We propose a probabilistic approach based on dynamic Bayesian networks for modelling non-homogeneous and non-linear dynamic gene-regulatory processes. The new approach, which is similar to the BGM model we have recently developed, is based on a change-point process and a mixture model, using latent variables to assign individual measurements to different components. The practical inference follows the Bayesian paradigm, and we use small synthetic dynamic network domains to demonstrate emprically that this new method reduces the susceptibility to spurious feedback loops. Finally we apply the new method to a real gene expression data set from Arabidopsis thaliana.