Deep Reinforcement Learning for Control of Probabilistic Boolean Networks

Probabilistic Boolean Networks (PBNs) were introduced as a computational model for the study of complex dynamical systems, such Gene Regulatory (GRNs). Controllability in this context is process making strategic interventions to state network order drive it towards some other that exhibits favourable biological properties. In paper we ability Double Deep Q-Network with Prioritized Experience Replay learning control strategies within finite number time steps PBN target state, typically an attractor. The method model-free and does not require knowledge network’s underlying dynamics, suitable applications where inference dynamics intractable. We present extensive experiment results on two synthetic PBNs constructed directly from gene-expression data metastatic-melanoma.