Identifying Disease Network Dysregulation Through Expression Mean, Variance, and Distribution Changes

Network-based approaches to identifying gene expression signatures of disease have been shown to be more reproducible, accurate, and biologically informative than single-gene markers. Differential expression statistics are most commonly used to rank and select nodes within biological networks. Disease heterogeneity and disease-specific mechanisms may create changes in gene expression distributions that are not captured by differential expression analysis. We implemented a disease network identification algorithm using four unique statistical metrics, based on changes in mean expression, expression variance, and overall expression distribution changes. Using three publicly available gene expression datasets, we evaluated the size, strength, and disease classification performance of the identified subnetworks in cancer and Posttraumatic Stress Disorder (PTSD) cohorts. Identified subnetworks for all three datasets had significant enrichment of disease signal over individual genes, based on
${p}$
-value. Further, novel and unique subnetworks were identified by each statistical metric, including changes in expression variance related to BRCA1 in breast cancer, and changes in blood coagulation in PTSD. Subnetwork-based biomarkers also slightly improved prediction performance over individual gene markers. This approach could provide a novel network-based framework for understanding and predicting gene expression signals of disease.