A Study on Modeling Tumor Evolution with Reinforcement Learning

Cancer is a disease which results from an evolutionary process that yields multiple sets of cells with different mutations. The heterogeneous nature of this process is one of the reasons that make cancer extremely difficult to be treated, not only because distinct subpopulations of cells can respond differently to treatments, but also because even if we consider the same cancer cohort, the evolution in patients can be significantly different. Therefore, the study of such a process is crucial to better deal with it and, consequently, find some possible remedy. An im­portant help can be given by the use of Artificial Intelligence techniques with which we would like to find recurrent patterns in cancer evolution processes, which are represented as phylo­genetic trees of clones. So far, many approaches have been proposed and this work focuses in particular on CloMu, a recent model that consists of a two­ layers neural network trained using reinforcement learning to predict the probability a mutation will be acquired by a clone, given the mutations already existing on it. The capability of determining such probabilities is then the basis for several other prediction tasks. This work proposes an improvement of CloMu, inspired by ensemble learning, which has recently been applied to many fields of Deep Learning and of­ten leads to better results. In our specific case, we start from the idea that, due to the intra­class heterogeneity of clonal evolution processes, training more than one model on different subsets of patients belonging to the same cancer cohort and then combining the outputs could be partic­ularly effective. We also analyze CloMu, focusing on the reasons why it stands out from other approaches and testing its performances, especially on the real datasets provided by the authors.