Dynamic Bayesian System Modeling and Structure Learning
The systems structure learning step integrates a selected set of significantly perturbed pathways and gene groups to construct a plausible system level model of the disease/condition under study. The selection and refinement is iterative in that input from the research biologists/scientist are essential in the selection process. The system model encompasses the whole time-course patterns and multi-conditional behaviors of a large group of genes/proteins ("omics" and other physiological responses). The system (disease) models can be used for more efficient comparative modeling, pattern recognition and simulations supporting "what-if" type of analyses. Two system model generation approaches are employed at the Seralogix. First is a top-down reconstruction method based on merging of pathways with known gene/protein relationships and the second is based on a bottom-up network learning approach based on the merger of gene/protein data and prior knowledge (i.e. known and/or predicted protein-protein interactions). The structure learning algorithms incorporate prior biological knowledge combined with multi-perturbation data (such as from knockout studies) to learn the system model network structure. Large amount of perturbation data provides better learning results. Figure 1 illustrates a simple merged model from the host response for a pathogen gene knockout study. Figure 2 is a magnified section of Figure 1 model. Only 3 pathways were selected as the seed pathways to constrain the network learning to a smaller set of genes. Figure 1b is an enlarged section of the system network at 36 hours post infection.