Abstract:Biological variables involved in a disease process often correlate with each other through for example shared metabolic pathways. In addition to their correlation, these variables contain complementary information that is particularly useful for disease classification and prediction. However, complementary information between variables is rarely explored. Therefore, establishing methods for the investigation of variable’s complementary information is very necessary. We propose a model population analysis approach that aggregates information of a number of classification models obtained with the help of Monte Carlo sampling in variable space for quantitatively calculating the complementary information between variables. We then assemble these complementary information to construct a variable complementary network (VCN) to give an overall visualization of how biological variables complement each other. Using a simulated dataset and two metabolomics datasets, we show that the complementary information is effective in biomarker discovery and that mutual associations of metabolites revealed by this method can provide information for exploring altered metabolic pathways. (The source codes for implementing VCN in MATLAB are freely available at: http://code.google.com/p/vcn2011/.)

Variable Complementary Network: a Novel Approach for Identifying Biomarkers and Their Mutual Associations