Bayesian sparse factor model for transcriptional regulatory networks inference

Uncovering transcription factor (TF) mediated regulatory networks from microarray expression data and prior knowledge is considered in this paper. Bayesian factor models that model direct TF regulation are formulated. To address the enormous computational complexity of the model in large networks, a novel, efficient basis-expansion factor model (BEFaM) has been proposed, where the loading (regulatory) matrix is modeled as an expansion using basis functions of much lower dimension. Great reduction is achieved with BEFaM as the inference involves estimation of expansion coefficients with much reduced dimensions. We also address the issue of incorporating the prior knowledge of TF regulation to constrain the factor loading matrix. A Gibbs sampling solution has been developed to estimate the unknowns. The proposed model was validated by simulation and then applied to breast cancer data to uncover the corresponding TF regulatory network and theirs protein levels.