Bayesian prognostic model for genomic discovery in bipolar disorder

Integrative approaches that incorporate multiple experiments have shown a potential application in the discovery of disease-related attributes. This study presents a unique, data-driven, integrative, Bayesian approach to merge gene expression data from various experiments into prognostic models and evaluate them for the discovery of bipolar-related attributes. Two prognostic models were constructed: a singly-structured Bayesian and a Bayesian multi-net model, which differentiated bipolar disease state at a higher level of abstraction. These prognostic models were evaluated to find the most common attributes responsible for the disease and their AUROC, using external cross-validation. The multi-net model achieved an AUROC of 0.907 significantly outperforming the single- structured model with an AUROC of 0.631. The study found six new genes and five chromosomal regions associated with the bipolar state. We anticipate this method and results will be used in the future to integrate information from multiple experiments for the same or related phenotypes of various diseases and also to predict the disease state earlier.