Improving nonhomogeneous dynamic Bayesian networks with sequentially coupled parameters (original) (raw)

In systems biology, nonhomogeneous dynamic Bayesian networks (NH-DBNs) have become a popular modeling tool for reconstructing cellular regulatory networks from postgenomic data. In this paper, we focus our attention on NH-DBNs that are based on Bayesian piecewise linear regression models. The new NH-DBN model, proposed here, is a generalization of an earlier proposed model with sequentially coupled network interaction parameters. Unlike the original model, our novel model possesses segment-specific coupling parameters, so that the coupling strengths between parameters can vary over time. Thereby, to avoid model overflexibility and to allow for some information exchange among time segments, we globally couple the segment-specific coupling (strength) parameters by a hyperprior. Our empirical results on synthetic and on real biological network data show that the new model yields better network reconstruction accuracies than the original model.