Transcriptomic Heterogeneity of Androgen Receptor Activity Defines a de novo low AR-Active Subclass in Treatment Naïve Primary Prostate Cancer

Although the genomic diversity and clinical relevance of androgen receptor (AR) activity (AR-A) is well-established in heavily pretreated metastatic castration-resistant prostate cancer (mCRPC), in treatment-naïve primary prostate cancer it is less well-characterized (1–3). Recently, The Cancer Genome Atlas (TCGA) demonstrated heterogeneity in AR and AR-A expression within 333 primary prostate cancer tumors. Although provocative, the biologic and clinical implications of these findings are unknown due to the small sample size, short follow-up (<36 months), and lack of information on metastatic progression (3). Nonetheless, the well documented interplay between AR signaling and numerous biological processes within mCRPC suggest that these interactions could potentially impact therapeutic response in even earlier disease states. In addition, it is unknown whether the variability in AR-A in mCRPC is solely treatment induced, or whether it may exist de novo in the treatment-naïve setting.

We hypothesized that understanding the implications of AR-A diversity within treatment-naïve primary prostate cancer would provide important prognostic and predictive information that could explain the observed clinical heterogeneity in response to standard treatments. In addition, we sought to explore whether the observed biologic phenotypes of heavily pretreated mCRPC tumors (e.g., increased neuroendocrine differentiation) could be appreciated in treatment-naïve primary prostate cancer. To test these hypotheses with a high degree of granularity, we performed the largest transcriptomic study to date, 40 times the size of the TCGA prostate adenocarcinoma dataset, wherein full transcriptomic profiles from 19,470 primary prostate cancer tumors were comprehensively analyzed, including patients with long-term detailed clinical follow-up.

Genome-wide expression profiles of prostate adenocarcinoma (small cell and neuroendocrine prostate cancer were excluded) from radical prostatectomy or biopsy tumor samples for a total of 19,470 patients with primary prostate cancer were used. This was comprised from two prospective population-based registry cohorts (discovery n = 5,239 and validation n = 12,728), six retrospective institutional cohorts with long-term clinical outcomes (n = 1,170), and the TCGA (n = 333). The discovery cohort was exploratory in nature, and the validation cohorts were used to independently confirm any findings. The prospective cohorts were comprised of anonymized genome-wide expression profiles of formalin-fixed, paraffin-embedded (FFPE) samples from clinical use of the Decipher test between February 2014 and August 2017 retrieved from the Decipher GRID (NCT02609269). Individual patient genomic and clinicopathologic data were gathered from each study after institutional review boards at the participating institutions approved the research protocol under which the data were collected. Informed consent was not necessary to conduct this study, and thus was not obtained.

Basic demographic and pathologic data, but not longitudinal clinical outcomes, were available. Data from the six retrospective cohorts (Table 1; Supplementary Table S1) of men treated with radical prostatectomy at Johns Hopkins Medical Institution (JHMI, n = 355), Mayo Clinic (n = 235), Thomas Jefferson University (n = 139), Durham VA (n = 117), Kaiser Permanente Northwest (n = 224), or external beam radiotherapy at Brigham & Women's Hospital (BWH, n = 100). Local institutional review boards approved all data collection. The TCGA PRAD dataset (n = 333) was also used and is publicly available (3).

For all cases (except TCGA), tumor RNA was extracted from FFPE blocks or unstained slides after macrodissection guided by histologic review of the tumor lesion by a genitourinary pathologist. All cases had central pathology review prior to sampling for the Decipher assay, at least 0.5 cm² of tumor with ≥60% tumor cellularity was required for RNA extraction and microarray hybridization (Human Exon 1.0 ST GeneChips), which were performed in a Clinical Laboratory Improvement Amendments–certified laboratory facility (GenomeDx; ref. 4). Quality control was performed using Affymetrix Power Tools, and normalization was performed using the Single Channel Array Normalization algorithm (5). This study was conducted in accordance with the International Ethical Guidelines for Biomedical Research Involving Human Subjects.

AR gene expression was determined by summarizing 72 intronic and exonic probe sets within the AR gene. This AR-A signature was derived from prior work (6). The finalized AR-A signature used in this study was defined a priori, and the AR and TMPRSS2 genes were excluded from the original gene model that was selected to allow comparison of AR-A to AR expression and to not bias AR-A with ERG status. AR-A score was taken as a weighted linear sum of nine canonical AR transcriptional target genes (KLK3, KLK2, FKBP5, STEAP1, STEAP2, PPAP2A, RAB3B, ACSL3, and NKX3-1; ref. 6). Gene weights were based on their distribution skewness in a subset of the prospective cohort using “robustbase” R package. Patients with outlier AR-A score [less than mean (AR-A) – 1 × SD (AR-A)] were classified as low AR-A. Using the prospective discovery cohort, low AR-A was defined and locked as a score of 11 or less and then applied to the prospective validation and retrospective institutional cohorts to define AR-A low. For TCGA, which used an RNAseq platform, the same methodology was utilized to as the discovery cohort to define AR-A low.

The Molecular Signatures Database (MSigDB) was queried for 37 oncology related hallmark gene sets (7). Hallmark gene set scores were computed by taking the mean expression of each gene in the set. For immune cell quantification, we used immunophenoscores to measure suppressor immune cells (T regulatory and myeloid-derived suppressor cells) infiltration from gene expression data (8), and CIBERSORT tool to measure immune infiltration of 22 immune cells from gene expression data as described previously (9). FGFR activity score was calculated using z-score method in GSVA R package using nine FGF and four FGFR genes. Decipher and cell-cycle activity scores were extracted from the GRID as described previously (10). We investigated the expression of 39 neuroendocrine prostate cancer markers (11) in our large cohort of histologically confirmed adenocarcinomas. Stromal infiltration score was calculated by averaging 141 stromal genes reported previously (12). In addition, tumor purity based on consensus purity scores within TCGA for both gene expression and IHC were calculated (13).

Radiation sensitivity score was calculated using a gene expression signature developed and validated to predict response to radiotherapy (14). Drug sensitivity was calculated using in vitro drug sensitivity and microarray data to generate gene signatures predicting tumor sensitivity to 89 oncology drugs (15, 16). For each drug, CellMiner tool (15) was used to identify drug response–related genes and their correlations to the IC₅₀ value. Most significantly correlated genes were selected, and the expression of the corresponding genes were extracted for drug response score (DRS) calculations. A patient-specific DRS was calculated using these correlation coefficients as weighting factors of the corresponding gene expression normalized by the sum of correlations.

P53 missense mutation was detected using IHC assay as described previously (17). Each tissue microarray spot containing tumor cells was visually dichotomously scored for presence or absence of nuclear p53 signal by a urologic pathologist blinded to the gene expression data (T.L. Lotan). ERG and PTEN IHC was performed as described previously (4, 18).

Pearson correlation was used to assess correlation coefficients. Euclidean distance and ward linkage function was used for hierarchical clustering. We used FDR to adjust for multiple testing when we looked at association between low AR-A and gene expression and signature activity. Recurrence was defined per TCGA dataset, and metastatic disease for the retrospective cohorts was defined by radiographic evidence of metastatic disease. Development of CRPC was assessed within the JHMI cohort, and was defined as radiographic or biochemical progression in the setting of castrate levels of testosterone. Cumulative incidence curves were constructed using Fine–Gray competing risks analysis (19) to estimate the risk of metastasis over time with deaths from other causes as a competing risk. In addition, Kaplan–Meier analyses with log-rank test for recurrence (TCGA) or metastasis (retrospective cohorts) was performed. Time to distant metastasis from initial local therapy was modeled using multivariable competing risks regression analysis. Univariable logistic regression was used to associate AR-A with CRPC endpoint. Statistical analyses were performed in R v3.3.1, and all tests were using a 5% significance level.

Cohort and tumor characteristics are summarized in Table 1 and how each cohort was used in Fig. 1. Leveraging the prospective discovery primary prostate cancer samples, we first characterized the population-level variability in the distribution of the expression of 72 AR exon and intron probe sets (Fig. 2A), demonstrating remarkable heterogeneity in AR expression. Next, we rank ordered the expression of nine canonical transcriptional AR targets (ref. 6; Fig. 2B). After generation of a composite AR-A score, unsupervised hierarchical clustering revealed a distinct subclass of tumors with low AR-A expression, which was confirmed in the validation cohort (Supplementary Fig. S1). Importantly, analyses were conducted and confirmed that the subset of low AR-A primary prostate cancer was not impacted by either stromal contamination or tumor purity (Supplementary Fig. S2).

Overall, low AR-A primary prostate cancer was uncommon (∼10% across all cohorts; Fig. 2C). The distribution of AR-A is a skewed distribution (Shapiro–Wilk normality test; P < 0.0001), rather than a normal distribution, with a significant tail that captures the low AR-A subclass (Fig. 2C). Notably, the AR-A signature utilized in this study is not unique to capture AR signaling and has a high correlation to other AR-A signatures, including Kumar and colleagues (ref. 2; r = 0.82) or AR response from Hallmarks of cancer signatures (ref. 7; r = 0.74; Supplementary Fig. S3). Interestingly, baseline pretreatment serum PSA is often suggested to be a clinical surrogate for AR-A. However, serum PSA had no correlation to either intratumoral AR-A (r = 0.06) or KLK3 expression (r = 0.01; Fig. 2D; Supplementary Fig. S4). Low AR-A tumors were enriched in higher grade tumors (10%, 14%, and 22% for Grade group 1–3, 4, and 5, respectively; Supplementary Fig. S5).

To better understand the biologic phenotype of low AR-A tumors, a series of gene expression and IHC analyses were conducted. Low AR-A tumors were more likely to be triple negative (ERG⁻, ETS⁻, and SPINK⁻), and resemble a basal (or nonluminal) subtype (Fig. 3A; Supplementary Fig. S6). High AR-A tumors were more likely to ERG⁺, which was confirmed by IHC for ERG staining in a subset of patients from the JHMI cohort. Low AR-A tumors had more p53 mutations in TCGA, which was confirmed in the JHMI cohort in that by IHC p53 staining assay (17) was found only in lower AR-A tumors (Supplementary Fig. S6).

MSigDB analyses demonstrated that high AR-A tumors were associated with androgen response, DNA repair, and cell cycle (Supplementary Fig. S7A). Low AR-A tumors were significantly associated with multiple immune response gene sets. CD3 expression, which is specific for effector immune cells, was significantly overexpressed in low AR-A tumors (Fig. 3B). In addition, immune suppressor regulatory T cells and myeloid-derived suppressor cell signature scores (8) were lower in low AR-A tumors. Low AR-A tumors showed higher activity of IFNα, IFNγ, and TNF signaling and higher activity of genes needed for immune cell recruitment and lower activity of cytosolic nucleic acid–sensing pathways (Supplementary Fig. S7B). Furthermore, Gene set enrichment analyses demonstrated increased enrichment of signatures for chemokine–chemokine interactions, PD1 signaling, and CD3 phosphorylation (Supplementary Fig. S8). In addition, using CIBERSORT, low AR-A tumors are estimated to have increased neutrophils, B cells, activated mast cells, gamma delta T cells, activated dendritic cells, eosinophils, and activated memory CD4 T cells. Collectively this data suggests low AR-A tumors may have enhanced immunogenicity.

Low AR-A tumors had significantly decreased DNA repair pathway expression, including individual genes of mismatch repair (PMS2 and MLH1) as well as mismatch repair pathway gene sets (Fig. 3B; Supplementary Fig. S9; P < 1e⁻²⁰). Low AR-A tumors also had significantly lower pathway expression of homologous recombination (P < 1e⁻²⁰; Fig. 3B).

Given the increased understanding and recognition of the spectrum from adenocarcinoma to neuroendocrine prostate cancer (20), we investigated the expression of 39 neuroendocrine prostate cancer markers (11) in the prospective cohorts of histologically confirmed adenocarcinomas. After adjusting for FDR, low AR-A tumors had significantly higher neuroendocrine marker expression of 32 of the 39 neuroendocrine biomarkers, including NCAM1, ENO2, and SYP (Fig. 3B; Supplementary Fig. S10; Supplementary Table S2; P < 1e⁻¹⁰). Recent work has also demonstrated that in AR-null/neuroendocrine prostate cancer–null mCRPC tumors, FGF pathway activation is utilized to bypass AR dependence (21). We found that FGF activity is significantly increased in low AR-A primary prostate cancer (P < 1e⁻¹⁰), suggesting this bypass event may begin in primary tumor before androgen ablation (Supplementary Fig. S11). In addition, low AR-A primary prostate cancer had significantly higher expression of alternative nuclear hormone receptors, including PGR, NR3C1, and ESR1 (P < 1e⁻¹⁰; Supplementary Fig. S12) consistent with observations in mCRPC (2).

Given low AR-A tumors were more likely to harbor a more aggressive biologic phenotype, we next assessed the prognostic difference of low versus high AR-A tumors. Across the prospective discovery, prospective validation, JHMI, and BWH cohorts, lower AR-A patients were significantly more likely to have higher Decipher scores, a biomarker of metastatic potential (Supplementary Fig. S13). To clinically validate this, we analyzed AR-A in the JHMI natural history cohort (22) for the cumulative incidence of metastases by both AR-A quartile and by a 4-tiered system of ≤10%, 10%–50%, >50%–90%, and >90% AR-A expression. Lower AR-A in both analyses was associated with worse metastatic outcome (P < 0.001), and the patients with tumors having the lowest decile of AR-A had the worst metastatic outcome (Supplementary Fig. S14). This was confirmed when analyzing patients with low AR-A tumors versus all others in three independent cohorts; TCGA [HR, 2.00; 95% confidence interval (CI), 1.02–3.57; P = 0.04; Fig. 4A], JHMI (HR, 1.82; 95% CI, 1.27–2.63; P < 0.001; Fig. 4B), and BWH (HR, 5.26; 95% CI, 1.75–14.29; P < 0.001; Fig. 4C). The median time to event for AR-A low patients in TCGA, JHMI, and BWH was 82, 96, and 76 months, respectively. The median time to event for AR-A high patients was not reached in any cohort.

These findings were further confirmed in a multivariable competing risks regression analysis adjusting for age, PSA, Gleason grade, surgical margin status, extracapsular extension, seminal vesicles invasion, lymph node invasion, and adjuvant and salvage treatment, and low AR-A remained significantly associated with an increased risk for developing metastatic disease (HR, 2.61; 95% CI, 1.22–5.60; P = 0.01; Fig. 5A). To account for any potential batch effects, batch was included in the model, and low AR-A remained significant (HR, 2.69; 95% CI, 1.21–5.96; P = 0.02; Supplementary Table S3).

The most common treatment for prostate cancer are targeted to the AR or elicit DNA damage using radiotherapy or chemotherapy, and thus treatment sensitivity may differ by AR-A. To test potential differing pharmacologic sensitivities, we used in vitro drug sensitivity and microarray data to generate drug response scores for 89 oncology drugs (Supplementary Fig. S15). Low AR-A tumors had significantly lower predicted sensitivity than AR-A high tumors for response to androgen deprivation therapy (ADT; OR, 0.61; 95% CI, 0.52–0.73; P < 0.001; Fig. 5B). This was then validated clinically using the JHMI cohort to determine in patients that developed metastatic disease who were receiving chemical castration whether low AR-A predicted development of CRPC (e.g., resistance to ADT). Again, low AR-A tumors were significantly less likely to be sensitive to ADT and more likely to develop CRPC (OR, 0.41; 95% CI, 0.21–0.80; P = 0.008; Fig. 5B; Supplementary Fig. S16). Furthermore, we demonstrate that clinically low AR-A primary prostate cancer tumors had similar AR-A scores as AR-independent tumors from more advanced or aggressive disease states (CRPC, neuroendocrine, and/or small-cell prostate cancer; Supplementary Fig. S17).

Within the prospective cohort, tumors with low AR-A had increased sensitivity scores for platinum chemotherapy (OR, 8.08; 95% CI, 6.36–10.29) and PARP inhibitors (OR, 5.24; 95% CI, 4.24–6.47), and predicted worse response to taxane chemotherapy (OR, 0.14; 95% CI, 0.11–0.18; Fig. 5C). Notably, the cisplatin and PARP inhibitor signatures have <5% gene overlap, and independently correlated with AR-A.

Finally, we investigated AR-A as a predictive biomarker of treatment response to radiotherapy. Utilizing a 24-gene signature developed by Zhao and colleagues that predicts radiation response (PORTOS; ref. 14), we demonstrated that low AR-A tumors have significantly higher PORTOS scores, denoting increased potential radiation sensitivity (P < 1e⁻²⁰; Supplementary Fig. S18).

In the largest full transcriptomic analysis of primary prostate cancer, comprising of nearly 20,000 patients, we have performed a detailed series of analyses to gain an understanding of the biological and clinical relevance of AR-A. We provide robust clinical results from a large prospective cohort, with multiple forms of independent validation, of a low AR-A subclass of prostate adenocarcinomas that has high metastatic potential and distinct therapeutic sensitivities. Furthermore, we identify a subset of treatment-naïve primary prostate cancer that has comparable AR-A with heavily pretreated mCRPC, and similar biological phenotypic characteristics (e.g., increased neuroendocrine prostate cancer markers and FGF signaling, decreased DNA repair expression, and predominately basal subtype). Our findings suggest that a portion of the observed differences in biology between primary prostate cancer and mCRPC may be an expansion of a preexisting subset of low AR-A tumors rather than solely therapy-induced changes, consistent with other work (23, 24). Our work suggests that these men can and should be identified prior to initial therapy.

The data presented here offer several additional novel insights. First, we demonstrate that there is marked interindividual heterogeneity in AR and AR-A expression in primary prostate cancer across multiple retrospective and prospective cohorts. Second, we demonstrate that serum PSA, a clinical biomarker often used to infer AR-A, had no correlation to intratumoral AR-A, or even the gene for PSA, KLK3. Thus, PSA should not be used in isolation to clinically assign AR-A in primary prostate cancer. We hypothesize that in primary prostate cancer, PSA is a marker of tumor burden and prostate size, and is unreliable to capture intrinsic AR-A.

Third, the diversity in AR signaling in primary prostate cancer represents important biological heterogeneity that is both prognostic and predictive of treatment response. ADT is commonly delivered as concurrent or adjuvant therapy to radical treatment based on multiple randomized trials demonstrating benefit in unselected populations (25, 26). Furthermore, more intensive combinatorial approaches are showing promise, including recent data from STAMPEDE from the addition of abiraterone to standard LHRH agonist therapy (27). These clinical trial results are consistent with our data, as only approximately 10% of primary prostate cancer would be classified as low AR-A, and thus the vast majority of patients would be predicted to be sensitive to AR-directed therapies. However, despite the general efficacy of combined modality therapy, approximately 10% of high risk men will develop distant metastases 10-years posttreatment (28), and we show these patients are more likely to harbor low AR-A tumors.

Currently, there are no prospectively validated predictive biomarkers in primary prostate cancer to help select men toward a specific therapeutic approach. Recently, luminal and basal subtyping of patients with prostate cancer using the breast cancer classifier, PAM50, has demonstrated the ability to predict which patients are most likely to respond to post-prostatectomy ADT (29). Our data shed insights into underlying biology for these findings, in that we demonstrate that low AR-A patients are most likely to be of basal subtype, which Zhao and colleagues have shown to have less sensitivity to ADT than luminal B tumors (which more often have higher AR-A). Our study builds upon these efforts by providing sound biologic rationale and provocative results of a distinct subclass with a poor prognosis to current standard-of-care therapies.

Our study is timely, given the greater understanding of DNA repair defects (e.g., ATM and BRCA2), microsatellite instability, and the spectrum of neuroendocrine prostate cancer, which has sparked interest in combined modality therapy with PARP inhibitors (NCT02324998), platinum chemotherapy (NCT03275857), and immunotherapy (NCT02787005). Our data demonstrates that low AR-A tumors are not only more resistant to ADT and potentially docetaxel, but also are more sensitive to radiotherapy and alternative nonstandard-of-care treatment options, including PARP inhibition and platinum chemotherapy. Furthermore, low AR-A tumors have increased immunogenicity and decreased expression of mismatch repair, both potential markers of an ideal population to investigate immunotherapeutic strategies on. Our study will require prospective validation to confirm our drug sensitivity predictions.

Our study has limitations. To minimize potential sources of bias or limitations, we performed robust validation of all analyses in at least one if not multiple independent cohorts. Analyses for contamination of stromal content, assessment of tumor purity, and correction for batch effects were performed, which are known potential confounders within gene expression analyses, and we were unable to demonstrate this as an unlikely source of bias. There are known differences in microarray and RNA-seq data, and thus TCGA was used to validate our findings, which showed similar distribution of AR-A and a negative prognostic impact of low AR-A status. We demonstrate that our AR-A score is highly correlated to other AR-A scores in the literature. However, AR signaling is complex and context specific, and alternative AR-A models could improve the utility of using AR-A to serve as a predictive biomarker. Finally, time to CRPC analyses was not performed and rather simply the development of CRPC as a binary event was used given exact dates for the formation of CRPC could not be collected.

In summary, our study establishes low AR-A primary prostate cancer as a clinically relevant subclass of treatment-naïve localized prostate adenocarcinoma that harbors biology more akin to mCRPC. This aggressive subtype of basal-like low AR-A tumors is more likely to develop resistance to ADT and be less responsive to docetaxel, and may have other distinct treatment sensitivities. Thus, dedicated biomarker enhanced clinical trials in earlier stages of the disease are warranted for these patients.

D.E. Spratt is an advisory board member/unpaid consultant for Blue Earth and Janssen. N. Fishbane holds ownership interest (including patents) in Decipher Biosciences. S.G. Zhao is an employee/paid consultant for PFS Genomics and holds ownership interest (including patents) in Decipher Biosciences. T.M. Morgan reports receiving other commercial research support from GenomeDX. A.P. Dicker is an advisory board member/unpaid consultant for Roche, Oncohost, Google, Dreamit Ventures, EMD Serono, Janssen, Cybrexa, ThirdBridge, and Wilson Socini. S.J. Freedland reports receiving commercial research grants from Genome DX. R.J. Karnes reports receiving commercial research grants from and holds ownership interest (including patents) in GenomeDX. E. Davicioni is an employee/paid consultant for and holds ownership interest (including patents) in Decipher Biosciences. A.E. Ross is an advisory board member/unpaid consultant for GenomeDx Biosciences. P.L. Nguyen is an employee/paid consultant for Janssen, Astellas, Ferring, Bayer, Nanobiotix, GenomeDX, Blue Earth, Augmenix, Boston Scientific, and Dendreon, and reports receiving commercial research grants from Janssen, Astellas, and Bayer. F.Y. Feng is an employee/paid consultant for Astellas, Bayer, Celgene, Janssen, Sanofi, Genentech, and EMD Serono, and holds ownership interest (including patents) in PFS Genomics and Nutcracker Therapeutics. T.L. Lotan reports receiving commercial research grants from GenomeDX and Ventana/Roche. No potential conflicts of interest were disclosed by the other authors.

Conception and design: D.E. Spratt, M. Alshalalfa, N. Fishbane, B.A. Mahal, S.G. Zhao, C. Speers, A.P. Dicker, E. Davicioni, R.B. Den, E.M. Schaeffer

Development of methodology: D.E. Spratt, M. Alshalalfa, N. Fishbane, B.A. Mahal, Y. Liu, S.G. Zhao, A.P. Dicker, E. Davicioni, E.M. Schaeffer

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): D.E. Spratt, M. Alshalalfa, R. Mehra, Y. Liu, A.P. Dicker, S.J. Freedland, R.J. Karnes, S. Weinmann, E. Davicioni, A.E. Ross, R.B. Den, P.L. Nguyen, T.L. Lotan, E.M. Schaeffer

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): D.E. Spratt, M. Alshalalfa, N. Fishbane, A.B. Weiner, R. Mehra, J. Lehrer, Y. Liu, S.G. Zhao, C. Speers, T.M. Morgan, R.J. Karnes, E. Davicioni, R.B. Den, P.L. Nguyen, F.Y. Feng, E.M. Schaeffer

Writing, review, and/or revision of the manuscript: D.E. Spratt, M. Alshalalfa, N. Fishbane, A.B. Weiner, R. Mehra, B.A. Mahal, Y. Liu, S.G. Zhao, C. Speers, T.M. Morgan, A.P. Dicker, S.J. Freedland, R.J. Karnes, S. Weinmann, E. Davicioni, A.E. Ross, R.B. Den, P.L. Nguyen, F.Y. Feng, A.M. Chinnaiyan, E.M. Schaeffer

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): D.E. Spratt, E. Davicioni, R.B. Den, A.M. Chinnaiyan, E.M. Schaeffer

Study supervision: D.E. Spratt, T.M. Morgan, E. Davicioni, E.M. Schaeffer

This work was funded in part by the Prostate Cancer Foundation Young Investigator Award (to D.E. Spratt), Prostate Cancer Foundation Challenge Award (to E.M. Schaeffer), the Department of Defense (to D.E. Spratt; W81XWH-16-1-0571), U01CA196390 (to E.M. Schaeffer), P50 CA186786 (to D.E. Spratt and A.M. Chinnaiyan), and generous philanthropic gifts from patients (SEH, MP, and PM) and the Ambrose Monell Foundation.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.