Clinical trial data science to advance precision oncology

Abstract

Clinical trials are the most expensive and highest-stakes experiments in biomedicine, and their results determine which therapies are ultimately used to treat patients. While a typical pivotal trial costs millions of dollars and enrolls hundreds of patients, its associated data will not be made public, limiting our ability to learn from its results. We reconstructed otherwise siloed clinical trial data by using image analysis on published trial figures. We then developed methods to re-analyze these results at scale and help design more effective clinical and preclinical studies. In Chapter 1, we describe how a probabilistic model can predict the success of combination therapies in oncology using single agent data alone and how this insight can be leveraged to improve survival outcomes in presently incurable cancers. In Chapter 2, we find that a parametric distribution accurately describes the shape of survival curves in oncology trials, and we show how historical results can inform distribution parameters to more precisely infer drug efficacy from small sample sizes. We also find that differences in curve shape correspond to a violation of the proportional hazards assumption typically used to assess trial results and, consequently, that a trial’s length has an underappreciated impact on its likelihood of success. In Chapter 3, we develop a more powerful approach to detect patient subgroups responding to new therapies as compared to standard methods used in oncology basket trials. Finally, in Chapter 4, we apply the lessons learned from clinical data to incorporate interpatient heterogeneity into drug and biomarker discovery efforts using patient-derived mouse xenografts. This body of work offers new tools to better understand drug-response data in human, animal, and cell line studies. We hope that these methods help take clinical findings from the bedside back to the bench and advance precision oncology.