663 Correlation between early endpoints and overall survival in non-small-cell lung cancer: a trial-level meta-analysis

Background In clinical trials that assess novel therapeutic agents in patients with non-small-cell lung cancer (NSCLC), early endpoints (e.g. progression-free survival [PFS] and objective response rate) are often evaluated as indicators of biological drug activity, and are used as surrogate endpoints for overall survival (OS). Compiling trial-level data could help to develop a predictive framework to ascertain correlation trends between treatment effects for early (e.g. odds ratio [OR] for PFS at 6 months) and late endpoints (e.g. hazard ratio [HR] OS). Methods A dataset was compiled, which included 81 randomized, controlled trials (RCTs; Phase II–IV) of NSCLC (Stages I–IV), with 35 drugs and 156 observations. The dataset was collected from multiple source databases, including Citeline, TrialTrove, clinicaltrials.gov, and PubMed. We applied random-effects meta-analysis to correlate a variety of treatment effects for early endpoints with HR OS. We performed meta-regression analyses across different data-strata, stratified by the mechanism of action (MoA) of the investigational product (programmed death protein-1/programmed death-ligand 1 [PD-1/PD-L1], epidermal growth factor receptor [EGFR], vascular endothelial growth factor receptor, and DNA damage response). Results Low (Spearman’s rho 0.3– 1, indicate benefit with the investigational product. Conclusions Using a comprehensive summary data set in the NSCLC space, we observed low-to-moderate correlations between treatment effects for early endpoints and HR OS across RCTs of agents with different MoAs, including trials of PD-1/PD-L1 checkpoint inhibitors. Exploration of additional endpoints, beyond RECIST, is required to identify other early indicators of efficacy that might predict HR OS. By incorporating additional trial-level parameters and building composite biomarkers using machine intelligence methods, in collaboration with innovative trial design efforts, we envisage to improve the prediction of HR OS from early endpoints.