Survival prediction of stage I lung adenocarcinomas by expression of 10 genes (original) (raw)
Strategy of the integrated approach. The general strategy of our approach is illustrated in Figure 1. Initially, we performed meta-analyses on 2 published expression datasets of lung adenocarcinomas, totaling 170 patients, from studies by Beer et al. (ref. 9; henceforth the Michigan cohort) and by Bhattacharjee et al. (ref. 10; henceforth the Harvard cohort). Patients (Supplemental Table 1; supplemental material available online with this article; doi:10.1172/JCI32007DS1) were divided into good- and poor-prognosis groups according to their clinical outcomes (see Methods). A number of patients, who did not fit the established prognostic criteria, were therefore excluded from the meta-analysis (see Methods). We refer to the datasets from the initial 170 patients as original datasets (Michigan, n = 86; Harvard, n = 84) and to those of the selected patients as reduced datasets or cohorts (Michigan, n = 41; Harvard, n = 60); each of these datasets included patients with stage I, II, and III tumors. The reduced Michigan and Harvard datasets were then analyzed to obtain lists of genes that were differentially expressed between good- and poor-prognosis patients. This led to the identification of a 49-gene prognostic model that exhibited good prognostic value on the Michigan and Harvard cohorts. More importantly, the 49-gene model was a good predictor of prognosis (Figure 1) in a third independent cohort, composed of 34 stage I lung adenocarcinomas (ref. 23; henceforth the Duke cohort).
Strategy of the study. Validation of models as good predictors of prognosis in the Duke and IFOM validation cohorts is indicated. For details, see Results and Methods.
To improve the model, we used a biased cancer signature of 28 genes derived from an experimental model that mimics important cancer-related pathways (25). This signature was by itself predictive of prognosis in the Duke cohort (Figure 1).
Finally, we combined genes from the 2 models. We also added 3 genes identified in the literature as individual prognostic markers for stage I lung adenocarcinoma (26–28). The resulting 80-gene model was tested in a real-time PCR–based approach on a fourth cohort of patients (henceforth the IFOM training cohort) to define a predictive model using a limited number of genes as well as a readily accessible technical platform (Figure 1). By the leave-one-out validation method, we refined the model to a final 10-gene model. The 10-gene model was tested on a fifth independent cohort of patients (henceforth the IFOM validation cohort) and on the Duke cohort (Figure 1).
Meta-analysis of 2 lung adenocarcinoma expression profile datasets. As a first approach, we assumed that a reliable list of genes that are differentially regulated in the good- versus poor-prognosis groups should be concomitantly found in independent analyses of the reduced Michigan and Harvard datasets (Figure 1). Therefore, we performed a class comparison test and identified 361 unique differentially expressed genes (P < 0.05, parametric Student’s t test) in the reduced Michigan cohort and 429 unique differentially expressed genes (P < 0.05, parametric Student’s t test) in the reduced Harvard cohort. Twenty genes were shared between the 2 lists (P < 0.05, parametric Student’s t test). The modest overlap could be explained, at least in part, by the fact that the tumors from the 2 cohorts were analyzed on different array platforms — carrying a substantially different number of genes (Harvard, 9,096; Michigan, 5,588; of which 5,249 common genes were present) — with different protocols. Moreover, individual genetic differences can have an enormous impact on genetic signatures. Thus, the inherent imbalance between conditions (hundreds of patients) and variables (thousands of genes) may generate different signatures. Indeed, a recent study proposed that, to reach an overlap of 50% between 2 lists of prognostic genes, expression profiling studies would need several thousands of patients (29).
In a complementary approach, we assumed that a reliable list of genes should not necessarily be shared by the 2 independent analyses (Figure 1). Hence, we searched for the most stably differentially expressed genes in each dataset, using a stringent P value cutoff (P < 0.001, parametric Student’s t test). We found 21 unique genes in the reduced Michigan cohort and 12 unique genes in the reduced Harvard cohort, for a total of 33 unique genes. In total, by combining the 2 approaches, we identified 49 unique genes (including 4 genes in common between the approaches), which we referred to as the 49-gene model.
Next, we analyzed the prognostic predictive accuracy of the 49-gene model. Predictive accuracy for the reduced datasets was 90% and 72% in the Michigan and Harvard cohorts, respectively (Supplemental Table 2). For the original datasets, predictive accuracy was 69% and 71% in the Michigan and Harvard cohorts, respectively (Supplemental Table 2). Of note, the 49-gene model performed well when compared with the 2 signatures derived by Beer et al. (9) from the Michigan cohort (Supplemental Table 2).
Finally, the performance of the 49-gene model was tested by Kaplan-Meier analysis on stage I adenocarcinomas (Figure 2). The 49-gene model was very effective in predicting overall survival in the stage I patients from both the Michigan and the Harvard cohorts (Michigan, n = 67; Harvard, n = 62; Figure 2 and Supplemental Figure 1A). In addition, when we tested a dataset from a third independent expression profile study, the Duke cohort (23), the 49-gene model proved remarkably effective in predicting prognosis (Supplemental Table 2) and overall survival (Figure 2 and Supplemental Figure 1B).
The 49-gene model predicts overall survival. The 49-gene model was used to predict overall survival in the stage I subset of lung adenocarcinomas from the Michigan (n = 67 of the original dataset’s 86), Harvard (n = 62 of the original dataset’s 84), and Duke (n = 34) cohorts. Data are shown as the probability of survival, in a Kaplan-Meier plot, as a function of a favorable (red line) or unfavorable (green line) signature.
Analysis of an in vitro–derived transcriptional signature. We have previously shown that a biased approach to cancer transcriptomes can lead to the identification of cancer signatures (25). In particular, a 28-gene biased signature was identified (Figure 1) by profiling terminally differentiated myotubes forced to reenter the cell cycle by the viral oncoprotein early region 1A (E1A). The expression of genes from this signature was frequently found to be altered in human neoplasia (25). Thus, we investigated whether the expression of these genes had predictive value in patients with stage I lung adenocarcinomas. We used the dataset from the Duke cohort, because it was the only one for which the expression data for all 28 genes was available. As shown in Figure 3A, the biased signature effectively predicted overall survival, further confirming that a biased approach can lead to the discovery of cancer-relevant signatures (see Supplemental Table 3).
The 28-gene biased signature and the 80-gene model predict overall survival. The 28-gene biased signature (A) and the 80-gene model (B) were used to predict overall survival in stage I lung adenocarcinomas of the Duke cohort (n = 34). Data are shown as the probability of survival, in a Kaplan-Meier plot, as a function of a favorable (red line) or unfavorable (green line) signature.
A 10-gene prognostic model in stage I lung adenocarcinomas. The next step in our experimental approach was to integrate models derived from unbiased and biased screenings. Thus, we combined the 49-gene model and the 28-gene biased signature. We also added 3 genes (SCGB3A1, TERT, and EIF3S6; see Methods and Figure 1) identified in the literature as individual prognostic markers for stage I lung adenocarcinoma (26–28). This set of 80 genes demonstrated excellent predictive power for overall patient survival in Kaplan-Meier analysis of the Duke cohort, the only one for which expression data for all 80 genes were available (Figure 3B and Supplemental Table 4).
The major goal of our efforts, however, was to identify a small number of genes, amenable to analysis using readily available technology (such as real-time PCR), that constitute a prognostic model that can be rapidly transferred to the clinical laboratory. Thus, we used TaqMan Low-Density Arrays (Applied Biosystems) to profile the IFOM training cohort, a set of 25 patients with stage I lung adenocarcinomas (Supplemental Table 1). At the time of our analysis, TaqMan Low-Density Arrays were available for 53 of the 80 genes (Supplemental Table 4). The results are summarized in Supplemental Table 4. From these results, we excluded a number of genes that did not show variability between the good- and poor-prognosis groups; 16 genes were therefore selected for further analysis using cutoff values of P ≤ 0.05 or fold change greater than 1.5 (Supplemental Table 4).
The final prognostic model was obtained by the leave-one-out cross-validation procedure, with independent gene selection (P < 0.05 as cutoff; parametric Student’s t test). We found that on the IFOM training cohort, a 10-gene model (Table 1) displayed a predictive accuracy of 84% (sensitivity, 90%; specificity, 80%) and a P value of 0.004 after 2,000 random permutations of class labels.
To confirm the robustness of this new prognostic model, we used it on the IFOM validation cohort, an independent cohort of 45 stage I lung adenocarcinomas (Supplemental Table 1). Univariate and multivariate analysis showed that the 10-gene model predicted survival of patients more accurately than did tumor stage (IA versus IB), grading, age, sex, or presence of mutated KRAS (Table 2). The 10-gene model was also independent of tumor histological subtype (bronchoalveolar cell carcinoma versus adenocarcinoma; Supplemental Table 5). Kaplan-Meier survival curves showed a significant difference in the survival rate of patients stratified according to the 10-gene prognostic model (P = 0.008, log-rank test; Figure 4). It is also of note that our 10-gene model showed very good predictive power in Kaplan-Meier analysis and multivariate analysis of the Duke cohort, for which microarray expression data for all 10 genes were available (Figure 4, Supplemental Figure 2, and Supplemental Table 6).
The 10-gene model predicts overall survival. The 10-gene model was tested to predict overall survival in the indicated cohorts of stage I (A) and stage IA (B) lung adenocarcinomas. Data are shown as the probability of survival, in a Kaplan-Meier plot, as a function of a favorable (red line) or unfavorable (green line) signature.
Univariate and multivariate analysis of various biological and biochemical parameters.
Importantly, the 10-gene model retained excellent predictive power also when patients with stage IA and IB disease were considered separately (Supplemental Figure 3). In particular, it was able to accurately predict prognosis in stage IA patients from both the IFOM and Duke cohorts. This is relevant, because the 5-year survival rate of stage IA NSCLC patients ranges from 67% to 77% (30–32) after surgery alone. Thus, in this group, molecular tools for prognostic prediction and patient stratification are greatly needed. On the other hand, the 10-gene model did not show predictive power on stage II–III adenocarcinomas (Supplemental Figure 2), possibly suggesting the existence of additional molecular mechanisms, occurring in more advanced lung carcinomas, that might influence the natural history of the tumor. Thus, the sum of our findings indicates that we have identified a prognostic signature specific for stage I lung adenocarcinoma.
One final question concerned the performance of our 10-gene model with respect to other prognostic models in NSCLC. Three prognostic models are described in the literature: a 5-gene model described by Chen et al. (33), and 50- and 100-gene models described by Beer et al. (9). These models were challenged against our 10-gene model on the independent Duke cohort of patients with stage I disease. While all of the models displayed good predictive accuracy, the 10-gene model displayed an overall better performance in terms of accuracy, sensitivity, specificity, and positive and negative predictive values (Table 3).
Comparison of the prognostic predictive accuracy of several prognostic models
A fourth model composed of 134 genes was previously described by Potti et al. (16). This 134-gene “lung metagenes” model is somewhat different from our 10-gene model and from the other models described above. It is composed of several metagenes that are used to partition the samples recursively into smaller groups and predict recurrence through binary classification-tree analysis (16). Consequently, a direct comparison of our 10-gene model with the 134-metagene model on an independent cohort was unfeasible. However, when we compared the prognostic power of the 134-metagene model (Table 3) as described by the authors (16), we found that its overall performance was similar to that of our model, which uses only 10 genes.