CONFIGURE: A pipeline for identifying context specific regulatory modules from gene expression data and its application to breast cancer

Background: Gene expression data is widely used for identifying subtypes of diseases such as cancer. Differentially expressed gene analysis and gene set enrichment analysis are widely used for identifying biological mechanisms at the gene level and gene set level, respectively. However, the results of differentially expressed gene analysis are difficult to interpret and gene set enrichment analysis does not consider the interactions among genes in a gene set. Results: We present CONFIGURE, a pipeline that identifies context specific regulatory modules from gene expression data. First, CONFIGURE takes gene expression data and context label information as inputs and constructs regulatory modules. Then, CONFIGURE makes a regulatory module enrichment score (RMES) matrix of enrichment scores of the regulatory modules on samples using the single-sample GSEA method. CONFIGURE calculates the importance scores of the regulatory modules on each context to rank the regulatory modules. We evaluated CONFIGURE on the Cancer Genome Atlas (TCGA) breast cancer RNA-seq dataset to determine whether it can produce biologically meaningful regulatory modules for breast cancer subtypes. We first evaluated whether RMESs are useful for differentiating breast cancer subtypes using a multi-class classifier and one-vs-rest binary SVM classifiers. The multi-class and one-vs-rest binary classifiers were trained using the RMESs as features and outperformed baseline classifiers. Furthermore, we conducted literature surveys on the basal-like type specific regulatory modules obtained by CONFIGURE and showed that highly ranked modules were associated with the phenotypes of basal-like type breast cancers. Conclusions: We showed that enrichment scores of regulatory modules are useful for differentiating breast cancer subtypes and validated the basal-like type specific regulatory modules by literature surveys. In doing so, we found regulatory module candidates that have not been reported in previous literature. This demonstrates that CONFIGURE can be used to predict novel regulatory markers which can be validated by downstream wet lab experiments. We validated CONFIGURE on the breast cancer RNA-seq dataset in this work but CONFIGURE can be applied to any gene expression dataset containing context information.