Modeling and predicting cancer clonal evolution with reinforcement learning

Overview of CloMu. (A) Using the independent clonal evolution assumption, our model determines a log rate of any clone c ∈ {0, 1}^m acquiring a mutation s ∈ [m]. (B) This in turn enables us to compute probabilities P = [p_i,s] that the next mutation to occur on a tree T is mutation s at node/clone c_i. The resulting Independent Clonal Evolution problem seeks model parameters θ that maximize the data probability of a cohort of trees for n patients. (C) CloMu represents using a low-parameter, two-layer neural network that is trained via reinforcement learning. We use the model for five prediction tasks: (D) tree selection for each patient, (E) determination of mutation fitness, (F) causality inference for pairs of mutations, (G) identification of interchangeable mutations, and (H) identification of their evolutionary pathways.