Integrated Classification of Prostate Cancer Reveals a Novel Luminal Subtype with Poor Outcome

Abstract

Prostate cancer is a biologically heterogeneous disease with variable molecular alterations underlying cancer initiation and progression. Despite recent advances in understanding prostate cancer heterogeneity, better methods for classification of prostate cancer are still needed to improve prognostic accuracy and therapeutic outcomes. In this study, we computationally assembled a large virtual cohort (n = 1,321) of human prostate cancer transcriptome profiles from 38 distinct cohorts and, using pathway activation signatures of known relevance to prostate cancer, developed a novel classification system consisting of three distinct subtypes (named PCS1-3). We validated this subtyping scheme in 10 independent patient cohorts and 19 laboratory models of prostate cancer, including cell lines and genetically engineered mouse models. Analysis of subtype-specific gene expression patterns in independent datasets derived from luminal and basal cell models provides evidence that PCS1 and PCS2 tumors reflect luminal subtypes, while PCS3 represents a basal subtype. We show that PCS1 tumors progress more rapidly to metastatic disease in comparison with PCS2 or PCS3, including PSC1 tumors of low Gleason grade. To apply this finding clinically, we developed a 37-gene panel that accurately assigns individual tumors to one of the three PCS subtypes. This panel was also applied to circulating tumor cells (CTC) and provided evidence that PCS1 CTCs may reflect enzalutamide resistance. In summary, PCS subtyping may improve accuracy in predicting the likelihood of clinical progression and permit treatment stratification at early and late disease stages. Cancer Res; 76(17); 4948-58. ©2016 AACR.

Figure 1

Integration of PC transcriptome and quality control. A. Schematic showing the process of collecting and merging PC transcriptomes. B. Clinical composition of 2,115 PC cases. C. MDS of merged expression profiles after MCQ or XPN correction in the DISC cohort. Dots with different colors represent different batches or datasets. D. Hierarchical clustering illustrates the sample distribution of uncorrected (upper), corrected by MCQ (middle), and corrected by XPN (lower). Different colors on ‘Batches’ rows represent different batches or datasets from the individual studies. E. MDS of pathway activation profiles in the DISC cohort shows distribution of the samples from same batches. Dots with different colors represent different batches or datasets.

Figure 2

Identification and validation of novel PC subtypes. A. Consensus matrix depicts robust separation of tumors into 3 subtypes. B. Changes of cophenetic coefficient and silhouette score at rank 2 to 6. C. Pathway activation profiles of 1,321 tumors defines 3 PC subtypes. D. Score plot of principal component analysis for benign and 3 subtypes. E and F. The 3 subtypes were recognized in 10 independent cohorts. G and H. Correlation of pathway activation profiles in 8 PC cell lines from the CCLE and 11 PC mouse models and probability from the pathway classifier.

Figure 3

Comparison of the PCS subtypes with previously described subtypes. A. Distribution of TCGA tumors (n=333) using the PCS subtypes compared to TCGA subtypes. B. Relationship between PCS subtyping and TCGA subtypes. C. Distribution of GRID tumors (n=1,626) using PCS categories compared to Tomlins subtypes. D. Relationship between PCS subtyping and Tomlins subtypes. E and F. Association of metastasis-free survival using Tomlins subtypes and using the PCS subtypes in the GRID tumors. G. Metastasis-free survival in tumors of GS≤7 (left) and GS≥8 (right).

Figure 4

Genes enriched in each of the 3 subtypes are associated with luminal and basal cell features. A. Relative gene expression (left) and pathway inclusion (right) of SEGs are displayed. B. Cellular processes enriched by each of the 3 SEGs (P<0.05). C. Expression of the luminal and basal markers in the 3 subtypes. D. Enrichment of basal stem cell signature. E. Correlation of pathway activities between samples from human and mouse prostate (left) and probability from the pathway classifier (right).

Figure 5

A 37-gene classifier employed in patient tissues and CTCs. A. Heatmap displays the mean expression pattern of the 37 gene panel in the 3 subtypes from the DISC cohort. B. Hierarchical clustering of 77 CTCs obtained from CRPC patients by gene expression of the 37-gene panel. Barplot in the bottom displays probability of PCS assignment from application of the classifier.