Control algorithms combined with microfluidic devices and microscopy have enabled in vivo real-time control of protein expression in synthetic gene networks. Most control algorithms rely on the a priori availability of mathematical models of the gene networks to be controlled. These models are typically black/grey box models, which can be obtained through the use of data-driven techniques developed in the context of systems identification. Data-driven inference of both model structure and parameters is the main focus of this paper. There are two main challenges associated with the inference of dynamical models for real-time control of gene regulatory networks in living cells. Since biological systems are typically evolving over time, the first challenge stems from the fact that model selection needs to be done online, which prevents the application of computationally expensive identification algorithms iterating through large amounts of streaming data. The second challenge consists in performing nonlinear model selection, which is typically too burdensome for Kalman filtering related techniques due the heterogeneity and nonlinearity of the candidate models. In this paper, we combine sparse Bayesian techniques with classic Kalman filtering techniques to tackle these challenges.