Validation of genomics-based prognostic tests in malignant pleural mesothelioma

Abstract

Purpose: Malignant pleural mesothelioma (MPM) is a highly lethal neoplasm with limited pretreatment prognostication strategies. In this report, we examine the accuracy of a previously proposed prognostic test in an independent cohort of MPM patients. This test uses simple ratios of gene expression levels to provide a novel prognostication scheme.

Experimental design: Gene expression data using high-density oligonucleotide microarrays (approximately 22,000 genes) were obtained for a new cohort of human MPM tumors from patients undergoing similar treatments (n = 39). The relative expression levels for specific genes were also determined using real-time quantitative reverse transcription-PCR. We also used a subset of these tumors associated with widely divergent patient survival (n = 23) as a training set to identify new treatment-specific candidate prognostic molecular markers and gene ratio-based prognostic tests. The predictive nature of these newly discovered markers and gene ratio-based prognostic tests were then examined in an independent group of tumors (n = 52) using microarray data and quantitative reverse transcription-PCR.

Results: Previously described MPM prognostic genes and gene ratio-based prognostic tests predicted clinical outcome in 39 independent MPM tumor specimens in a statistically significant manner. Newly discovered treatment-specific prognostic genes and gene ratio-based prognostic tests were highly accurate and statistically significant when examined in an independent group of 52 tumors from patients undergoing similar treatment.

Conclusions: The data support the use of gene ratios in translating gene expression data into easily reproducible, statistically validated clinical tests for the prediction of outcome in MPM.

Fig. 1

Kaplan-Meier survival predictions for MPM patients. Genes associated with favorable prognosis in MPM were examined in multiple contexts. In one experiment (A-C), previously described MPM prognostic genes and ratio-based tests (12) were examined in the 39 samples of the current study with linked clinical data using microarray data and quantitative RT-PCR. In a separate analysis (D-F), prognostic genes were identified in a training set of samples (Table 2) using current microarray data and examined in additional 26 independent samples with previously published microarray data and linked clinical data. Select ratio-based prognostic tests developed using these genes were subsequently validated in 23 of these 26 samples plus an additional 29 independent samples (i.e., the test set; see Materials and Methods) using quantitative RT-PCR only. A, overall survival for the 39 samples of the current study that were profiled using microarrays. The estimated median survival for this cohort (14 months) is consistent with that expected for MPM patients in general. B, overall survival of patient subsets defined based on the results of supervised clustering using 20 previously described MPM prognostic genes and the 39 samples of the current study. The overall survival of both subsets were found to differ significantly (P = 0.00102). C, overall survival in the 39 current samples for good-outcome (top line, median survival = 33 months) and poor-outcome (bottom line, median survival = 12 months) sample subsets as defined by a previously described four gene expression ratio model were found to differ significantly (P = 0.037). D, overall survival for the test set of 52 samples. The estimated median survival for this cohort (9 months) is consistent with that expected for MPM patients in general though moderately lower than in (A). E, overall survival in patient subsets defined based on the results of supervised clustering using microarray data for 26 independent samples and the prognostic genes from Table 3. The overall survival of both subsets differed significantly (P = 0.0013) and the estimated median survivals differed by nearly 5-fold. F, overall survival in the entire test set of 52 samples for good-outcome (top line, median survival = 12 months) and poor-outcome (bottom line, median survival = 5 months) sample subsets as defined by the same four-ratio model were found to differ significantly (P = 0.0096). Hash marks, censored data.