A predictive model for assessing prognostic risks in gastric cancer patients using gene expression and methylation data

The role(s) of epigenetic reprogramming in gastric cancer (GC) remain obscure. This study was designed to identify methylated gene markers with prognostic potential for GC. Five datasets containing gene expression and methylation profiles from GC samples were collected from the GEO database, and subjected to meta-analysis. All five datasets were subjected to quality control and then differentially expressed genes (DEGs) and differentially expressed methylation genes (DEMGs) were selected using MetaDE. Correlations between gene expression and methylation status were analysed using Pearson coefficient correlation. Then, enrichment analyses were conducted to identify signature genes that were significantly different at both the gene expression and methylation levels. Cox regression analyses were performed to identify clinical factors and these were combined with the signature genes to create a prognosis-related predictive model. This model was then evaluated for predictive accuracy and then validated using a validation dataset. This study identified 1565 DEGs and 3754 DEMGs in total. Of these, 369 were differentially expressed at both the gene and methylation levels. We identified 12 signature genes including VEGFC, FBP1, NR3C1, NFE2L2, and DFNA5 which were combined with the clinical data to produce a novel prognostic model for GC. This model could effectively split GC patients into two groups, high- and low-risk with these observations being confirmed in the validation dataset. The differential methylation of the 12 signature genes, including VEGFC, FBP1, NR3C1, NFE2L2, and DFNA5, identified in this study may help to produce a functional predictive model for evaluating GC prognosis in clinical samples.

Many genetic and epigenetic events have been linked to carcinogenesis. The major hallmarks of any epigenetic events include alterations at the promoter CpG sites within the gene or changes to the histone at the chromatin level, and the most widely studied and best characterised epigenetic events include differential methylation of tumour suppressors and oncogenes [4].
In GC, multiple epigenetic modifications have been linked to disease progression, and these alterations may contribute to the identification of biomarkers for early diagnosis [4]. Using epigenome wide and gene-specific DNA methylation analyses, a DNA methylation biomarker panel, which includes IRF4, ELMO1, CLIP4, and MSC, has been found to link GC and gastritis, and this panel has been shown to be useful in endoscopic biopsies allowing for the earlier detection of GC in these samples [5]. In a study of the Chinese population COX-2 methylation levels were shown to be reduced in the anti-Helicobacter pylori intervention group, compared with the placebo group, indicating that this differential methylation might be a useful indicator of chemoprevention efficacy for GC [6]. The aberrant methylation of the tumour suppressor gene, FAT4, in peripheral blood leukocytes has been linked to increased GC risk [7]. While decreased expression of HOXB13, caused by methylation of its promoter, is a reliable marker for poor prognosis in GC [8]. Increased methylation of the LINE1 and IGF2 differentially methylated regions (DMRs) correlates with more aggressive GC phenotypes and thus are considered potential biomarkers for GC progression [9].
Despite these encouraging findings, the underlying gene methylation mechanisms used in GC remain obscure. In addition, evaluation of methylation in samples collected in previous studies remains relatively rare. Therefore, we searched GC-related gene and methylation expression profiles found in public databases, and combined these data using meta-analysis, to enlarge the sample size and enhance statistical power. Using a series of bioinformatics tools and survival analysis, we could reveal the association between gene methylation and GC prognosis identifying several novel prognostic biomarkers.

Data resource and sample classification Dataset for meta-analysis
The gene expression and methylation profile datasets were selected from the GEO database (http://www.ncbi. nlm.nih.gov/geo/), using the keywords "gastric cancer" and "Homo sapiens". The inclusion criteria for the datasets were as follows: (1) the dataset must include both gastric cancer tumour samples and normal tissue samples; and (2) have at least 50 samples in the dataset.
Five eligible gene expression and two methylation profile datasets were identified and downloaded. GSE26942 was from the GPL6947 platform and consisted of 205 gastric tumour tissue samples (GC samples) and 12 Table 1. Detailed clinical information for each of the samples in these datasets was collected from the data derived in the GEO database.

Predictive modelling dataset
Relative gastric cancer gene expression and methylation profiles were downloaded from The Cancer Genome Atlas (TCGA, https ://gdc-porta l.nci.nih.gov/) database. Then, the gene expression and gene methylation profiles were matched. This created a single dataset containing a total of 398 matched tumour samples. Of these, 360 samples had complete prognostic information. These data were then used as the training dataset for the predictive model for prognosis developed in our study. Another set of gastric cancer-related gene expression profiles, GSE62254, was then downloaded from the GEO database (https ://www.ncbi.nlm.nih.gov/geo/query /acc. cgi?acc=GSE62 254). This dataset was from the GPL570 [HG-U133_Plus_2] Affymetrix Human Genome U133 Plus 2.0 Array platform and included the data from.300 gastric cancer tumour tissue samples, and was used as an independent validation dataset. Patients' clinical characteristics are listed in Table 2. The mean age of the patients was 64.9 years in the TCGA dataset and 61.9 years in the validation dataset. The majority of patients in the TGCA dataset had stage II and III disease while in the validation dataset had stage II to IV disease.

Data normalisation and consistency selection Data used for meta-analysis
Three of the five datasets used for meta-analysis, GSE29727, GSE54129, and GSE64951 were from the Affymetrix platform. The raw data from these three datasets was downloaded in the CEL format and were then transformed into gene symbols, their missing values were filled in using the median method, and then subjected to background correction using the minimal sets algorithm method and normalised using the quantiles method [10]. All of these methods were included in the oligo package from R (version 3.4.1, http://www.bioco nduct or.org/ packa ges/relea se/bioc/html/oligo .html). The other two datasets, GSE26942 and GSE65801, were from the Illumina and Agilent platforms, respectively. These datasets were downloaded in TXT format and gene annotation was performed using the probe information provided by the platform. This data was then subjected to a logarithmic transformation and normalised using the quantiles method. These steps were performed using the limma package from R (version 3.4.1, https ://bioco nduct or.org/ packa ges/relea se/bioc/html/limma .html).
In the case of the gene methylation datasets, GSE25869 and GSE30601, the corresponding chromosome locations and methylated beta values were evaluated and assigned using the Methylation Module in GenomeStudio [PMID: 22498030] [11].

Quality control and consistency selection
Given the fact that these datasets were all derived from different platforms, meta-analysis was used to combine consistent data from across these datasets into a single larger dataset generating better statistical power and improving the reliability of the results. To eliminate potential bias, produced by differences in the platforms used to generate this data, all of the datasets were subjected to quality control using the criteria established in the MetaQC package from R (version 3.4.1, https ://cran.r-proje ct.org/web/packa ges/MetaQ C/index .html). A total of five parameters, internal quality control (IQC), external quality control (EQC), accuracy quality control (AQC), consistency quality control (CQC), and standardised mean rank score (SMR), were calculated and evaluated. Then the reliable datasets were further analysed using the MetaDE.ES package (https ://  cran.r-proje ct.org/web/packa ges/MetaD E) which selected the differentially expressed genes (DEGs) and differentially expressed methylation genes (DEMGs) when comparing GC and normal tissue samples. In brief, we performed a heterogeneity test for the expression of each gene on different platforms using tau 2 , Q value and Q pval as the measures. Then, we performed a heterogeneity test on the differential expression patterns for a gene in the integrated dataset across the different sample groups. Using this analysis we were able to determine the false discovery rate (FDR) which was validated via multiple-testing correction and we identified a value of < 0.05 as the significance threshold value for DEGs and DEMGs between different sample groups [12]. To ensure each signature gene exhibited consistent expression across different datasets, the thresholds for the homogeneity test were set as tau 2 = 0 and Q pval > 0.05.

Correlation analysis between gene expression and methylation
The datasets containing DEGs and DEMGs were selected and compared. Overlapping genes, those that were both differentially expressed and exhibited altered methylation levels, were identified using the cor function in R (version 3.4.1, http://127.0.0.1:19124 /libra ry/stats /html/cor. html). The Pearson coefficient (CC) for gene expression and methylation levels was calculated, and the genes with significant associations with the methylation data were selected as candidate signature genes. These signature genes were then subjected to gene oncology (GO) biological functional enrichment and Kyoto Encyclopedia of Genes and Genomes (KEGG, https ://www.kegg.jp/) pathway enrichment analyses, using the Database for Annotation, Visualization and Integrated Discovery (DAVID, version 6.8, https ://david .ncifc rf.gov/) by a hypergeometric distribution [13].

Screening genetic prognostic biomarkers and clinical factors
Combining the identified tumour signature genes with the corresponding clinical factor information, we evaluated their correlation using univariate and multivariate cox regression analyses in the R survival package (version 3.4.1, http://bioco nduct or.org/packa ges/survi valr/). The threshold for significance was P < 0.05 when subjected to a logrank test.

Construction and validation of a predictive model of prognostic risk Construction and validation of a tumour signature gene-based predictive model
Based on the prognostic information identified in the previous step, we produced a tumour signature gene-based predictive model which was then used to calculate the prognosis index (PI) value for each sample. The median PI score was set as the cut-off for classifying samples as high-or low-risk in the training dataset. Then, Kaplan-Meier (KM) survival curves were constructed using the survival package from R (version 3.4.1, http:// bioco nduct or.org/packa ges/survi val/) and used to evaluate the correlations between the predictive model and clinical outcome [14]. Meanwhile, these correlations were validated using the validation dataset. The area under the receiver operating characteristic (ROC) curve (AUROC) was used to determine the predictive accuracy of this model for both the training and validation datasets. The closer the AUROC value to 1.0 the higher the accuracy of the predictive model.

Construction of a prognostic clinical factor-based predictive model
We used a cox regression analysis to use the prognostic clinical factor information from each dataset to generate a clinical factor-based predictive model. In this model, the PI for each sample was calculated and the median value was set as the cut-off for the high-and low-risk groups from the training dataset. Likewise, the KM survival curve was used to assess the relationships between the predictive model and clinical prognosis. These results were also assessed using the validation dataset.

Construction of the predictive model integrating signature genes and clinical factors
We created a novel integrated prognostic model for GC by combining the prognostic results from the signature gene-based model with those from the clinical factorbased model. A new PI value was then calculated for each and the samples in the training dataset were divided into high-and low-risk groups using these new median values. KM survival curves were then used to evaluate the predictive value of these PI values and the results were evaluated in the validation dataset.

Selection of DEGs and DEMGs
After data normalisation, quality control of the datasets was conducted, and values for IQC, EQC, AQC, CQC, and SMR were calculated. The quality results indicated these datasets were all eligible for meta-analysis. Finally, we identified 1565 DEGs and 3754 DEMGs with a high degree of consistency between GC and normal samples, using MetaDE. The evaluation and identification process is described in Fig. 1.

Correlation between gene expression and methylation
The DEGs and DEMGs were compared and matched, identifying 396 overlapping sequences that were differentially expressed at both the gene expression and methylation levels. We analysed the correlations between these values for each of the 396 genes identified from TCGA and GSE30601-GSE15460 (methylation profile with the matched gene profile) datasets. These evaluations indicated that overall gene expression was negatively associated with overall methylation in both TCGA (CC = − 0.5145, P = 0.004) and GSE30601-GSE15460 (CC = − 0.72704, P < 0.001) datasets. Given this, we then calculated the correlation values for gene expression and methylation for each gene, and genes with negative relationships, that is those genes with high degrees of methylation and low expression or vice versa, were retained. This evaluation narrowed our gene cohort to roughly 274 genes which were then evaluated as tumour signature genes.
Using GO functional and KEGG pathway enrichment analyses these 274 genes were found to be significantly enriched in 14 biological process categories including steroid metabolic process (NR3C1), fructose metabolic process (FBP1), regulation of cell migration (VEGFC), vitamin metabolic process (ACADM); and five pathway terms such as glycolysis/Gluconeogenesis (FBP1) ( Table 3).

Prognostic gene biomarkers and clinical factors
These 274 genes were then subjected to univariate and multivariate cox regression analyses to identify the prognostic genes and clinical factors. We finally selected 12 genes (SLC5A5, SLC7A6, NFE2L2, DFNA5,  VEGFC, MUM1, TRIB2, MCOLN1, FBP1, ACADM, WDR37, and NR3C1) that demonstrated a significant correlation with clinical prognosis (Table 4), and five independent clinical factors (age, pathologic_N, pathologic_T, targeted molecular therapy, and new tumour) for our predictive models ( Table 5). The KM survival curves for each are shown in Fig. 2.
We developed a clustering heatmap showing the expression and methylation status of all 12 of our tumour signature genes and then combined these with their corresponding prognostic clinical factors (Fig. 3). This analysis revealed that four clinical factors, age, pathologic_N, targeted molecular therapy, and new tumour, were significantly associated with the prognosis of patients in the two clusters (P < 0.01).

Construction and validation of an integrated prognostic risk prediction model The signature gene-based risk predictive model
After obtaining the prognostic gene information using the cox regression algorithm, we constructed a prognostic risk prediction model using our 12 signature genes. The samples in the training dataset were classified as high-or low-risk with their cut-off set to the median PI values.
In the training dataset, the survival analysis indicated that patients in the low-risk group had a significantly longer median overall survival (OS) (22.1 m vs. 15.1 m, P < 0.001, Fig. 4a) and median disease free survival (DFS) (22.1 m vs. 14.5 m, P < 0.001, Fig. 4b), than those patients in the high-risk group. The AUROC for the OS and DFS curves were 0.997 and 0.906, respectively (Fig. 4e), suggesting that both had a high predictive accuracy.
Similarly, in the validation dataset GSE62254, patients in the low-risk group had better survival rates compared with those in the high-risk group as evidenced by a prolonged median OS (55.  AUROCs for these two outcomes were 0.995 and 0.912, respectively (Fig. 4e), indicating that this model created values with high predictive accuracy.

The clinical factor-based prognostic risk model
Five clinical factors were identified using a cox regression algorithm, and based on the weight of their coefficients, samples containing all five factors were selected   those of the high-risk group. The AUROC was determined to be 0.923 and 0.921, respectively (Fig. 5e).
In the GSE62254 validation dataset, only three clinical factors, age, pathologic_N, and pathologic_T, were available. Thus, we adapted the three clinical factorbased predictive model to build a clinical factor-based model, to validate the results derived in TCGA training dataset. We again showed that patients in the lowrisk group had significantly prolonged OS (58.9 m vs. 37.2 m, P < 0.0001, Fig. 5c) and DFS (44.4 m vs. 24.3 m, P < 0.0001, Fig. 5d) compared to patients in the high-risk group. The AUROC values for OS and DFS were 0.897 and 0.882, respectively (Fig. 5e).

Building an integrated predictive model incorporating both clinical factors and signature gene expression
The integrated predictive model was constructed by combining the weight coefficient from the 12 signature genes and five clinical factors. Then, the PI of each sample was re-calculated, and then reassigned as low-or high-risk. The patients in TGCA dataset low-risk group had a significantly prolonged OS (20.3 m vs. 15.8 m, P < 0.001, Fig. 6a) and DFS (19.7 m vs. 14.6 m, P = 0.003, Fig. 6b) compared with those in the high-risk group. The AUROCs of the two outcomes were 0.985 and 0.939, respectively (Fig. 6e).
To evaluate the GSE62254 validation dataset, the clinical factors were reduced from five to three based on data availability and new PI values were calculated. Patients were then categorised as high or low risk and then their survival was evaluated. Patients in the low-risk group were shown to have significantly longer survival rates than patients in the high-risk group (OS: 62.3 m vs. 38.0 m, P < 0.0001, Fig. 6c; DFS: 47.9 m vs. 24.6 m, P < 0.0001, Fig. 6d). The AUROC values for OS and DFS were 0.942 and 0.921, respectively (Fig. 6e).

Discussion
Here, we used a series of bioinformatics analyses to construct a predictive model for evaluating the prognosis of GC samples using 12 signature genes and five clinical factors. All 12 signature genes were also differentially methylated and could be used to split patients into high and low risk groups. These models were all validated using a validation set.
When we evaluated our 12 signature genes, four were identified as particularly interesting, vascular endothelial growth factor c (VEGFC), nuclear receptor subfamily 3 group c member 1 (NR3C1), nuclear factor, erythroid 2 like 2 (NFE2L2), and fructose-1,6-bisphosphatase-1 (FBP1). VEGFC has been reported to be a key regulator in GC progression and its encoded protein facilitates angiogenesis and endothelial cell growth. In addition, oxidised low-density lipoprotein (oxLDL) is a risk factor in the pathogenesis of cancers linked to its roles in abnormal lipid metabolism, and has been shown to promote lymphatic metastasis of GC via the up-regulated expression and secretion of VEGFC [15]. microRNA (miR)-27b acts as a potential tumour suppressor in GC and targets VEGFC expression [16], while miR-101 promotes cisplatin (DPP)-induced apoptosis partly via its targeting of VEGFC in DDP-resistant GC cells [17]. In addition, VEGFC expression is associated with the GC prognosis, as survival is significantly poorer in VEGFC-positive GC patients, when compared to VEGFC-negative patients [18]. Moreover, decreased VEGFC was shown to correlate with an increased risk of tumour progression [19]. Here, we identified VEGFC as one of the 12 signature genes for evaluating GC prognosis and this gene was enriched in the 'regulation of cell migration' functional category which suggests that VEGFC methylation may be related to GC prognosis via its regulation of cell migration. However, this regulatory relationship needs to be further validated in vitro and across large populations.
FBP1 protein is a gluconeogenesis regulatory enzyme associated with metabolic acidosis. Snail is an important mediator in cancer and has been shown to be increased in GC inducing the glucose metabolism via the downregulated expression of FBP1 [20] indirectly regulating the epithelial-mesenchymal transition (EMT). Decreased FBP1 serves as a positive factor in the metastasis of GC and is an indicator of poor prognosis in patients [21]. In GC cell lines, FBP1 is downregulated and its promoter is hypermethylated, resulting in increased carcinogenesis. Moreover, the methylation of FBP1 at its promoter has been independently associated with GC prognosis [22]. This was consistent with our findings that FBP1 was identified as one of the 12 signature genes having some predictive value for GC prognosis. Additionally, this gene was enriched in glycometabolism-related functions and pathways. When taken collectively these data indicated that DNA methylation of FBP1 may be associated with GC prognosis via the differential regulation of the glycometabolism.
The NR3C1 gene encodes a glucocorticoid receptor. NR3C1 is important in the carcinogenesis of GC and has been used as a marker to identify primary GC [23,24]. The high degree of methylation within the NR3C1 promoter was also implicated in the initiation of GC progression, and four SNPs at this locus have been shown to be strongly associated with increased risk for GC in a Chinese population [20]. Here, we confirmed the link between NR3C1 methylation and GC prognosis, and suggest that NR3C1 methylation may be a reliable prognostic indicator for GC.
NFE2L2, also known as NRF2, encodes a transcription factor (TF) known to participate in GC development, and its overexpression is a predictive marker for the prognosis and 5-FU resistance in GC [25]. GC patients positive for NRF2 expression are known to exhibit significantly poorer OS rates when compared to NRF2-negative patients [26]. Deafness associated tumour suppressor (DFNA5) is inactivated in GC via methylation, and this methylation is found in half of all patients with primary GC [27]. Here, we propose that there is a relationship between the methylation status of NFE2L2 and NRF2 and the prognosis of GC.
Finally, our analysis suggests that the predictive models produced in this study were relatively precise probably as a result of the increased sample size resulting from our meta-analysis. In addition, to relatively high AUROC values, our predictive models provided reliable results in our validation datasets. These encouraging results shed lights on potential regulatory mechanisms on methylation genes in GC prognosis. In addition, this pilot bioinformatics analysis will lay the foundation of exploratory biomarker analysis, which could facilitate to the prediction or indication of patients with a low risk of death and a good survival outcome. Importantly, by the identification of these sensitive methylation gene markers and the methylation patterns, we might have a deeper understanding on this malignancy progression and might develop novel targeted therapies, which could improve the survival outcomes of the patients with GC. However, several limitations remain. The expression and methylation of these signature genes should be validated in vitro and in vivo with substantial cell lines and animal samples. Moreover, perspective studies are warranted using larger clinical cohorts to validate the prognostic values of these genes before being adopted in diagnostic and prognostic settings, and we will perform these studies in future.

Conclusion
In conclusion, methylation of 12 signature genes, including VEGFC, FBP1, NR3C1, NFE2L2, and DFNA5, may be associated with the prognosis of GC, and these genesbased risk models may be a useful tool in predicting prognostic outcomes for patients at earlier stages of disease. However, these results require validation in larger patient cohorts before they can be confidently applied in a clinical setting.