Biostatistics Seminar Series - Denise Scholtens, PhD

Title:Jointly Interpreting Graphical Models and Correlation Networks for Omics Data AnalysesAbstract:Construction of undirected biological networks from continuous cross-sectional data for multiple features (nodes) is dominated by approaches that rely on either calculation of pairwise marginal Pearson correlations between features (i.e. correlation networks) or pairwise partial correlations that condition on values all other features in the network (i.e. graphical models). This presentation addresses two fundamental challenges in the use of these methods for the analysis of biological data. The first challenge is the characterization of the relationship between these two methods that is both precise and biologically meaningful. The second challenge is jointly interpreting the results of correlation and graphical model analyses. With the goal of finer-scale interpretation in mind, we leverage the relationship between graphical models and correlation networks to propose the pair path subscore (PPS), a method of scoring individual network paths in a graphical model based on their relative importance in determining the overall network-level correlation between their endpoints. The PPS is demonstrated using metabolomics data from the Hyperglycemia and Adverse Pregnancy Outcome (HAPO) study and observations confirm well-documented biological relationships among the metabolites.