Senior Scientist, Drug Development Training Program Genentech Antioch, California, United States
Background: To develop and validate a tumor growth inhibition-overall survival (TGI-OS) model to predict OS using tumor size data in melanoma patients. Methods: TGI metrics were estimated from a biexponential model [1] using pooled longitudinal tumor size data from three Phase III studies in patients with advanced or metastatic melanoma; two studies evaluating the effectiveness of atezolizumab (IMspire150 and IMspire170) and one evaluating Cobimetinib (coBRIM) [2,3,4]. The TGI-OS model was built using parametric survival regression by including all significant prognostic factors from the Cox univariate analysis and followed by backward elimination. Seventy percent of the pooled data was used for model development and 30% was used for model validation. Results: Patients who had at least two tumor size measurements (one at baseline and one post-baseline) were considered evaluable and included in model development/validation. From a total 1,439 patients, 1,314 (91.3%) were evaluable. Predicted and observed tumor growth trajectories were well aligned. An inverse correlation was observed between tumor growth rate constant and overall survival based on Kaplan-Meier analysis. Based on covariate analysis, tumor growth rate was identified as the most significant predictor of OS. Alanine Transaminase, Lactate Dehydrogenase, albumin count and Eastern Cooperative Oncology Group (ECOG) performance score were also identified as significant covariates. Overall, the test data (30% by arm) falls within the prediction interval of the model, suggesting that the aggregated data generalizes well to predict tumor growth progression and overall survival in melanoma patients. Considering test data, predicted hazard ratio (HR) for IMspire150 (predicted HR 1.2, 95% PI 0.70-1.9 vs. 0.8 observed HR), IMspire170 (predicted HR 1.1, 95% PI 0.69-1.6 vs. 1.2 observed HR) and coBRIM (predicted HR 0.92, 95% PI 0.53-1.6 vs. 0.85 observed HR) were within the 95%PI. Conclusion: The final TGI-OS model was able to predict treatment effect in melanoma patients, and could be used to support design and analysis of future melanoma studies.
