Seventy-nine tissue samples from patients who were surgically treated for primary gastric carcinoma were procured at the Beijing Cancer Hospital (Peking University, School of Oncology) from 1999 to 2003. No patient received chemotherapy or radiotherapy before surgery. All patients were treated with curative surgical resection, which was, in some cases, followed by second-line treatment at the time of recurrence. Macroscopic and microscopic evaluations were conducted by pathologist according to the general rules for gastric cancer. Follow-up was performed every three month for the first two years, and every three to six months thereafter. Stage of gastric cancer was classified according to 2002 tumor-node-metastasis (TNM) classification system recommended by the American Joint Committee on Cancer.

Overall survival was calculated from the date of primary surgery to the date of last follow-up or to the date of death due to cancer relapse or metastasis. All tumor samples were obtained at the surgery, followed by fresh freezing in liquid nitrogen and stored at -80°C. Informed consent was obtained from each patient for the collection and storage of tissue samples in a tissue bank for future research. This investigation was performed after approval by Ethics Committee of Peking University.

Total RNA was purified from clinical samples using TRIzol reagent (GibcoBRL, Grand Island, New York, USA). And mRNA was linearly amplified by in vitro transcription using T7 RNA polymerase (MEGAscript T7 kit, Ambion, Inc, USA). The quality and integrity of total and amplified mRNA (cRNA) was monitored by both spectrophotometry (OD UV 260/280 ratio > 1.8) and agarose gel electrophoresis.

Gene-expression profiling was performed using Illumina HumanWG-6 BeadChip, which contains 47 296 transcripts. BeadChips were scanned with a BeadStation 500 GX and data are available at Gene Expression Omnibus (GSE21983). http://www.ncbi.nlm.nih.gov/geo).

Average normalization in BeadStudio software was conducted for probe level average normalization and background correction. A detection P value was used in BeadChip to calculate probability to see a certain signal level without specific probe-target hybridization. All genes and probes with P value > 0.01 were filtered and removed from the analysis. Among 47 296 transcripts, 18 819 were expressed with P values < 0.01.

The supervised principal components method was used for survival profiling[11]. In the training set, we calculated the modified univariate Cox proportional-hazard scores for all genes (n = 18 819), which were measured to identify genes with their expression correlated to the duration of survival. We selected a set of genes whose absolute Cox score exceeded a threshold using cross-validation. For each iteration of the complete cross-validation, 10% of the cases were omitted, and principal components derived from the remaining 90% of the cases were included in a Cox model to predict the survival in 10% of the cases. By repeating the iteration process for 10 times, we found that a threshold of 2.6 yielded the highest average partial log-likelihood ratio. Principal component analysis (PCA) was then performed using 10 transcripts whose absolute Cox score equalled or exceeded the threshold for all cases in the training data set. Kaplan-Meier survival curves were then plotted to predict overall survival. All analysis and plotting were conducted using R package superpc (http://www-stat.stanford.edu/~tibs/superpc).

Based on the transcript level in the 10 transcripts and the weight assigned to each transcript from the training set, a discrete risk score (the supervised principal components risk score) was then calculated for each patient in the validation dataset.

Multivariate analysis was conducted to evaluate the prediction accuracy of our survival profile in comparison with the standard clinicopathological covariates by Cox proportional hazards regression using SPSS software.

Functional gene set enrichment analysis was performed to find the pathways associated with prolonged and poor survivals. A total of 249 sets of canonical pathways (Gene Set Enrichment Analysis-Molecular Signatures Database) were analysed to indicate their correlations with overall survival to a greater degree than expected by chance[12].

Totally, 79 gastric cancer patients treated with surgical resection were recruited in this study. Samples were randomly separated into two batches with no significant differences between the two sets with respect to age, sex and other clinicopathological features. Microarray was conducted in all samples, and the data of batch one served as the training dataset for marker discovery and data of batch two as the validation dataset. In batch one, microarray Quality Control (QC) removed 7 samples due to failure in hybridization or failure to meet the analysis criteria, resulting in a total of 39 samples included in the training set. All of 33 samples in the second batch passed QC and were used in validation phase. The characteristics of the 72 patients are summarized in Table 1. The median overall survival time of all samples was 31 mo, ranging from 4.2 to 73.6 mo, and the 5-year overall survival was 33%.

The “semi-supervised” learning approach was used to identify the gene expression profile related to the overall survival in the training dataset[12] (Figure 1). A total of 18 819 expression signals passed QC. First, we calculated the Cox scores of all 18 819 genes based on the survival times versus the expression levels obtained in 39 training observations. To choose the genes with the best prediction power, the threshold of Cox scores was calculated by 10-fold cross-validation. The expression profile of 10 transcripts whose Cox score equalled or exceeded the threshold was obtained (Table 2). Next, we performed PCA on the entire training set. For each case, a risk score that represents the sum of the weighted expression levels of the 10 prognostic transcripts was computed by supervised component analysis in a regression model. As shown in the Kaplan-Meier survival curves in Figure 2A and B, the patients were categorized into two groups based on their scores above or below the median risk of death. The low-risk group (n = 20) had a median survival of 42.1 mo, whereas the high-risk group (n = 19) had a median survival of only 26.5 mo. The correlation of the risk score and the survival status (P = 0.000047, log-rank) indicates that this transcriptional pattern was associated with the patient outcomes. A similar classification was also seen by TNM staging (P = 0.00016, log-rank; Figure 3A).

Open in New Tab   Full Size Figure   Download Figure

Figure 1 Overview of the strategy used for the development and validation of prognostic markers.

Open in New Tab   Full Size Figure   Download Figure

Figure 2 Overall survival curves and the expression profile of the ten-gene prognostic marker in the training dataset. A: The gene expression pattern of the ten-gene prognostic marker. Nine genes were associated with the prolonged survival and one gene with poor survival. Red, high expression; green, low expression; B: Kaplan-Meier survival curves based on the expression profile of the ten-gene prognostic marker; C: Overall survival curves according to the tumor-node-metastasis stages.

Open in New Tab   Full Size Figure   Download Figure

Figure 3 Overall survival curves and the expression profile of the ten-gene prognostic marker in the validation dataset. A: The gene expression pattern of the ten-gene prognostic marker. Red, high expression; green, low expression; B: Kaplan-Meier survival curves based on the expression profile of the ten-gene prognostic marker; C: Overall survival curves according to the tumor-node-metastasis stages.

Among the ten-gene prognostic markers, high expression levels of 9 genes were associated with poor survival (Figure 2B). SEC61G, a subunit of the heteromeric SEC61 complex, was the only gene with its high expression associated with prolonged survival. Interestingly, six out of the 10 genes in this profile are either identified or predicted ribosomal proteins, including RPLP2, RPS12, RPS8, RPS19, RPS2P12, and RPS15P4. The involvement of numerous ribosomal genes with survival times suggested either their regulation by tumor suppressor or oncogenes, or their direct participation in certain pathways other than protein synthesis. Furthermore, RPS8 and RPS12 were previously reported as cancer related markers in colorectal tumor and cervical squamous cell carcinoma[13,14]. EIF3S6, a member of eukaryotic translation initiation factor 3, was identified as a prognostic factor in Stage I non-small cell lung cancers[15]. To test if the survival categories relate to known pathways, we applied gene set enrichment analysis (GSEA) to microarray data of 18 819 transcripts using 249 canonical gene sets collected by MsigDB. Glucocorticoid receptor (GCR) pathway gene set, referring to glucocorticoid receptor-related inhibition of inflammatory response, was significantly associated with the overall survival (FDR = 0.15, P = 0.004).

Next we evaluated the ten-gene prognostic marker in an independent dataset containing 33 cancer samples. As shown in Figure 3A and B, based on the expression of these 10 genes, the patients were classified into either a “low-risk group” or “high-risk group” with significantly different survival times (P = 0.0009, log-rank). The low-risk group patients had a median survival of 31.7 mo whereas the high-risk group one had a median survival of 21.4 mo. The expression patterns of these 10 genes in validation set were also similar to the observation in the data training set (Figure 3B). The results in validation dataset showed the consistency of our ten-gene prognostic marker in survival prediction. For pathway analysis, no significant association was observed by GSEA.

As certain clinicopathological parameters, especially TNM staging, have been used as prognosis indicators, we also compared our ten-gene prognostic marker with the clinicopathological characteristics in the validation dataset in order to assess the impact of clinicopathological factors on overall survival. First we examined the distribution of prognostic factors as a function of risk assignment based on our ten-gene prognostic marker (Table 3). Certain variation such as age was seen between high- and low-risk groups, while gender, tumor location and differentiation grade, the depth of wall invasion, and metastasis showed no significant difference between the two groups except for the TNM staging (P = 0.0023).

Since the TNM staging is used most widely in clinical prognosis prediction, we compared the predictive power of our survival markers with the TNM staging. All the TNM stage II patients were categorized into the low-risk group and the entire high-risk group patients were in stage III or IV. A consistency between TNM staging and the staging was found by the ten-gene marker. However, two IIIb patients of the low-risk group had a survival of 63.7 and 22.5 mo at last follow-up, respectively. By Kaplan-Meier survival plots and log-rank tests, we assessed the patient survival status predicted by our prognosis candidates and TNM staging. Relatively more accurate predictions were shown by the ten-gene prognostic marker in both sample groups (TNM, P = 0.00016 and P = 0.0023; survival markers P = 0.000047 and P = 0.0009, Figure 2B and C, Figure 3B and C). As a result, in multivariate analysis, our ten-gene marker was the only independent indicator in prognosis prediction with statistical significance (P = 0.007; Hazard ratio 0.13; 95% CI: 0.29-0.56, Table 4).

Gastric cancer is the second leading cause of cancer related death in China and worldwide. The 5-year survival rate of patients with advanced disease is very poor. Currently, treatment plan and prognosis prediction for gastric cancer mainly depend on the clinicopathological staging. However, prognosis varies among patients with the same clinicalpathological stage. An individualized expression test for selected markers in biopsy and surgical samples will complement the current staging system, especially for prognosis prediction.

The gene expression profiling has enabled researchers to quantify the biological states and consequently to uncover the subtle phenotypes in cancer. Such analyses have provided unique opportunities to develop various profiles that can distinguish, identify, and classify discrete subsets of disease, predict the disease outcome, and even predict the response to therapy.

In this study, based on the whole genome expression profiling, the authors identified and validated a ten-gene set that can be further developed as clinical prognosis markers to predict overall survival of gastric cancer patients. This marker set showed consistent prognosis results with the traditional Tumor-node-metastasis (TNM) staging system. The findings in this study also provided new clues about the possible association between the deregulation of ribosome related gene expression and survival status of the patients after surgery.

Based on the whole genome expression profiling, a ten-gene prognostic marker set for overall survival prognosis of gastric cancer patients may be applied in combination with the TNM staging system as a parallel and complementary approach. However, the predominance of ribosome protein genes in these molecular prognostic markers awaits for further research on their roles in cancer progression.

TNM: The TNM system is one of the most widely used staging systems in tumor classification. The system is based on the extent of the tumor (T), the extent of spread to the lymph nodes (N), and the presence of distant metastasis (M). A number is added to each letter to indicate the size or extent of the primary tumor and the extent of cancer spread. Principal component analysis (PCA): A mathematical tool used to reduce the number of variables while retaining the original variability of the data. The first principal component accounts for as much of the variability in the data as possible, and each succeeding component accounts for as much of the remaining variability as possible. Gene set enrichment analysis (GSEA): A computational method that determines whether a prior identified set of genes shows statistically significant, concordant differences between two biological states. It is a method which focuses on the analysis at the level of functional related gene sets instead of a single gene. It helps biologists to interpret the DNA microarray data by their previous biological knowledge of the genes in a gene set. GSEA has been shown to efficiently identify gene sets containing known disease-related genes in the real experiments.

A ten-gene prognostic marker, including 6 ribosomal proteins, for overall survival prognosis of gastric cancer were identified and validated based on whole genome expression profiling. By comparing with the traditional TNM staging system, this ten-gene prognostic marker showed consistent prognosis results, which may complement the predictive value of current TNM staging system.