1186PA NOVEL GENE EXPRESSION SIGNATURE THAT ROBUSTLY PREDICTS THE OUTCOME OF PATIENTS DIAGNOSED WITH STAGE 1 NSCLC

Abstract Aim: Although the management of metastatic lung adenocarcinoma has been profoundly modified by the identification of actionable molecular aberrations, the decision making for early-stage NSCLC still relies on the tumor stage (TNM) only. Stage 1 patients are considered of good prognosis and do not receive adjuvant therapy. However a significant proportion of those patients recur within 5 years after tumor resection. The identification of those high-risk patients through high performing molecular signature would enable to treat them preventively in order to avoid or reduce the rate of recurrence and premature death. Methods: Datasets were selected from public repositories (Director's challenge (DC) n = 402, Rousseaux n = 200, Hou n = 90, Zhu n = 143; Raponi: n = 120; Bhattacharjee; n = 125) upon the following criteria: NSCLC, stage IA-IB, no adjuvant therapy, R0 tumor resection, min 3-year follow-up. Gene expression data from various data sets were normalized using RMA and rescaled. To generate robust predictors of 3-year overall survival, an iterative feature selection framework based on predictive multivariate rules was applied on DC and on a pool of data sets composed of Hou, Zhu and Raponi (Pool) (Development sets). Only highly predictive rules in both datasets were then included in a subsequent predictive model and finally assessed for external validation in Bhattacharjee and Rousseaux (Validation sets). The predictive performances of the models were evaluated by ROC-AUC and by Kaplan Meier curves. Results: Our best model was robust across three independent datasets (all AUC > 0.50, p