e22065 Background: Cancer patients with wild type EGFR respond to treatment with erlotinib at a lower rate than patients with EGFR mutations. It would be relevant to predict which EGFR wild type patients benefit from erlotinib. We have developed a response predictor based on NCI60 cancer cell lines measurements of erlotinib effect and gene expression measurements. Clinical relevance was improved by filtering the model through expression profiles from more than 3,000 clinical samples.The erlotinib response predictor consists of 94 sensitivity genes and 10 resistance genes and is completely defined before independent validation on patients treated with erlotinib. Methods: Gene expression data from pre-treatment core biopsies from 25 patients with refractory NSCLC (clinicaltrials.gov: NCT00409968) later treated with erlotinib (The BATTLE trial; Kim, ES et al. Cancer Discov. 2011; 1:44-53) were downloaded from Gene Expression Omnibus with accession number GSE33072. PFS after erlotinib treatment was recorded. The erlotinib response predictor was applied retrospectively to the gene expression measurements.Patients with a predicted sensitivity above the 0.15 quantile were categorized as sensitive, patients with a predicted sensitivity below were categorized as resistant.The median was also tested as a cutoff, and correction for multiple testing was taken into consideration. Results: In a Kaplan-Meier analysis of PFS comparing predicted sensitive and predicted resistant patients, the predicted sensitive patients survived longer (PFS 2.3 months (1.9-2.7)) than predicted resistant patients (PFS 1.2 months (0.5-1.8)). A log-rank test for PFS found the difference significant (p=0.005). Hazard ratio 4.6 (1.4-15.2). Conclusions: The small clinical data set validated the cell line derived erlotinib response predictor. It identified a subset of 16% of patients with no apparent benefit from erlotinib. We have performed similar validations for 24 other response predictors in 24 clinical trials. In 20 out of the 24 clinical trials, the agreement between prediction and clinical outcome was statistically significant. Thus, our cell line based method represents a general method for predicting clinical response to cytotoxic or cytostatic agents.
