A Novel Stratification Framework for Predicting Outcome in Patients with Prostate Cancer

Unsupervised learning methods such as Hierarchical Cluster Analysis are commonly used for the analysis of genomic platform data. Unfortunately, such approaches ignore the well documented heterogeneous composition of cancer samples. Our aim is to use more sophisticated analytical approaches to deconvolute the structure of prostate cancer transciptome data providing novel clinically actionable information for human prostate cancer.

Methods: We apply an unsupervised model called Latent Process Decomposition (LPD), which can handle heterogeneity within cancer samples, to genome-wide expression data from eight prostate cancer clinical series including 1,785 malignant samples with the clinical endpoints of PSA failure and metastasis.

Results: We show that PSA failure is correlated with the level of an expression signature called DESNT (HR = 1.52, 95% CI = [1.36, 1.7], P = 9.0x10-14, Cox model) and that patients with a majority DESNT signature have an increased metastasic risk (X2-test, P = 0.0017, and P = 0.0019). Additionally, we develop a stratification framework that incorporates DESNT and identifies three novel molecular subtypes of prostate cancer.