Prognostic prediction of a 12-methylation gene-based risk score system on pancreatic adenocarcinoma

Abstract

Pancreatic adenocarcinoma (PAAD) accounts for ~85% of all pancreatic cancer cases and is associated with a less favorable prognosis. Aberrant DNA methylation may influence the progression of PAAD by inducing abnormal gene expression. Methylation data of PAAD samples with prognosis information were obtained from The Cancer Genome Atlas (training set) and European Bioinformatics Institute Array Express databases (validation sets). Using the limma package, the differentially methylated genes in the training dataset were screened. Combined with the Weighted Gene Co-expression Network Analysis package, the co-methylated genes in key modules were identified. Then, a cor.test function in R software was applied to explore the functions of key the methylated genes. Correlation analyses of the expression levels and methylation levels of key methylated genes were performed, followed by identification of methylated genes associated with prognosis using Univariate Cox regression analysis. The optimal combination of prognosis related methylated genes was determined using a Cox-Proportional Hazards (Cox-PH) model. Subsequently, the risk score prognostic prediction system was constructed by combining the Cox-PH prognosis coefficients of the selected optimized genes. Based on the constructed risk score system, samples in all datasets were divided into high and low risk samples and the survival status was compared using survival curves. Furthermore, the correlation between independent prognostic factors and the risk score system was determined using the survival package. A total of 50 genes associated with prognosis of PAAD and a 12-gene optimal combination were obtained, including: CCAAT/enhancer binding protein α, histone cluster 1 H4E, STAM binding protein-like 1, phospholipase D3, centrosomal protein 55, ssDNA binding protein 4, glutamate AMPA receptor subunit 1, switch-associated protein 70, adenylate-cyclase activating polypeptide 1 receptor 1, yippee-like 3, homeobox C4 and insulin-like growth factor binding protein 1. Subsequently, a risk score prognostic prediction system of these 12 genes was constructed and validated. In addition, pathological N category, radiotherapy and risk status were identified as independent prognostic factors. Overall, the risk score prognostic prediction system constructed in the present study may be effective for predicting the prognosis of patients with PAAD.

Introduction

Pancreatic cancer (PC) is primarily associated with diabetes, obesity, smoking and genetic conditions, such as germline pathogenic variants and somatic pathogenic variants in DNA damage repair (DDR) genes (1,2). Jaundice, weight reduction, back or abdominal pain, deep-colored urine, pale-colored stools and anorexia are the typical symptoms (3). However, at diagnosis, the cancer has usually metastasized (4,5). The mortality rate of PC is high and there were 411,600 PC-associated deaths globally in 2015 (6). Pancreatic adenocarcinoma (PAAD) is a common type of PC, accounting for ~85% of all PC cases (7). The survival rate of PAAD is very low and the 5-year survival rate was 5% in 2015 (8). A few prognostic indicators are now available for PC, such as C-reactive protein/albumin ratio and neutrophil/lymphocytes ratio (9,10). Therefore, it is of great importance to investigate the prognostic factors of PAAD for improved prediction.

It has been hypothesized that DNA methylation may provide a link between environmental factors contributing to cancer development (11). The stability of the genome and gene expression levels are primarily maintained by a pre-determined pattern of DNA methylation (12). It has been reported that DNA hypermethylation has prognostic value and acts as independent predictor of survival in other cancer types, such as head and neck cancer (13). A previous study reported an association between abnormal methylation of the Reprimo gene and genetic instability and poor survival following surgical resection in patients with PC (14). Hypermethylation of Cyclin D2 is also frequently observed in PC (15). Meanwhile, another study reported a significant difference in the hypermethylation frequency of ALX4, BNC1, HIC1, SEPT9V2,SST, TFPI2 and TAC1 between PAAD samples of stage I, II, III and IV, and these genes are significantly associated with distant metastasis of PAAD (16).

A number of clinical markers of PAAD survival have been recognized, including stage at diagnosis, grading and performance status and the treatment received, such as resection versus no resection and chemotherapy vs. no chemotherapy (17–20). It has also been reported that obesity and smoking are associated with a less favorable prognosis of PC (21). Cigarette smoking is associated with the development of ~20% of PC cases and is therefore a consistent risk factor (22). Bioinformatic databases may serve as a valuable tool to further our understanding of the molecular mechanisms underpinning the prognosis of patients with cancer.

The present study aimed to explore the aberrant methylation of genes associated with prognosis of patients with PAAD. The methylation data of genes associated with PAAD were obtained from The Cancer Genome Atlas (TCGA) database and screened for differentially methylated genes (DMGs) associated with prognosis. Subsequently, these data were used to construct a risk score system, which may be effective in predicting the prognosis of patients with PAAD.

Materials and methods

Datasets

The methylation data for the training dataset was obtained from the TGCA database (accessed on 5^th June 2018; cancer.gov/tcga), which were based on the Illumina Infinium Human Methylation 450 BeadChip platform. There were a total of 184 samples, 168 of which included prognostic information [mean age, 64.89±11.24 years (range: 40–88); male: female, 94/75; average overall survival time, 17.09±15.22 months; death: survival, 88/80]. The methylation data for the validation training set was obtained from the European Bioinformatics Institute ArrayExpress database (ebi.ac.uk/arrayexpress/), specifically the E-MTAB-5008 and E-MTAB-5571 datasets. The E-MTAB-5008 dataset consisted of 29 PAAD samples with prognostic information and the E-MTAB-5571 dataset consisted of 24 PAAD samples which prognostic information. Both of these datasets were sequenced on the platform of Illumina Infinium Human Methylation 450 BeadChip.

Screening of DMGs

To screen genes associated with PAAD prognosis, samples in the TCGA dataset were divided into less favorable prognosis (defined as a survival time <6 months and death) and more favorable prognosis (defined as survival time >24 months or alive) groups based on a previously described grouping method (23). The methylation loci of genes associated with PAAD prognosis were annotated and combined with the platform annotation information on the Illumina Infinium Human Methylation 450 BeadChip platform and the loci within CpG islands of the genes were selected and used for the following analysis. Using the limma package in R (version 3.34.7) (24), the DMGs between the less favorable prognosis and more favorable prognosis groups were screened according to the following criteria: |log fold change| >0.263 and false discovery rate (FDR) <0.05. Then, the Kernel density curve of the DMGs was generated by calculating the Log 2 (FC).

Identification of co-methylated genes based on Weighted Gene Co-expression Network Analysis (WGCNA)

Co-methylation analysis using the WGCNA package (version 1.63) (25) in R was performed on genes located in CpG islands to identify differentially methylated CpG genes (DMCpGs). The sets of CpG genes with highly related methylation levels under the same biological process or in different tissues were considered as modules. The modules which had a significant association with the methylation levels were identified. The DMCpGs were mapped to the modules and the significant enrichment parameters and fold enrichment were calculated using a hypergeometric test (26). The DMCpGs enriched modules were screened under the following criteria: P<0.05 and a fold enrichment value of >1 was considered to indicate a statistically significant difference. Genes in DMCpGs enriched modules were recognized as key methylation genes and were analyzed using Gene Ontology (GO) enrichment analysis (27) using the Database for Annotation, Visualization and Integrated Discovery (version 6.8) (28).

To meet scale-free network distribution, the weighting parameter ‘power’ in WGCNA algorithm was explored. When the square of the correlation coefficient between log(k) and log[p(k)] reached 0.9, the corresponding value of parameter ‘power’ (power=7) was selected. Under power=7, the mean connectivity of genes was calculated to be 1. Subsequently, the adjacency matrix elements were serialized, and the topological overlap matrix was calculated to evaluate the correlation of gene methylation levels and obtain a system clustering tree. According to the standards of hybrid dynamic shear tree, pruning height (cutHeight) and the minimum number of module genes (minSize) separately were set as 0.95 and 50.

Correlation analysis for the expression levels and methylation levels of key methylated genes

The expression and methylation levels of key methylation genes in matched training PAAD samples were collected and correlation analysis was performed. Pearson correlation coefficients (PCCs) were calculated and the cor.test function (stat.ethz.ch/R-manual/R-devel/library/stats/html/cor.test.html) in R Software was used (29,30). PCCs of single genes were also calculated and the genes with a negative correlation between expression level and methylation level were selected as key methylation genes for subsequent analyses. P<0.05 was considered to indicate a statistically significant difference.

Identification of the methylated genes associated with prognosis

Key methylation genes with a negative correlation between expression level and methylation level were further analyzed for prognosis associated genes. Univariate Cox regression analysis was performed to identify prognosis associated methylation genes using the survival package (version 2.41–1) in R (31). P<0.05 was considered to indicate a statistically significant difference.

Construction of risk score prognostic prediction system

The optimal combination of prognosis related methylation genes was screened using the Cox-Proportional Hazards (Cox-PH) model using the penalized package (version 0.9–50) in R (32). The optimal parameter of ‘lambda’ in the Cox-PH model was calculated through 1,000 cross-validation likelihoods (cvl) (33).

The risk score prognostic prediction system was constructed combining the Cox-PH prognosis coefficients and methylation levels of the selected optimized genes. The resultant formula was: Risk score=∑coef _gene × Methylation _gene where Coef_gene and Methylation_gene represented regression coefficient and gene methylation levels, respectively.

The risk scores of samples in the TCGA, E-MTAB-5008 and E-MTAB-5571 datasets were calculated and stratified into high and low risk groups according to the median risk scores. Kaplan-Meier (KM) survival curves (34) of the high and low samples were plotted using the survival package, which were compared with the prognosis of all samples. The area under the receiver operating characteristic (ROC) curve (AUC) was compared, also using the survival package.

Correlation analysis between independent prognostic factors and risk score prognostic prediction system

Using the Cox regression analysis in the survival package (31), independent clinical prognostic factors were screened. Next, the relations between collected factors and the risk score prognostic prediction system were analyzed using KM curves.

Results

Screening of DMGs

Median survival time of samples in the training datasets was 17.09±15.22 months, which is consistent with the time reported in PC (35). There were 13,903 methylation sites containing CpG islands in the training dataset. In TCGA training dataset, based on the predefined method for grouping, each less favorable and more favorable prognosis group had 19 samples, and a total of 1,067 DMGs between the two groups were identified (Fig. 1).

Figure 1.

Volcano plot, density curve and cluster heatmap of DMGs. (A) Volcano plot showing the FC of each identified gene. Red dots, DMGs; horizontal dashed line, FDR <0.05; vertical dashed lines, |log FC| >0.263. (B) Kernel density curve showing that 74.98% of the DMGs are significant hypomethylated in the good prognostic group. (C) Bidirectional hierarchical cluster heatmap. Orange and purple samples strips separately represent the samples in good and bad prognostic groups, respectively; color changes from green to red represent methylation level changes from low to high, respectively. DMGs, differentially methylated genes; FDR, false discovery rate; FC, fold change.

As shown in the Log2 Kernel density curve of the DMGs, 74.98% (800/1,067) of DMGs were significantly hypomethylated and 25.02% (267/1,067) were significantly hypermethylated in the good prognostic group (Fig. 1B). The cluster heatmap of the DMGs suggested that the samples with different prognoses in the TCGA dataset exhibited different methylation levels (Fig. 1C). Furthermore, as the feature factors had different weights in the calculation process for heatmap analysis, there was a slight crossover between good and bad prognosis samples.

Among the CpGs in the 1,067 DMGs, 309, 321, 118, 44, 185 and 90 CpGs were separately located in transcription start site areas, body areas, 5′untranslated regions (UTR), 3′UTR regions, promoter regions and the first exon regions (data not shown). The top 20 DMGs with smaller FDR values were screened and presented in Table I.

Table I.

Top 20 differentially methylated genes with significance.

Methylation loci	Chr.	Position	Genes	Location	β-bad	β-good	Effect	P-value	FDR
cg24902557	chr14	104786509	BTBD6	Promoter	0.3277	0.0926	−1.8239	5.24×10−7a	4.25×10−6
cg27106909	chr16	30014398	YPEL3	Promoter	0.3150	0.1325	−1.2489	8.88×10−7a	7.20×10−6
cg25207224	chr6	34319167	HMGA1	Body	0.5645	0.7547	0.4188	5.80×10−6a	4.70×10−5
cg27543607	chr16	88289602	CDK10	3′UTR	0.6878	0.5288	−0.3791	6.59×10−6a	5.34×10−5
cg27058889	chr1	55237413	BSND	1stExon	0.5441	0.7574	0.4772	6.71×10−6a	5.44×10−5
cg21197594	chr9	139776802	EHMT1	Body	0.7238	0.4881	−0.5683	9.65×10−6a	7.82×10−5
cg26789779	chr5	37875700	GDNF	TSS200	0.3598	0.1482	−1.2801	9.92×10−6a	8.04×10−5
cg23066982	chr6	26312442	HIST1H4E	Promoter	0.2446	0.0759	−1.6886	1.12×10−5a	9.08×10−5
cg27575890	chr22	35237104	EIF3D	Body	0.3592	0.2527	−0.5075	1.16×10−5a	9.44×10−5
cg24864161	chr22	48870409	MOV10L1	TSS200	0.2022	0.0689	−1.5531	1.21×10−5a	9.84×10−5
cg24460144	chr5	149550073	SLC6A7	5′UTR	0.3093	0.1822	−0.7637	1.43×10−5a	1.16×10−4
cg27576241	chr1	153301194	ADAM15	Body	0.5530	0.4002	−0.4667	1.50×10−5a	1.21×10−4
cg23930825	chr6	43721343	RSPH9	Body	0.3502	0.2102	−0.7365	1.74×10−5a	1.41×10−4
cg26432961	chr6	41625040	FOXP4	5′UTR	0.7897	0.5080	−0.6365	1.79×10−5a	1.45×10−4
cg27626790	chr7	150305686	KCNH2	Body	0.2627	0.1002	−1.3909	1.80×10−5a	1.45×10−4
cg09730123	chr16	1767949	SPSB3	Body	0.4173	0.3468	−0.2670	1.81×10−5a	1.47×10−4
cg25499067	chr8	23615933	NKX2-6	Body	0.2310	0.1187	−0.9606	1.97×10−5a	1.60×10−4
cg25902939	chr19	18405350	SSBP4	Body	0.4202	0.2254	−0.8986	2.01×10−5a	1.63×10−4
cg27390009	chr17	18091226	FLII	Body	0.5073	0.3645	−0.4768	2.05×10−5a	1.66×10−4
cg24542766	chr22	38043478	RPL3	Body	0.3924	0.2856	−0.4580	2.10×10−5a	1.70×10−4

^aP<0.001. cg indicates methylation ID of a gene; chr., chromosome; Position, position of the methylation sites in chromosomes; Location, location of CpG islands; β-bad, the average methylation level β in bad prognostic groups; β-good, the average methylation level β in good prognostic groups; Effect, log₂ fold change; FDR, false discovery rate; TSS, transcription start site; UTR, untranslated region.

Identification of co-methylated genes based on WGCNA

A total of 10 modules were identified (Fig. 2A) and the detailed information of each module is listed in Table II. CpG island genes in 9 modules showed significant a association with methylation levels (P<0.05; correlation coefficients, 0.226–0.729; average correlation coefficient, 0.7742; Table II). The identified DMGs were mapped into each module and their distribution in the modules is shown in Fig. 2B. Two modules were identified as differentially expressed CpG gene enriched modules, black module (comprised of 90 DMGs) and the turquoise module (comprised of 394 DMGs), in which the CpG genes were significantly associated with methylations. The DMGs in these two modules were significantly enriched in 18 GO_Biology Process (BP; such as ‘neuron differentiation’), 7 GO_cellular component (CC; such as ‘integral to plasma membrane’), and 9 GO_molecular function (MF; such as ‘sequence-specific DNA binding’) terms (Table III) and were predominantly associated with transcriptional regulation.

Figure 2.

Results of Weighted Gene Co-expression Network Analysis. (A) Module partition tree. (B) Mapping results of the DMGs into the identified modules. Numbers in the brackets represent genes in this module. (C) Histogram of fold enrichment of the modules. (D) Scatter diagram showing the overall correlation between the methylation levels and expression levels of the 484 DMGs. Red line, trend line of point distribution. *Represents the modules with significantly enriched methylated genes (P<0.05). DMGs, differentially methylated genes; cor., correlation coefficient.

Table II.

A total of 10 modules identified by weighted gene co-expression network analysis.

Modules	Count of CpGs	Correlation	Pcorr	DMGs	Enrichment fold (95% CI)	Phyper
Black	159	0.703	1.01×10−13	90	3.481 (2.622–4.599)	2.20×10−16
Blue	427	0.587	1.68×10−4	63	0.908 (0.676–1.202)	5.37×10−1
Brown	364	0.551	2.23×10−3	10	0.169 (0.0799–0.317)	4.83×10−13
Green	352	0.616	6.05×10−2	12	0.209 (0.107–0.374)	3.49×10−11
Grey	574	0.226	1.32×10−1	48	0.514 (0.371–0.701)	5.31×10−6
Magenta	90	0.729	9.47×10−16	23	1.572 (0.941–2.529)	7.38×10−2
Pink	138	0.679	1.44×10−8	6	0.267 (0.0961–0.601)	2.88×10−4
Red	191	0.308	6.69×10−7	1	0.0322 (0.000814–0.182)	3.26×10−11
Turquoise	1344	0.687	1.72×10−6	394	1.803 (1.564–2.077)	4.11×10−16
Yellow	359	0.609	4.27×10−6	3	0.0514 (0.0105–0.152)	2.20×10−16

DMGs, differentially methylated genes; CI, confidence interval; P_corr, P-value for correlation; P_hyper, P-value for hypermethylation.

Table III.

Functional terms enriched for the 484 differentially methylated genes involved in black or turquoise modules.

A, Biological process
Term	Count	P-value	FDR
GO:0030182-neuron differentiation	40	1.79×10−10a	3.13×10−7
GO:0006355-regulation of transcription, DNA-dependent	95	4.27×10−10a	7.45×10−7
GO:0051252-regulation of RNA metabolic process	96	6.29×10−10a	1.10×10−6
GO:0007423-sensory organ development	27	1.46×10−9a	2.55×10−6
GO:0007389-pattern specification process	28	8.94×10−9a	1.56×10−5
GO:0045165-cell fate commitment	20	1.20×10−8a	2.10×10−5
GO:0048568-embryonic organ development	22	1.59×10−8a	2.77×10−5
GO:0048598-embryonic morphogenesis	28	1.66×10−7a	2.90×10−4
GO:0007267-cell-cell signaling	40	9.79×10−7a	1.71×10−3
GO:0006928-cell motion	34	1.66×10−6a	2.90×10−3
GO:0045449-regulation of transcription	111	2.63×10−6a	4.60×10−3
GO:0003002-regionalization	20	3.02×10−6a	5.27×10−3
GO:0007610-behavior	33	3.54×10−6a	6.18×10−3
GO:0048666-neuron development	27	3.76×10−6a	6.57×10−3
GO:0031328-positive regulation of cellular biosynthetic process	42	4.11×10−6a	7.17×10−3
GO:0009891-positive regulation of biosynthetic process	42	5.85×10−6a	1.02×10−2
GO:0016477-cell migration	23	1.16×10−5a	2.02×10−2
GO:0019226-transmission of nerve impulse	26	1.97×10−5a	3.45×10−2
B, Cellular component
Term	Count	P-Value	FDR
GO:0005887-integral to plasma membrane	59	9.12×10−7a	1.21×10−3
GO:0031226~intrinsic to plasma membrane	59	1.89×10−6a	2.50×10−3
GO:0044459-plasma membrane part	90	2.93×10−6a	3.88×10−3
GO:0034703-cation channel complex	15	6.78×10−6a	8.98×10−3
GO:0034705-potassium channel complex	12	1.11×10−5a	1.48×10−2
GO:0008076-voltage-gated potassium channel complex	12	1.11×10−5a	1.48×10−2
GO:0034702-ion channel complex	18	1.99×10−5a	2.64×10−2
C, Molecular function
Term	Count	P-value	FDR
GO:0043565-sequence-specific DNA binding	58	4.65×10−16a	6.55×10−13
GO:0003700-transcription factor activity	74	4.42×10−15a	6.52×10−12
GO:0030528-transcription regulator activity	85	3.53×10−10a	5.18×10−7
GO:0003677-DNA binding	109	2.11×10−8a	3.09×10−5
GO:0022836-gated channel activity	25	6.95×10−6a	1.02×10−2
GO:0005261-cation channel activity	23	1.01×10−5a	1.48×10−2
GO:0015267-channel activity	29	1.41×10−5a	2.07×10−2
GO:0022803-passive transmembrane transporter activity	29	1.48×10−5a	2.18×10−2
GO:0005216-ion channel activity	27	3.37×10−5a	4.94×10−2

^aP<0.001. GO, gene ontology; FDR, false discovery rate.

Correlation analysis of the expression levels and methylation levels of key methylated genes

Overall, the methylation levels and expression levels of the 484 DMGs in the black and turquoise modules were significantly negatively correlated (Cor.=−0.478, P=8.169×10⁻⁵; Fig. 2D). A total of 192 DMGs exhibited negative correlation between the expression levels and methylation levels (Table SI).

Construction of the risk score system

A total of 50 genes among the 192 DMGs were found to be associated with prognosis. Following this, a Cox-PH model was utilized to screen the optimal gene combination. When λ=1.1389, the maximum value of cvl was obtained as −458.1914 (Fig. 3A). Using λ=1.1389, a 12-gene optimal combination was acquired: CCAAT/enhancer binding protein α (CEBPα); histone cluster 1 H4E (HIST1H4E); STAM binding protein-like 1, (STAMBPL1) phospholipase D3 (PLD3); centrosomal protein 55 (CEP55); ssDNA binding protein 4 (SSBP4); glutamate AMPA receptor subunit 1 (GRIA1); switch-associated protein 70 (SWAP70); adenylate-cyclase activating polypeptide 1 receptor 1 (ADCYAP1R1); yippee-like 3 (YPEL3); homeobox C4 (HOXC4); and insulin-like growth factor binding protein 1 (IGFBP1) (Fig. 3B; Table IV). Combined with the prognostic coefficients of these 12 optimal genes, the following risk score system was constructed (cg is the methylation ID for corresponding genes.).

Figure 3.

Selection curve and coefficient distribution diagram. (A) Selection curve of lambda (the junction indicates that the maximum value of cvl is −458.1914 when λ=1.1389) (B) The coefficient distribution diagram of the genes implicated in the optimal gene combination.cvl, cross validation likelihood; prof$ in Y-axis indicated the p value using the predictive model, and in X-axis indicated the lambda value using the predictive model; CEBPA, CCAAT/enhancer binding protein α; HIST1H4E, histone cluster 1 H4E; STAMBPL1, STAM binding protein-like 1; PLD3, phospholipase D3; CEP55, centrosomal protein 55; SSBP4, ssDNA binding protein 4; GRIA1, glutamate AMPA receptor subunit 1; SWAP70, switch-associated protein 70; ADCYAP1R1, adenylate-cyclase activating polypeptide 1 receptor 1; YPEL3, yippee-like 3; HOXC4, homeobox C4; IGFBP1, insulin-like growth factor binding protein 1.

Table IV.

Information of the 12 optimal genes.

Methylation ID	Gene	Chr.	Position	Location	Coefficient	Hazard ratio	P-value
cg22250546	CEBPA	chr19	38483210	Promoter	−0.4701559	0.197	2.01×10−2a
cg23066982	HIST1H4E	chr6	26312442	Promoter	1.461097	1.228	4.11×10−2a
cg23264429	STAMBPL1	chr10	90631983	5′UTR	−0.1543761	0.369	4.53×10−2a
cg25509871	PLD3	chr19	45563397	5′UTR	0.2124921	1.082	4.40×10−2a
cg25827255	CEP55	chr10	95246749	Promoter	−0.7063513	0.129	2.04×10−2a
cg25902939	SSBP4	chr19	18405350	Body	0.1268035	1.139	4.45×10−2a
cg26343183	GRIA1	chr5	152988914	Body	−2.4642526	0.276	1.33×10−2a
cg26645401	SWAP70	chr11	9643090	Body	−0.4583647	0.272	2.02×10−2a
cg27076139	ADCYAP1R1	chr7	31058243	TSS1500	0.0652014	1.014	1.35×10−2a
cg27106909	YPEL3	chr16	30014398	Promoter	0.2566994	1.294	4.32×10−2a
cg27138204	HOXC4	chr12	52732367	5′UTR	0.7707755	1.679	1.58×10−2a
cg27447599	IGFBP1	chr7	45894465	TSS200	0.172493	1.193	2.18×10−2a

^aP<0.01. chr, chromosome; coef, correlation coefficient; UTR, untranslated region; TSS, transcription start site.

Risk score=(−0.4701559) × Methylation _cg22250546 + (1.461097) × Methylation _cg23066982 + (−0.1543761) × Methylation _cg23264429 + (0.2124921) × Methylation _cg25509871 + (−0.7063513) × Methylation _cg25827255+ (0.1268035) × Methylation _cg25902939 + (−2.4642526) × Methylation _cg26343183+ (−0.4583647) × Methylation _cg26645401 +(0.0652014) xMethylation _cg27076139+ (0.2566994) × Methylation _cg27106909 + (0.7707755) × Methylation _cg27138204 + (0.172493) × Methylation _cg27447599.

According to the median of the risk scores of the samples in the TCGA dataset, the samples were divided into high and low risk groups. For the TCGA dataset, the comparison between the actual overall survival and risk score system predicting survival of the risk groups was performed and the AUC was 0.976 (Fig. 4A). Moreover, the risk score system was validated in the E-MTAB-5008 (Fig. 4B) and E-MTAB-5571 (Fig. 4C) datasets and the AUCs were 0.919 and 0.924, respectively. The TCGA, E-MTAB-5008 and E-MTAB-5571 datasets had consistent results and all the samples in low risk group had improved survival.

Figure 4.

KM curves and AUC under the ROC curve. (A) KM curve (left) and ROC curve (right) for TCGA dataset. (B) KM curve (left) and ROC curve (right) for E-MTAB-5008 dataset. (C) KM curve (left) and ROC curve (right) for E-MTAB-5571 dataset. KM, Kaplan-Meier; HR, hazard ratio; TCGA, The Cancer Genome Atlas; ROC, receiver operating characteristic; AUC, area under the curve.

Correlation analysis between independent prognostic factors and risk groups

The clinical information of 168 samples in the TCGA dataset was statistically analyzed, and pathological tumor-node-metastases (TNM) staging system (36), radiotherapy and risk status were identified as independent prognostic factors (Table V). Survival status of high and low risk groups with different pathological N stages (N0 vs. N1) and treatment with radiotherapy (treatment with radiotherapy vs. without radiotherapy) were compared (Fig. 5).

Table V.

The results for screening independent prognostic factors.

		Univariate cox			Multivariate cox
Clinical characteristics	TCGA, n=168	HR	95% CI	P-value	HR	95% CI	P-value
Age, years, mean ± sd	64.89±11.24	1.029	1.008–1.049	5.83×10−3b	1.019	0.997–1.0413	9.46×10−2
Sex		0.919	0.605–1.394	6.89×10−1	Unknown	Unknown	Unknown
Male	94
Female	75
Pathologic M		1.491	0.457–4.858	5.05×10−1	Unknown	Unknown	Unknown
M0	79
M1	4
Unknown	85
Pathologic N		2.156	1.276–3.643	3.31×10−3b	2.12	1.139–3.947	1.19×10−2a
N0	45
N1	119
Unknown	4
Pathologic T		1.711	1.116–2.622	1.35×10−2a	1.221	0.576–2.585	6.03×10−1
T1	8
T2	19
T3	136
T4	4
Unknown	1
Pathologic stage		1.488	1.040 – 2.129	3.12×10−2a	0.996	0.473–2.096	9.92×10−1
I	20
II	138
III	5
IV	4
Unknown	1
Pathologic grade		1.538	1.148–2.061	3.76×10−3b	1.366	0.993–1.879	5.56×10−2
1	30
2	88
3	47
4	2
Unknown	1
History of alcohol usage		1.099	0.688–1.757	6.92×10−1	Unknown	Unknown	Unknown
Yes	96
No	56
History of alcohol usage		1.099	0.688–1.757	6.92×10−1	Unknown	Unknown	Unknown
Unknown	16
Tobacco history		1.117	0.874–1.427	3.75×10−1	Unknown	Unknown	Unknown
Never	61
Reform	56
Current	19
Unknown	32
Chronic pancreatitis history		1.101	0.524–2.314	7.99×10−1	Unknown	Unknown	Unknown
Yes	117
No	13
Unknown	38
Diabetes history		0.903	0.509–1.602	7.27×10−1	Unknown	Unknown	Unknown
Yes	36
No	99
Unknown	33
Radiotherapy		0.529	0.302–0.928	2.39×10−2a	0.502	0.279–0.906	3.78×10−2a
Yes	38
No	112
Unknown	18
Recurrence		1.558	0.951–2.552	7.59×10−2	Unknown	Unknown	Unknown
Yes	40
No	102
Unknown	26
Risk status		2.453	1.578–3.815	3.93×10−5c	2.146	1.289–3.571	3.31×10−3b
High	84
Low	84
Dead		Unknown	Unknown	Unknown	Unknown	Unknown	Unknown
Death	88
Alive	80
Overall survival time, months, mean ± sd)	17.09±15.22	Unknown	Unknown	Unknown	Unknown	Unknown	Unknown

^aP<0.05

^bP<0.01

^cP<0.001. TCGA, The Cancer Genome Atlas; HR, Hazard Ratio; CI, confidence interval; sd, standard deviation; T, tumor stage; N, node stage; M, metastasis stage.

Figure 5.

KM curves for high- and low-risk groups with different TNM stages and radiotherapy treatment status. (A) KM curves showing the correlations of pathological N0 (left) and pathological N1 (right) with risk groups. (B) KM curves showing the correlations between without radiotherapy (left) or with radiotherapy (right) and risk groups. KM, Kaplan-Meier; HR, hazard ratio; T, tumor stage; N, node stage; M, metastasis stage.

Discussion

The mechanisms underlying tumor development and progression of PAAD are complex, and influencing factors include epigenetic regulation of gene expression, epigenetic silencing of genes, oncogenic/tumor suppressor gene mutation, telomere alteration, genomic instability and DNA methylation (37–40). The present study aimed to identify potential important key methylated genes in PAAD. A total of 1,067 DMGs were identified in the less favorable prognosis and more favorable prognosis groups. From the 10 modules identified by WGCNA, the black (involving 90 DMGs) and turquoise (involving 394 DMGs) modules, in which the CpG genes were significantly associated with methylations, were selected for further analysis. For the 484 DMGs involved in the two key modules, 18 GO_BP, 7 GO_CC and 9 GO_MF terms were enriched. There were 192 DMGs associated with prognosis (less favorable or more favorable). Correlation analysis indicated that the expression levels and methylation levels of 192 DMGs among the 484 DMGs were negatively correlated. Furthermore, 50 prognosis-associated genes were further screened from the 192 DMGs. After a 12-gene optimal combination (CEBPA, HIST1H4E, STAMBPL1, PLD3, CEP55, SSBP4, GRIA1, SWAP70, ADCYAP1R1, YPEL3, HOXC4 and IGFBP1) was identified, the risk score system was constructed and validated in the TCGA, E-MTAB-5008 and E-MTAB-5571 datasets. In addition, pathological N category, radiotherapy and risk status were found to be independent prognostic factors.

It is demonstrated that upregulation of CCAAT/enhancer binding protein (C/EBP) beta (C/EBPβ), encoded by CEBPB, could restore the anti-cancer functions of Menin in PC (41) The inadequate cytoplasmic localization and abnormal silencing of C/EBP results in its dysfunction, and thus, C/EBP may serve as a novel suppressor in PC cells (42). Downregulated C/EBPα induced by lysine (K)-specific demethylase 6B (KDM6B) promotes the aggressiveness of pancreatic ductal adenocarcinoma (PDAC) cells, indicating that the KDM6B-C/EBPα axis is associated with the progression of PDAC (43,44). Histone H3 modification affects the gene expression and promoter methylation of MUC2, which may be critical for the prognostic prediction of patients with PC (45). The mRNA expression levels of histone H4 is lowered by polyamide-chlorambucil conjugate (1R-Chl) in the MIA PaCa-2 PC cell line and histone H4 genes have elevated histone acetylation in tumor cells (46). Therefore, CEBPA and HIST1H4E may be critical for the survival of PAAD patients.

STAMBPL1 affects the activation of NF-κB through mediating the stability and localization of Tax (47) and NF-κB blockade can inhibit the oncogenicity and metastasis of PC cells (48). Overexpression of CEP55 can promote PC cell aggressiveness via activation of the NF-κB pathway; therefore, CEP55 may be a prognostic factor and therapeutic target for patients with PC (49). Downregulated expression of YPEL1 in PAAD samples is associated with perineural invasion and survival prognosis, thus YPEL1 may serve a role in the malignant transformation of pancreatic tissues (50). Low IGFBP1 plasma levels have a more notable influence in non-smoking patients with PC and predicts an increased risk of PC (51). These data suggest that STAMBPL1, CEP55, YPEL3 and IGFBP1 may be associated with the prognosis of patients with PAAD.

Although PLD3, SSBP4, GRIA1, SWAP70 and ADCYAP1R1 do not have reported associations with PAAD to the best of our knowledge, their influence on other types of human cancer have been reported. Elevated expression and activity of PLD is detected in multiple types of cancer, such as gastric, colorectal, renal, stomach, lung and breast cancers (52), and PLD serves a role in mediating cell proliferation, cell transition, survival signaling and tumor progression (53). SSBP2 is a tumor suppressor gene and the disruption of SSBP2-associated pathways may be involved in the malignant transformation of various tissues (54). GRIA1 is involved in glutamate receptor signaling, which is an epigenetic marker for overall mortality rate of basal-like urothelial carcinomas (55). The oncogene SWAP70 functions in regulating actin rearrangement in basal-like bladder cancer (55) and serves a role in the transformation-associated signaling pathway (56). The promoter hypermethylation level of ADCYAP1 is associated with cervical cancer development and is considered as a promising methylation marker for the early detection of cervical cancer (57). Therefore, PLD3, SSBP4, GRIA1, SWAP70 and ADCYAP1R1 may also be associated with the prognosis of patients with PAAD.

In the present study comprehensive bioinformatics analysis of PAAD samples was performed to identify prognosis-associated genes and to construct a risk score prognostic prediction system. All findings were obtained from relatively small-sized cohorts and thus require further experimental validation. Additionally, the patient cohort samples in the three datasets may exert different clinical features, such as disease stage, historical treatment and demographics, which should be carefully compared in further studies.

In conclusion, 1,067 DMGs were identified and a 12-gene optimal combination consisting of CEBPA, HIST1H4E, STAMBPL1, PLD3, CEP55, SSBP4, GRIA1, SWAP70, ADCYAP1R1, YPEL3, HOXC4 and IGFBP1 was obtained. This 12-gene risk score prognostic prediction system may be valuable for predicting the prognosis of patients with PAAD.

Funding

No funding was received.

Availability of data and materials

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

Authors' contributions

DD conceived the study design and performed manuscript drafting, SY was responsible for data collection and analysis and JZ performed data interpretation and manuscript writing. All authors read and approved the final manuscript.

Ethical approval and consent to participate

Not applicable.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.