Statistical Methods for Identification and Evaluation of Predictive Biomarkers in Cancer

Project Summary Relapsed/refractory cancer is a principal cause of cancer-related death. Targeted therapies, which represent a new generation of cancer therapies, have advanced the treatment for relapsed/refractory patients. However, treatment effects are still heterogeneous. Only a fraction of patients who are treated with these new therapies experience clinically beneficial outcomes. Therefore, it is critical to identify new predictive biomarkers that can further stratify patient into subgroups that are most likely to yield a favorable or unfavorable treatment effect. Analysis of non-randomized phase II clinical trial data to identify predictive biomarkers is particularly important because such information is crucial to guide efficient subsequent randomized phase II or enriched phase III trials and improve the success rate of clinical drug development. In non-randomized phase II trials, progression-free survival (PFS) has been increasingly considered as an important clinical endpoint. As these trials do not have an independent control arm, the PFS on the most recent prior treatment on which the patient had experienced progression has been suggested to serve as the patient-specific control. The ratio of paired PFSs on the new versus prior treatments is used to evaluate treatment efficacy. The PFS ratio has become an important endpoint in the era of precision oncology. However, using paired PFS data to identify and evaluate predictive biomarkers from non-randomized phase II trials has been hampered due to major challenges in statistical methods. First, the identification of predictive biomarkers is typically achieved by testing the interaction effects in multivariable models, which usually requires large sample sizes. As phase II trials usually have small sample sizes, detecting interaction effects is challenging. Second, it is challenging to deal with high- dimensional candidate biomarkers. Third, the PFS ratio endpoint is dependently censored, which creates a challenge for accurate statistical inference because traditional methods for censored data require independent censoring assumption. Fourth, there is a lack of clinically meaningful statistical measures to evaluate and compare the accuracy of predictive biomarkers. To address these challenges, we propose to a) develop novel semiparametric statistical models to identify and combine predictive biomarkers; and b) develop new clinically meaningful statistical measures to evaluate and compare the accuracy of predictive biomarkers based on paired PFS data from non-randomized phase II trials. We will implement the statistical methods into an R package as well as a web-based application. We will also apply these new methods to three precision medicine clinical cohorts. Our new methods will establish a systematic and effective framework to advance the predictive biomarker analysis based on paired PFS data from non-randomized phase II trials, which will have direct impact on drug development by facilitating more informed design for further validation in randomized trials.