Development and Validation of a Six-Gene Recurrence Risk Score Assay for Gastric Cancer

Gastric cancer is the third leading cause of cancer-related death worldwide (1), and surgery is the mainstay of treatment (2, 3). However, surgery alone benefits primarily patients with relatively early-stage disease. Multimodality therapies, consisting of surgery with adjuvant chemotherapy, radiotherapy, or both, have been established to improve the survival rates of patients (4–7).

Two recent phase III trials showed that more than 40% of patients with stage II or III gastric cancer experience a recurrence within 5 years after a curative D2 gastrectomy without adjuvant chemotherapy (6, 7). When chemotherapy was administered in conjunction with a D2 gastrectomy, the absolute increase in 3-year recurrence-free survival (RFS) or disease-free survival was about 13% to 15% (6, 7). Given the very high clinical burden of gastric cancer worldwide, tools are clearly needed for predicting outcomes in patients after gastrectomy. Although conventional approaches using clinicopathologic characteristics have been highly useful for prognostication, gene/genomics–based approaches will very likely give us better prognostication and prediction of response to treatment (8–11). Understanding of the biological characteristics associated with the inherent heterogeneity of gastric cancer and identification of molecular markers reflecting that heterogeneity would significantly improve patient care.

One of the promising candidate markers of cancer recurrence might be YAP1. When Hippo signaling, a tumor suppressor pathway that is well conserved among species (12, 13), is reduced or absent, YAP1 enters the cell nucleus and increases transcriptional activation of genes involved in proliferation and survival of cancer cells (13, 14). The fact that a high expression level and amplification of YAP1 have been observed in many cancers, including gastric cancer (15–21), supports the idea that YAP1 is an oncogene.

In the current study, we found that activated YAP1 is significantly associated with poor prognosis in gastric cancer. We used a systematic multistep strategy to develop and validate a robust recurrence risk score (RRS) assay that, by reflecting underlying biological differences among tumors, can help identify patients at a high risk of disease recurrence after surgery.

Tumor tissues for the training cohort of gastric cancer patients (n = 267) had been obtained surgically between 1999 and 2006 at the Korea University Guro Hospital (Seoul, Korea), Kosin University College of Medicine (Busan, Korea), and Yonsei University Severance Hospital (Seoul, Korea). All patients underwent a D2 gastrectomy. Of the 267 patients in training cohort, 155 had received adjuvant chemotherapy [either single-agent 5-fluorouracil (5-FU) or a combination of 5-FU and cisplatin/oxaliplatin, doxorubicin, or paclitaxel]. Tumor tissues for the validation cohort (n = 317) had been obtained surgically between 2000 and 2010 at the Seoul National University Bundang Hospital (Seongnam, Korea) and Yonsei University Severance Hospital (Seoul, Korea; the validation cohort did not overlap with the training cohort). All patients in validation cohort also underwent a D2 gastrectomy and received adjuvant chemotherapy [either fluoropyrimidine alone (5-FU or S-1) or a combination of 5-FU and cisplatin/oxaliplatin, or doxorubicin]. Because any pretreatments can have a significant effect on genome characteristics of tumors, pretreated tissues were not included in training and validation cohorts. Patients were followed up once every 3 to 6 months for first 3 years after surgery and once every 6 months after 3 years of surgery. Informed consent for sample collection had been obtained from all patients. The study protocol for retrospective analysis of tumor samples and clinical information was approved by the Institutional Review Boards of The University of Texas MD Anderson Cancer Center (Houston, TX) and the five Korean hospitals where the tumor tissues had been obtained. Overall survival (OS) was defined as the time from surgery to death, and RFS was defined as the time from surgery to the first confirmed recurrence. Data were censored when a patient was alive without recurrence at last contact. Samples from the training cohort were fresh-frozen tissues; samples from the validation cohort were formalin-fixed, paraffin-embedded (FFPE) tissues. Gene expression data from the 267 patients in the training cohort and gastric cancer cell lines MKN45 were generated as described in Supplementary Methods. Gene expression data from cell line are available from Gene Expression Omnibus database (GSE41387). Cell line authentication was performed in 2011 using STR analysis conducted by Yonsei University Severance Hospital DNA analysis facility. All of experiments and analyses were carried out at MD Anderson Cancer Center.

BRB-ArrayTools software v4.3 was used for all statistical analyses (22). Gene expression differences were considered statistically significant if the P value was less than 0.005. Cluster analysis was performed with Cluster and TreeView v3.0 (23). Prediction of a patient's class was made as described previously (24–28) and in the Supplementary Methods. Multivariate Cox proportional hazards regression analysis was used to evaluate independent prognostic factors associated with survival; the RRS (defined below), tumor stage, and other clinicopathologic characteristics were included as covariates. A P < 0.05 was considered statistically significant. All statistical analyses were conducted in the R language environment (http://www.r-project.org).

All expression analyses were carried out as described in the Supplementary Methods.

A multistep strategy was used to develop the RRS from the reference-normalized expression measurements. Six cancer genes were selected and grouped into three categories based on correlation between the results of microarray and qRT-PCR experiments and nonoverlapping biological characteristics as described in Results: stress response (GADD45B), cellular signaling (IGFBP4 and SFRP4), and interaction with the microenvironment (SPOCK1, SULF1, and THBS). To calculate the RSS, first, the expression level of each gene was multiplied by the Cox coefficient value derived from RFS data of the training cohort, generating a GENE_cox value. Second, category scores were calculated by multiplying the total GENE_cox value for each category by a contribution factor as follows: stress response score = GADD45B_cox × 0.4, cell signaling score = (IGFBP4_cox + SFRP4_cox) × 0.2, and microenvironment score = (SPOCK4_cox + SULF4_cox + THBS4_cox) × 0.1. The contribution factors were set so that the three categories had approximately equal contributions to the RRS. Third, to generate a dynamic score range, each category score was transformed by applying a natural exponential function (e^{category score}), and RRS_raw was generated by summation of the transformed scores. Fourth, to generate a range of 0 to 100, the RRS_raw was rescaled as follows: RRS = (RRS_raw − 10) × 10, with RRS = 0 if RRS < 0 and RRS = 100 if RRS > 100. RRSs were categorized as follows: <20, low risk; 20 to 40, intermediate risk; >40, high risk as indicated in Supplementary Methods.

As YAP1 is frequently activated in many cancers, including gastric cancer (29–32), we first examined the clinical relevance of YAP1 activation in gastric cancer by generating gene expression data from gastric cancer cells overexpressing YAP1 (Supplementary Fig. S1). Further analysis of gene expression data identified 88 genes whose expression was significantly associated with YAP1 activation (Supplementary Fig. S2A). We next tested the clinical significance of YAP1 activation in human gastric cancer by comparing the YAP1 signature with gene expression data from gastric cancer tissue samples. Gene expression data from the 267 samples in the training cohort (Table 1) were used for this analysis. A support vector machine (SVM) algorithm (33) was applied to calculate the probability of YAP1 activation in each tumor sample (Fig. 1A). Briefly, YAP1 signature expression data from cell lines were used to generate an SVM classifier for estimating the likelihood that a particular gastric tumor belonged to the subgroup in which the YAP1 signature was present (active YAP1 subgroup) or the subgroup in which the signature was absent (inactive YAP1 subgroup). When the tumors were thus categorized, a third of the patients (n = 88) were predicted to be classified into the active YAP1 gastric cancer subgroup. Kaplan–Meier plots for the training cohort showed significant differences in RFS and OS (P < 0.001) between the two subgroups (Fig. 1B and C), strongly indicating that YAP1 may play an important role in the clinical course of gastric cancer. Further analysis showed that YAP1 signature is independent of stages (Supplementary Fig. S3).

Because all gene expression data were generated from fresh-frozen tissues, which are not easily available in the clinic, and because the prediction model used a complex algorithm with long gene list from microarray data, which is not practical in the clinic, we sought to identify a small number of genes predictive of gastric cancer recurrence by using easily available FFPE tissues. We used a two-step strategy to identify candidate genes for development of an easy-to-use prognostic scoring system (Fig. 2). First, we identified candidate genes whose expression was significantly associated with RFS, OS, and activation of YAP1 in the training cohort; of the 824, 793, and 652 genes identified, respectively, 156 were associated with all three variables (Supplementary Fig. S2B). Second, we identified biologically nonredundant genes by an informatics approach. Because many of the 156 candidate genes could be involved in similar biological processes, we tried to identify a few genes that would represent diverse biological process. To avoid overrepresentation of particular biological processes or characteristics in the final gene set, we mapped these genes to biological processes in the Gene Ontology database (34) and selected top-ranked genes for each biological process (Supplementary Table S1). Of the 156 genes, 40 unique genes with minimum overlap in biological functions were selected (Supplementary Fig. S5).

We next assessed the reproducibility of gene expression measurements in our microarray study by carrying out qRT-PCR experiments with RNAs that were randomly selected from the training cohort (48 samples). Reproducibility was high for the vast majority of genes. In particular, intermeasurement correlation was extremely high for 26 genes (r > 0.7 and P < 1.0 × 10⁻⁷). On the basis of nonoverlapping biological characteristics, we selected 6 genes: IGFBP4 and SFRP4, cellular signaling; SPOCK1, SULF1, and THBS, interaction with the microenvironment; and GADD45B, stress response (Fig. 3; Supplementary Table S2). To generate RRSs in the range of 0 to 100 from the reference-normalized expression measurements, a 4-step strategy was used (Supplementary Fig. S6A). As expected, the three RRS categories, low, intermediate, and high risk, were significantly associated with RFS (P = 0.02) and OS (P = 0.006; Supplementary Fig. S6B and S6D). The combined low-intermediate risk category was also significantly associated with RFS (P = 0.005) and OS (P = 0.001; Supplementary Fig. S6C and S6E). ROC analysis showed that overall performance of RRS for prognostication of patients is better than individual genes (Supplementary Fig. S8).

We next sought to validate the RRS assay in an independent cohort by using RNAs from FFPE tissues. We found that the continuous RRS was significantly associated with risk of recurrence as evidenced by their linear correlation with probability of recurrence within 5 years postsurgery (Supplementary Fig. S9A). When patients in validation cohort were stratified according to RRS, the 5-year RFS rate was 82.5% for low-risk, 70.5% for intermediate-risk, and 58.1% for high-risk patients (P = 0.001; Fig. 4A). Likewise, two subgroup stratification showed significant association with RFS (P = 0.001; Fig. 4B). RRS was also significantly associated with OS (Fig. 4C and D). Furthermore, when patients with stage II/III disease (n = 264), who are routinely treated with adjuvant chemotherapy as standard treatment, were only included in analysis, association of RRS with prognosis remained significant (Supplementary Fig. S9B–S9E). Taken together, these results clearly demonstrate the robustness of the RRS regardless of differences in the source of RNAs (fresh-frozen or FFPE tissues) and its clinical relevance.

To evaluate the prognostic value of the RRS in combination with other clinical variables, we next carried out univariate and multivariate Cox proportional hazards regression analyses. Like tumor–node–metastasis stage, which is a well-known predictor of RFS, RRS was a statistically significant indicator of RFS [HR, 1.92; 95% confidence interval (CI), 1.02–3.6; P = 0.04; univariate analysis; Supplementary Table S3]. In multivariate analysis, which included all clinicopathologic variables with moderate association (P < 0.2), RRS remained a significant prognostic factor (HR, 1.6; 95% CI, 1.02–2.4; P = 0.03).

Because stage II gastric cancer is considered to be heterogeneous as reflected in lower recurrence rate than stage III but higher recurrence rate than stage I after D2 gastrectomy (6, 7, 35), we carried out a subset analysis in stage II disease. In multivariate analyses, RRS was a significant predictor of RFS (Table 2), providing independent predictive value beyond tumor stage and number of nodes examined. Taken together, our data strongly suggest that the RRS model measures underlying biological characteristics that are predictive of clinical outcomes and that are not captured by traditional clinical and pathologic indicators.

In this study, we showed that YAP1 activation is significantly associated with poor prognosis in gastric cancer. Using a systematic multistep strategy, we developed a 6-gene–based RRS assay that can identify patients at high risk for cancer recurrence after surgery by measuring underlying biological differences. The robustness of the RRS assay was validated in an independent cohort of 317 patients. Because the RRS system uses qRT-PCR on RNAs from FFPE tissues, it can be easily implemented in the clinic.

Multigene assays are currently being investigated for their abilities to improve prognostication and prediction of treatment outcomes for various cancers. For example, the Oncotype DX Breast Cancer Assay (Genomic Health) uses 21 genes (16 recurrence risk–associated genes and 5 reference genes) that were identified from 250 candidate genes using tumor recurrence data and FFPE tumor tissues (8). In this system, the mRNA expression level of each gene measured by qRT-PCR is used to generate a recurrence score. In a retrospective analysis of data from two clinical trials, the assay not only quantified risk of recurrence in patients with hormone receptor–positive, axillary lymph node–negative breast cancer (8) but also predicted the magnitude of the chemotherapy benefit (9). Another assay system, the Oncotype DX Colon Cancer Assay (Genomic Health), uses 7 recurrence risk–associated and 6 chemotherapy benefit–associated genes that were identified from 761 candidate genes (10). Clinical validation using stage II/III colon cancer clinical trial data showed that a recurrence score based on 7 genes was prognostic for recurrence, but a treatment score based on 6 genes was not predictive of benefit from adjuvant chemotherapy (11). All these recurrence score validation analyses were retrospectively conducted and thus could not give definitive conclusion. Prospective clinical trials to evaluate the clinical usefulness of these assays are under way.

To the best of our knowledge, no assay system that can potentially provide prognostic or predictive information to assist decision making for gastric cancer in clinics has been tested in clinical trials. In most existing multigene assays for gastric cancer, genes were selected using predefined candidate genes from the literature (36) or small-scale screening of candidate genes (37). To overcome our previous analysis with smaller number of patients that did not well reflect biological characteristics of gastric cancer (37), we applied multistep systematic strategy that may reflect underlying biology of gastric cancer and prognostic significance. In the current study, we found the robustness of the YAP1 signature in predicting gastric cancer recurrence after resection, which was in agreement with previous observations in other cancers (38–41). Therefore, we decided to develop an easy-to-use RRS assay that reflects underlying biological differences (YAP1 activation) in gastric cancer. We found that 156 genes were significantly associated with RFS, OS, and YAP1 activation. From the 156 candidate genes, we identified 6 recurrence risk–associated genes on the basis of biological function and data reproducibility of gene expression measurements in qRT-PCR and microarray experiments. Notably, one of the 6 genes, GADD45B, was previously identified as one of the recurrence risk–associated genes in colon cancer (10), suggesting that this gene may play a critical role in progression of gastrointestinal cancers in general.

The continuous RRS was linearly correlated with probability of recurrence, allowing classification of risk as low, intermediate, or high. The 6-gene RRS assay was validated in an independent patient cohort, and the prognostic relevance of RRS was independent of other clinical variables. Given that the rate of recurrence for stage II gastric cancer after a D2 gastrectomy alone is approximately 30% compared with >50% for stage III (6, 7, 35), the decision on whether adjuvant chemotherapy should be administered is more challenging in stage II than in stage III. The subset analysis clearly showed that the 6-gene RRS assay can be used to stratify stage II patients according to recurrence risk. These findings suggest that the 6-gene RRS assay might help clinicians decide whether to administer adjuvant chemotherapy to patients after gastrectomy. Because our RRS assay uses qRT-PCR of RNA from FFPE tissues, we expect that it will be easily applicable in clinical practice.

The limitations of our study should be acknowledged. First, the study was performed retrospectively. Second, the sample of patients was heterogeneous and included patients with stage I or IV disease, who were not candidates for adjuvant chemotherapy. However, vast majority of patients were in stage II or III. In subset analysis with patients with stage II/III disease, who are candidates for adjuvant chemotherapy, association of RRS with prognosis remained significant. Third, the association between the RRS and benefit from adjuvant treatment could not be investigated because all patients in validation cohort received adjuvant chemotherapy. The potential of RRS in predicting benefit from adjuvant therapy after gastrectomy needs to be validated in future prospective studies. Fourth, because all patients in the current study were from Korea, the applicability of our RRS system to patients with different ethnicities needs to be further investigated.

In conclusion, our robust RRS assay can help facilitate rational design of clinical studies by stratifying patients according to their risk of cancer recurrence. Prospective studies will need to be conducted to validate the clinical usefulness of the assay.

No potential conflicts of interest were disclosed.

Conception and design: K.-W. Lee, H.-S. Lee, S.H. Lee, W. Jeong, J.Y. Cho, J.-S. Lee

Development of methodology: J.-E. Hwang, S.H. Lee, J.-S. Lee

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K.-W. Lee, S.S. Lee, J.-E. Hwang, H.-J. Jang, S.C. Oh, S.H. Lee, J.-H. Cheong, J.Y. Lim, Y.-K. Kang, S.H. Noh

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K.-W. Lee, S.S. Lee, J.-E. Hwang, H.-J. Jang, H.-S. Lee, S.H. Lee, B.H. Sohn, S.B. Kim, J.J. Shim, M. Cha, E.S. Park, S.-C. Kim, J.A. Ajani, J.-S. Lee

Writing, review, and/or revision of the manuscript: K.-W. Lee, S.S. Lee, J.-E. Hwang, S.H. Lee, Y.-K. Kang, J.A. Ajani, J.-S. Lee

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S.H. Lee, J.J. Shim, S.H. Noh, J.A. Ajani, J.-S. Lee

Study supervision: S.H. Lee, J.-S. Lee

This research was supported by grants from the NIH (CA127672, CA129906, CA138671, CA 172741, and CA150229); 2016 Institutional Research Grant (IRG) and 2016 Sister Institute Network Fund (SINF) grant from The University of Texas MD Anderson Cancer Center; Bio & Medical Technology Development Program (M10642040002-07N4204-00210); Scientific Research Center Program (2012R1A5A1048236); Korea National Research Foundation (No. 2014M3C1A3051981); and Research Initiative grant from Korean Research Institute of Bioscience and Biotechnology (KRIBB).

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.