Modeling Dependence in Omics Association Analysis via Structured Co-Expression Networks to Improve Power and Replicability

Association analysis (e.g., differential expression analysis) and co-expression analysis are two major classes of statistical methods for omics data. While association analysis identifies individual features linked to health conditions, co-expression analysis examines dependencies among features to uncover functional modules and regulatory interactions. However, these approaches are often conducted separately, potentially leading to statistical inference with reduced sensitivity and replicability. To address this, we propose CoReg, a new statistical framework that integrates co-expression network analysis and factor models into the covariance modeling of multivariate regression. By accounting for the dependencies among omics features, CoReg enhances the power and sensitivity of association analysis while maintaining a well-controlled false discovery rate, thereby improving replicability across omics studies. We developed computationally efficient algorithms to implement CoReg and applied it to extensive simulation studies and real-world omics data analyses. Results demonstrate that CoReg improves statistical inference accuracy and replicability compared to conventional methods.