Bayesian blind source separation applied to the lymphocyte pathway

In many biological applications one observes a multivariate mixture of signals, where both the mixing process and the signals are unknown. Blind source separation can extract such source signals. Often the data have additional structure, i. e. the variables (e. g. genes) are linked by an interaction network. Recently, we developed the probabilistic method emGrade that explicitly uses this network structure as a Bayesian network and thus performs a more appropriate separation of the data than standard methods. Here, we consider the application of emGrade to gene expression data together with a literature-derived pathway. Thanks to the probabilistic modeling, we can use model selection criteria and demonstrate the relevance of the pathway information for explaining the data. We further use estimates of missing observations to identify the most appropriate microarray probe sets for two genes that were not uniquely annotated after standard filtering. Finally, we identify genes relevant for the dynamics underlying the data; these genes were not detected without the network information.