Dealing with data of a specific temporal or spatial structure is well established in blind source separation. However, in biology one often faces more complex network structures. The recently published GraDe-algorithm addresses such structures; it separates sources with respect to a given network in an analytical manner. We formulate corresponding assumptions and assign them to a very flexible Bayesian model. This allows us to include for instance missing observations and use prior parameter knowledge. Technically, we propose a Gaussian graphical model with latent variables to include all structural information from the data. The parameters and latent variables are estimated using expectation maximization, where we exploit the restrictions given by the separation assumptions. In a large scale application we consider gene expression data, where the dependence structure is given by a gene regulatory network. We demonstrate how the model indeed identifies relevant biological processes.     «

Dealing with data of a specific temporal or spatial structure is well established in blind source separation. However, in biology one often faces more complex network structures. The recently published GraDe-algorithm addresses such structures; it separates sources with respect to a given network in an analytical manner. We formulate corresponding assumptions and assign them to a very flexible Bayesian model. This allows us to include for instance missing observations and use prior parameter kn...     »