A robust multigene expression assay to predict clinical response to chemoradiotherapy

B38

Background: The development of in vitro diagnostics to predict response to therapeutic agents is a central goal of molecular medicine. In this study we validate a multi-gene expression model of radiosensitivity in three prospectively-collected datasets of patients treated with concurrent radiochemotherapy.
 Methods: The model, which was developed in a 48 cell line database is a rank-based linear regression algorithm based on the expression of 10 hub genes. It predicts a continuous radiosensitivity index (RSI), that is based on the survival fraction at 2Gy (SF2), measured for the cell lines in the database. Thus, the index is directly proportional to radioresistance (high index=radioresistance). The linear coefficients were established in the cell lines and were unchanged in the validation experiments. The three independent clinical datasets in this study include: 1. Head and neck cancer (HNC) cohort (n=92) treated with concurrent radiochemotherapy (cisplatin-based) at the Neatherlands Cancer Institute (NKI) 2. Rectal cancer cohort (n=14) treated with preoperative radiochemotherapy at Moffitt Cancer Center (MCC), 3. Esophageal cancer cohort (n=12), treated with preoperative radiochemotherapy (MCC). Gene expression profiles were generated for all three cohorts (Affymetrix U133 Plus in cohorts 2 and 3, NKI array in cohort 1). The clinical endpoint for cohort 1 was recurrence free survival (RFS). In cohorts 2 and 3 the endpoint was pathological response. This was defined by a decrease of a least two T stages (esophageal) or one T stage (rectal) in the primary tumor between the pretreatment clinical stage by endorectal ultrasound and the pathological stage. These definitions were based on the distribution of responses within both cohorts.
 Results: First, we applied the model to the prediction of response to concurrent radiochemotherapy (cohort 2 and 3). The model significantly separated responders (R) from non-responders (NR) (mean RSI, R vs. NR 0.34 vs. 0.48, p=0.002). Importantly, the model was accurate in both disease cohorts despite the small number of patients (rectal cancer, mean RSI, R vs. NR 0.32 vs. 0.46, p=0.03) (esophageal cancer, mean RSI, R vs. NR 0.37 vs. 0.50, p=0.05). To further describe the assay, we generated an ROC curve. Using a threshold RSI value of 0.46, the predictor has a sensitivity and specificity of 80% and 82% respectively, and a positive predictive value (PPV) of 86%. Importantly, 6/8 complete responders had a predicted RSI below the threshold. Finally, we evaluated the model, as a prognostic marker in HNC patients treated with concurrent radiochemotherapy. Interestingly, the average RSI prediction was lower (RSI=0.06) in these patients when compared with rectal (RSI=0.39) and esophagus (RSI=0.43). Importantly, RSI was of prognostic value in this HNC dataset. The predicted radiosensitive group had a strong trend towards improved 2 year Relapse-Free survival (2yr RFS 86% vs. 62%, p=0.06)
 Conclusions: This study provides clinical support for a novel assay to predict response to concurrent chemoradiation. The model was validated in three different disease sites, totaling 118 patients suggesting that it is applicable in multiple disease sites. We argue that this model may play a central role in developing biologically-guided personalized treatment strategies in radiation oncology.