Confirmation of Gene Expression–Based Prediction of Survival in Non–Small Cell Lung Cancer

Purpose: It is a critical challenge to determine the risk of recurrence in early stage non–small cell lung cancer (NSCLC) patients. Accurate gene expression signatures are needed to classify patients into high- and low-risk groups to improve the selection of patients for adjuvant therapy.

Experimental Design: Multiple published microarray data sets were used to evaluate our previously identified lung cancer prognostic gene signature. Expression of the signature genes was further validated with real-time reverse transcription-PCR and Western blot assays of snap-frozen lung cancer tumor tissues.

Results: Our previously identified 35-gene signature stratified 264 patients with NSCLC into high- and low-risk groups with distinct overall survival rates (P < 0.05, Kaplan-Meier analysis, log-rank tests). The 35-gene signature further stratified patients with clinical stage 1A diseases into poor prognostic and good prognostic subgroups (P = 0.0007, Kaplan-Meier analysis, log-rank tests). This signature is independent of other prognostic factors for NSCLC, including age, sex, tumor differentiation, tumor grade, and tumor stage. The expression of the signature genes was validated with real-time reverse transcription-PCR analysis of lung cancer tumor specimens. Protein expression of two signature genes, TAL2 and ILF3, was confirmed in lung adenocarcinoma tumors by using Western blot analysis. These two biomarkers showed correlated mRNA and protein overexpression in lung cancer development and progression.

Conclusions: The results indicate that the identified 35-gene signature is an accurate predictor of survival in NSCLC. It provides independent prognostic information in addition to traditional clinicopathologic criteria.

Lung cancer is the leading cause of cancer-related deaths, and non–small cell lung cancer (NSCLC) accounts for almost 80% of deaths (1, 2). Currently, surgery is the major treatment option for patients with stage I NSCLC. However, 35% to 50% of stage I NSCLC patients will relapse within 5 years (3), indicating that a subgroup of these patients might benefit from adjuvant chemotherapy (4). On the other hand, patients with clinical stage IB, IIA, IIB, or IIIA NSCLC receive adjuvant chemotherapy, and some may unnecessarily receive potentially toxic chemotherapeutic treatment (5). It is a critical and unsolved challenge for clinicians to precisely estimate the risk of recurrence in individual patients for appropriate personalized therapy.

The emerging use of biomarkers may enable clinicians to make treatment decisions based on the specific characteristics of individual patients and their tumors (6). There have been significant advances in refining the prognosis of NSCLC by gene expression signatures (3, 7–14), most notably the 5-gene signature from Chen et al. (4) and the 133-gene signature from Potti et al. (5). Gene expression-based diagnosis of lung adenocarcinomas (15) has already been incorporated in clinical settings to treat this deadly disease.

The major histologic types of NSCLC include adenocarcinoma and squamous cell carcinoma. Our previous analysis identified a prognostic 35-gene signature from microarray profiles of 170 lung adenocarcinomas (16). The present study further validates that the 35-gene signature quantifies overall survival in 264 patients with lung adenocarcinoma and squamous cell lung cancer. These microarray profiles and associated clinical information were obtained from three previously published patient cohorts (8, 9, 11). The association between the expression-defined risk groups and lung cancer prognostic factors, including patient age, sex, tumor grade, clinical stage, and tumor differentiation was assessed. The expression of the signature genes was further confirmed by real-time reverse transcription-PCR (RT-PCR) and Western blot analysis of lung cancer tumor specimens.

Patient samples and microarray profiles. The microarray profiles of the patients analyzed in this study included a training set from Beer et al. (n = 86; ref. 3), and three validation sets from Bild et al. (n = 111; ref. 8), Garber et al. (n = 24; ref. 9), and Raponi et al. (n = 129; ref. 11). The histologic groups of the studied patient cohorts include lung adenocarcinoma and squamous cell lung cancer. The clinical characteristics of these patient cohorts are described in Supplementary Table S3. Genes screened on different microarray platforms were matched by their Unigene Cluster IDs or gene names with the interactive MatchMiner (17) Web site interface.⁵

The lung adenocarcinoma tumor specimens used in Western blot analysis were obtained from the Cooperative Human Tissue Network (Ohio State University Tissue Bank). Tumor tissues were collected in surgical resections and were snap-frozen at −80°C until used for protein extraction. Histologic preparations of tumor sections were examined by pathologists. This study was approved with an Institutional Review Board exemption from West Virginia University.

Nearest centroid classification method. The raw microarray data from Bild et al. (8) were obtained from the Duke Web site.⁶

As these microarray data were measured on different platforms, a two-step normalization method was used to convert these data sets into comparable scales. First, the raw microarray data were quantile-normalized with dChip (18). Second, the signature genes were sample-wise- normalized to have a mean value of 0 and a SD of 1. Specifically, for each patient sample, the gene expression g(x) was normalized to [g(x) − mean(x)]/sd(x), where mean(x) is the mean of all the genes measured on this sample in the quantile-normalized microarray data, and sd(x) is the SD of all the genes measured on this sample. After the normalization, the signature genes in the validation sets were identified. In the training set from Beer et al. (3), patients who survived 5 y constitute the good prognosis group (centroid). The average expression value for each signature gene in the good prognosis centroid was computed. In the validation sets, Pearson's correlation coefficient was determined between each tumor sample and the good prognosis centroid in the training set. The cutoff value for patient stratification was determined from Garber's cohort (n = 24; ref. 9). Each tumor sample was classified into the good prognosis group if the correlation coefficient was >0.32; otherwise, it was classified into the poor prognosis group. The same prognostic categorization scheme was applied to Bild's cohort (n = 111; ref. 8).

The cohort from Raponi et al. (n = 129; ref. 11) was retrieved from the Gene Expression Omnibus Web site (GDS2373). The data were randomly partitioned into a training set (n = 65) and a test set (n = 64). In the training set, each tumor sample was classified into the good prognosis group if the correlation coefficient was >−0.15; otherwise, it was classified into the poor prognosis group. The same cutoff was applied to the test set in patient stratification.

Statistical methods. Kaplan-Meier analysis was used to assess the probability of overall survival of two prognostic groups in the studied patient cohorts. To evaluate the association between gene expression–defined risk groups and clinicopathologic parameters in the studied patient cohorts, χ² tests or Fisher's exact tests (two-sided) were used. Differential gene expression was assessed by using t tests. All statistical testing was done with software R.

Evaluation of other NSCLC prognostic signatures. The DNA microarray data from Bild et al. (8), Garber et al. (9), and Raponi et al. (11) were retrieved from the Gene Expression Omnibus Web site. The 5-gene signature from Chen et al. (4) contains DUSP6, MMD, STAT1, ERBB3, and LCK; and the control gene is TBP. To make the levels of gene expression from the microarrays and from RT-PCR comparable, the microarray data were log-transformed to a base-2 scale after assigning a value of 1.1 to intensity values of <1.1. After log transformation, the levels of expression of the five genes were divided by the level of expression of the control gene TBP to calculate the relative level of expression. The decision-tree model was described in the Supplementary Fig. S1 from Chen et al. (4). The overall survival rates of the high- and low-risk groups defined by the 5-gene signature were assessed by Kaplan-Meier analysis.

The probe sets of the 133-gene signature (listed in Supplementary Table S2) from Potti et al. (5) were identified from Bild et al. (8) and Raponi et al. (11). A total of 123 signature genes were identified in each of the validation sets. Garber's cohort was not used in the validation because the samples size (n = 24) is too small compared with the number of signature genes. The mean value of multiple probes was computed to obtain a unique expression value for each gene. There was a convergence problem when fitting these genes in a Cox hazard proportional model. The convergence problem could be caused by overfitting and/or correlated covariates. To solve the problem, we computed the correlation coefficient of all possible pairs of the signature genes. Highly correlated genes were randomly removed from the Cox model of overall survival. Specifically, genes with correlation coefficients of >0.44 (P < 1E−6) were dropped from the Cox model of Bild's cohort; and genes with correlation coefficients of >0.46 (P < 1E−6) were dropped from the Cox model of Raponi's cohort. The Cox model could be fit with =71 genes for Bild's cohort (8) and with =88 genes for Raponi's cohort (11).

RNA extraction. Total RNA was extracted from frozen lung tissue using the RNeasy mini kit according the manufacturer's protocol (Qiagen). RNA was eluted in 30 μL of RNase-free water and stored at −80°C. The quality and integrity of the total RNA was evaluated on the 2100 Bioanalyzer (Agilent Technologies).

Reverse transcription. From each sample, 1 μg of RNA was used to generate complementary DNA using the High Capacity cDNA kit according to manufacturer's protocol (Applied Biosystems).

Real-time RT-PCR. An endogenous control gene TaqMan Low Density Array card (Applied Biosystems) was run on the ABI PRISM 7900HT Sequence Detection System for eight clinical samples to choose a gene that had the most relatively constant expression in the different tissue samples. Three control genes, namely 18S, UBC, and POLR2A had constant expression in the different tissue samples. Constant expression of mRNA for the 18S and UBC genes was also confirmed for all lung tissue samples using the individual TaqMan Gene Expression Assays.

Expression of mRNA for 35 signature genes was measured in each of the lung tissues by real-time PCR using TaqMan Gene Expression Assays on ABI PRISM 7500 HT Sequence Detection System (Applied Biosystems). On each plate, one no-template control was also run. Total RNA samples ran on an Agilent 2100 Bioanalyzer RNA 6000 Nano LabChip.

Protein expression validation using Western blot analysis. Anti-TAL2 antibodies were obtained from Santa Cruz Biotech and anti-ILF3 antibodies were obtained from Abcam.⁷

The protein extraction kit was ordered from EMD. Western blot analysis was done according to the methods described previously (19).

Gene expression-based prediction of lung cancer overall survival. In a previous DNA microarray study (16), we identified a 37-gene prognostic signature from 86 lung adenocarcinomas from Beer et al. (3), and validated the signature in 84 adenocarcinomas from Bhattacharjee et al. (7). Several previously unknown genes now have functional annotations, and two unknown genes are removed from the signature (Supplementary Table S1). In this study, we sought to explore whether this gene signature could also predict overall survival in other major cell types of NSCLC. Three additional cohorts were obtained from Bild et al. (n = 111; ref. 8), Garber et al. (n = 24; ref. 9), and Raponi et al. (n = 129; ref. 11). These cohorts include lung adenocarcinomas and squamous cell carcinomas. The clinical information provided in the patient cohorts includes age, smoking status, tumor stage, grade, differentiation, and overall survival. The signature genes were identified from various microarray platforms with MatchMiner (17).

To substantiate the prognostic prediction of the 35-gene signature, a nearest centroid classification method was used to stratify patients into high- and low-risk groups. In the cohorts from Bild et al. (n = 111; ref. 8) and Garber et al. (n = 24; ref. 9), patients were stratified based on the correlation between the gene expression signature in the tumor sample and the good prognosis centroid in the training set from Beer et al. (3). The cutoff value for patient stratification was determined from Garber's cohort (n = 24; ref. 9). Each tumor sample was classified into the good prognosis group if the correlation coefficient was >0.32; otherwise, it was classified into the poor prognosis group. The same prognostic categorization scheme was applied to Bild's cohort (n = 111; ref. 8). In both validation sets, the gene expression–defined high- and low-risk groups had distinct overall survival (P < 0.05, log-rank tests) in Kaplan-Meier analysis (Fig. 1A and B).

In this study, the validation sets were generated on heterogeneous DNA microarray platforms and the RNA was extracted according to different experimental protocols. The same cutoff as described above did not generate significant patient stratification in Raponi's cohort (n = 129; ref. 11). To avoid overfitting in the prognostic validation, Raponi's cohort was randomly partitioned into a training set (n = 65) and a test set (n = 64). In the training set, each tumor sample was classified into the good prognosis group if the correlation coefficient was >−0.15; otherwise, it was classified into the poor prognosis group. In the training set, the high- and low-risk groups had distinct (log-rank P < 0.01) overall survival in Kaplan-Meier analysis (Fig. 2A). The same cutoff was applied to the test set in patient stratification. In the test set, the high- and low-risk groups had distinct (log-rank P < 0.03) overall survival in Kaplan-Meier analysis (Fig. 2B). These results indicate that the 35-gene signature could stratify patients into high- and low-risk groups in multiple independent cohorts.

The 35-gene signature identified poor and good prognostic subgroups in stage 1A NSCLC. Clinical stage (20) is the most important prognostic factor in lung cancer treatment. In the current practice, surgery is the major treatment option for patients with stage 1 lung cancer. A 5-year survival rate of stage 1 NSCLC was in the range of 40% to 67%, with better results in patients in stage 1A (1). Refined prognostic models are needed to select specific patients in this stage who are at high risk of tumor recurrence for adjuvant chemotherapy.

We sought to explore whether the 35-gene signature could further stratify stage 1A NSCLC into poor and good prognostic groups. In three combined validation cohorts from Bild et al. (8), Garber et al. (9), and Raponi et al. (11), all patients with stage 1A were analyzed. Based on the gene expression–defined prognostic classification as described above, the overall survival rates of the high- and low-risk groups defined in the nearest centroid classification were estimated using Kaplan-Meier analysis. The 35-gene signature was able to further stratify stage 1A NSCLC into high- and low-risk groups with distinct overall survival (P = 0.0007, log-rank tests; Fig. 3). The results indicate that the 35-gene signature provides additional prognostic information for stage 1A NSCLC.

Association between expression-defined risk groups and clinicopathologic parameters. To test whether the 35-gene prognostic signature is independent of traditional criteria, the association of the expression-defined risk groups and clinicopathologic parameters was assessed with χ² tests or Fisher's exact tests (Supplementary Table S7). Based on the available information, there was no significant association between the prognostic signature and patient age (>60 years), tumor differentiation, tumor grade, tumor stage, or sex. These results indicate that the 35-gene signature is independent of traditional clinicopathologic factors in lung cancer prognosis. It should be noted that several clinical parameters are missing in these validation cohorts.

Comparison with other lung cancer prognostic gene signatures. Previous research has established two prognostic gene signatures for lung cancer, namely, the 5-gene signature from Chen et al. (4) and the 133-gene signature from Potti et al. (5). The performance of these two signatures was compared with our identified 35-gene signature in three validation sets. The 5-gene signature was evaluated with the decision-tree model as described in Chen et al. (details provided in Materials and Methods; ref. 4;). In Kaplan-Meier analysis on Raponi's cohort, the difference between the overall survival of gene expression–defined high- and low-risk groups reached borderline significance (log-rank P = 0.06). Different prognostic groups stratified by the 5-gene model did not have distinct overall survival rates in Bild's or Garber's cohorts (Fig. 4).

There were insufficient details about the weighted classification-tree model for the 133-gene signature from Potti et al. (5). In addition, the survival information was not available in the data provided in Potti et al. (5). A total of 123 signature genes were identified from both validation sets in Bild et al. (8) and Raponi et al. (11). To validate the prognostic power of Potti's gene signature, the signature genes were fitted in a Cox proportional hazard ratio. There was a convergence problem when fitting these genes into a Cox proportional hazard model. This problem was solved after correlated genes were removed from the Cox model, suggesting that the 133-gene signature caused an overfitting problem due to correlated or redundant biomarkers within the gene signature. Based on the fitted Cox model, a survival risk score was generated for each patient. The median of these risk scores was used as a cutoff in the prognostic stratification. Patients with a risk score greater than median were defined as high-risk, whereas those with a smaller risk score were defined as low-risk. This stratification partitioned patients into different prognostic groups with distinct (P < 0.0001, log-rank tests; Fig. 5) overall survival in Kaplan-Meier analysis. Due to the lack of details of the classification model in Potti et al. (5), a strict separate training-validation scheme on the 133-gene signature could not be done in this study.

Validation of gene expression and protein expression in lung adenocarcinoma. Having established the clinical relevance of the identified prognostic signature in NSCLC using publicly available DNA microarray data, we further confirmed the expression of the signature genes using real-time RT-PCR assays of eight snap-frozen lung cancer tissue samples (details provided in Materials and Methods as well as Supplementary Materials). We then sought to confirm the protein expression of several signature genes of interest because protein products ultimately play an essential role in cancer development and progression. Tissue lysates of both lung cancer tissues and the adjacent normal tissues were subjected to Western blot analysis, which was used to measure the protein expression of the identified 35 genes. Among the identified 35 genes, only 15 genes have commercially available antibodies to detect their protein products. We first examined the specificity of these antibodies using the cell lysates from a normal lung epithelial cell line (BEAS-2B) and a lung cancer cell line (A549), and found that only six antibodies were able to detect their corresponding protein products (CHD4, GHRHR, ILF3, TAL2, CREB3, and MSX2) in the cell lysates (data not shown). In further analysis with these six antibodies, we found that four of six genes overexpressed their proteins in a lung cancer tumor tissue sample (ILF3, TAL2, CREB3, and MSX2). To validate the specificities of these identified six proteins, each protein was further probed with a different specific antibody that targets a different epitope. The results confirmed that the TAL2 and ILF3 proteins were overexpressed in lung adenocarcinoma tumor tissues compared with the adjacent normal tissues (Fig. 6). Both TAL2 and ILF3 are oncogenic proteins.

Gene expression levels may not be necessarily correlated with protein expression levels. We further investigated the differential gene expression and qualitative correlation with protein expression for the identified lung cancer signature genes. Our previous study (16) analyzed differential expression patterns of the identified signature genes by using two bioinformatics tools, ONCOMINE (21) and SAGE (22). Specifically, the expression level of ILF3 was significantly higher in lung cancer tissues compared with normal tissues (7, 9, 23), higher in metastasis than in primary lung cancer (P < 0.002; ref. 7), and higher in poorly differentiated lung cancer tumors than in well differentiated tumors (P < 0.002; ref. 3). In this study, we further analyzed the differential expression patterns of the signature genes in high-risk versus low-risk groups in the studied lung cancer cohorts (details provided in Supplementary Materials). TAL2 had consistent overexpression in high-risk groups in Beer et al. (P < 0.005; ref. 3), Bild et al. (8), and Larsen et al. (Supplementary Table S10; ref. 14). These results show that several members of our identified gene signature are correlated to the protein expression profiles in lung cancer.

Lung cancer is a dynamic and diverse disease associated with numerous somatic mutations, deletion, and amplification events. Patients with the same stage of disease can have markedly different clinical outcomes. Traditional diagnostic and prognostic factors may stratify patients with molecularly distinct diseases into the same group based on morphologic assessments. It is a critical issue to reliably identify specific patients at high risk of recurrence and metastasis of lung cancer. Molecular prediction is a necessary step in the future direction of personalized cancer care. The Food and Drug Administration has recently approved the first cancer gene test, MammaPrint (24), for treating early stage breast cancer patients with negative lymph node status. Oncotype is another clinically applied gene test for predicting recurrence of tamoxifen-treated, node-negative, and estrogen receptor–positive breast cancer (25). To date, there has been no clinically applied gene test for predicting lung cancer recurrence.

Research in transforming molecular diagnostic and prognostic models into predictive and preventive medicine has become increasingly important and needs to meet important recommendations. According to an alternative REMARK (REporting recommendations for tumor MARKer) system (26, 27), cancer prognostic studies must show whether tumor markers provide information independent of traditional criteria or provide prognostic information within subgroups defined by traditional criteria. In this study, we used previously published DNA microarray data to validate a 35-gene prognostic signature for NSCLC. A nearest centroid classification method based on the correlation of gene expression patterns was used in patient stratification. The overall survival rates of the classified high- and low-risk groups were assessed with Kaplan-Meier analysis and log-rank tests. The patient stratification scheme determined on Garber's cohort (n = 24; ref. 9) also classified patients into distinct (log-rank P < 0.05, Kaplan-Meier analysis) prognostic groups on Bild's cohort (n = 111; ref. 8). Because the validated DNA microarray data were generated using heterogeneous experimental protocols and platforms, a different cutoff value was used in patient stratification on Raponi's cohort (11). To avoid overfitting in the validation, Raponi's cohort was randomly partitioned into a training set (n = 65) and a test set (n = 64). The cutoff value identified in the training set generated significant patient stratification in both training (log-rank P < 0.01, Kaplan-Meier analysis) and test (log-rank P < 0.03, Kaplan-Meier analysis) sets. The results indicate that the 35-gene signature provides useful information for patient stratification in clinical decision-making. The analysis showed that this gene signature could further stratify patients in stage 1A NSCLC into subgroups with distinct overall survival (log-rank P < 0.0007, Kaplan-Meier analysis). The relation of the gene signature to standard prognostic values was investigated. There was no significant association between the 35-gene prognostic signature and patient age, sex, tumor differentiation, tumor grade, or tumor stage. It should be noted that many clinical parameters were not available in the studied cohorts. This is acknowledged as a limitation of the current study.

The 35-gene signature was compared with two other most notable lung cancer gene signatures, namely, the 5-gene signature from Chen et al. (4) and the 133-gene signature from Potti et al. (5). The 5-gene signature was validated based on the decision-tree model reported in Supplementary Fig. S1 in Chen et al. (4). The 5-gene model did not stratify patients into distinct prognostic groups in any of the three validation cohorts. In comparison, the 35-gene signature generated significant (log-rank P < 0.05, Kaplan-Meier analysis) patient stratification on the same data sets, indicating that the 35-gene signature outdid the 5-gene signature from Chen et al. (4). The 133-gene signature (as listed in Supplementary Table S2) from Potti et al. (5) could not be validated because insufficient details were provided for their prognostic model, including patient survival time and the weighted decision-tree algorithm. When fitted in a Cox regression hazard function, the 133-gene signature caused a nonconvergence problem. This problem was solved after the removal of significantly correlated signature genes, suggesting that Potti's gene signature contains correlated and/or redundant biomarkers. In this study, the 133-gene signature from Potti et al. (5) could not be evaluated with other lung cancer signatures because the details of the prognostic model were not provided either through the publication (5) or personal communication.

After we validated the clinical prognostic value of the lung cancer gene signature using public DNA microarray data, we validated the expression of the signature genes using real-time RT-PCR analysis of snap-frozen lung cancer tissue samples. Furthermore, we sought to determine the relevance of our identified signature genes at the protein level. Because the commercially available antibodies to the identified 35 genes are very limited, we have thus far only validated the protein expression for TAL2 and ILF3. Western blot results show that both TAL2 and ILF3 are overexpressed in lung cancer tumor tissues compared with adjacent normal lung tissues. To prove the specificity of the identified proteins, each protein was probed with two different antibodies that recognize different epitopes of the same protein. Both TAL2 and ILF3 are oncogenes, and their protein products are likely involved in the cancer process. Our results indicated that some signature genes have correlated protein expression in lung cancer.

Microarray technologies promise the discovery of novel biomarkers in genome-scale association analysis. Nevertheless, there are several disadvantages that have limited their application into routine clinical tests: (a) microarray tests are labor intensive, time consuming, and expensive; (b) they require a specific system to do the assays, which is not available in many clinical centers; (c) the results can be influenced by each step of the complex assay, ranging from array manufacturing to sample preparation (extraction, labeling, hybridization) and image analysis, which raises the issue of reproducibility; and (d) they usually include a high number of genes and their predictive value needs to be substantially improved before being accepted for routine clinical tests (28–31). Compared with microarrays, real-time RT-PCR is more efficient and consistent. It requires only a small amount of samples and can quantify gene expression in paraffin-embedded tissues. The combined use of real-time RT-PCR with microarray analysis can overcome the inherent biases of the microarray technique and is emerging as the optimal method of choice for genome-scale gene expression analysis (32).

In this study, the 35-gene signature is validated as an independent prognostic factor for NSCLC. The transcriptional profiles analyzed in this study were generated on DNA microarrays, and the gene expression of the identified biomarkers was further confirmed using real-time RT-PCR assays of snap-frozen lung cancer tumor tissues. Several protein products in the signature were also validated in both lung cancer cell lines and lung adenocarcinoma tumor tissues. Our future research will use quantitative RT-PCR techniques to validate the gene expression profiles on a separate patient cohort. A clinical protocol will be developed based on RT-PCR assays for the management of NSCLC patients. Further, more signature genes will be validated in proteomic assays to explore whether this signature is associated with clinical outcome or contributive to tumor development and progression.

No potential conflicts of interest were disclosed.

We thank Dr. Lisa McShane from the National Cancer Institute (Bethesda, MD) for her help in statistical analysis and Drs. Mitch Raponi and Jack Yu from Johnson & Johnson (Radnor, PA) for providing the original microarray data.