Modelling non-stationary gene regulatory process with hidden Markov Dynamic Bayesian Network

Abstract

Dynamic Bayesian Network (DBN) has been widely used to infer gene regulatory network from time series gene expression dataset. The standard assumption underlying DBN is based on stationarity, however, in many cases, the gene regulatory network topology might evolve over time. In this paper, we propose a novel non-stationary DBN based network inference approach. In this model, for each variable, a specific HMM implicitly well handles the transition of the stationary DBNs along timesteps. Furthermore, we present a criterion, named as BWBIC score. This criterion is an approximation to the EM objective term, which can reasonably and easily evaluate hmDBN Towards BWBIC score, a greedy hill climbing based structural EM algorithm is proposed to efficiently infer the hmDBN model. We respectively apply our method on synthetic and real biological data. Compared to the recent proposed methods, we obtained better prediction accuracy on both datasets.