UCLA

Background

Prostate cancer (PCa) is a clinically heterogeneous disease. The creation of an expression-based subtyping model based on prostate-specific biological processes was sought.

Methods

Unsupervised machine learning of gene expression profiles from prospectively collected primary prostate tumors (training, n = 32,000; evaluation, n = 68,547) was used to create a prostate subtyping classifier (PSC) based on basal versus luminal cell expression patterns and other gene signatures relevant to PCa biology. Subtype molecular pathways and clinical characteristics were explored in five other clinical cohorts.

Results

Clustering derived four subtypes: luminal differentiated (LD), luminal proliferating (LP), basal immune (BI), and basal neuroendocrine (BN). LP and LD tumors both had higher androgen receptor activity. LP tumors also had a higher expression of cell proliferation genes, MYC activity, and characteristics of homologous recombination deficiency. BI tumors possessed significant interferon γactivity and immune infiltration on immunohistochemistry. BN tumors were characterized by lower androgen receptor activity expression, lower immune infiltration, and enrichment with neuroendocrine expression patterns. Patients with LD tumors had less aggressive tumor characteristics and the longest time to metastasis after surgery. Only patients with BI tumors derived benefit from radiotherapy after surgery in terms of time to metastasis (hazard ratio [HR], 0.09; 95% CI, 0.01-0.71; n = 855). In a phase 3 trial that randomized patients with metastatic PCa to androgen deprivation with or without docetaxel (n = 108), only patients with LP tumors derived survival benefit from docetaxel (HR, 0.21; 95% CI, 0.09-0.51).

Conclusions

With the use of expression profiles from over 100,000 tumors, a PSC was developed that identified four subtypes with distinct biological and clinical features.

Plain language summary

Prostate cancer can behave in an indolent or aggressive manner and vary in how it responds to certain treatments. To differentiate prostate cancer on the basis of biological features, we developed a novel RNA signature by using data from over 100,000 prostate tumors-the largest data set of its kind. This signature can inform patients and physicians on tumor aggressiveness and susceptibilities to treatments to help personalize cancer management.