Gene Expression Signature–Based Prognostic Risk Score in Gastric Cancer

Gastric cancer is the second leading cause of cancer-related death in the world (1). Surgery remains the gold standard in the treatment of gastric cancer (2, 3). In the United States, however, only a small fraction of patients with gastric cancer who undergo curative resection have early-stage disease, and the prognosis for patients with more advanced stage (II or III) remains poor because of the high rate of relapse after gastrectomy (4, 5). Preoperative staging techniques, including laparoscopy and noninvasive imaging systems (i.e., endoscopic ultrasonography and positron emission tomography), have a relatively low sensitivity for discriminating patients with favorable clinical biology versus poor clinical biology (6). Moreover, the outcomes of patients considered to have a similar clinical or pathologic stage remain unpredictable, especially when patients are treated similarly. For example, not all patients with stage III tumor survive 5 years even after successful curative resection (7, 8), and the outcomes remain uneven whether or not preoperative or postoperative therapy is administered. This inherent clinical heterogeneity is most likely due to the diverse molecular profile of gastric cancer. Thus, identifying the diversity in the molecular profile of gastric cancer that governs the clinical behavior of tumors could lead to new and more effective clinical strategies. Microarray technologies have been successfully used to predict clinical outcomes and survival, as well as classify different types of cancer (9–13). Recent studies in gastric cancer have identified genes that differ according to histologic factors and age, as well as those for gastric cancer prognosis prediction (14–16). However, these studies have failed to create molecular prognostic tests that could be practical for gastric cancer patients.

In the present study, we characterized tumor transcriptome at the systems level to identify potential markers that could be used to divide patients into distinct subclasses that have not been recognized by current staging system.

Tumor specimens and clinical data from 213 gastric cancer patients undergoing gastrectomy as a primary treatment option were obtained from Yonsei University Severance Hospital, Seoul, Korea. Sixty-five surgically removed frozen gastric adenocarcinoma tissues, with 19 normal surrounding tissue samples, from gastric cancer patients were used for microarray experiments [Yonsei gastric cancer (YGC) cohort]. In addition, 6 frozen tissue samples from gastrointestinal stromal tumor (GIST) patients were included in the microarray experiments as reference for distinct tumors resided in gastric tissues. To validate gene expression patterns found by microarray analysis, quantitative reverse transcriptase PCR (qRT-PCR) experiments were done with RNA from 96 formalin-fixed, paraffin-embedded (FFPE) tissues from a separate gastric adenocarcinoma patient group from Yonsei Gangnam Severance Hospital (GSH1 cohort). For validation of risk score, FFPE tissues of independent patient group from Yonsei Gangnam Severance Hospital (GSH2 cohort, n = 52) were used for qRT-PCR. Tissue specimens used in microarray and qRT-PCR were obtained from the surgical specimens. All samples were collected after obtaining written informed consent from patients, and the study was approved by the Institutional Review Board of The University of Texas M. D. Anderson Cancer Center (Houston, TX), the Yonsei University Severance Hospital (Seoul, Korea), and Yonsei Gangnam Severance Hospital (Seoul, Korea). Clinical data also were obtained retrospectively. All of the experiments and analyses were done in the Department of Systems Biology at M. D. Anderson Cancer Center.

Total RNA was extracted from the fresh-frozen tissues by using a mirVana RNA Isolation labeling kit (Ambion, Inc.). Five hundred nanograms of total RNA was used for labeling and hybridization, according to the manufacturer's protocols (Illumina). The microarray data were normalized using the quantile normalization method in the LIMMA (Linear Models for Microarray Data) package in the R language environment (17). The expression level of each gene was transformed into a log₂ base before further analysis. Primary microarray data are available in NCBI's GEO (Gene Expression Omnibus) database (http://www.ncbi.nlm.nih.gov/geo/; microarray platform, GPL6884; microarray data, accession number GSE13861).

BRB-ArrayTools were primarily used for all statistical analysis (18, 19). Gene expression differences were considered statistically significant if the P value was less than 0.001. Cluster analysis was done with Cluster and Treeview (20). Kaplan–Meier plots and the log-rank test were used to estimate patient prognosis. Multivariate Cox proportional hazards regression analysis was used to evaluate independent prognostic factors associated with survival, and gene signature, tumor stage, and pathologic characteristics were used as covariates. A P value of less than 0.05 was considered to indicate statistical significance, and all tests were 2-tailed.

To select the candidate genes during the course of validation experiments, we used receiver operating characteristic (ROC) curves from censored relapse-free survival (RFS) data using the nearest neighbor estimation method, with a cutoff value of 36 months and under the curve (AUC) were calculated with 95% CIs (21). First ROC models were constructed by using gene expression data from microarray experiments. From 2,755 gene features identified from microarray experiments, top 27 genes with highest AUC values (14 genes) or lowest AUC values (13 genes) were selected for validation with qRT-PCR experiments in the GSH1 cohort. Using qRT-PCR–based gene expression data from GSH1 cohort, we constructed second ROC curves to further select genes with AUCs of more than 0.55 as risk genes and less than 0.45 as protective genes. Of 27 candidate genes, only 6 genes were within the range of selection cutoff.

Total RNA was extracted from the FFPE sections according to the manufacturer's instruction manual (RecoverAll Total Nucleic Acid Isolation; Ambion, Inc.). Real-time RT-PCR amplification was done using the 7900HT Fast Real-Time PCR System with a 384-well block module (Applied Biosystems). Cycling conditions were 45°C for 10 minutes and 95°C for 10 minutes, followed by 40 cycles of 95°C for 15 seconds and 60°C for 60 seconds. Relative amounts of mRNA were calculated from the threshold cycle (C_t) number using expression of cyclophilin A (PPIA) as an endogenous control. All PCR experiments were duplicated and the values averaged.

To generate a risk score with 6 genes, we adopted a previously established strategy by using the Cox regression coefficient of each gene (22, 23). The risk score of each patient was derived from sum of multiplication of reference-normalized expression level of the gene by its corresponding coefficient: Risk score = (0.097 × CTNNB1 value) + (0.141 × EXOCS3 value) + (0.148 × TOP2A value) + (−0.0898 × LBA1 value) + (−0.0985 × CCL5 value) + (−0.0618 × LZTR1 value). GSH1 patients were dichotomized into both a high-risk group and a low-risk group, using the 50th percentile (median) cutoff of the risk score as the threshold value. Both the coefficient and the threshold value derived from the GSH1 cohort were directly applied to the gene expression data from the exploration data set (YGC cohort) and an independent test sets (GSH2 patient cohorts).

We applied hierarchical clustering analysis to gene expression data from 65 human primary tumor tissue samples (YGC cohort in Table 1) and 6 GIST tissue samples. Unsupervised clustering revealed 3 distinctive subtypes with clear differences in overall gene expression patterns (Fig. 1A). Most of the gastric cancer tissues were subdivided into 1 of 2 subgroups (C1 and C2). Intriguingly, a few gastric cancer tumors (C3) were coclustered with the GIST tissues, indicating that a small percentage of gastric cancers may acquire sarcomatoid features during progression. When clinical relevance was examined, RFS (systemic) was found to differ significantly between the 2 major clusters (C1 and C2). Kaplan–Meier plots and log-rank tests indicated that C2 patients had a significantly better RFS than C1 or C3 patients (P = 0.001 by the log-rank test; Fig. 1B). When only patients with stage III tumors were considered for analysis, the differences in prognosis between C1 and C2 patients were still significant (P = 0.005 by the log-rank test; Fig. 1C), indicating that the molecular features of these tumors reflected in gene expression patterns might be strong independent predictors of clinical outcomes. Because the number of gastric cancer patients in group C3 was too small (n = 5), these patients were removed from further analysis.

Because the C1 subgroup was strongly associated with poor prognosis, we next sought to identify genes whose expression is unique to the C1 subgroup by cross-comparing gene lists from different statistical tests. We first generated 2 different gene lists by applying 2-sample t tests (P < 0.001). Gene list A represents the genes that were differentially expressed between C1 and C2. Gene list B represents the genes that were differentially expressed between C2 and normal gastric tissues (Supplementary Fig. S1). When gene expression patterns of all tissues were compared together, 3 different patterns were observed: A not B (2,755 genes), A and B (241 genes), and B not A (1,437 genes). Genes in the A not B category displayed a poor prognostic C1-specific gene expression pattern and are potential markers for predicting RFS (Fig. 2). Because the use of a complex algorithm with a long gene list from microarray data may not be practical in the clinic, we tried to identify a small number of genes whose expression patterns can still reliably predict RFS. Of 2,755 genes in A not B category, we further selected candidate genes on the basis of ROC model analysis. Top 27 genes with highest AUC values (14 genes) or lowest AUC values (13 genes) were selected for validation with qRT-PCR experiments. We next tested, using qRT-PCR, whether expression of these genes or their subsets could predict the RFS in an independent cohort (GSH1).

We again constructed another ROC models, using gene expression data from qRT-PCR experiments with tissues from GSH1 cohort, to assess the prognostic relevance of gene expression in 96 GSH1 patients. Considering AUC over 0.55 and under 0.45 to be significant, we found that the expression of 6 genes (CTNNB1, EXOSC3, TOP2A, LBA1, LZTR1, and CCL5) had a nontrivial correlation with RFS. We next tested whether we could use expression of these 6 genes as a prognostic signature in the GSH1 cohort. When hierarchical clustering analysis was applied to the 6-gene expression data, the 96 patients were divided into 2 subgroups with significantly different RFS (P = 0.017 by the log-rank test; Supplementary Fig. S2).

Because the use of hierarchical clustering analysis methods in clinical practice has proven to be difficult (24), we developed risk score methods by using the Cox regression coefficient of each gene (Supplementary Table S1; refs. 22, 23). Patients in the GSH1 cohort were dichotomized according to their risk score, and the RFS rate was significantly lower in the patient group with the high risk score (P = 0.048 by the log-rank test; Fig. 3). Gene expression data from the YGC cohort were reanalyzed with the 6-gene–based risk score. With direct application of the Cox regression coefficient from the GSH1 cohort and the 50th percentile cutoff threshold, RFS in 2 patient groups differed significantly (P = 0.04 by the log-rank test; Supplementary Fig. S3).

In the GSH1 cohort, the prognostic association between our newly identified 6-gene signature and other known clinical and pathologic risk factors for gastric cancer progression was assessed by univariate and multivariate analyses. As expectedly, in addition to stage and lymph node status, which are already well-known risk factors, the 6-gene signature was a significant risk factor for shorter RFS in univariate analysis (Table 2). Multivariate analysis that included all relevant pathologic variables revealed that the gene signature remained an independent prognostic risk factor for RFS.

We next sought to validate risk score in another independent cohort, especially in stage III patients (GSH2 cohort, n = 52; Supplementary Table S2). Expression data of 6 genes from FFPE tissues were obtained by applying qRT-PCR and used to generate risk score. When patients in the GSH2 cohort were dichotomized according to their risk score, both of RFS and overall survival rate were significantly lower in the patient group with the high risk score (P = 0.028 and P = 0.032, respectively, by the log-rank test; Fig. 4A and D).

To test whether the 6-gene–based risk score is independent of American Joint Committee on Cancer (AJCC) stage, patients in GSH2 cohort were restratified according to substage and risk score. As expected, prognosis of stage IIIB is significantly worse than that of stage IIIA (Supplementary Fig. S4A and C). When the risk score was applied to stage IIIA and IIIB separately, the risk score successfully identified a population of high-risk patients in both subgroups (Fig. 4B, C, E, and F). In fact, when all stratifications were combined together, the risk score identified gastric cancer patients in IIIA subgroup whose risk of relapse was similar to or worse than that of IIIB (Supplementary Fig. S4B and D). Because nodal stage is best known predictor of relapse, we assessed the utility of the risk score in T3 patients who differed only with N stage (T3N1 and T3N2). Nodal stage is well associated with RFS (Supplementary Fig. S5A). Within these groups, the risk score clearly identified high-risk patients (Supplementary Fig. S5B–D). Taken together, these data indicate that the risk score provides information on the risk of relapse independent of current staging systems and shows that the risk score captures biological differences among gastric cancers that are not encompassed in the current staging criteria.

We identified and validated, using a series of independent experiments and complementary data analyses (Supplementary Fig. S6), robust prognostic markers in gastric cancer and developed a prognostic risk score that can be easily translated into the clinic. First, we used microarray technology to uncover potential prognostic subgroups of gastric cancer patients and identify candidate genes for prognostic markers. The association of gene expression patterns with prognosis was significant (Fig. 1), which suggests that our gene expression signature well reflects clinical differences between subgroups of patients with gastric cancer. However, the difficulty of acquiring fresh-frozen tissues from patients and the complexity of data analysis make it hard to use this approach in the clinic. To overcome this limitation, we switched from microarray-based technology to qRT-PCR technology, simpler and easily accessible technology in clinics, to measure gene expression and identified a small number of genes (CTNNB1, EXOSC3, TOP2A, LBA1, LZTR1, and CCL5) whose expression patterns can reliably predict the prognosis of gastric cancer patients. The robustness of the prognostic gene expression signature was validated in an independent cohort by using the reduced gene set. For easy translation of our finding to the clinic, we developed a risk score for relapse after curative resection of tumor (Fig. 3). Finally, the robustness of our risk score was validated in an independent cohort, especially in stage III patients who show most heterogeneous clinical outcome (Fig. 4 and Supplementary Fig. S5).

The unique molecular characteristics of each subgroup of gastric cancer may lead to new therapeutic strategies. CTNNB1 is a key mediator of the WNT signaling pathway that regulates cell-fate decisions and cell proliferation during gut development (25). Activated mutation of CTNNB1 was reported in gastric cancer, and abnormal expression of CTNNB1 in gastric cancer was significantly associated with poorer survival (26, 27), supporting the notion that our gene expression data may well recapitulate the molecular abnormality of gastric carcinogenesis. TOP2A encodes a DNA topoisomerase II, an enzyme that controls the topologic state of DNA during transcription. This gene is currently the target of several anticancer agents, and a variety of its mutations have been associated with the development of drug resistance (28–30).

The development of the 6-gene–based risk score has strong clinical implications. We could identify gastric cancer patients at higher risk by using simple qRT-PCR technology and paraffin-embedded tissues, which are routinely acquired at diagnosis. One of limitations of our study is its retrospective character. Thus, to validate its true clinical relevance of the risk score, it will be necessary that 6-gene–based risk score is integrated into prospective randomized trials in the form of a biologic stratification criterion. In addition, because this new approach has not been applied to samples from small pretreatment biopsies, the reliability of the new approach should be extensively tested before its use in clinical trials.

Prognostic characteristics of the risk score may not be sufficient to change current clinical practice because they provide only information on probable course of the disease and not the probable response to treatments. However, biomarker study in breast cancer showed that 21-gene prognostic marker could also be used as predictive marker for standard adjuvant chemotherapy (23). Thus, in future study, it will be interesting to test whether the risk score can also be good predictive marker for response to adjuvant chemotherapy.

No potential conflicts of interest were disclosed.

This study was supported by NIH grants (CA127672, CA129906, CA138671, and CA150229); Multidisciplinary Research Program grants from The University of Texas M. D. Anderson Cancer Center; Institute of Personalized Cancer Treatment grant from The University of Texas M. D. Anderson Cancer Center faculty startup fund from The University of Texas M. D. Anderson Cancer Center; Dallas, Cantu, Smith, Park, and Myer Families; the Kevin Fund; the River Creek Foundation; and 2009 Academic faculty research fund of Yonsei University College of Medicine.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.