Designing combination therapies with modeling chaperoned machine learning

Chemotherapy resistance is a major challenge to the effective treatment of cancer. Thus, a systematic pipeline for the efficient identification of effective combination treatments could bring huge biomedical benefit. In order to facilitate rational design of combination therapies, we developed a comprehensive computational model that incorporates the available biological knowledge and relevant experimental data on the life-and-death response of individual cancer cells to cisplatin or cisplatin combined with the TNF-related apoptosis-inducing ligand (TRAIL). The model’s predictions, that a combination treatment of cisplatin and TRAIL would enhance cancer cell death and exhibit a “two-wave killing” temporal pattern, was validated by measuring the dynamics of p53 accumulation, cell fate, and cell death in single cells. The validated model was then subjected to a systematic analysis with an ensemble of diverse machine learning methods. Though each method is characterized by a different algorithm, they collectively identified several molecular players that can sensitize tumor cells to cisplatin-induced apoptosis (sensitizers). The identified sensitizers are consistent with previous experimental observations. Overall, we have illustrated that machine learning analysis of an experimentally validated mechanistic model can convert our available knowledge into the identity of biologically meaningful sensitizers. This knowledge can then be leveraged to design treatment strategies that could improve the efficacy of chemotherapy.Author summary: Combination chemotherapy is frequently used in the fight against cancer as treatment with multiple chemotherapy drugs of different molecular mechanisms reduces the chance of resistance. The complex mechanisms involved makes it essential to develop a comprehensive computational model that comprehends experimental data and biological knowledge to facilitate design of combination therapies. As computational models grow and capture more and more molecular events governing the chemotherapy response, it becomes harder to explore the treatment space efficiently and systematically. To facilitate the extraction of unbiased solutions from complicated models, we have conducted systematic analysis using a series of machine learning methods including Partial Least Squares regression (PLS), Random forest (RF), Logistic Regression (LR) and Support Vector Machine (SVM). The results of these different methods were cross-validated to reduce the chance of overfitting or bias by any single method. Overall, we propose a novel computational pipeline, where machine learning analysis of experimentally validated models is used to generate unbiased predictions of novel chemotherapy targets.