Bayesian Network Approaches for Refining and Expanding Cellular and Immunological Pathways.

Abstract

This thesis focuses on computational analysis of cellular and immune pathways of living cells in response to molecular signals using Bayesian networks (BN). Although Bayesian networks have been applied to the reconstruction and expansion of gene regulatory and protein signaling pathways using existing biological data, the results generated from existing BN methods show high false positive and false negative rates. To resolve these issues, two major Bayesian network approaches were developed to allow refinement and expansion of known biological pathways to identify new interactions and molecular entities participating in the pathway. How to refine existing Bayesian networks to identify the best-supported interactions predicted using underlying biological data was explored initially. A posterior probability-based EdgeClipper refinement algorithm was developed to identify well-supported interaction hypotheses in distributions of saved BNs. EdgeClipper incorporates posterior weighting to prioritize and clip interactions. This approach identified many known interactions in synthetic and Escherichia coli reactive oxygen species (ROS) pathways as well as novel interactions and improved specificity with decreasing sensitivity. Second, an expansion approach called BN+1 was introduced to identify unknown though potentially novel pathway members which likely influence biological pathways. BN+1 was applied to the expansion of several synthetic, prokaryotic, and eukaryotic pathways. Major findings included the identification of genetic interactions between genes gadX and uspE and their direct regulation of biofilm activities in E.coli, which was verified experimentally. Finally, the expansion and refinement algorithms were combined to recover a known acid fitness island and new putative acid fitness regulators using E.coli ROS pathway members, and later applied towards understanding Jak/Stat pathway regulation during human progressive kidney disease in glomerular and tubule compartments. The Jak/Stat pathway showed relatively low overlap in supported interactions for the two compartments, though recovered BN+1 genes reflected relevant biological functions and stages of disease progression for the respective kidney compartments. Overall, the results demonstrate that it is possible to refine and expand protein-level signaling pathways using transcriptional microarray data and the introduced expansion and refinement algorithms. The methods are applicable to other biological and computational systems, and are available as publicly-accessible software tools.