Using prior knowledge to improve genetic network reconstruction from microarray data

The use of Bayesian Network methods to recover transcriptional regulatory networks from static microarray data is an active area of bioinformatics research. However, early work in this area lacked realistic analysis of the effects of data set size on learning performance and ignored the potentially immense benefits of using prior biological knowledge. More recent work which has utilized such information has tended to focus on qualitative descriptions of the results. In this paper, we construct a detailed, realistic model for glucose homeostasis and use this model to generate static, synthetic gene expression data. We then use a Bayesian Network method to reconstruct this genetic network from the synthetic microarray data utilizing various amounts and types of prior knowledge. By quantitatively analyzing the effects of data set size and the incorporation of different types of prior biological knowledge on our ability to reconstruct the original network, we show that characteristic portions of genetic networks can be reconstructed from microarray data. Incorporating prior knowledge into the learning scheme greatly reduces the data required, allowing these reverse engineering techniques to be used to learn regulatory interactions from microarray data sets of realistic size.