Adverse prognosis gene expression patterns in metastatic castration-resistant prostate cancer

Abstract

Metastatic castration-resistant prostate cancer (mCRPC) is a heterogeneous disease. Several studies have identified transcriptional subtypes of mCRPC, but comprehensive analysis of prognostic gene expression pathways has been limited. Therefore, we aggregated a cohort of 1012 mCRPC tissue samples from 769 patients and investigated the association of gene expression-based pathways with clinical outcomes and intrapatient and intratumor heterogeneity. Survival data were obtained for 272 patients. Pathway-level enrichment was evaluated using gene set variation analysis. scRNA-seq datasets from mCRPC tissue biopsies and circulating tumor cells were used to investigate heterogeneity of adverse pathways. We identified five pathway clusters: (a) Immune response/WNT/TGF-beta signaling, (b) AR signaling/luminal signatures, (c) mTOR signaling and glycolysis, (d) cell proliferation, and (e) neuroendocrine differentiation. Proliferation, AR signaling loss, and glycolysis/mTOR signaling were independently prognostic. Adverse prognostic pathway scores decreased on treatment with AR signaling inhibitors, but not at progression, suggesting failure to permanently target these pathways. scRNA-seq datasets from mCRPC tissue biopsies and circulating tumor cells were used to investigate heterogeneity of adverse pathways. Our results suggest loss of AR signaling, high proliferation, and a glycolytic phenotype as adverse prognostic pathways in mCRPC that could be used in conjunction with clinical factors to prognosticate for treatment decisions.

Keywords: biomarker; gene expression; metastatic castration‐resistant prostate cancer; precision medicine; prognosis.

Fig. 1

Composition of the combined mCRPC transcriptional profiling dataset. (A) mCRPC RNA‐sequencing data were collected from five large clinical cohorts comprising 1012 tissue samples from 769 patients across mCRPC disease stages, including the Prostate Cancer Medically Optimized Genome‐Enhanced Therapy (PROMOTE) study cohort, the Stand Up 2 Cancer/Prostate Cancer Foundation (SU2C/PCF) East Coast Dream Team (ECDT), the SU2C/PCF West Coast Dream Team (WCDT) mCRPC cohort, a neuroendocrine prostate cancer‐enriched cohort from Weill Cornell Medicine (WCM), and the Fred Hutchinson Cancer Research Center (FHCRC) autopsy cohort. (B) Distribution of samples from each cohort among all samples with DNA sequencing available. (C) Distribution of samples from each cohort among all adenocarcinoma and NEPC, (D) Distribution of samples from each cohort among the subset of patients with overall survival data, PSA response data, and matched pre‐/post‐ASI samples.

Fig. 2

Correlation of hallmark and prostate associated pathways across the combined dataset and association with prognosis. Pearson's correlation of hallmark pathway and prostate‐related gene signatures across the combined cohort (n = 1012 samples) reveals five signaling clusters including (1) basal, mesenchymal and immune signatures, (2) luminal and androgen response signatures, (3) metabolic signatures, (4) proliferation and tumor suppressor loss signatures, and (5) neuroendocrine signatures. In univariate cox proportional hazards models of the survival subset (n = 272 patients) adjusted for biopsy site and stratified by cohort, signature scores associated with worse prognosis are labeled in yellow, while signature scores associated with improved prognosis are labeled in green, along with significance for each association (univariate cox proportional hazards model *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, corrected for multiple testing).

Fig. 3

Prognostic signatures and histologic subtype. (A) Multivariate cox proportional hazards analysis in the survival subset (n = 272 patients) of combination prognostic signatures created from signatures associated with prognosis in each signature cluster including proliferation and tumor suppressor loss cluster (Proliferation), neuroendocrine cluster (NEPC), metabolism cluster (mTOR_glycolysis), and androgen response/luminal cluster (Luminal_AR), demonstrating independent prognostic value for the Proliferation (P = 0.003), mTOR_glycolysis (P = 0.032) and Luminal_AR (P = 0.001) cluster scores. Model is adjusted for biopsy site and cohort. The NEPC (B) and Proliferation (C) scores are increased in histologically defined NEPC (n = 70) compared to adenocarcinoma samples (n = 660), while the Luminal_AR (D) and mTOR_glycolysis (E) scores are decreased. Box and whiskers plots are shown with box representing interquartile range (IQR) split at median, and whiskers extending to minimum and maximum values within 1.5 × IQR; points represent all samples in each group.

Fig. 4

Expression signatures of adverse prognosis pathways per metastatic site, genomic alteration, and examples of intrapatient homogeneity and heterogeneity. (A) Association of expression signatures with metastatic site (Primary site n = 30, Bone n = 321, Liver n = 124, Lung n = 46, Lymph node n = 316, Other n = 87) (B) Association of expression signatures with genomic alterations (AR mutation or amplification n = 582, MYC amplification n = 498, PTEN 2‐copy loss n = 224, RB1 2‐copy loss n = 108, TP53 2‐copy loss n = 187) (B). Expression signatures from different tissue sites in a single patient can be concordant (C, D) or discordant (E, F). For this analysis, pathway scores were rescaled from 0% to 100% over all 1012 samples. For A and B, median pathway score was taken across all samples from a metastatic site or genomic alteration. If a sample had more than one genomic alteration the sample was included in more than one group.

Fig. 5

Adverse prognosis pathways show homogeneity and heterogeneity at the single‐cell level. Single‐cell RNA‐seq from an adenocarcinoma bone metastatic tumor (n = 338 single cells) (A) and a liver neuroendocrine metastatic tumor (n = 166 single cells) (B) show distinct patient specific patterns. Single‐cell RNA‐seq from a lymph node sample before (n = 159 single cells) (C) and after (n = 265 single cells) (D) enzalutamide show more cells with high proliferation and NEPC scores after enzalutamide. Considerable interpatient heterogeneity of adverse pathways was observed in CTCs both before (n = 9 CTCs) (E) and after (n = 12 CTCs) (F) enzalutamide. Pathway scores were rescaled from 0% to 100% over all prostate cancer cells per study (tissue single‐cell or CTC single‐cell RNA‐seq).

Fig. 6

Prognostic signatures are decreased on treatment with ASI. Pre‐ versus early on‐treatment biopsies in the PROMOTE cohort (n = 52) demonstrate a significant decrease in Luminal_AR, mTOR_glycolysis, and Proliferation scores (A–C) on ASI treatment, while no significant change is seen in NEPC signatures (D). In contrast, only the Luminal_AR score is significantly decreased between matched pretreatment and post‐ASI progression biopsies in the WCDT cohort (n = 23) (E–H). Wilcoxon signed‐rank test is used to compare groups. Box and whiskers plots are shown with box representing interquartile range (IQR) split at median, and whiskers extending to minimum and maximum values within 1.5 × IQR; points represent all samples in each group and lines connect samples from the same patient.

Fig. 7

A combined transcriptional adverse feature score is prognostic for overall survival. (A) Patients with 3–4 adverse transcriptional features defined as (1) Luminal_AR score in the lowest tertile (2) mTOR_glycolysis score in the highest tertile, (3) Proliferation score in the highest tertile and (4) NEPC score in the highest tertile have significantly shorter survival than patients with 1–2 adverse features, while patients with no adverse features have improved survival (log‐rank P < 0.0001). (B) transcriptional adverse feature score is prognostic of survival independent of tumor site and genomic RB1 loss status in a multivariate cox proportional hazards model. (C) In patients who received an ASI postbiopsy and for whom PSA50 response was available (n = 142), number of transcriptional adverse features is significantly associated with likelihood of PSA response. Cochran‐Armitage test for trend is used to compare groups.