Gene Expression Profile Predicts Patient Survival of Gastric Cancer After Surgical Resection
http://www.100md.com
《临床肿瘤学》
the Department of Surgery, Traumatology, Graduate Institute of Clinical Research, National Taiwan University Hospital and College of Medicine
Laboratory of Molecular and Cellular Toxicology, Institute of Toxicology, College of Medicine
Angiogenesis Research Center, National Taiwan University
Department of Applied Statistics and Information Science, Ming Chuan University, Taipei
the Institutes of Biomedical Sciences and Molecular Biology, National Chung Hsing University, Taichung, Taiwan
ABSTRACT
PURPOSE: This study was conducted to characterize gene expression profile of survival in patients with surgically curable gastric cancer by using an in-house membrane microarray and developing a survival prediction model.
MATERIALS AND METHODS: Data of cDNA microarrays were obtained from 18 pairs of cancerous and noncancerous gastric tissues. Nine patients who survived > 30 months were identified as good survival, and the other nine, who survived < 12 months, were identified as poor survival. Supervised analysis was performed to identify a gene expression profile by good and poor survival. Semiquantitative reverse transcriptase polymerase chain reaction (RT-PCR) was used to confirm the microarray data in 10 patients with sufficient RNA. Using these 10 patients and another 10 patients selected randomly from 40 newly enrolled patients as the training group, the RT-PCR status of the confirmed genes was used for predicting good versus poor survival. Finally, the prediction model was tested in the remaining 30 newly enrolled gastric cancer patients.
RESULTS: A survival prediction model consisting of three genes (CD36, SLAM, PIM-1) was developed. This model could correctly predict poor or good survival in 23 (76.7%) of 30 newly enrolled patients, and yielded a specificity of 80% and a sensitivity of 73.3%. The survival rate of the patients predicted to have good survival was significantly higher than that of those predicted to have poor survival in the test group as a whole (N = 30; P = .00531) and in stage III patients (n = 16; P = .04467).
CONCLUSION: The semiquantitative RT-PCR gene expression profiling of three genes extracted from microarray study can accurately predict surgery-related outcome in gastric cancer patients.
INTRODUCTION
Gastric cancer is one of the most frequent cancers in the world, and it is the fourth most common in Taiwan. Early-stage disease tends to be diagnosed in countries in which endoscopic screening is more common; however, there still exist patients whose tumor is advanced at the time of diagnosis.1 Patients with stage I disease have a good prognosis, and those with stage IV disease show a poor prognosis. Bewilderingly, the prognosis varies widely in patients with stage II or III disease for as of yet undetermined biologic reasons.2
Traditional clinicopathologic factors and several interesting molecules, including cell cycle regulation factors such as p27 or cyclin E,3,4 cell adhesion molecules such as E-cadherin,5 angiogenic factors such as vascular endothelial growth factor and placenta growth factor,6,7 oncogenes such as c-erbB2 and c-myc,8 and tumor suppressor genes such as p53,8 have been reported to correlate to the prognosis of gastric cancer patients. However, there exists inconsistency among different studies, and the reported parameters provided limited information about prognosis of individual patients because of complex biology of the disease.9 The cellular and molecular heterogeneity of gastric cancers and the large number of genes potentially involved in the multi-step process of gastric cancer pathogenesis emphasize the importance of studying multiple genetic alterations in concert.
Recent advances in the cDNA microarray technique that can investigate gene expression systematically enable us to visualize gene expression profiles in human tumors, and those gene expression profiles can help to identify gene activity patterns that can distinguish subclasses of gastric tumors. Gene profiling studies were also used to better stratify and select patients for adjuvant therapies who may be at higher risk for recurrence.10-13 Recently, Gordon et al14 have shown that simple patterns of gene expression levels, using as few as four to six genes selected from microarray, are highly accurate in the outcome prediction of methothelioma.
Unfortunately, practical methods for identification of individuals with gastric cancer who are at risk for recurrence after surgical resection are not currently available. To explore the possibility of using a limited gene analysis as a predictor of outcome in resected gastric cancer and to find out genomic information related to the survival of gastric cancer patients, an in-house nylon membrane cDNA microarray with colorimetric detection system, which includes 328 genes known as cell cycle regulation factors, matrix proteinases, cell adhesion molecules, angiogenic factors, immunity related factors, oncogenes, and tumor suppressor genes, was constructed to characterize discriminative gene expression for survival, and to develop a quantitative model consisting of selected gene profile which is capable of predicting outcome of gastric cancer patients after curative resection.
MATERIALS AND METHODS
Summary of Experimental Design
Eighteen tumor and healthy pairs of gastric tissue samples were obtained from 18 patients with gastric cancer who underwent D2 gastrectomy without gross residual tumor at the National Taiwan University Hospital (Taipei, Taiwan) between 1995 and 2002. The tumor stage ranged from stage I to stage IV. Nine patients who died as a result of tumor recurrence within 12 months after surgery were identified as poor survival, and the other nine patients, who survived beyond 30 months after surgery were identified as good survival. The poor-survival group included two patients with stage II disease, four with stage III, and three with stage IV. The good-survival group included three patients with stage I disease, two with stage II, and four with stage III. There was a balance of stage II and III between survival groups. There was no stage I patient in poor-survival group, and no stage IV patient in the good-survival group. All patients did not receive postoperative chemotherapy and radiotherapy. Pair samples of tumor and healthy tissues of these 18 patients were dissected and frozen within 30 minutes of removal in a liquid nitrogen tank. Healthy mucosa samples were taken from areas of grossly normal mucosa located at least 3 cm away from tumor border. RNAs of these 18 pairs of the specimens were extracted and were used to perform microarray hybridizations; microarray data of these 18 pairs of samples were used to select gene profiles supervised by poor and good survival through a statistical algorithm. After genes characterizing survival were selected, semiquantitative reverse transcriptase polymerase chain reaction (RT-PCR) analysis of the selected genes was used to verify microarray data in 10 patients with sufficient RNA; thereafter, the genes with RT-PCR results consistent with microarray data were used for further data training and prediction model selection. In addition to these 10 patients, RT–PCR analyses of the selected genes characterizing survival were also performed in 40 newly enrolled patients. We assigned the RT-PCR status of the selected genes into categoric variables that were used in a logistic regression model to predict survival based upon 20 patients (10 patients with good survival; 10 patients with poor survival). To avoid the overfitting problem, these 20 patients (training samples) included 10 patients who had sufficient RNA to confirm microarray data and another 10 patients who were selected randomly from 40 newly enrolled patients. To assess the robustness, 1,000 possible sets of 20 training samples were generated. A stepwise approach, based on Akaike's information criterion (AIC), was used for model selection. A candidate prediction model was selected by using the smallest AIC of the stepwise process in each set of generated training samples. Then, the final effective prediction model, selected from 1,000 candidate prediction models, was the one with the largest AIC among 1,000 smallest AIC. RT-PCR analyses of the selected genes characterizing survival were then performed in the 30 remaining newly enrolled patients (ie, the test group), which included seven patients with stage I, five with stage II, 16 with stage III, and two with stage IV. The RT-PCR status of 30 patients of the test group was translated into the aforementioned categoric variables to test the predictive value of the model extracted from training group. There was no postoperative chemotherapy and/or radiotherapy for patients of either training or test group in this study. This is our standard practice for managing gastric cancer patients who underwent D2 gastrectomy without gross residual tumor in this hospital. Studies using human tissues were approved by and conducted in accordance with the policies of the institutional review board at National Taiwan University Hospital.
Microarray
We used an in-house nylon membrane cDNA microarray made of nylon membrane containing 384 spots and aligned by 16 spots in each row and 24 spots in each column with a spot diameter of 250 μM. It contains sequence-verified cDNA clones of 328 selected known human genes that are considered to be cancer related. These include oncogenes, tumor suppressor genes, apoptosis-related genes, matrix proteinase genes, angiogenesis-related genes, immune-related genes, and so on. The internal control genes included 16 plant genes and GAPDH. This cDNA microarray system and membranes (Boehringer, Mannheim, Germany) were prepared as previously described.15 These 328 sequence verified cDNA clones served as the hybridization targets.
RNA Extraction, Biotinylated Probe Preparation and Microarray Hybridization
Total RNAs were extracted with Trizol reagent (Invitrogen Life Technologies Inc, Carlsbad, CA). Thirty micrograms of the total RNAs derived from each gastric cancer tumor tissue and corresponding nontumor part were reversed transcribed and labeled with biotin. The microarray carrying the double-stranded cDNAs was prehybridized in 1 mL of hybridization buffer (5x RNA extraction standard saline citrate [lsqb]SSC[rsqb], 0.1% N-lauroylsarcosine, 0.1% sodium dodecyl sulfate [lsqb]SDS[rsqb], 1% blocking reagent mixture manufactured by Roche Molecular Biochemicals (Mannheim, Germany), and salmon-sperm DNA [50 μg/mL]) at 63°C for 1.5 hour. The biotin-labeled cDNA probes and hybridization solution (13 μL) containing human COT-1 DNA instead of salmon-sperm DNA were sealed with a microarray in a hybridization bag, and the bag was incubated at 63°C for 10 hours. The membrane was then washed with 2x SSC containing 0.1% SDS for 5 minutes at room temperature followed by three washes with 0.1x SSC containing 0.1% SDS at 63°C for 15 minutes each. After hybridization, the color reaction was initiated by incubating the membrane for 1 hour in 1 mL containing alkaline phosphatase-conjugated streptavidin (diluted 1:1,000; Roche Molecular Biochemicals), 4% polyethylene glycol (molecular weight, 8,000; Sigma Chemical Co, St Louis, MO), and 0.3% bovine serum albumin (BSA) in 1x phosphate-buffered saline (PBS) buffer. The substrate used for color development is BCIP/NBT (5-bromo-4-chloro-3-indolyl-phosphate/ nitro blue tetrazolium) substrate (Promega, Madison, WI). Color reaction was stopped with 1x PBS buffer that contains 20 μM EDTA.
Image Processing
After color development, the membrane was scanned and images taken by using a flat-bed scanner (UMAX Fremont, CA MagicScan at 3,000 dpi). This scanner provided 3,000-dpi resolution and was suitable for our arrays with 384 elements and the resulting database was stored in a tagged image file format (Tiff). To quantify the expression levels of the genes, the digitized images were measured by GenePix Pro software program (Axon Instruments, Foster City, CA).
Microarray Data and Statistical Analysis
Eighteen pairs of patient samples with clinical information were submitted for microarray studies. After the experimental processes, the color of a spot yielded by the enzymatic reaction was converted to gray levels, of which the brightness value was assigned to each pixel of that spot ranging from 0 to 256 (from black, through shades of gray, to white). The spots' expression values of the microarray generated by the GenePix Pro 2.0 software by scanning the microarray images with UMAX 6,000 were collected as an Excel file (Microsoft, Redmond, WA). Expression level of each gene on the microarray was defined as the log ratio (base 2) of the spot expression value of each gene of tumor tissue to that of nontumor tissue. Eighteen patients consisting of poor- and good-survival groups were used specifically to characterize gene expression of survival. To identify genes whose expression levels could be used to discriminate patients with good and poor survival, a three-step supervised classification method was used to extract the most significantly differentially expressed genes between these two survival groups (Fig 1). In the first step, to avoid the systematic error at each microarray sample, the nonlinear LOWESS method,16,17 which fits a curve to the log ratios using robust locally weighted regression, was used to normalize the log ratios of the 328 genes to a LOWESS curve fitting through the MA plot. Second, to extract the significantly regulated genes among all 18 microarray samples, the fold-change method18 which defined the threshold of the significance for expression microarrays, was used to define the regulated genes whose fold-change (normalized log ratios) are greater in magnitude than one at each microarray sample. Among all 18 microarray samples, the significantly regulated genes whose fold-changes are greater in magnitude than one for at least two samples among 18 samples were extracted. Third, to extract the most significant differentially expressed genes between the two survival groups, the multiple permutation test17,19 was used to test simultaneously all the significantly regulated genes filtered at step two, and the adjusted P value for each gene was obtained. The permutation test, reflecting the actual random assignment, is to obtain the original P value of each gene by resampling and recomputing the test statistic several times and comparing the test statistics to the initial test statistics. The multiple test is a method to control the probability of producing incorrect test conclusions (false positives and false negatives) for the whole gene set from original P value to the adjusted P value by the modified statistical algorithm. To assess the internal consistency of the 18 samples, the leave-one-out cross validation (CV) method13 was used to generate the 18 CV samples, and the differentially expressed genes were extracted whose adjusted P value was < .05 family-wise error rate for all 18 CV samples.
RT-PCR Analysis for Selected Genes Confirmation
To verify our microarray results and to further clarify the difference in the expression of the selected genes, we carried out semiquantitative RT-PCR for selected gene confirmation20 by using RNA stock of 10 pairs of samples that had been subjected to microarray study, and the remaining 8 had no RNA left. Two micrograms total RNAs was reverse transcribed using Moloney Murine Leukemia Virus Reverse Transcriptase, Random Primer, and other kit reagents (Promega), followed by PCR. The primers for RT-PCR amplification for examining expression of selected marker genes follow. PCR conditions we used were as follows: for CD36 (40 seconds at 94°C, 35 seconds at 53°C, and 40 seconds at 72°C; 30 cycles), forward, 5'ATGTAACCCAGGACGCTGAG3', and reverse 5'TGGGTTTTCAACTGGAGAGG3'; for signaling lypmphocytic activation molecule (SLAM) (40 seconds at 94°C, 35 seconds at 54°C, and 40 seconds at 72°C; 30 cycles), forward, 5'CCTTCGTGCTGTTTCTCTCC3', and reverse 5'GCTCACGGTGCAGATGTAGA3'; for IGF-1 (40 seconds at 94°C, 35 seconds at 59°C, and 35 seconds at 72°C; 25 cycles) forward, 5'CATTGCTCTCAACATCTCCC 3', and reverse 5'ACGAACTGAAGAGCATCCAC3'; for PIM-1 (40 seconds at 94°C, 35 seconds at 57°C, and 37 seconds at 72°C; 25 cycles) forward, 5'GCTCGGTCTACTCAGGCATC3', and reverse 5'GTCCGTGTAGACGGTGTCCT3'; for transcription factor AP-2 alpha (TFAP; 40 seconds at 94°C, 35 seconds at 51°C, 35 seconds at 72°C, 30 cycles) forward, 5'TGGACCCTGGAAAGATTTTG3', and reverse, 5'CCAGGAAACGGAGGTTGTAG3'; for TIMP-4 (40 seconds at 94°C, 35 seconds at 53°C, and 40 seconds at 72°C; 30 cycles) forward, 5'TGCTGTCAAACCACCTTCTG3', and reverse 5'TGGGCAACAAAGGACTATGA3'. PCR products were separated by electrophoresis on 1.5% agarose gels and visualized under UV light after ethidium bromide staining. We determined the mean band densities using National Institutes of Health Image 1.62 software, and we calculated levels of selected genes relative to ?-actin gene. The relation between microarray expression ratio and RT-PCR results of six selected genes were determined.
Translation of Semiquantitative RT-PCR Data Into Survival Prediction Model
To establish a survival prediction model for gastric cancer patients, the differentially expressed genes whose consistent rates between microarray and RT-PCR results were greater than 60% or Spearman rank correlation coefficient showed significant and P < .05 were selected for prediction model training. Then, we assigned one of four categories to RT-PCR expression data for each sample. The frequencies of the four categoric RT-PCR results at each predictor in the training group were used to establish the prediction model. In the testing samples, each predictor was assigned the same frequencies of the four categoric PCR results as those obtained in the training group. Next, we applied the logistic regression together with stepwise model selection to find the effective prediction model by the AIC.21
Finally, we applied this prediction model to a test group of 30 pair samples and assessed the accuracy of the model by evaluating the sensitivity, specificity, positive prediction value, and negative prediction value on the test group as a whole and the stage III patients. The Kaplan-Meier survival model was also used to estimate the survival of patients stratified by this prediction model. The log-rank test was used to statistically assess significance between survival curves. All differences were considered statistically significant if the P value was < .05.
RESULTS
Gene Selection, Confirmation, and Prediction Model Selection
The three-step supervised classification method was applied to the 18 patients consisting of poor- and good-survival groups to search for gene expression profile of survival (Fig 2). At the normalization step, we used 328 genes to normalize the log ratios at each sample. One hundred forty-one genes are extracted with fold-change method after normalization of log ratios. Finally, the multiple permutation test together with CV was used to select the six most significant differentially expressed genes. The selected genes included CD36 antigen, SLAM, TFAP, insulin-like growth factor 1 (IGF-1), PIM-1 oncogene, and tissue inhibitor of metalloproteinase-4 (TIMP-4).
Semiquantitative RT-PCR was performed to test the expression of these six genes in 10 pairs of tumor and nontumor tissues. The results of RT-PCR were compared with those of microarray. Four of the six aforementioned genes (CD36, SLAM, TFAP, PIM-1), whose consistent rates were > 60% and Spearman rank correlation coefficient showed significance and P < .05 (Fig 3), were picked up for model training.
Duplication of Microarray Experiments in Four Patients
For quality control, we also tested the consistency of microarray experiments in four patients. These four patients were chosen randomly for duplicate study from 18 patients to test the reproducibility of this in-house nylon microarray. Total RNAs of these four patients were hybridized to two different microarray nylon filters in different time, and the Pearson correlation coefficients across hybridizations was calculated and were all > 0.75 (P < .05), which meant good repetition of the experiments.
Translation of Expression of CD36, SLAM, TFAP, and PIM-1 Into Prediction Model of Survival
Four types of paired RT-PCR status for each selected genes could be identified in all blotting results. We assigned one of the following four categories to RT-PCR expression data for each sample: Tumor Healthy, Healthy Tumor, Tumor and Healthy equally positive, and Tumor and Healthy equally negative (Fig 4). The frequencies of the four categoric RT-PCR results at each predictor in the training group were used to establish the prediction model. In the testing samples, each predictor will be assigned the same frequencies of the four categoric PCR results as those obtained in the training group. The logistic regression together with stepwise model selection was used to select the prediction model by the AIC criteria, and the most effective logistic model consisting of three genes (CD36, SLAM, and PIM-1) was obtained as follows:
where is the probability of "poor survival status." Good survival (defined as survival time > 30 months) was predicted when 0.5. Poor survival (defined as survival time < 12 months) was predicted when >0.5. The SEs of the logistic regression coefficients are 0.411 for CD36, 0.436 for SLAM, and 0.173 for PIM-1.
Survival Prediction for an Independent Test Group of 30 Gastric Cancer Patients
The prediction model consisting of CD36, SLAM, and PIM-1 was applied to an independent test group of 30 patients. Survival of 23 patients (76.7%) was correctly predicted, and yielded a specificity of 80%, a sensitivity of 73.3%, a positive prediction value of 75%, and a negative prediction value of 78.57%. The frequency distribution is shown in Table 1. This reveals that this prediction model showed highly predictive power in the independent test group. The survival rate of the patients predicted to have good survival was significantly higher than that of the patients predicted to have poor survival (P = .00531; Fig 5A).
Of the seven stage I patients, the survival of six was correctly predicted by this model. One patient was predicted as having poor survival, and died of multiple liver metastases in 12 months. Of the other six patients, the survival of five was predicted correctly, and the frequency distribution is shown in Table 2. Of the five stage II patients, the survival of three was predicted correctly. Two of three patients predicted to have poor survival died of disease in 12 months, and the frequency distribution is shown in Table 3. Two stage IV patients were predicted correctly by this model, and the frequency distribution is shown in Table 4.
Survival Prediction for the Patients With Stage III Gastric Cancer
The prediction model was applied to 16 patients with stage III disease, and the frequency distribution of accuracy is shown in Table 5. Twelve patients (75%) were predicted correctly, and yielded a specificity of 100%, a sensitivity of 63.6%, a positive prediction value of 100%, and a negative prediction value of 55.6%. The survival rate of the patients predicted to have good survival was significantly higher than that of patients predicted to have poor survival (P = .04467; Fig 5B).
DISCUSSION
cDNA microarray is a powerful tool to monitor global transcription profile of hundreds to tens of thousands of genes at once.10-12 This technology has a strong impact not only on the basic biology but also on clinical practice, and it provides a powerful tool to discover new potential targets against carcinogenesis and metastasis.22,23 It is also anticipated to render more accurate disease diagnosis and more appropriate therapy. Several studies have reported that DNA microarray could identify previously undetected and clinically significant subtypes, and also been used to formulate a molecular predictor of survival after chemotherapy for diffuse large-B-cell lymphoma.10,24 cDNA microarray can also provide a strategy for selecting patients who would benefit from adjuvant therapy and gene expression profile might outperform all currently used clinical parameters in predicting disease outcome in breast cancer.13 Subsequently, these investigators have successfully tested outcome predictor models in independent samples,13,24,25 but those studies continue to be restricted in their clinical applicability because they rely on data gathered from relatively large numbers of genes, costly data acquisition platforms (eg, microarrays), and sophisticated algorithms and/or software and are unable to analyze a sample independently and without reference to other samples. In this study, an in-house nylon membrane mini-microarray containing 328 known genes with colorimetry detection was used to seek predictive gene expression profile for the survival of patients with gastric cancer. We demonstrated that gene expression patterns could more accurately predict survival of gastric cancer patients after curative resection than traditional staging while successfully avoiding many of the shortcomings that preclude the use of microarray techniques in clinical applications.10,26
Initially, six genes were selected from microarrays in this study. They were SLAM, CD36, PIM-1 oncogene, TFAP, IGF-1, and TIMP-4. After confirmation by using semiquantitative RT-PCR, SLAM, CD36, PIM-1 and TFAP were included in prediction model training because of their high consistency rate with microarray data. Finally, the most effective prediction logistic regression model defined by SLAM, CD36, and PIM-1, which can predict the survival of gastric cancer patients with reasonable efficacy, was achieved. Among those three genes, SLAM is a CD2-related surface receptor expressed by activated T cells, B cells, and dendritic cells. Th0/Th1 immunity, which is usually impaired in gastric cancer patients, could also be induced by SLAM to enhance proliferation and cytotoxic ability of CD8+ tumor-specific lymphocytes.27-29 Although the real role of SLAM in tumor-associated immunity of gastric cancer patients remains unclear, it seems to have potential effects on anti-tumor immunity and deserves further investigation. CD36 is a trans-membrane receptor that regulates apoptosis and angiogenesis in response to its ligand thrombospondin-1 (TSP-1). TSP-1 is localized to tumor-associated extracellular matrix, and CD36 expressed on surface of tumor cells.30 The regulation of CD36 expression in tumor cells may play an important role in tumor growth, metastasis and angiogenesis.31 CD36 overexpression, which decreased stromal vascularization by induction of apoptosis of vascular endothelial cells, is correlated with better prognosis of colon cancer.32 However, there existed conflicting reports in the association of CD36 expression with angiogenesis and patient survival in head and neck cancer.33 PIM-1, a product of serine/threonine kinase which can be induced in gastric epithelial cells by Heliobacter pylori infection, may be involved in gastric carcinogenesis.34 PIM-1 also plays an important roles in proliferation, differentiation and maturation of T cells, which may be associated with tumor immunity.35 PIM-1 induced by hypoxia is involved in drug resistance and tumorigenesis of solid tumor cells and leads to genomic instability.36 It is also required in endothelial cells for vascular endothelial growth factor-A–dependent proliferation and migration.37 Recently, the expression of PIM-1 has been shown to correlate significantly with measures of clinical outcome in prostate cancer.38 The three genes selected in our prediction model may be, in some way, involved in tumor angiogenesis and tumor immunity, which is closely related to patient survival. However, a wide range of other biologic processes and complex interactions among different cell types in cancer tissue should also modulate cancer biology and result in heterogeneity of disease. None of these genes has been reported to correlate with clinical outcome in gastric cancer. Further functional and clinical studies of these genes are worthwhile, and it is more likely that these genes in isolation have only limited predictive power, which highlights the need for a genomic approach based on the expression of multiple genes in our attempts of disease subtyping.
To our knowledge, this is the first study to use gene expression profiling techniques to develop a survival prediction model for human gastric cancer. In our study, a data acquisition platform of semiquantitative RT-PCR was developed and used to validate this model in an independent cohort. The use of RT-PCR–based gene expression patterns to predict outcome in gastric cancer patients overcomes several major obstacles that hinder the clinical use of microarray. Unlike other widely accepted supervised learning techniques with similar predictive accuracy,10,13,24-26 this prediction model utilizes data from RT-PCR status to create categories of expression among the three genes in paired samples. Because of its independence of microarray platform after gene selection, our methodology requires only small quantities of RNA (as little as 2 μg when using RT-PCR) and can be performed easily in a common laboratory. In this study, microarray data normalized with the LOWESS method to avoid the systematic error within each microarray sample that might be overly normalized. Therefore, verification of selected genes with RT-PCR is essential in microarray study. The overfitting problem is a crucial issue in the selection process of the prediction model. We used the randomly generated samples to overcome this potential pitfall and established the three-gene prediction model, which had better sensitivity and specificity than the models with one or two genes.
The prognostic tool described herein could dramatically influence the current clinical treatment of gastric cancer patients with stage III disease by allowing the identification of the patients who are unlikely to respond to conventional surgical treatment only. Adjuvant chemotherapy has been reported to have marginal effect for overall gastric cancer patients with D2 gastrectomy (gastrectomy with D2 lymphadenectomy). If survival outcome of gastric cancer patients can be reasonably predicted, adjuvant therapies may help those patients who have high probability of poor survival, whereas those who have high probability of good survival can be spared from the adverse effects of adjuvant therapies. Although whether the patients who are predicted to have high probability of poor outcomes really benefit from adjuvant therapies remains unknown, the results of this work, if confirmed prospectively in a larger patient population, could be helpful in the development of meaningful clinical trials of new compounds or regimens in advanced gastric cancer patients after surgical resection, especially in those with stage III disease, by helping to select suitable candidates.
In the coming era of personalized medicine, we believe that molecular profiling attempts like this study will be the basis of individualized therapy of gastric cancer. In conclusion, this report demonstrated that the survival of individual gastric cancer patients can be predicted with reasonable accuracy by using a simple RT-PCR–based gene expression profile and a logistic regression model. Our methodology is one of the applicable models and might be important to both patients and their doctors in terms of life planning and therapeutic strategies.
Authors' Disclosures of Potential Conflicts of Interest
The authors indicated no potential conflicts of interest.
Acknowledgment
We thank Professor Chien C. Chang for his helpful discussion and comprehensive suggestion in preparing this manuscript. We also thank research assistant Su-I Chen for her excellent technique.
NOTES
Supported by grant from National Science Council and Department of Industrial Technology, Ministry of Economic Affairs, Taipei, Taiwan.
F.J.H. and K.J.C. are joint senior authors.
Terms in blue are defined in the glossary, found at the end of this issue and online at www.jco.org.
Authors' disclosures of potential conflicts of interest are found at the end of this article.
REFERENCES
Akoh JA, Macintyre IM: Improving survival in gastric cancer—Review of 5-year survival rates in English language publications from 1970. Br J Surg 79:293-299, 1992
Allgayer H, Heiss MM, Schildberg FW: Prognostic factors in gastric cancer. Br J Surg 84:1651-1664, 1997
Mori M, Mimori K, Shiraishi T, et al: p27 expression and gastric carcinoma. Nat Med 3:593, 1997
Akama Y, Yasui H, Yokozaki H, et al: Frequent amplification of the cyclin E gene in human gastric cancer. Jpn J Cancer Res 86:617-621, 1995
Graziano F, Mandolesi A, Ruzzo A, et al: Predictive and prognostic role of E-cadherin protein expression in patients with advanced gastric carcinomas treated with palliative chemotherapy. Tumour Biol 25:106-110, 2004
Tanigawa N, Amaya H, Matsumura M, et al: Correlation between expression of vascular endothelial growth factor and tumor vascularity, and patient outcome in human gastric carcinoma. J Clin Oncol 15:826-832, 1997
Chen CN, Hsieh FJ, Cheng YM, et al: The significance of placenta growth factor in angiogenesis and clinical outcome of human gastric cancer. Cancer Lett 213:73-82, 2004
Sanz-Ortega J, Steinburg SM, Moro E, et al: Comparative study of tumor angiogenesis and immunohistochemistry for p53, c-erbB2, c-myc and EGFr as prognostic factors in gastric cancer. Histol Histopathol 15:455-462, 2000
Zheng L, Wang L, Ajani J, et al: Molecular basis of gastric cancer development and progression. Gastric Cancer 7:61-77, 2004
Golub TR, Slonim DK, Tamayo P, et al: Molecular classification of cancer: Class discovery and class prediction by gene expression monitoring. Science 286:531-537, 1999
Alizadeh AA, Eisen MB, Davis RE, et al: Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling. Nature 403:503-511, 2000
Okabe H, Satoh S, Kato T, et al: Genome-wide analysis of gene expression in human hepatocellular carcinomas using cDNA microarray: Identification of genes involved in viral carcinogenesis and tumor progression. Cancer Res 61:2129-2137, 2000
van't Veer LJ, Dai H, van de Vijver MJ, et al: Gene expression profiling predicts clinical outcome of breast cancer. Nature 415: 530-536, 2002
Gordon GJ, Jensen RV, Hsiao LL, et al: Using gene expression ratios to predict outcome among patients with mesothelioma. J Natl Cancer Inst 95:598-605, 2003
Chen JJ, Wu R, Yang PC, et al: Profiling expression patterns and isolating differentially expressed genes by cDNA microarray system with colorimetry detection. Genomics 51:313-324, 1998
Cleveland WS: Robust locally weighted regression and smoothing scatterplots. J Am Stat Assoc 74:829-836, 1979
Yang YH, Buckley MJ, Dudoit S, et al: Comparison of methods for image analysis on cDNA microarray data. J Comput Graph Statist 11:108-136, 2002
White KP, Rifkin SA, Hurban P, et al: Microarray analysis of Drosophila development during metamorphosis. Science 286:2179-2184, 1999
Westfall PH, Young SS: Resampling-Based Multiple Testing: Examples and Methods for P-Value Adjustment. New York, NY, Wiley, 1993
Iizuka N, Oka M, Yamada-Okabe H, et al: Comparison of gene expression profiles between hepatitis B virus-and hepatitis C virus-infected hepatocellular carcinoma by oligonucleotide microarray data on the basis of a supervised learning method. Cancer Res 62:3939-3944, 2002
Akaike H: A new look at statistical model identification. IEEE Transactions on Automatic Control AU-19:716-722, 1974
Bertucci F, Salas S, Eysteries S, et al: Gene expression profiling of colon cancer by DNA microarrays and correlation with histoclinical parameters. Oncogene 23:1377-1391, 2004
Shih JY, Yang SC, Hong TM, et al: Collapsin response mediator protein-1 and the invasion and metastasis of cancer cells. J Natl Cancer Inst 93:1932-1400, 2001
Rosenwald A, Wright G, Chan WC, et al: The use of molecular profiling to predict survival after chemotherapy for diffuse large-B-cell lymphoma. N Engl J Med 346:1937-1947, 2002
Beer DG, Kardia SL, Huang CC, et al: Gene-expression profiles predict survival of patients with lung adenocarcinoma. Nat Med 8:816-824, 2002
Shipp MA, Ross KA, Tamayo P, et al: Diffuse large B-cell lymphoma outcome prediction by gene expression profiling and supervised machine learning. Nat Med 8:68-74, 2002
Henning G, Kraft MS, Derfuss T, et al: Signaling lymphocytic activation molecule (SLAM) regulates T cellular cytotoxicity. Eur J Immunol 31:2741-2750, 2001
Nakayama H, Kitayama J, Muto T, et al: Characterization of intracellular cytokine profile of CD4(+) T cells in peripheral blood and tumor-draining lymph nodes of patients with gastrointestinal cancer. Jpn J Clin Oncol 30:301-305, 2000
Lee KY, Goedegebuure PS, Linehan DC, et al: Immunoregulatory effects of CD4+ T helper subsets in human melanoma. Surgery 117:365-372, 1995
Albo D, Shinohara T, Tuszynski GP: Up-regulation of metalloproteinase 9 by thrombospondin 1 in gastric cancer. J Surg Res 108:51-60, 2002
Chen M, Pych E, Corpron C, et al: Regulation of CD36 expression in human melanoma cells. Adv Exp Med Biol 507:337-342, 2002
Tsuchida T, Kijima H, Tokunaga T, et al: Expression of the thrombospondin 1 receptor CD36 is correlated with decreased stromal vascularisation in colon cancer. Int J Oncol 14:47-51, 1999
Wang TN, Qian XH, Granick MS, et al: The effect of thrombospondin on oral squamous carcinoma cell invasion of collagen. Am J Surg 170:502-505, 1994
Sepulveda AR, Tao H, Carloni E, et al: Screening of gene expression profiles in gastric epithelial cells induced by Helicobacter pylori using microarray analysis. Aliment Pharmacol Ther 16:145-157, 2002 (suppl 2)
Wang Z, Bhattacharya N, Weaver M, et al: PIM-1: A serine/threonine kinase with a role in cell survival, proliferation, differentiation and tumorigenesis. J Vet Sci 2:167-179, 2001
Roh M, Gary B, Song C, et al: Overexpression of the oncogenic kinase PIM-1 leads to genomic instability. Cancer Res 63:8079-8084, 2003
Dhanasekaran SM, Barrette TR, Ghosh D, et al: Delineation of prognostic biomarkers in prostate cancer. Nature 412:822-826, 2002
Valdman A, Fang X, Pang ST, et al: PIM-1 expression in prostatic intraepithelial neoplasia and human prostate cancer. Prostate 60:367-371, 2004(Chiung-Nien Chen, Jen-Jen)
Laboratory of Molecular and Cellular Toxicology, Institute of Toxicology, College of Medicine
Angiogenesis Research Center, National Taiwan University
Department of Applied Statistics and Information Science, Ming Chuan University, Taipei
the Institutes of Biomedical Sciences and Molecular Biology, National Chung Hsing University, Taichung, Taiwan
ABSTRACT
PURPOSE: This study was conducted to characterize gene expression profile of survival in patients with surgically curable gastric cancer by using an in-house membrane microarray and developing a survival prediction model.
MATERIALS AND METHODS: Data of cDNA microarrays were obtained from 18 pairs of cancerous and noncancerous gastric tissues. Nine patients who survived > 30 months were identified as good survival, and the other nine, who survived < 12 months, were identified as poor survival. Supervised analysis was performed to identify a gene expression profile by good and poor survival. Semiquantitative reverse transcriptase polymerase chain reaction (RT-PCR) was used to confirm the microarray data in 10 patients with sufficient RNA. Using these 10 patients and another 10 patients selected randomly from 40 newly enrolled patients as the training group, the RT-PCR status of the confirmed genes was used for predicting good versus poor survival. Finally, the prediction model was tested in the remaining 30 newly enrolled gastric cancer patients.
RESULTS: A survival prediction model consisting of three genes (CD36, SLAM, PIM-1) was developed. This model could correctly predict poor or good survival in 23 (76.7%) of 30 newly enrolled patients, and yielded a specificity of 80% and a sensitivity of 73.3%. The survival rate of the patients predicted to have good survival was significantly higher than that of those predicted to have poor survival in the test group as a whole (N = 30; P = .00531) and in stage III patients (n = 16; P = .04467).
CONCLUSION: The semiquantitative RT-PCR gene expression profiling of three genes extracted from microarray study can accurately predict surgery-related outcome in gastric cancer patients.
INTRODUCTION
Gastric cancer is one of the most frequent cancers in the world, and it is the fourth most common in Taiwan. Early-stage disease tends to be diagnosed in countries in which endoscopic screening is more common; however, there still exist patients whose tumor is advanced at the time of diagnosis.1 Patients with stage I disease have a good prognosis, and those with stage IV disease show a poor prognosis. Bewilderingly, the prognosis varies widely in patients with stage II or III disease for as of yet undetermined biologic reasons.2
Traditional clinicopathologic factors and several interesting molecules, including cell cycle regulation factors such as p27 or cyclin E,3,4 cell adhesion molecules such as E-cadherin,5 angiogenic factors such as vascular endothelial growth factor and placenta growth factor,6,7 oncogenes such as c-erbB2 and c-myc,8 and tumor suppressor genes such as p53,8 have been reported to correlate to the prognosis of gastric cancer patients. However, there exists inconsistency among different studies, and the reported parameters provided limited information about prognosis of individual patients because of complex biology of the disease.9 The cellular and molecular heterogeneity of gastric cancers and the large number of genes potentially involved in the multi-step process of gastric cancer pathogenesis emphasize the importance of studying multiple genetic alterations in concert.
Recent advances in the cDNA microarray technique that can investigate gene expression systematically enable us to visualize gene expression profiles in human tumors, and those gene expression profiles can help to identify gene activity patterns that can distinguish subclasses of gastric tumors. Gene profiling studies were also used to better stratify and select patients for adjuvant therapies who may be at higher risk for recurrence.10-13 Recently, Gordon et al14 have shown that simple patterns of gene expression levels, using as few as four to six genes selected from microarray, are highly accurate in the outcome prediction of methothelioma.
Unfortunately, practical methods for identification of individuals with gastric cancer who are at risk for recurrence after surgical resection are not currently available. To explore the possibility of using a limited gene analysis as a predictor of outcome in resected gastric cancer and to find out genomic information related to the survival of gastric cancer patients, an in-house nylon membrane cDNA microarray with colorimetric detection system, which includes 328 genes known as cell cycle regulation factors, matrix proteinases, cell adhesion molecules, angiogenic factors, immunity related factors, oncogenes, and tumor suppressor genes, was constructed to characterize discriminative gene expression for survival, and to develop a quantitative model consisting of selected gene profile which is capable of predicting outcome of gastric cancer patients after curative resection.
MATERIALS AND METHODS
Summary of Experimental Design
Eighteen tumor and healthy pairs of gastric tissue samples were obtained from 18 patients with gastric cancer who underwent D2 gastrectomy without gross residual tumor at the National Taiwan University Hospital (Taipei, Taiwan) between 1995 and 2002. The tumor stage ranged from stage I to stage IV. Nine patients who died as a result of tumor recurrence within 12 months after surgery were identified as poor survival, and the other nine patients, who survived beyond 30 months after surgery were identified as good survival. The poor-survival group included two patients with stage II disease, four with stage III, and three with stage IV. The good-survival group included three patients with stage I disease, two with stage II, and four with stage III. There was a balance of stage II and III between survival groups. There was no stage I patient in poor-survival group, and no stage IV patient in the good-survival group. All patients did not receive postoperative chemotherapy and radiotherapy. Pair samples of tumor and healthy tissues of these 18 patients were dissected and frozen within 30 minutes of removal in a liquid nitrogen tank. Healthy mucosa samples were taken from areas of grossly normal mucosa located at least 3 cm away from tumor border. RNAs of these 18 pairs of the specimens were extracted and were used to perform microarray hybridizations; microarray data of these 18 pairs of samples were used to select gene profiles supervised by poor and good survival through a statistical algorithm. After genes characterizing survival were selected, semiquantitative reverse transcriptase polymerase chain reaction (RT-PCR) analysis of the selected genes was used to verify microarray data in 10 patients with sufficient RNA; thereafter, the genes with RT-PCR results consistent with microarray data were used for further data training and prediction model selection. In addition to these 10 patients, RT–PCR analyses of the selected genes characterizing survival were also performed in 40 newly enrolled patients. We assigned the RT-PCR status of the selected genes into categoric variables that were used in a logistic regression model to predict survival based upon 20 patients (10 patients with good survival; 10 patients with poor survival). To avoid the overfitting problem, these 20 patients (training samples) included 10 patients who had sufficient RNA to confirm microarray data and another 10 patients who were selected randomly from 40 newly enrolled patients. To assess the robustness, 1,000 possible sets of 20 training samples were generated. A stepwise approach, based on Akaike's information criterion (AIC), was used for model selection. A candidate prediction model was selected by using the smallest AIC of the stepwise process in each set of generated training samples. Then, the final effective prediction model, selected from 1,000 candidate prediction models, was the one with the largest AIC among 1,000 smallest AIC. RT-PCR analyses of the selected genes characterizing survival were then performed in the 30 remaining newly enrolled patients (ie, the test group), which included seven patients with stage I, five with stage II, 16 with stage III, and two with stage IV. The RT-PCR status of 30 patients of the test group was translated into the aforementioned categoric variables to test the predictive value of the model extracted from training group. There was no postoperative chemotherapy and/or radiotherapy for patients of either training or test group in this study. This is our standard practice for managing gastric cancer patients who underwent D2 gastrectomy without gross residual tumor in this hospital. Studies using human tissues were approved by and conducted in accordance with the policies of the institutional review board at National Taiwan University Hospital.
Microarray
We used an in-house nylon membrane cDNA microarray made of nylon membrane containing 384 spots and aligned by 16 spots in each row and 24 spots in each column with a spot diameter of 250 μM. It contains sequence-verified cDNA clones of 328 selected known human genes that are considered to be cancer related. These include oncogenes, tumor suppressor genes, apoptosis-related genes, matrix proteinase genes, angiogenesis-related genes, immune-related genes, and so on. The internal control genes included 16 plant genes and GAPDH. This cDNA microarray system and membranes (Boehringer, Mannheim, Germany) were prepared as previously described.15 These 328 sequence verified cDNA clones served as the hybridization targets.
RNA Extraction, Biotinylated Probe Preparation and Microarray Hybridization
Total RNAs were extracted with Trizol reagent (Invitrogen Life Technologies Inc, Carlsbad, CA). Thirty micrograms of the total RNAs derived from each gastric cancer tumor tissue and corresponding nontumor part were reversed transcribed and labeled with biotin. The microarray carrying the double-stranded cDNAs was prehybridized in 1 mL of hybridization buffer (5x RNA extraction standard saline citrate [lsqb]SSC[rsqb], 0.1% N-lauroylsarcosine, 0.1% sodium dodecyl sulfate [lsqb]SDS[rsqb], 1% blocking reagent mixture manufactured by Roche Molecular Biochemicals (Mannheim, Germany), and salmon-sperm DNA [50 μg/mL]) at 63°C for 1.5 hour. The biotin-labeled cDNA probes and hybridization solution (13 μL) containing human COT-1 DNA instead of salmon-sperm DNA were sealed with a microarray in a hybridization bag, and the bag was incubated at 63°C for 10 hours. The membrane was then washed with 2x SSC containing 0.1% SDS for 5 minutes at room temperature followed by three washes with 0.1x SSC containing 0.1% SDS at 63°C for 15 minutes each. After hybridization, the color reaction was initiated by incubating the membrane for 1 hour in 1 mL containing alkaline phosphatase-conjugated streptavidin (diluted 1:1,000; Roche Molecular Biochemicals), 4% polyethylene glycol (molecular weight, 8,000; Sigma Chemical Co, St Louis, MO), and 0.3% bovine serum albumin (BSA) in 1x phosphate-buffered saline (PBS) buffer. The substrate used for color development is BCIP/NBT (5-bromo-4-chloro-3-indolyl-phosphate/ nitro blue tetrazolium) substrate (Promega, Madison, WI). Color reaction was stopped with 1x PBS buffer that contains 20 μM EDTA.
Image Processing
After color development, the membrane was scanned and images taken by using a flat-bed scanner (UMAX Fremont, CA MagicScan at 3,000 dpi). This scanner provided 3,000-dpi resolution and was suitable for our arrays with 384 elements and the resulting database was stored in a tagged image file format (Tiff). To quantify the expression levels of the genes, the digitized images were measured by GenePix Pro software program (Axon Instruments, Foster City, CA).
Microarray Data and Statistical Analysis
Eighteen pairs of patient samples with clinical information were submitted for microarray studies. After the experimental processes, the color of a spot yielded by the enzymatic reaction was converted to gray levels, of which the brightness value was assigned to each pixel of that spot ranging from 0 to 256 (from black, through shades of gray, to white). The spots' expression values of the microarray generated by the GenePix Pro 2.0 software by scanning the microarray images with UMAX 6,000 were collected as an Excel file (Microsoft, Redmond, WA). Expression level of each gene on the microarray was defined as the log ratio (base 2) of the spot expression value of each gene of tumor tissue to that of nontumor tissue. Eighteen patients consisting of poor- and good-survival groups were used specifically to characterize gene expression of survival. To identify genes whose expression levels could be used to discriminate patients with good and poor survival, a three-step supervised classification method was used to extract the most significantly differentially expressed genes between these two survival groups (Fig 1). In the first step, to avoid the systematic error at each microarray sample, the nonlinear LOWESS method,16,17 which fits a curve to the log ratios using robust locally weighted regression, was used to normalize the log ratios of the 328 genes to a LOWESS curve fitting through the MA plot. Second, to extract the significantly regulated genes among all 18 microarray samples, the fold-change method18 which defined the threshold of the significance for expression microarrays, was used to define the regulated genes whose fold-change (normalized log ratios) are greater in magnitude than one at each microarray sample. Among all 18 microarray samples, the significantly regulated genes whose fold-changes are greater in magnitude than one for at least two samples among 18 samples were extracted. Third, to extract the most significant differentially expressed genes between the two survival groups, the multiple permutation test17,19 was used to test simultaneously all the significantly regulated genes filtered at step two, and the adjusted P value for each gene was obtained. The permutation test, reflecting the actual random assignment, is to obtain the original P value of each gene by resampling and recomputing the test statistic several times and comparing the test statistics to the initial test statistics. The multiple test is a method to control the probability of producing incorrect test conclusions (false positives and false negatives) for the whole gene set from original P value to the adjusted P value by the modified statistical algorithm. To assess the internal consistency of the 18 samples, the leave-one-out cross validation (CV) method13 was used to generate the 18 CV samples, and the differentially expressed genes were extracted whose adjusted P value was < .05 family-wise error rate for all 18 CV samples.
RT-PCR Analysis for Selected Genes Confirmation
To verify our microarray results and to further clarify the difference in the expression of the selected genes, we carried out semiquantitative RT-PCR for selected gene confirmation20 by using RNA stock of 10 pairs of samples that had been subjected to microarray study, and the remaining 8 had no RNA left. Two micrograms total RNAs was reverse transcribed using Moloney Murine Leukemia Virus Reverse Transcriptase, Random Primer, and other kit reagents (Promega), followed by PCR. The primers for RT-PCR amplification for examining expression of selected marker genes follow. PCR conditions we used were as follows: for CD36 (40 seconds at 94°C, 35 seconds at 53°C, and 40 seconds at 72°C; 30 cycles), forward, 5'ATGTAACCCAGGACGCTGAG3', and reverse 5'TGGGTTTTCAACTGGAGAGG3'; for signaling lypmphocytic activation molecule (SLAM) (40 seconds at 94°C, 35 seconds at 54°C, and 40 seconds at 72°C; 30 cycles), forward, 5'CCTTCGTGCTGTTTCTCTCC3', and reverse 5'GCTCACGGTGCAGATGTAGA3'; for IGF-1 (40 seconds at 94°C, 35 seconds at 59°C, and 35 seconds at 72°C; 25 cycles) forward, 5'CATTGCTCTCAACATCTCCC 3', and reverse 5'ACGAACTGAAGAGCATCCAC3'; for PIM-1 (40 seconds at 94°C, 35 seconds at 57°C, and 37 seconds at 72°C; 25 cycles) forward, 5'GCTCGGTCTACTCAGGCATC3', and reverse 5'GTCCGTGTAGACGGTGTCCT3'; for transcription factor AP-2 alpha (TFAP; 40 seconds at 94°C, 35 seconds at 51°C, 35 seconds at 72°C, 30 cycles) forward, 5'TGGACCCTGGAAAGATTTTG3', and reverse, 5'CCAGGAAACGGAGGTTGTAG3'; for TIMP-4 (40 seconds at 94°C, 35 seconds at 53°C, and 40 seconds at 72°C; 30 cycles) forward, 5'TGCTGTCAAACCACCTTCTG3', and reverse 5'TGGGCAACAAAGGACTATGA3'. PCR products were separated by electrophoresis on 1.5% agarose gels and visualized under UV light after ethidium bromide staining. We determined the mean band densities using National Institutes of Health Image 1.62 software, and we calculated levels of selected genes relative to ?-actin gene. The relation between microarray expression ratio and RT-PCR results of six selected genes were determined.
Translation of Semiquantitative RT-PCR Data Into Survival Prediction Model
To establish a survival prediction model for gastric cancer patients, the differentially expressed genes whose consistent rates between microarray and RT-PCR results were greater than 60% or Spearman rank correlation coefficient showed significant and P < .05 were selected for prediction model training. Then, we assigned one of four categories to RT-PCR expression data for each sample. The frequencies of the four categoric RT-PCR results at each predictor in the training group were used to establish the prediction model. In the testing samples, each predictor was assigned the same frequencies of the four categoric PCR results as those obtained in the training group. Next, we applied the logistic regression together with stepwise model selection to find the effective prediction model by the AIC.21
Finally, we applied this prediction model to a test group of 30 pair samples and assessed the accuracy of the model by evaluating the sensitivity, specificity, positive prediction value, and negative prediction value on the test group as a whole and the stage III patients. The Kaplan-Meier survival model was also used to estimate the survival of patients stratified by this prediction model. The log-rank test was used to statistically assess significance between survival curves. All differences were considered statistically significant if the P value was < .05.
RESULTS
Gene Selection, Confirmation, and Prediction Model Selection
The three-step supervised classification method was applied to the 18 patients consisting of poor- and good-survival groups to search for gene expression profile of survival (Fig 2). At the normalization step, we used 328 genes to normalize the log ratios at each sample. One hundred forty-one genes are extracted with fold-change method after normalization of log ratios. Finally, the multiple permutation test together with CV was used to select the six most significant differentially expressed genes. The selected genes included CD36 antigen, SLAM, TFAP, insulin-like growth factor 1 (IGF-1), PIM-1 oncogene, and tissue inhibitor of metalloproteinase-4 (TIMP-4).
Semiquantitative RT-PCR was performed to test the expression of these six genes in 10 pairs of tumor and nontumor tissues. The results of RT-PCR were compared with those of microarray. Four of the six aforementioned genes (CD36, SLAM, TFAP, PIM-1), whose consistent rates were > 60% and Spearman rank correlation coefficient showed significance and P < .05 (Fig 3), were picked up for model training.
Duplication of Microarray Experiments in Four Patients
For quality control, we also tested the consistency of microarray experiments in four patients. These four patients were chosen randomly for duplicate study from 18 patients to test the reproducibility of this in-house nylon microarray. Total RNAs of these four patients were hybridized to two different microarray nylon filters in different time, and the Pearson correlation coefficients across hybridizations was calculated and were all > 0.75 (P < .05), which meant good repetition of the experiments.
Translation of Expression of CD36, SLAM, TFAP, and PIM-1 Into Prediction Model of Survival
Four types of paired RT-PCR status for each selected genes could be identified in all blotting results. We assigned one of the following four categories to RT-PCR expression data for each sample: Tumor Healthy, Healthy Tumor, Tumor and Healthy equally positive, and Tumor and Healthy equally negative (Fig 4). The frequencies of the four categoric RT-PCR results at each predictor in the training group were used to establish the prediction model. In the testing samples, each predictor will be assigned the same frequencies of the four categoric PCR results as those obtained in the training group. The logistic regression together with stepwise model selection was used to select the prediction model by the AIC criteria, and the most effective logistic model consisting of three genes (CD36, SLAM, and PIM-1) was obtained as follows:
where is the probability of "poor survival status." Good survival (defined as survival time > 30 months) was predicted when 0.5. Poor survival (defined as survival time < 12 months) was predicted when >0.5. The SEs of the logistic regression coefficients are 0.411 for CD36, 0.436 for SLAM, and 0.173 for PIM-1.
Survival Prediction for an Independent Test Group of 30 Gastric Cancer Patients
The prediction model consisting of CD36, SLAM, and PIM-1 was applied to an independent test group of 30 patients. Survival of 23 patients (76.7%) was correctly predicted, and yielded a specificity of 80%, a sensitivity of 73.3%, a positive prediction value of 75%, and a negative prediction value of 78.57%. The frequency distribution is shown in Table 1. This reveals that this prediction model showed highly predictive power in the independent test group. The survival rate of the patients predicted to have good survival was significantly higher than that of the patients predicted to have poor survival (P = .00531; Fig 5A).
Of the seven stage I patients, the survival of six was correctly predicted by this model. One patient was predicted as having poor survival, and died of multiple liver metastases in 12 months. Of the other six patients, the survival of five was predicted correctly, and the frequency distribution is shown in Table 2. Of the five stage II patients, the survival of three was predicted correctly. Two of three patients predicted to have poor survival died of disease in 12 months, and the frequency distribution is shown in Table 3. Two stage IV patients were predicted correctly by this model, and the frequency distribution is shown in Table 4.
Survival Prediction for the Patients With Stage III Gastric Cancer
The prediction model was applied to 16 patients with stage III disease, and the frequency distribution of accuracy is shown in Table 5. Twelve patients (75%) were predicted correctly, and yielded a specificity of 100%, a sensitivity of 63.6%, a positive prediction value of 100%, and a negative prediction value of 55.6%. The survival rate of the patients predicted to have good survival was significantly higher than that of patients predicted to have poor survival (P = .04467; Fig 5B).
DISCUSSION
cDNA microarray is a powerful tool to monitor global transcription profile of hundreds to tens of thousands of genes at once.10-12 This technology has a strong impact not only on the basic biology but also on clinical practice, and it provides a powerful tool to discover new potential targets against carcinogenesis and metastasis.22,23 It is also anticipated to render more accurate disease diagnosis and more appropriate therapy. Several studies have reported that DNA microarray could identify previously undetected and clinically significant subtypes, and also been used to formulate a molecular predictor of survival after chemotherapy for diffuse large-B-cell lymphoma.10,24 cDNA microarray can also provide a strategy for selecting patients who would benefit from adjuvant therapy and gene expression profile might outperform all currently used clinical parameters in predicting disease outcome in breast cancer.13 Subsequently, these investigators have successfully tested outcome predictor models in independent samples,13,24,25 but those studies continue to be restricted in their clinical applicability because they rely on data gathered from relatively large numbers of genes, costly data acquisition platforms (eg, microarrays), and sophisticated algorithms and/or software and are unable to analyze a sample independently and without reference to other samples. In this study, an in-house nylon membrane mini-microarray containing 328 known genes with colorimetry detection was used to seek predictive gene expression profile for the survival of patients with gastric cancer. We demonstrated that gene expression patterns could more accurately predict survival of gastric cancer patients after curative resection than traditional staging while successfully avoiding many of the shortcomings that preclude the use of microarray techniques in clinical applications.10,26
Initially, six genes were selected from microarrays in this study. They were SLAM, CD36, PIM-1 oncogene, TFAP, IGF-1, and TIMP-4. After confirmation by using semiquantitative RT-PCR, SLAM, CD36, PIM-1 and TFAP were included in prediction model training because of their high consistency rate with microarray data. Finally, the most effective prediction logistic regression model defined by SLAM, CD36, and PIM-1, which can predict the survival of gastric cancer patients with reasonable efficacy, was achieved. Among those three genes, SLAM is a CD2-related surface receptor expressed by activated T cells, B cells, and dendritic cells. Th0/Th1 immunity, which is usually impaired in gastric cancer patients, could also be induced by SLAM to enhance proliferation and cytotoxic ability of CD8+ tumor-specific lymphocytes.27-29 Although the real role of SLAM in tumor-associated immunity of gastric cancer patients remains unclear, it seems to have potential effects on anti-tumor immunity and deserves further investigation. CD36 is a trans-membrane receptor that regulates apoptosis and angiogenesis in response to its ligand thrombospondin-1 (TSP-1). TSP-1 is localized to tumor-associated extracellular matrix, and CD36 expressed on surface of tumor cells.30 The regulation of CD36 expression in tumor cells may play an important role in tumor growth, metastasis and angiogenesis.31 CD36 overexpression, which decreased stromal vascularization by induction of apoptosis of vascular endothelial cells, is correlated with better prognosis of colon cancer.32 However, there existed conflicting reports in the association of CD36 expression with angiogenesis and patient survival in head and neck cancer.33 PIM-1, a product of serine/threonine kinase which can be induced in gastric epithelial cells by Heliobacter pylori infection, may be involved in gastric carcinogenesis.34 PIM-1 also plays an important roles in proliferation, differentiation and maturation of T cells, which may be associated with tumor immunity.35 PIM-1 induced by hypoxia is involved in drug resistance and tumorigenesis of solid tumor cells and leads to genomic instability.36 It is also required in endothelial cells for vascular endothelial growth factor-A–dependent proliferation and migration.37 Recently, the expression of PIM-1 has been shown to correlate significantly with measures of clinical outcome in prostate cancer.38 The three genes selected in our prediction model may be, in some way, involved in tumor angiogenesis and tumor immunity, which is closely related to patient survival. However, a wide range of other biologic processes and complex interactions among different cell types in cancer tissue should also modulate cancer biology and result in heterogeneity of disease. None of these genes has been reported to correlate with clinical outcome in gastric cancer. Further functional and clinical studies of these genes are worthwhile, and it is more likely that these genes in isolation have only limited predictive power, which highlights the need for a genomic approach based on the expression of multiple genes in our attempts of disease subtyping.
To our knowledge, this is the first study to use gene expression profiling techniques to develop a survival prediction model for human gastric cancer. In our study, a data acquisition platform of semiquantitative RT-PCR was developed and used to validate this model in an independent cohort. The use of RT-PCR–based gene expression patterns to predict outcome in gastric cancer patients overcomes several major obstacles that hinder the clinical use of microarray. Unlike other widely accepted supervised learning techniques with similar predictive accuracy,10,13,24-26 this prediction model utilizes data from RT-PCR status to create categories of expression among the three genes in paired samples. Because of its independence of microarray platform after gene selection, our methodology requires only small quantities of RNA (as little as 2 μg when using RT-PCR) and can be performed easily in a common laboratory. In this study, microarray data normalized with the LOWESS method to avoid the systematic error within each microarray sample that might be overly normalized. Therefore, verification of selected genes with RT-PCR is essential in microarray study. The overfitting problem is a crucial issue in the selection process of the prediction model. We used the randomly generated samples to overcome this potential pitfall and established the three-gene prediction model, which had better sensitivity and specificity than the models with one or two genes.
The prognostic tool described herein could dramatically influence the current clinical treatment of gastric cancer patients with stage III disease by allowing the identification of the patients who are unlikely to respond to conventional surgical treatment only. Adjuvant chemotherapy has been reported to have marginal effect for overall gastric cancer patients with D2 gastrectomy (gastrectomy with D2 lymphadenectomy). If survival outcome of gastric cancer patients can be reasonably predicted, adjuvant therapies may help those patients who have high probability of poor survival, whereas those who have high probability of good survival can be spared from the adverse effects of adjuvant therapies. Although whether the patients who are predicted to have high probability of poor outcomes really benefit from adjuvant therapies remains unknown, the results of this work, if confirmed prospectively in a larger patient population, could be helpful in the development of meaningful clinical trials of new compounds or regimens in advanced gastric cancer patients after surgical resection, especially in those with stage III disease, by helping to select suitable candidates.
In the coming era of personalized medicine, we believe that molecular profiling attempts like this study will be the basis of individualized therapy of gastric cancer. In conclusion, this report demonstrated that the survival of individual gastric cancer patients can be predicted with reasonable accuracy by using a simple RT-PCR–based gene expression profile and a logistic regression model. Our methodology is one of the applicable models and might be important to both patients and their doctors in terms of life planning and therapeutic strategies.
Authors' Disclosures of Potential Conflicts of Interest
The authors indicated no potential conflicts of interest.
Acknowledgment
We thank Professor Chien C. Chang for his helpful discussion and comprehensive suggestion in preparing this manuscript. We also thank research assistant Su-I Chen for her excellent technique.
NOTES
Supported by grant from National Science Council and Department of Industrial Technology, Ministry of Economic Affairs, Taipei, Taiwan.
F.J.H. and K.J.C. are joint senior authors.
Terms in blue are defined in the glossary, found at the end of this issue and online at www.jco.org.
Authors' disclosures of potential conflicts of interest are found at the end of this article.
REFERENCES
Akoh JA, Macintyre IM: Improving survival in gastric cancer—Review of 5-year survival rates in English language publications from 1970. Br J Surg 79:293-299, 1992
Allgayer H, Heiss MM, Schildberg FW: Prognostic factors in gastric cancer. Br J Surg 84:1651-1664, 1997
Mori M, Mimori K, Shiraishi T, et al: p27 expression and gastric carcinoma. Nat Med 3:593, 1997
Akama Y, Yasui H, Yokozaki H, et al: Frequent amplification of the cyclin E gene in human gastric cancer. Jpn J Cancer Res 86:617-621, 1995
Graziano F, Mandolesi A, Ruzzo A, et al: Predictive and prognostic role of E-cadherin protein expression in patients with advanced gastric carcinomas treated with palliative chemotherapy. Tumour Biol 25:106-110, 2004
Tanigawa N, Amaya H, Matsumura M, et al: Correlation between expression of vascular endothelial growth factor and tumor vascularity, and patient outcome in human gastric carcinoma. J Clin Oncol 15:826-832, 1997
Chen CN, Hsieh FJ, Cheng YM, et al: The significance of placenta growth factor in angiogenesis and clinical outcome of human gastric cancer. Cancer Lett 213:73-82, 2004
Sanz-Ortega J, Steinburg SM, Moro E, et al: Comparative study of tumor angiogenesis and immunohistochemistry for p53, c-erbB2, c-myc and EGFr as prognostic factors in gastric cancer. Histol Histopathol 15:455-462, 2000
Zheng L, Wang L, Ajani J, et al: Molecular basis of gastric cancer development and progression. Gastric Cancer 7:61-77, 2004
Golub TR, Slonim DK, Tamayo P, et al: Molecular classification of cancer: Class discovery and class prediction by gene expression monitoring. Science 286:531-537, 1999
Alizadeh AA, Eisen MB, Davis RE, et al: Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling. Nature 403:503-511, 2000
Okabe H, Satoh S, Kato T, et al: Genome-wide analysis of gene expression in human hepatocellular carcinomas using cDNA microarray: Identification of genes involved in viral carcinogenesis and tumor progression. Cancer Res 61:2129-2137, 2000
van't Veer LJ, Dai H, van de Vijver MJ, et al: Gene expression profiling predicts clinical outcome of breast cancer. Nature 415: 530-536, 2002
Gordon GJ, Jensen RV, Hsiao LL, et al: Using gene expression ratios to predict outcome among patients with mesothelioma. J Natl Cancer Inst 95:598-605, 2003
Chen JJ, Wu R, Yang PC, et al: Profiling expression patterns and isolating differentially expressed genes by cDNA microarray system with colorimetry detection. Genomics 51:313-324, 1998
Cleveland WS: Robust locally weighted regression and smoothing scatterplots. J Am Stat Assoc 74:829-836, 1979
Yang YH, Buckley MJ, Dudoit S, et al: Comparison of methods for image analysis on cDNA microarray data. J Comput Graph Statist 11:108-136, 2002
White KP, Rifkin SA, Hurban P, et al: Microarray analysis of Drosophila development during metamorphosis. Science 286:2179-2184, 1999
Westfall PH, Young SS: Resampling-Based Multiple Testing: Examples and Methods for P-Value Adjustment. New York, NY, Wiley, 1993
Iizuka N, Oka M, Yamada-Okabe H, et al: Comparison of gene expression profiles between hepatitis B virus-and hepatitis C virus-infected hepatocellular carcinoma by oligonucleotide microarray data on the basis of a supervised learning method. Cancer Res 62:3939-3944, 2002
Akaike H: A new look at statistical model identification. IEEE Transactions on Automatic Control AU-19:716-722, 1974
Bertucci F, Salas S, Eysteries S, et al: Gene expression profiling of colon cancer by DNA microarrays and correlation with histoclinical parameters. Oncogene 23:1377-1391, 2004
Shih JY, Yang SC, Hong TM, et al: Collapsin response mediator protein-1 and the invasion and metastasis of cancer cells. J Natl Cancer Inst 93:1932-1400, 2001
Rosenwald A, Wright G, Chan WC, et al: The use of molecular profiling to predict survival after chemotherapy for diffuse large-B-cell lymphoma. N Engl J Med 346:1937-1947, 2002
Beer DG, Kardia SL, Huang CC, et al: Gene-expression profiles predict survival of patients with lung adenocarcinoma. Nat Med 8:816-824, 2002
Shipp MA, Ross KA, Tamayo P, et al: Diffuse large B-cell lymphoma outcome prediction by gene expression profiling and supervised machine learning. Nat Med 8:68-74, 2002
Henning G, Kraft MS, Derfuss T, et al: Signaling lymphocytic activation molecule (SLAM) regulates T cellular cytotoxicity. Eur J Immunol 31:2741-2750, 2001
Nakayama H, Kitayama J, Muto T, et al: Characterization of intracellular cytokine profile of CD4(+) T cells in peripheral blood and tumor-draining lymph nodes of patients with gastrointestinal cancer. Jpn J Clin Oncol 30:301-305, 2000
Lee KY, Goedegebuure PS, Linehan DC, et al: Immunoregulatory effects of CD4+ T helper subsets in human melanoma. Surgery 117:365-372, 1995
Albo D, Shinohara T, Tuszynski GP: Up-regulation of metalloproteinase 9 by thrombospondin 1 in gastric cancer. J Surg Res 108:51-60, 2002
Chen M, Pych E, Corpron C, et al: Regulation of CD36 expression in human melanoma cells. Adv Exp Med Biol 507:337-342, 2002
Tsuchida T, Kijima H, Tokunaga T, et al: Expression of the thrombospondin 1 receptor CD36 is correlated with decreased stromal vascularisation in colon cancer. Int J Oncol 14:47-51, 1999
Wang TN, Qian XH, Granick MS, et al: The effect of thrombospondin on oral squamous carcinoma cell invasion of collagen. Am J Surg 170:502-505, 1994
Sepulveda AR, Tao H, Carloni E, et al: Screening of gene expression profiles in gastric epithelial cells induced by Helicobacter pylori using microarray analysis. Aliment Pharmacol Ther 16:145-157, 2002 (suppl 2)
Wang Z, Bhattacharya N, Weaver M, et al: PIM-1: A serine/threonine kinase with a role in cell survival, proliferation, differentiation and tumorigenesis. J Vet Sci 2:167-179, 2001
Roh M, Gary B, Song C, et al: Overexpression of the oncogenic kinase PIM-1 leads to genomic instability. Cancer Res 63:8079-8084, 2003
Dhanasekaran SM, Barrette TR, Ghosh D, et al: Delineation of prognostic biomarkers in prostate cancer. Nature 412:822-826, 2002
Valdman A, Fang X, Pang ST, et al: PIM-1 expression in prostatic intraepithelial neoplasia and human prostate cancer. Prostate 60:367-371, 2004(Chiung-Nien Chen, Jen-Jen)