Validation of a Proliferation-Based Expression Signature as Prognostic Marker in Early Stage Lung Adenocarcinoma

Despite the introduction of adjuvant chemotherapy, survival rates in early stage non–small cell lung cancer (NSCLC) remain considerably lower than in early stages of any other major cancer type. Surgical resection is the standard treatment of patients with early stage NSCLC. Adjuvant chemotherapy has been recommended as standard of care for stage II and III patients based on results of several randomized studies and most recently summarized in the Lung Adjuvant Cisplatin Evaluation meta-analysis of multiple cisplatin-based trials (1–3). In contrast, no significant treatment benefit could be demonstrated for patients with stage I disease (3, 4), leaving surgery alone as the only curative treatment option. Nevertheless, 5-year survival rates in stage I NSCLC patients are as low as 50% in stage IB and reach only 70% in IA patients (5, 6), indicating a considerable high-risk subpopulation that could benefit from more aggressive therapy. Advances in surgical techniques, for example video-assisted thoracic surgery (VATS) are expected to increase the number of patients with improved performance status and the ability to undergo additional postsurgical treatment options (7–9). In addition, low-dose computed tomographic screening for high-risk populations will lead to increased numbers of patients with early stage diagnoses (10).

Many clinical and molecular factors have been evaluated as possible markers to improve staging and treatment decisions in stage I patients. This includes the use of pathological parameters, such as tumor size, to guide adjuvant treatment in stage IB (4). Histologic (11), immunohistochemistry (12), and genetic markers (13) have been suggested as possible factors for patient stratification. Microarray studies have produced a large number of tumor RNA expression signatures associated with survival and outcome in NSCLC (reviewed in refs. 14 and 15). However, most of these expression profiles, developed from fresh frozen samples, have never been rigorously tested in formalin-fixed and paraffin-embedded (FFPE) tissue samples and their performance on clinical samples remains unknown. Recently, a 14-gene profile was developed on a quantitative PCR platform and tested in 2 large cohorts of nonsquamous, resectable NSCLC showing prognostic separation of low- and high-risk patients (16, 17). However, neither the 14-gene signature nor any other expression profile has gained sufficient acceptance to become standard of care.

In contrast, tumor RNA signatures have been highly successful as prognostic tools in breast cancer. A careful evaluation and comparison of prognostic breast RNA profiles revealed a common component of cell-cycle–regulated mRNAs that contains the major prognostic power of each expression profile (18–21). The expression levels of cell-cycle progression (CCP) genes measure tumor growth irrespective of the underlying genetic aberrations. The prognostic power of proliferation in lung cancer was first shown in early studies with the mitotic index (22) as well as the examination of single gene expression data (23–25). More recently, proliferation has been the focus of 1 of 14 methods applied to the microarray data in the Director's Consortium (26). In addition, a proliferation signature derived from meta-analysis of breast cancer data was successfully applied to lung cancer (20) and the prognostic signature derived by Park and colleagues (27) has a strong cell-cycle–related component. We therefore decided to test an independently developed signature of cell-cycle genes in lung adenocarcinoma. The CCP score was first applied to prostate cancer where it is a strong predictor of death from prostate cancer in biopsies (28) and of biochemical recurrence in post-prostatectomy samples (29). The CCP signature was established as a quantitative PCR assay to accommodate FFPE clinical specimens and implemented in a Clinical Laboratory Improvements Amendments (CLIA)-certified laboratory.

This study investigates the utility of the CCP score as a predictor of survival in patients with early stage lung adenocarcinoma. To focus on the group of patients with the most possible clinical utility, we selected stage I and stage II patients with the goal of identifying low-risk patients who may forego adjuvant treatment and high-risk patients with increased need for more aggressive treatment options. The CCP score was tested retrospectively in 3 large, independent patient sets for its ability to predict death from lung cancer.

Clinical and pathological characteristics of the different patient sets with microarray data tested are detailed in Table 1. A patient set of 442 resectable lung adenocarcinomas was collected at 4 US institutions through the Director's Consortium for the development and validation of gene expression signatures in lung adenocarcinoma (26). Expression values, clinical, and pathology data for the dataset were retrieved from the NCI website https://array.nci.nih.gov/caarray/project/details.action?project.id=182. Patient inclusion criteria, sample processing, and clinical data collection have been described in detail (26). For the complete dataset, pathological stage of each patient according to IASLC 6th edition criteria was determined from available tumor-node-metastasis (TNM) staging variables resulting in 371 stage I and stage II patients. From this subset, 92 patients were excluded due to incomplete clinical data. A total of 19 patients were omitted for unknown progression status, 2 patients for incomplete resection, and a further 2 patients for death within 30 days from surgery. The remaining 256 patients were used for analysis (Supplementary Fig. S1A).

CEL files and patient data of an adenocarcinoma cohort collected at the University of Tokyo are available under GSE31210 from the GEO database. Patient selection, sampling, and processing have been described previously (30). All patients in the GSE31210 set were chemo-naive. From the total dataset of 226 stage stage I and stage II samples, 22 patients with incomplete resection were excluded, leaving 204 patients for analysis (Supplementary Fig. S1A).

Archived FFPE samples from surgically resected lung adenocarcinomas were obtained from 2 institutions: The University of Texas MD Anderson Cancer Center (MDACC) and European Institute of Oncology (IEO). Clinical and pathological characteristics of the different patient sets with FFPE tumors examined are detailed in Table 1. The MDACC patients had been ascertained between 1997 and 2008, were surgically treated, and had complete resection. Initial patient selection criteria were based on staging according to IASLC 6th edition guidelines. Subsequently, tumors were restaged using the current IASLC 7th edition criteria. The total cohort of 294 samples yielded 234 stage I and stage II lung adenocarcinoma patients. Fifteen samples were excluded for reasons related to clinical data: 9 patients were excluded for missing data related to tumor size, adjuvant treatment, and/or pleural invasion; 3 patients were excluded for having less than 1 month of follow-up; and 3 patients were excluded for having synchronous nonlung tumors. Of the remaining 219 samples, 207 produced CCP scores and were included in the analysis. The IEO provided 186 archived FFPE samples from 185 patients diagnosed with stage I lung adenocarcinoma ascertained between 1999 and 2005. All patients had complete surgical resection. Tumor staging followed the 7th edition of IASLC guidelines. One patient was excluded due to death 12 days after surgery and 2 patients were omitted for missing clinical variables. Of the remaining set, 174 samples produced expression scores and were included in the analysis (Supplementary Fig. S1B).

For each dataset, microarrays were subjected to robust multiarray analysis (RMA) for normalization of gene expression. CCP scores were calculated as the average RMA normalized signal of 31 cell-cycle genes. Where multiple probes per gene were present, all probes per gene were averaged first, followed by averaging across 31 genes.

Clinical specimens were analyzed for the expression of 31 cell-cycle and 15 housekeeping genes by quantitative PCR according to a previously published protocol (29) summarized in Supplementary Materials. The performance of housekeeper genes is used to monitor and ensure adequate RNA quality. A list of genes can be found in Supplementary Tables S1 and S2. The success rate for obtaining CCP scores from FFPE samples was 96% in the IEO cohort and 95% in the MDACC patient set.

The CCP score is the unweighted average of 31 cell-cycle genes normalized by the average of 15 housekeeping genes. Each unit change is equivalent to a 2-fold change in expression levels. Details of the CCP score calculation as well as failure criteria for housekeeping genes and cell-cycle genes were described previously (29) and have been summarized in the Supplementary Materials. Calculation for the lung cancer CCP score was identical except centering according to a set of 98 commercial lung adenocarcinomas (Proteogenex and Asterand).

All analyses were carried out with the R (version 2.15.1; The R Foundation for Statistical Computing, Vienna, Austria, http://R-project.org).

In each cohort, clinical data were required to include age of patient at surgery, gender, smoking status, TNM staging, and adjuvant treatment status. Adjuvant treatment was defined as chemotherapy and/or radiation. FFPE cohorts had additional pathological data on tumor size and pleural invasion that were incorporated into modeling. To account for different effects of adjuvant treatment by stage, we included an interaction term for treatment and stage, where applicable. Details on modeling of clinical variables by cohort are provided in Supplementary Materials.

We evaluated the prognostic value of the CCP score using univariate and multivariate Cox proportional hazards models. The primary endpoint was disease-related death within 5 years of surgery. Disease-related death was defined as death following recurrence. Patients who were lost to follow-up or died of other causes were censored at the last observation.

Univariate P-values were based on the partial likelihood ratio. Multivariate P-values were based on the partial likelihood ratio for the change in deviance from a full model (which included all relevant covariates) versus a reduced model (which included all covariates except for the covariate being evaluated, and any interaction terms involving the covariate being evaluated). All reported P-values are two-sided. HRs for the CCP score are per interquartile range of the CCP distribution.

A proportional hazards violation related to processing site was detected in the Director's Consortium cohort, and a proportional hazards violation related to stage was detected in the GSE31210 cohort. To address these violations, univariate and multivariate Cox proportional hazards models were stratified by processing site for the Director's Consortium cohort and by stage for the GSE31210 cohort (Supplementary Materials).

To assess the association of the CCP score with absolute benefit from adjuvant treatment, we used the Zhang and Klein (31) method to construct confidence bands for the difference of 2 survival curves under the Cox proportional hazards model. In addition, we constructed Cox proportional hazards models with contrast coding to test for increased benefit from adjuvant treatment among patients with high CCP scores relative to patients with low CCP scores. CCP scores were categorized as high or low using the median as the cutoff point (Supplementary Methods).

We tested the ability of the CCP score to predict 5-year death from disease after resection with curative intent in public expression microarray data. Because the CCP score had been defined previously, no further gene selection or training in lung cancer was necessary. A total of 256 stage I and stage II lung adenocarcinoma patients with complete clinical data from the Director's Consortium cohort (26) and 204 stage I and stage II patients from a Japanese study on lung adenocarcinoma (30) were analyzed independently. In both cohorts, patients had undergone complete resections and RNA from frozen surgery specimens had been subjected to gene expression analysis on Affymetrix expression microarrays. Twenty-eight percent of the Director's Consortium cohort had died at 5 years, providing 71 events for analysis. Within 5 years from surgery 25 deaths were observed in the GSE31210.

In univariate Cox proportional hazards analysis, the CCP score was significantly predictive of 5-year outcome in both the Director's Consortium cohort (P = 0.00014) and the Japanese patient set (P = 0.0010). The HR per interquartile range of the CCP score was 2.08 (95% CI, 1.43–3.02) in the Director's Consortium cohort and 2.25 (95% CI, 1.42–3.56) in the Japanese dataset. As expected, stage was highly prognostic (P < 0.0001) in the Director's Consortium patients. As shown in Table 2, among clinical variables, only adjuvant treatment in the Director's Consortium cohort was significantly associated with outcome. However, adjuvant treatment did not reduce the risk of death (HR = 2.78, 95% CI, 1.69–4.58). In the Japanese cohort, EGFR and KRAS mutation and EML4-ALK translocation status were available. Neither of the mutation classes nor a combination of mutations contributed prognostic information in this cohort.

The prognostic utility of the CCP score after adjustment for clinical parameters was evaluated in multivariate Cox proportional hazards regression. The CCP score remained the most significant predictor of 5-year disease survival in both microarray datasets (P = 0.0022 for the Director's Consortium cohort, P = 0.0026 for GSE31210) with HRs per interquartile range of 2.02 (95% CI, 1.29–3.17) in the Director's Consortium data and 2.16 (95% CI, 1.32–3.53) in GSE31210. Results from univariate and multivariate Cox proportional hazards analysis are summarized in Table 2.

We tested for an interaction between the CCP score and any of the clinical variables by introducing an interaction term into the model. None of these interaction terms reached significance at the 5% level. Scaled Schoenfeld residuals versus untransformed time were used to evaluate the appropriateness of the proportional hazards assumption. No evidence was found supporting time dependence for the HR of the CCP score. To evaluate the possibility that CCP score might have a nonlinear effect, second- and third-order polynomials for CCP score were tested in Cox proportional hazards models but were not significant at the 5% level.

Kaplan–Meier curves visualize the separation of low- and high-risk patients according to CCP score (Fig. 1A and B). For illustration purposes, the patient sets were divided into equally sized groups based on terciles of the CCP score. For these low, intermediate, and high CCP score patient groups, 5-year survival rates were 84%, 68%, and 56%, respectively, in the Director's Consortium cohort and 97%, 92%, and 70% in the Japanese dataset. Thus, the lowest tercile of CCP score consistently identifies a low-risk subgroup of early stage adenocarcinoma.

To validate the prognostic use of the CCP score in FFPE specimens, surgically resected tumors were obtained from MDACC and IEO (Table 1). The MDACC cohort comprised 204 stage I and stage II specimens with a median follow-up time for patients alive at the date of last follow-up of 132 months. The IEO cohort consisted of 174 stage I patients with a median follow-up time for patients alive at the date of last follow-up of 80 months. At 5 years from surgery, 34 (16%) of patients in the MDACC cohort had died of disease and 28 (16%) deaths had occurred in the IEO patient set. A statistical comparison of the 2 cohorts found no significant differences in the distribution of pathological or clinical parameters (Supplementary Methods). Thus, to improve statistical power, the cohorts were combined for survival analysis.

Each unit of CCP score represents a 2-fold change in mRNA expression. The median CCP score was 0.020, and the interquartile range extended from −0.87 to 0.91. A significant variation in CCP scores was observed in all stages, in particular stage IA and stage IB, with only a minor shift toward higher CCP scores in stage II.

Results from univariate and multivariate Cox proportional hazards analysis are summarized in Table 3. In univariate analysis, the CCP score was the most significant predictor of outcome (P = 0.00033). An increase in CCP score resulted in increased risk of dying from lung cancer with an HR of 2.10 (95% CI, 1.39–3.17). Several clinical parameters also were prognostic, including stage (P = 0.0043), gender (P = 0.0019), tumor size (P = 0.0069), pleural invasion (P = 0.011), and age (P = 0.031). When added to clinical variables in multivariate analysis, the CCP score remained the dominant prognostic factor (P = 0.0071) with an HR of 1.92 (95% CI, 1.18–3.10). Among the clinical factors, only pleural invasion (P = 0.0090) and gender (P = 0.012) retained significance in multivariate analysis. None of the clinical variables showed a significant interaction with the CCP score. No time dependence for the HR of the CCP score or nonlinear effects of the CCP score was observed.

To illustrate the results, a Kaplan–Meier curve using the same terciles of the CCP score as used for the analysis of the microarray data is presented in Fig. 1C. Five-year lung cancer–specific survival for the low CCP patient group was 92%, decreasing to 79% and 73% in the intermediate and high CCP patients. Heterogeneity analysis between the 3 datasets (Director's Consortium, GSE31210 and MDACC/IEO) did not identify any significant cohort differences. This is supported by the consistent HRs for the CCP score observed the 3 datasets (Fig. 1D).

Low CCP scores are clearly associated with improved outcome and the predicted probability of 5-year death from lung cancer increases as a function of the CCP score. Figure 2A shows the relationship between the risk of lung cancer–related death and the CCP score measured by quantitative PCR in the FFPE cohort. The CCP score provides additional prognostic information within each stage group. This is visualized in Fig. 2A by the wide distribution of CCP scores within each stage and supported by the HR in multivariate Cox proportional hazards regression.

To verify the utility of the CCP score in patients with stage IA and stage IB tumors, we conducted a subanalysis in stage I samples. This subset consisted of 163 patients from the MDACC set and 174 patients from the IEO cohort. Each patient set was divided into 2 equally sized groups based on the median CCP score. A statistically significant separation was observed in Kaplan–Meier survival curves of 5-year lung cancer–related survival between the low and high CCP groups (Supplementary Fig. S2). Cancer-specific survival rates at 5 years were 90% for patients with CCP scores below the median and 77% for patients with scores higher than the median CCP. The effect was similar in the MDACC cohort (low CCP 90%, high CCP 79%) and the IEO patients (low CCP 90%, high CCP 75%).

To explore the possibility of using the CCP score as a predictive tool, we examined the relationship between CCP score and absolute benefit from adjuvant treatment in the MDACC cohort. The 207 patients included 46 patients who had received adjuvant therapy. Most importantly, the treated patient set from the MDACC cohort showed significant improvement (P = 0.030, HR = 0.32) in 5-year survival (Kaplan–Meier estimate 92.25%, 95% CI, 77.70–97.46%) compared to patients not receiving adjuvant treatment (Kaplan–Meier estimate 77.56%, 95% CI, 69.46–83.76%). The number of treated patients in the IEO dataset was insufficient to provide statistical power and adjuvant therapy provided no measurable survival benefit. Therefore, association with treatment benefit was restricted to the MDACC cohort. To evaluate the relationship between CCP score and absolute treatment benefit, we categorized CCP scores as being high or low using the median CCP value as the cutoff point. Cox proportional hazards analysis with contrast coding indicated significantly greater absolute treatment benefit for patients with high CCP scores compared to patients with low CCP scores (P = 0.0060; Fig. 2B). The association between CCP score and absolute treatment benefit maintained significance after adjusting for clinical variables (P = 0.024).

A major goal of biomarker development is directed at improving patient-specific treatment decisions. Tools assisting with the final determination of treatment, especially in “cured” patients have been utilized in diseases such as breast cancer (Oncotype DX or Mammaprint; refs. 32 and 33). Treatment of early stage lung cancer, specifically stage I, has been challenging for physicians as these patients are “cured” and have not demonstrated significant benefit from adjuvant cytotoxic chemotherapy, which carries sometimes significant toxicity. In this article, we tested a previously developed and validated 31-gene expression signature in 3 independent datasets as a prognostic tool in early stage lung adenocarcinoma. Multiple aspects distinguish this study from previously published lung signatures: (i) the testing of a single, predefined gene profile; (ii) the use of cohorts with adequate statistical power; (iii) inclusion of all relevant and available clinical variables in the survival analysis; (iv) the presentation of solid evidence for the utility of the expression score in addition to standard clinical parameters; and (v) the validation of the expression score on a quantitative PCR-based platform suitable for the analysis of routine clinical FFPE specimens.

The data presented here establish the CCP score as a reliable prognostic marker in early stage lung adenocarcinoma. The CCP score effectively identified patient groups of reduced or increased risk of death after surgical resection and added significant prognostic information not contained in clinical variables. Importantly, none of the datasets described in this study had been used in the development of the gene profile. The composition of the signature centers on a characteristic trait of tumor cells. In contrast, other prognostic profiles aggregate the highest ranked genes in survival analysis from a single training dataset into a prognostic score. Inevitably, these lists contain genes that seem correlated with outcome due to statistical noise rather than true biological association. Not surprisingly, many of these signatures lack robustness when applied to additional datasets. In contrast, the CCP signature was previously shown to be a superior prognostic tool in prostate cancer (28, 29).

Expression of CCP genes is elevated in dividing cells and the CCP score likely reflects the proportion of actively dividing cells in the tumor. The CCP score thus provides a measure of the inherent aggressiveness of the tumor that is related to outcome. Proliferation associated genes have been correlated with outcome in lung cancer by individual gene expression (23–25) or immunohistochemistry (22, 34). However, single genes lack robustness as expression markers in clinical specimens where nucleic acid quality is compromised by formalin fixation and may lead to the loss of individual signals. We addressed this issue by creating a signature that uses redundancy to minimize technical variability. In addition, signal averaging reduces biological variability from copy number variations.

Outcomes after resection in stage I and stage II NSCLC remain considerably worse compared to other solid tumor types (35). Multiple randomized trials have shown limited improvement in survival from adding adjuvant platinum-containing systemic therapy after surgical treatment and benefit of adjuvant therapy is largely restricted to stage II and stage IIIA patients (3). No improvement in survival could be shown for stage I patients in a large meta-analysis of 5 trials (3), nor in a randomized controlled trial of stage IB disease treated with adjuvant paclitaxel and carboplatin (4). Thus, clinical TNM staging remains the main determinant of postsurgery treatment (e.g., chemotherapy) for early stage NSCLC and stage I patients remain largely untreated despite 5-year survival rates of less than 70% (6). Most of these deaths are likely due to occult disease at the time of surgery and those cases would be expected to benefit from systemic therapy. Heterogeneity of patients for whom surgery was curative likely contributes to the failure to show a treatment benefit in stage I patients. This highlights the need for prognostic markers to identify patients with an increased risk of death not captured by current clinical evaluation tools. The development of such markers has been the goal of many microarray studies in the lung cancer field and several signatures have been described as prognostic tools in early stage NSCLC. As highlighted in a recent review (14), many of these signatures are not independently validated, failed to show independence from clinical variables, or were not established on platforms suitable for clinical use. A notable exception is the recently published Lung RS score (16), which uses quantitative PCR to examine the expression levels of 14 genes in formalin-fixed tissue. The score was shown to predict overall survival in nonsquamous, NSCLC, separating patients into groups with low, intermediate, and high risk of 5-year mortality in cohorts of mixed stages and in a subanalysis of tumors <2 cm (17). The Lung RS study supports the idea that expression signatures can be established for use in formalin-fixed lung tissue and that they may have similar importance in modifying treatment decisions in lung adenocarcinoma as pioneered in estrogen receptor positive breast cancer. This study distinguishes itself from the 14 gene score by focusing on a gene set that is directly related to a well-established feature of tumor cells. In addition, the choice of outcome measure is more relevant to the intended clinical application. For the purpose of informing treatment decisions in resectable lung cancer, association with overall survival is less informative than prediction of lung cancer–specific survival. In a patient population that is often of advanced age and frequently burdened with comorbidities, potential treatment benefits need to be weighed against therapy side effects in the context of both noncancer-related and cancer-specific mortality.

Adjuvant chemotherapy in NSCLC has limited effectiveness and the estimated absolute benefit in a large meta-analysis was a rather modest at approximately 5% (1, 3). Therefore, it is desirable not only to identify patients at high risk of lung cancer-related death, but also those patients with the greatest likelihood of therapeutic benefit. Patients not likely to benefit could be directed to clinical trials or avoid the treatment and its toxicities altogether. Showing the association of a biomarker with treatment benefit requires a cohort with one treated and one untreated arm and a measurable treatment benefit between both arms. In many cohorts, treatment fails to improve outcome. For example, in the Director's Consortium dataset, those patients treated fare worse than untreated patients (14). The dataset from MDACC included patients who had received adjuvant therapy and, more importantly, the treated patients showed a significant treatment benefit. We observed a statistically significant association between treatment benefit and CCP score. This observation requires further confirmation in other datasets, but it suggests that the CCP score may provide predictive information about therapy benefit in addition to its prognostic utility. We recognize that for clinical purposes a risk threshold that guides treatment decisions is useful. However, in this first study in lung adenocarcinoma, we avoided setting predefined thresholds because treatment decisions in resectable NSCLC are subject to other important considerations, most notably stage, but also age, performance status, and preexisting conditions. All of these factors influence the balance between risk of cancer-related death, risk of noncancer-related death, and expected therapy benefit. A clinically meaningful risk threshold for recommending treatment should be drawn when these factors are appropriately integrated. We will address this issue with a combined score in further prospectively collected studies.

In summary, our study shows convincingly that the CCP score is a significant prognostic marker in lung adenocarcinoma that stratifies early stage patients more effectively than conventional clinical parameters. Its CLIA-certified quantitative PCR platform is compatible with clinical practice patterns and may help differentiate patient populations with early stage disease to improve treatment selection.

I.I. Wistuba has commercial research grant in Myriad Genetics, Inc. S. Wagner has ownership interest (including patents) in Myriad Genetics, Inc. E. Hughes is employed (other than primary affiliation; e.g., consulting) as a Research Biostatistician in Myriad Genetics, Inc. J. Reid is employed (other than primary affiliation; e.g., consulting) as an Associate Director Biostatistics in Myriad Genetics, Inc. A. Gutin is employed (other than primary affiliation; e.g., consulting) as a SVP of Bioinformatics in Myriad Genetics, Inc. Z. Sangale has an expert testimony from Myriad Genetics, Inc. J.S. Lanchbury is employed (other than primary affiliation; e.g., consulting) as a Chief Scientific Officer in Myriad Genetics, Inc. No potential conflicts of interest were disclosed by the other authors.

Conception and design: I.I. Wistuba, S. Wagner, S.G. Swisher, A. Gutin, J.S. Lanchbury, M. Barberis, E.S. Kim

Development of methodology: I.I. Wistuba, L. Spaggiari, Z. Sangale, J.S. Lanchbury, E.S. Kim

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): I.I. Wistuba, C. Behrens, F. Lombardi, J. Fujimoto, M.G. Raso, D. Galetta, R. Riley, N. Kahlor, J.S. Lanchbury, M. Barberis, E.S. Kim

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): I.I. Wistuba, F. Lombardi, S. Wagner, D. Galetta, E. Hughes, J. Reid, A. Gutin, J.S. Lanchbury, E.S. Kim

Writing, review, and/or revision of the manuscript: I.I. Wistuba, S. Wagner, J. Fujimoto, L. Spaggiari, D. Galetta, E. Hughes, Z. Sangale, S.G. Swisher, C. Moran, A. Gutin, J.S. Lanchbury, E.S. Kim

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): I.I. Wistuba, F. Lombardi, J. Fujimoto, E.S. Kim

Study supervision: I.I. Wistuba, L. Spaggiari, J.S. Lanchbury

Other: Z. Sangale provided pathologic evaluation of tumor samples.

Myriad Genetics, Inc. paid the costs of shipping the paraffin-embedded samples and performing all qPCR assays. This study was supported in part by the UT Lung Specialized Programs of Research Excellence grant (P50CA70907; to I.I. Wistuba), and the MDACC Support Grant CA016672.

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