Weighted overlapping group lasso for integrating prior network knowledge into gene set analysis.

Background: Gene set analysis aims to identify gene sets containing differentially expressed genes between two different experimental conditions. A representative example of gene sets is a gene regulatory network where multiple genes are linked with each other for regulation of gene expression. Most of statistical methods for gene set analysis were designed to capture group-based association signals, ignoring a genetic network structure. Consequently, they often fail to identify gene sets where the number of differentially expressed genes are only a few and they have sparse association signals.

Results: We propose a new computational method to utilize prior network knowledge for gene set analysis. The proposed method is essentially combines the coefficient estimates of network-based regularization into overlapping group lasso. Network-based regularization can boost association signals among linked genes while overlapping group lasso performs selection of gene sets including differentially expressed genes. In our extensive simulation study, the performance of the proposed method has been evaluated, compared with the existing methods. We also applied it to gene expression data of The Cancer Genome Atlas Breast Invasive Carcinoma Collection (TCGA-BRCA). We were able to identify cancer-related pathways that were missed by the existing methods.

Conclusion: Overlapping group lasso is a regularization method for group selection allowing overlapping variables. Network-based regularization is a variable selection method utilizing graph information among variables. The proposed weighted overlapping group lasso (wOGL) adopts the coefficient estimates of network-based regularization for the weight of overlapping group lasso. Consequently, it can identify gene sets containing differentially expressed genes, utilizing prior network knowledge.