Robust gene expression-based classification of cancers without normalization

Binary classification using gene expression data is commonly used to stratify cancers into molecular subgroups that may have distinct prognoses and therapeutic options. A limitation of many such methods is the requirement for comparable training and testing data sets. Here, we describe and demonstrate a self-training implementation of probability ratio-based classification prediction score (PRPS-ST) that facilitates the porting of existing classification models to other gene expression data sets. We demonstrate its robustness through application to two binary classification problems in diffuse large B-cell lymphoma using a diverse variety of gene expression data types and normalization methods.