Methodological challenges in translational drug response modeling in cancer: A systematic analysis with FORESEE

Translational models directly relating drug response specific processes that can be observed in vitro to their in vivo role in cancer patients constitute a crucial part of the development of personalized medication. Unfortunately, current studies often focus on the optimization of isolated model characteristics instead of examining the overall modeling workflow and the interplay of the individual model components. Moreover, they are often limited to specific data sets only. Therefore, they are often confined by the irreproducibility of the results and the non-transferability of the approaches into other contexts. In this study, we present a thorough investigation of translational models and their ability to predict the drug responses of cancer patients originating from diverse data sets using the R-package FORESEE. By systematically scanning the modeling space for optimal combinations of different model settings, we can determine models of extremely high predictivity and work out a few modeling guidelines that promote simplicity. Yet, we identify noise within the data, sample size effects, and drug unspecificity as factors that deteriorate the models’ robustness. Moreover, we show that cell line models of high accuracy do not necessarily excel in predicting drug response processes in patients. We therefore hope to motivate future research to consider in vivo aspects more carefully to ultimately generate deeper insights into applicable precision medicine.Author summary: In the context of personalized medicine, finding genomic patterns in a cancer patient that can predict how a specific drug will affect the patient’s survival is of great interest. Translational approaches that directly relate drug response specific processes observed in cell line experiments to their role in cancer patients have the potential to increase the clinical relevance of models. Unfortunately, existing approaches are often irreproducible in other applications. In order to address this irreproducibility aspect, our work comprises a thorough investigation of a diverse set of translational models. In contrast to other approaches that focus on one isolated model characteristic at a time, we examine the overall workflow and the interplay of all model components. Additionally, we validate our models in multiple patient data sets and identify differences between cell line and patient models. While we can establish models of high predictive performance, we also expose the deceptive potential of optimizing methods to a specific use case only by showing that those models do not necessarily depict biological processes. Thus, this study serves as a guide to interpret new approaches in a broader context to avoid the dissemination of noise-driven models that fail to serve in everyday applications.