Fig. 1B). Two-way hierarchical clustering analysis based on the methylation levels of these DMGs revealed that the subjects in the TCGA set were clustered into two different groups (Fig. 1C). Furthermore, out of the CpGs of the 742 identified DMGs, 447 were located in gene bodies, 155 in 5′ untranslated regions (UTRs), 98 in 3′UTRs and 42 in promoters. The top 20 DMGs were chosen according to their FDR value and shown in ascending order in Table II.

Figure 1.

Analysis of DMGs between samples with good and bad prognosis. (A) Volcano plot of effect size log₂(FC) against -log₁₀(FDR) of DMGs. Red spots represent DMGs with FDR<0.05 (horizontal red dashed line) and |log₂FC| >0.1 (two vertical red dashed lines). (B) Kernel density plotting of log₂(FC) displays 289 DMGs of hypermethylation (61.05%) and 453 DMGs of hypomethylation (38.95%). (C) Two-way hierarchical clustering analysis of the samples in the The Cancer Genome Atlas set, using a heatmap based on the methylation level of the DM CpGs. The color bar represents CpG methylation level (green to red, low to high level). FC, fold changes; FDR, false discovery rate; DMGs, differentially methylated genes.

Table II.

Top 20 significant differentially methylated genes.

Methylated loci	Chromosome	Gene	Position	Location	β-good	β-bad	Effect	Pnominal	FDR
cg21644316	3	QTRTD1	115260371	5′UTR	0.7249	0.8314	0.1978	3.79×10−7	2.77×10−5
cg10753610	17	ITGB3	42690202	Body	0.2511	0.3604	0.5216	8.45×10−7	6.17×10−5
cg02960016	3	HRH1	11167068	5′UTR	0.7498	0.6936	−0.1125	2.15×10−6	1.57×10−4
cg12240358	15	HOMER2	81410528	Body	0.5326	0.4059	−0.3918	2.52×10−6	1.84×10−4
cg07684796	10	DKK1	53744216	1stExon	0.1152	0.1554	0.4317	2.78×10−6	2.03×10−4
cg18440692	10	FAM53B	126298025	3′UTR	0.4420	0.3205	−0.4636	3.17×10−6	2.31×10−4
cg23894219	1	LOC441869	1346429	Body	0.5987	0.4876	−0.2961	4.50×10−6	3.28×10−4
cg23581793	14	SCARNA13, SNHG10	95069529	Body	0.3519	0.4572	0.3778	5.54×10−6	4.05×10−4
cg01531333	7	TMEM184A, MAFK	1548749	3′UTR	0.7444	0.8366	0.1685	5.97×10−6	4.36×10−4
cg01289421	19	B3GNT3	17769247	5′UTR	0.6009	0.6740	0.1657	6.24×10−6	4.56×10−4
cg07428439	17	LOC284023	7759563	Promoter	0.1277	0.1120	−0.1892	6.39×10−6	4.67×10−4
cg01798157	1	BTG2	201543219	Promoter	0.5384	0.3961	−0.4427	7.28×10−6	5.32×10−4
cg00259834	3	RFTN1	16528504	Promoter	0.3736	0.2582	−0.5332	7.45×10−6	5.44×10−4
cg04481715	13	ATP4B	113351540	3′UTR	0.9121	0.8259	−0.1432	7.89×10−6	5.76×10−4
cg26511108	10	COMMD3, BMI1	22648958	3′UTR	0.4687	0.5726	0.2886	8.86×10−6	6.47×10−4
cg03366285	15	FSD2	81271781	5′UTR	0.724321	0.651058	−0.15384	9.07×10−6	6.63×10−4
cg03087607	6	TCF21	134252361	body	0.427617	0.327662	−0.38411	9.66×10−6	7.06×10−4
cg07097184	4	FHDC1	154119463	3′UTR	0.433799	0.359003	−0.27303	1.17×10−5	8.58×10−4
cg02538597	12	ARPC3	109372526	Promoter	0.099173	0.090652	−0.12961	1.35×10−5	9.89×10−4
cg24272980	3	CIDECP	10040206	Body	0.752599	0.816553	0.117666	1.36×10−5	9.92×10−4

UTR, untranslated region; FDR, false discovery rate.

LUAD-associated co-methylation modules

For the purpose of detecting LUAD-associated co-methylation modules, a weighted gene co-methylation network was constructed for the identified DMGs. As depicted in Fig. 2A and B, the DMGs with significantly correlated methylation level (P<0.05) were assigned into 13 methylation modules (Table III). The DMGs without significant correlation in their methylation level were grouped into the grey module.

Figure 2.

WGCNA methylation modules. (A) A hierarchical clustering dendrogram generated by WGCNA analysis. A total of 14 modules were identified and depicted in different colors. The colored row underneath shows the module assignment determined by the Dynamic Tree Cut. (B) The number of differentially methylated genes mapped to different modules. (C) Fold enrichment of each module. Fold enrichment represents enrichment of differentially methylated CpGs in each module. *P<0.05, **P<0.01 and ***P<0.001. WGCNA, weighted correlation network analysis.

Table III.

Methylation modules from weighted correlation network analysis.

Module color	Count of CpGs	Correlation efficient	Pcorr	Count of DM CpGs	Enrichment fold (95% CI)	Phyper
Black	468	0.502	0.0024	68	2.818 (2.106–3.728)	<0.0001
Blue	368	0.609	0.0033	22	1.159 (0.709–1.811)	0.4730
Brown	410	0.488	0.0047	10	0.473 (0.223–0.889)	0.0176
Green	302	0.495	0.0027	18	1.156 (0.668–1.886)	0.5110
Yellow	228	0.486	0.0023	8	0.681 (0.288–1.379)	0.3560
Grey	3363	0.407	0.6550	117	0.674 (0.541–0.837)	0.0002
Magenta	341	0.532	0.0007	17	0.967 (0.550–1.594)	0.9680
Pink	299	0.469	0.0005	27	1.752 (1.119–2.644)	0.0095
Purple	194	0.517	0.0068	10	0.999 (0.468–1.901)	1.0000
Red	127	0.539	0.0005	1	0.153 (0.038–0.872)	0.0214
Salmon	138	0.428	0.0039	26	3.653 (2.275–5.673)	<0.0001
Tan	127	0.471	0.0045	8	1.222 (0.512–2.509)	0.5450
Turquoise	343	0.811	0.0029	15	0.848 (0.465–1.439)	0.6160
Yellow	256	0.518	0.0007	12	0.909 (0.459–1.637)	0.8850

Count of CpGs, the number of CpGs included in one module; DM, differentially methylated; count of DM CpGs, the number of DM CpGs mapped to one module; P_corr, P-value for correlation coefficient; P_hyper, P-value for enrichment analysis; CI, confidential interval.

The enrichment of the DMGs in each module was analyzed using a hypergeometric-based test. Three modules with fold enrichment >1 and P<0.05 were identified as LUAD-associated methylation modules (black module, size=68; pink module, size=27; salmon module, size=26; Fig. 2C). Moreover, GO enrichment analysis found that the total of 121 DMGs in the three modules were highly enriched in 16 biological processes, primarily associated with cell death, cytoskeleton organization and cell junction organization (Table IV).

Table IV.

Significant GO terms for the three LUAD-associated modules.

GO term	Count of genes	Fold enrichment	P-value
Cytoskeleton organization	8	2.821	0.0110
Response to DNA damage stimulus	7	2.885	0.0165
Spindle organization	3	10.248	0.0169
Positive regulation of cytoskeleton organization	3	10.248	0.0169
Regulation of transcription from RNA polymerase II promoter	10	2.115	0.0218
Cell junction organization	3	8.091	0.0260
Anterior/posterior pattern formation	4	4.392	0.0307
Regulation of cellular component biogenesis	4	4.330	0.0318
Microtubule cytoskeleton organization	4	4.183	0.0345
Positive regulation of transcription, DNA-dependent	7	2.256	0.0433
Positive regulation of RNA metabolic process	7	2.237	0.0446
Cell death	9	1.924	0.0454
Apoptosis	8	2.043	0.0462
Death	9	1.911	0.0468
Programmed cell death	8	2.013	0.0490
Positive regulation of organelle organization	3	5.556	0.0498

GO, Gene Ontology; LUAD: lung adenocarcinoma; Count of genes, the number of genes significantly enriched in one GO term.

Development and validation of a prognostic scoring model based on DMGs

The methylation levels of the 121 DMGs clustered in the three LUAD-associated modules were negatively correlated with their overall gene expression levels (PCC=−0.5008, P=1.618×10⁻⁸; Fig. 3). Furthermore, the negative correlation between the methylation level of the individual gene and its gene expression level was observed in 94 genes, of which 58 DMGs were found to be significantly associated with survival via univariate Cox regression analysis (P<0.05, Table V). Consequently, a LASSO Cox PH model was fitted with the methylation level of the 58 prognosis-associated DMGs. When the cross-validated likelihood (CVL) reached the maximal value of 1.537, the optimal λ value was −691.265. As a result, the most powerful prognostic panel of genes was selected, including C20orf56, BTG anti-proliferation factor 2 (BTG2), C13orf16, deoxyribonuclease 1 Like 1 (DNASE1L1), zinc finger DHHC-type containing 3 (ZDHHC3), FH2 domain containing 1 (FHDC1), ADP ribosylation factor 6 (ARF6), integrin subunit beta 3 (ITGB3) and intercellular adhesion molecule 4 (ICAM4) (Table VI).

Figure 3.

A scatter diagram for correlation analysis between the gene methylation and expression levels in the three lung adenocarcinoma-associated modules. Black dots represent the genes. The red line represents the trend line of the data points. Cor, Pearson's correlation coefficient.

Table V.

Genes associated with survival according to Cox regression analysis.

ID	Gene	P-value
cg03395898	TGFB3	6.0×10−5
cg01798157	BTG2	4.2×10−4
cg07097184	FHDC1	5.1×10−4
cg02334643	DHX40	8.4×10−4
cg05003322	COL16A1	8.6×10−4
cg27500918	FLYWCH2	9.1×10−4
cg00933153	C20orf56	1.0×10−3
cg20541456	CYFIP2	1.1×10−4
cg05898928	YPEL1	1.2×10−4
cg16194253	C14orf21	1.5×10−4
cg21494776	ICAM4	1.6×10−4
cg11013977	QRICH1	1.7×10−4
cg06620210	AP1M1	2.1×10−4
cg02016545	MICA	2.3×10−4
cg03595580	TECPR1	2.5×10−4
cg18050194	SEC22C	4.0×10−3
cg08517562	PTPN1	5.6×10−4
cg02156071	C10orf84	5.7×10−4
cg06877599	SUPT5H	5.7×10−4
cg04902921	EDEM3	6.3×10−4
cg16740905	SEC1	6.9×10−4
cg04432377	ZDHHC3	7.1×10−4
cg10753610	ITGB3	7.9×10−4
cg05714082	POLR1A	8.5×10−4
cg04446303	TFAP4	9.4×10−4
cg02282317	AATF	0.011
cg08100565	SLC25A36	0.011
cg01889020	MEGF11	0.014
cg02181309	MRPL52	0.014
cg01498883	SNRPB	0.014
cg06677352	STAG3L3	0.014
cg24122247	CIDECP	0.015
cg24135606	PFN1	0.016
cg24612420	ACLY	0.017
cg19759282	GPR155	0.017
cg13939431	MEAF6	0.017
cg01963754	C13orf16	0.018
cg04324559	DNASE1L1	0.019
cg24951800	MEF2A	0.019
cg00567190	C1orf97	0.020
cg12811419	TMEM214	0.024
cg00399374	CHMP4C	0.025
cg01779512	IFT88	0.027
cg16931807	KIAA0195	0.032
cg15599146	ZDHHC14	0.033
cg07459266	RNF213	0.034
cg08125503	PIGL	0.035
cg07057042	RAB5A	0.035
cg21994174	ETFB	0.036
cg06421633	LUC7L3	0.036
cg03270167	RAMP1	0.037
cg05006947	SLC38A7	0.040
cg10156217	ARF6	0.042
cg07097417	LPGAT1	0.042
cg02146453	PROSC	0.043
cg17286258	SF3B1	0.043
cg18085070	PSRC1	0.044
cg27229100	C20orf199	0.046

Table VI.

A panel of nine differentially methylated genes predictive of prognosis.

ID	Gene	Chromosome	Position	Location	Coefficient	Hazard ratio	P-value
cg00933153	C20orf56	20	22498129	Body	0.8529	3.177	0.0010
cg01798157	BTG2	1	201543218	Promoter	0.9931	2.606	0.0004
cg01963754	C13orf16	13	110775249	Body	−0.5670	0.499	0.0180
cg04324559	DNASE1L1	X	153290455	5′UTR	−0.1602	0.712	0.0190
cg04432377	ZDHHC3	3	44992301	Promoter	0.6870	2.894	0.0071
cg07097184	FHDC1	4	154119462	3′UTR	1.0356	5.175	0.0005
cg10156217	ARF6	14	49432104	Promoter	−0.9612	0.260	0.0420
cg10753610	ITGB3	17	42690201	Body	−0.4113	0.489	0.0079
cg21494776	ICAM4	19	10258780	Body	0.4918	2.065	0.0016

The association of the methylation level of the nine prognostic genes and survival was investigated. All samples in the TCGA set were dichotomized into hypermethylated group and hypomethylated group, based on the median methylation level of each prognostic gene, separately (Fig. 4). For C20orf56, BTG2, ZDHHC3, FHDC1 and ICAM4, the hypomethylated samples showed better prognosis compared with the hypermethylated samples (P<0.05). By contrast, DNASE1L1, ARF6, ITGB3 and C13orf16 had significantly longer OS time in the hypermethylated samples relative to the hypomethylated samples.

Figure 4.

Kaplan-Meier curves of dichotomized patients, based on the methylation level of each gene in the training set. According to the methylation level of one gene, all patients in the training set were separated into a hypermethylation and a hypomethylation group.

The association between the gene expression level of the nine predictive genes and prognosis was also underscored. Specifically, all samples in the training set were divided into a high- and low-expression group, according to the median expression level of each gene, separately. High expression of C20orf56, BTG2, ZDHHC3, FHDC1 and ICAM4 showed significantly improved prognosis in comparison with low expression samples (Fig. 5). Concerning DNASE1L1, ARF6, ITGB3 and C13orf16, improved survival was reported in the low-expression samples relative to high-expression samples (Fig. 5). For prognosis stratification, a prognostic scoring system was developed as follows: Risk score = (0.8529) × methylation_cg00933153 + (0.9931) × methylation_cg01798157 + (−0.5670) × methylation_cg01963754 + (−0.1602) × methylation_cg04324559 + (0.6870) × methylation_cg04432377 + (1.0356) × methylation_cg07097184 + (−0.9612) × Methylation_cg10156217 + (−0.4113) × methylation_cg10753610 + (0.4918) × methylation_cg21494776.

Figure 5.

Kaplan-Meier curves showing patients, classified based on the expression level of each gene in the training set. Based on the gene expression level of one gene, all patients in the training set are classified into a high and low expression group.

The risk score was calculated for each individual patient in the TCGA set accordingly. All patients in the TCGA set were classified into a high- and low-risk group by risk score using the median risk score of all patients as the cut-off value. As shown in Fig. 6A, low-risk patients had significantly longer OS time compared with the high-risk patients (P=8.992×10⁻⁷), with an AUC value of 0.966. The nine-gene methylation signature was also applied to predict the patients' probability of recurrence in the training set. The low-risk patients had visibly longer RFS time in comparison with the high-risk patients, with an AUC value of 0.928 (Fig. 6B). In order to test the capability of the nine-gene methylation signature, all samples in the validation set were classified by risk score into high- and low-risk groups. Similarly, low-risk patients had a dramatically longer RFS time relative to high-risk patients. The AUC value was 0.920 (Fig. 6C), confirming the prognostic power of the nine-gene methylation signature.

Figure 6.

Kaplan-Meier curves (left panel) and ROC curves (right panel) demonstrating patients assigned to two risk groups, based on the nine-gene methylation signature in the (A and B) training set and the (C) validation set. (A) Overall survival time and (B) recurrence-free survival time of the high- and low-risk groups of the training set were estimated using Kaplan-Meier curves. The AUC of ROC reveals the capability of the nine-gene methylation signature to predict survival or recurrence. ROC, receiver operating characteristic; AUC, area under the curve.

Discussion

LUAD is clinicopathologically and molecularly heterogeneous, making the prediction of patient outcome a necessity (22,23). Gene silencing at the epigenetic level by DNA methylation was acknowledged as an important mechanism underlying tumorigenesis (24). The present study placed a special focus on DNA methylation alterations in LUAD and their implications for prognosis. A total of 742 DMGs were identified, showing significantly different methylations in CpGs between the samples designated as good and bad prognosis. There were three LUAD-associated DNA co-methylation modules mined by WGCNA. The 121 DMGs in the 3 modules were significantly associated with 16 GO terms. Notably, three GO terms were associated with cytoskeleton organization, two GO terms with transcription and four GO terms with cell apoptosis or death. These findings provided some insight into the underlying mechanisms of DNA methylation alterations in LUAD.

The present study identified a nine-gene methylation signature from the genes included in the three LUAD-associated DNA methylation modules. Moreover, risk score derived from the sum of each candidate methylation marker multiplied by the corresponding regression coefficient successfully classified patients into two risk groups, with significantly different OS or RFS time. The prognostic performance of this nine-gene signature was successfully verified for RFS time in the validation set. These findings indicate that the nine genes were valuable methylation markers for prognostic evaluation in LUAD patients.

The nine novel candidate methylation markers for prognosis prediction included C20orf56, BTG2, C13orf16, DNASE1L1, ZDHHC3, FHDC1, ARF6, ITGB3 and ICAM4. Moreover, the present study found that hypomethylation/high expression of ICAM4, ZDHHC3, C20orf56, BTG2 and FHDC1 was associated with significantly improved survival outcome compared with hypermethylation/low expression. In contrast, hypermethylation/low expression of DNASE1L1, ARF6, ITGB3 and C13orf16 corresponded with significantly improved survival in comparison with when these genes were hypomethylated/highly expressed.

BTG2 is involved in various biological activities, such as cell differentiation, proliferation and apoptosis, and has been acknowledged as a tumor suppressor in several types of cancer in humans, including laryngeal carcinoma, gastric cancer, clear cell renal cell carcinoma and breast carcinoma (25–29). More importantly, there is evidence that BTG2 overexpression suppressed the growth and proliferation of lung cancer cells (30). It can be speculated that aberrant BTG2 methylation participated in the regulation of genes involved in LUAD tumor growth. In addition, the present study implicated two possible functions for BTG2: Response to DNA damage stimulus and anterior/posterior pattern formation in LUAD. These findings offer useful information concerning the role of BTG2 through the regulation of methylation in LUAD.

ARF6 is a member of the human ARF gene family and plays a role in vesicular trafficking. Increasing evidence demonstrated an association between ARF6 and tumor cell invasion (31,32). Moreover, in the present study, GO analysis showed that ARF6 was significantly implicated in the positive regulation of cytoskeleton organization, regulation of cellular component biogenesis, cell death, apoptosis and positive regulation of organelle organization. Thus, aberrant ARF6 methylation may play a role in the regulation expression of genes associated with these biological processes in LUAD.

ITGB3, also known as CD61, is a protein encoded by ITGB3 and participates in cell adhesion and signaling mediated by the cell surface (33,34). The present study found that ITGB3 was significantly enriched in cell junction organization in LUAD. Based on these observations, it can be inferred that alteration of ITGB3 methylation exerted an effect on the expression of genes involved in cell junction organization, thus modulating cell adhesion and cell surface-mediated signaling. ICAM4 is a member of the ICAMs family and is critical for inflammation and immune responses (35). DNASE1L1 is an enzyme encoded by DNASE1L1, a member of the human DNase family. To the best of our knowledge, there are few reports regarding the function of ZDHHC3, C13orf16, C20orf56, FHDC1, ICAM4 and DNASE1L1 in LUAD.

The limitations of the present study include minimal information on the OS time in the validation set, as other DNA methylation dataset of LUAD with survival information could not be located in NCBI GEO. Furthermore, only correlations between DNA methylation and the gene expression levels were investigated, based on data of LUAD downloaded from TCGA. Thus, protein expression of the nine novel candidate methylation markers for prognostic prediction should be also studied in the future. The aim of the present study was to provide novel prognostic DNA methylation biomarkers of LUAD, since the modern precise medicine for the treatment of LUAD required additional biomarkers and novel therapeutic targets of interest. However, the findings of the present study require validation in prospective clinical trials before the prognostic multigene methylation signature can be applied. Therefore, Chinese-population-based validation could be considered in the future.

In conclusion, the present study focused on the DNA methylation changes associated with LUAD and identified a prognostic nine-gene methylation signature for LUAD. The findings shed light on the DNA methylation landscape in LUAD and its implications on the development of optimized and individualized management of this condition.

Glossary

Abbreviations

LUAD

lung adenocarcinoma

DMGs

differentially methylated genes

WGCNA

weighted correlation network analysis

OS

overall survival

RFS

recurrence-free survival

NSCLC

non-small-cell lung cancer

TCGA

The Cancer Genome Atlas

PH

proportional hazards

TOM

topological overlap matrix

PCC

Pearson correlation coefficient

Funding

No funding was received.

Availability of data and materials

The datasets used and/or analyzed during the present study are available from the corresponding author on reasonable request.

Authors' contributions

RW and HZ analyzed and interpreted the sequencing data. MXY conceived and designed the study. CRZ checked the data analysis results and was a major contributor in writing the manuscript. All authors read and approved the final manuscript.

Ethics approval and consent to participate

Not applicable.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.