Identification of progression markers for prostate cancer.

Abstract

TGFβ functions as a tumor suppressor or promoter, depending on the context, making TGFβ a useful predictive biomarker. Genes related to TGFβ signaling and Aurora kinase were tested for their ability to predict the progression risk of primary prostate tumors. Using data from The Cancer Genome Atlas (TCGA), we trained an elastic-net regularized Cox regression model including a minimal set of gene expression, copy number (CN), and clinical data. A multi-step feature selection and regularization scheme was applied to minimize the number of features while maintaining predictive power. An independent hold-out cohort was used to validate the model. Expanding from prostate cancer, predictive models were similarly trained on all other eligible cancer types in TCGA. AURKA, AURKB, and KIF23 were predictive biomarkers of prostate cancer progression, and upregulation of these genes was associated with promotion of cell-cycle progression. Extending the analysis to other TCGA cancer types revealed a trend of increased predictive performance on validation data when clinical features were complemented with molecular features, with notable variation between cancer types and clinical endpoints. Our findings suggest that TGFβ signaling genes, prostate cancer related genes and Aurora kinases are strong candidates for patient-specific clinical predictions and could help guide personalized therapeutic decisions.