Novel RB1-Loss Transcriptomic Signature Is Associated with Poor Clinical Outcomes across Cancer Types

RB1 is a tumor suppressor that has been implicated in the pathogenesis of numerous cancer types. In addition to causing pediatric retinoblastoma, RB1 alterations have been shown to play a major role in the progression of osteosarcoma (1), lymphoma (2), and breast (3–5), lung (6, 7), and prostate (8, 9) malignancies. Moreover, recent studies have highlighted RB1 loss as an important clinical prognostic factor in specific cancer types. For example, RB1 loss has been shown to be associated with poor overall survival (OS) in osteosarcoma (1), glioblastoma (10), and lung cancers (11) and has been shown to predict resistance or sensitivity to various small-cell lung cancer (7), pancreatic cancer (12), and breast cancer therapies (3, 13).

In order to study the clinical implications of Rb-pathway disruption, one must first be able to confidently assess RB1 status. Next-generation DNA-sequencing approaches are well suited for identifying mutations, copy-number alterations, and structural variants. However, there is often uncertainty as to whether a DNA alteration truly inactivates the affected allele. Moreover, other mechanisms of gene inactivation exist that may not be captured by DNA-sequencing techniques (e.g., epigenetic, posttranscriptional, or posttranslational modifications). An alternative approach to assessing gene inactivation is to examine the sequelae of genomic alterations by assessing the resulting expression of related, downstream genes.

There exist a few RB1 gene sets (genes theorized to be collectively indicative of RB1 status) and two gene signatures (combinatorial expression pattern of the genes in a gene set) for predicting RB1 loss (14, 15). However, they all share the key limitation that they consist largely of cell-cycle genes (whose expression is not specific to RB1 loss). Moreover, because these gene sets and signatures were primarily developed using breast cancer data, their generalizability to different cancer types has not been validated. Our first aim was to develop a novel pan-cancer RB1 biallelic loss gene signature (RBS) that outperformed existing RB1-loss signatures and accurately predicted biallelic RB1 loss across cancer types.

After generating and validating RBS, we then sought to use it to assess RB1 loss as a prognostic factor across all major cancer types using The Cancer Genome Atlas (TCGA) Pan-Cancer database (N = 11,007). Because RB1 loss was known to be clinically important in metastatic prostate cancer (not included in the TCGA Pan-Cancer dataset), we examined the prognostic significance of RBS in an independent metastatic castration-resistant prostate cancer (mCRPC) cohort.

We defined “RB1 loss” in our article as predicted biallelic loss of RB1. For the purposes of training and testing our RB1-loss classifier (RBS), ground-truth labels of RB1 status for each tumor were assigned based on the number of DNA alterations (i.e., nonsilent exonic mutations, copy-number loss, and inactivating structural variants) observed in RB1. For these ground-truth labels, RB1 loss was defined as presence of at least two DNA alterations in RB1.

Taking an unbiased approach to selecting genes indicative of RB1 loss, we leveraged microarray log₂-normalized RPKM gene expression data of 951 pan-cancer cell lines from the Cancer Cell Line Encyclopedia (CCLE; ref. 16). We extracted GISTC2.0 (17) and whole-exome sequencing (WES)–based mutation calls from UCSC Xena Browser to annotate RB1 copy number (CN) and mutation calls (18). Cell lines with GISTC score < -0.8 were annotated as deep (two-copy) deletion (CN-2), and cell lines with GISTC score between -0.8 and -0.4 were annotated as shallow (single-copy) deletion (CN-1). The remaining cell lines were annotated as two-copy intact (CN-0). To build an mRNA classifier to predict RB1 functional loss, we defined the tumor cell lines with predicted biallelic loss (i.e., deep deletion, shallow deletion with additional DNA mutation, or 2+ DNA mutations) as the RB1-loss group and remaining cell lines as the RB1-intact group. To identify differentially expressed genes between the two groups, we used the Wilcoxon Mann–Whitney test with an adjusted P value threshold of P < 1 × 10⁻¹⁰.

We then used a nearest shrunken centroid approach (PAM; ref. 19) to generate our gene signature based on the expression pattern of the genes selected as described above. We trained the model by applying PAM to CCLE expression data, using posterior class probabilities for RB1-loss class predictions. The model was trained using 10-fold cross validation to optimize the PAM shrinkage parameter.

RBS was then validated on the TCGA Pan-Cancer RNA-sequencing (RNA-seq) expression dataset of 11,007 tumor samples spanning 33 cancer types, downloaded from UCSC Xena Browser using the Synapse platform (syn4976369). RB1 copy-number calls and mutation data for these samples were obtained from UCSC Xena Browser, and the same GISTIC2.0 copy-number thresholds and mutation criteria as used in the CCLE training set were applied to the validation set. Final RB1-loss annotations were defined based on the number of RB1 DNA alterations observed: 2 alterations (deep deletion, shallow deletion with one mutation, or 2+ mutations), 1 alteration (shallow deletion with no mutations or one mutation with no deletion), 0 alteration (no deletions or mutations). Model accuracy was assessed based on area under the ROC curve (AUC), benchmarked against the leading existing RB1-loss signature (14).

The EnrichR web tool was used to identify genomic pathways enriched in the RBS gene set. Candidate gene sets were defined as all pathways in the KEGG, Reactome, WikiPathways, and BioCarta databases. Pathways were considered significantly enriched if their adjusted P values were less than a predetermined significance level of 0.05.

Differential expression analysis between CN-0 tumors with no mutations and CN-0 tumors with two or more mutations was performed to identify genes that were differentially expressed in tumors with 2+ RB1 mutations. Given that there were far fewer tumors with 2+ mutations than there were with no mutations, we randomly subsampled a set of CN-0 tumors with no mutations equal in size to the subset of tumors with 2+ mutations. We then performed a differential expression analysis between the tumors with 2+ mutations and the tumors with no mutations using the Wilcoxon Mann–Whitney test with an adjusted P-value threshold of P < 0.001. For statistical robustness, we performed a bootstrapped analysis with 1,000 different subsamples. Genes were considered significantly differentially expressed if, in >95% of all comparisons, they demonstrated the same directionality of over- versus underexpression and had adjusted P values below the predetermined significance level of 0.001.

To assess the utility of RBS in detecting gene silencing due to methylation, we downloaded TCGA Pan-Cancer methylation data for 49 RB1 methylation probes from the UCSC Xena Browser. We first filtered out probes that were previously identified to be of low quality (20). We then computed Spearman correlation coefficients between RBS score and Illumina DNA methylation 450K array beta values for each RB1 methylation probe. To test whether the correlations between RBS score and methylation probe values were significant in the RB1 promoter and gene body regions, we generated a null model by computing the correlation between RBS score and methylation in the promoter and gene body regions of 20 other random tumor suppressors not known to be related to RB1. For this analysis, a large set of tumor suppressors (N = 1,217) was downloaded from the Tumor Suppressor Gene Database (21) and those not located on the same chromosome as RB1 (i.e., not on chromosome 13) and not included in RBS were used as candidate genes for the null model. Spearman correlation coefficients computed between RBS and each methylation probe in the promoter region of a gene (defined as ± 1.5kb of the transcription start site; ref. 22) were then modeled as a distribution. The distribution of correlations between RBS and RB1 promoter methylation probes was compared with the distribution of correlations between RBS and non-RB1 promoter methylation probes using the Kolmogorov–Smirnov test with a two-sided significance level of 0.05. Analogous analyses were performed for the gene body region, where gene body was defined as the region 1.5 kb downstream of the transcription start site to the transcription terminator. Transcription start sites and terminators were defined using the “biomaRt” R package (23).

Clinical outcomes data [progression-free survival (PFS), OS, and disease-specific survival (DSS)] were obtained from the TCGA Pan-Cancer Clinical Data Resource (24). All patients with available log₂-normalized RPKM RNA-seq data and clinical outcomes data were included in survival analyses. Microarray expression data were log₂-normalized and scaled prior to RBS analysis. Data for the mCRPC cohort were obtained from a previously published study (25). This cohort consisted of 101 patients with deep whole-genome sequencing, whole-transcriptome RNA-seq, and clinical outcomes data available. The mCRPC RNA-seq data were log₂-normalized FPKM values. The clinical endpoint examined was OS, with time of study entry defined as date of mCRPC diagnosis.

The threshold of RBS score used to assign binary RB1-impaired versus RB1-intact status in both cohorts was determined by using the Youden index (computed using the “OptimalCutpoints” R package; ref. 26) to select a threshold that maximized prediction accuracy in the CCLE training dataset. Cox proportional hazard models were used to model time-to-event data. All survival analyses were performed using R version 3.5.0.

To define our RB1-loss gene set, we identified genes that were differentially expressed between CCLE cell lines that demonstrated RB1 loss and cell lines that had intact RB1. Note that 951 of the 995 total cell lines had both copy-number and microarray expression data available. Of these 951, 126 were identified as having biallelic RB1 loss (99 harbored two-copy deletions, 23 harbored single-copy deletions with an additional mutation, and 11 harbored 2+ mutations) and 797 were identified as RB1 intact. Our unbiased approach to defining an RB1-loss gene set using CCLE data identified a final set of 186 genes that were indicative of RB1 loss (Supplementary Table S1A). Of note, only 7 of the 186 genes overlapped with genes in the existing RB1-loss signature (14).

To assess the potential utility of our 186-gene signature for predicting RB1 loss, we first performed t-SNE dimensionality reduction on the CCLE training data (N = 951). Visualization of the t-SNE embedding revealed that cell lines with 2+ DNA alterations in RB1 tended to map to similar parts of the embedding, suggesting that these cell lines had similar 186-gene expression profiles (Fig. 1A). This finding supported the hypothesis that the 186 genes were useful in differentiating between RB1-loss and RB1-intact samples.

The expression values of the 186 genes nominated as described above were then used in a supervised learning approach (PAM) to compute an RBS score for predicting RB1 loss. The model was trained using the gene expression profiles of CCLE cell lines with known RB1 status (i.e., RB1-loss vs. RB1-intact). The model identified 144 genes whose expression values were most predictive of RB1 status—these genes were used to compute the final RBS score (Fig. 1B; Supplementary Table S1B). RB1 and CCND1 were among the genes expressed at relatively low levels in RB1-loss samples, whereas CDKN2A was among the genes expressed at relatively high levels in RB1-loss samples. This was consistent with prior studies which found a high ratio of CDKN2A to CCND1 expression to be associated with RB1 loss in multiple cancer types (27, 28). Because we noticed that prior RB1 gene sets and gene signatures largely consisted of cell proliferation genes, we assessed the association between RBS and a previously published cell proliferation activity score (29). Although a previously published RB1-loss signature (14) was found to be strongly correlated with the cell proliferation score (r = 0.93), we found that RBS was only weakly correlated with the cell proliferation score (r = 0.03). These findings suggested that RBS was not a surrogate marker for cell proliferation and was potentially more specific to RB1 loss than existing signatures. Moreover, EnrichR pathway enrichment analysis revealed that RBS was enriched for genes not only in the cyclin D–CDK4/6 and cell-cycle–related pathways but also in the DNA damage response and TP53 signaling pathways (Supplementary Table S2). Altogether, these results were consistent with recent literature that suggests RB1 may play a role in processes other than cell-cycle control (30).

To validate RBS as an accurate model for predicting biallelic RB1 loss, we used the TCGA Pan-Cancer atlas expression dataset containing RNA expression data for 11,007 tumors spanning 33 cancer types with known mutation and copy-number annotations. Note that 698 of these samples were annotated as having two or more RB1 DNA alterations [559 had deep deletion (CN-2), 89 had shallow deletion with mutation (CN-1/mut), and 50 had two or more mutations with no deletions], 1,514 as having one RB1 alteration [1,332 with shallow deletion and no mutation (CN-1/no-mut), and 182 with one mutation and no deletions (CN-0/mut)], and 7,727 as having no RB1 DNA alterations. RBS achieved an AUC of 0.89 for predicting RB1 biallelic loss in this validation set—far superior to an AUC of 0.66 achieved by applying the leading existing RB1-loss signature (14) to the same dataset (Fig. 2A and B). RBS also outperformed a predictive model based solely on the ratio of CDKN2A to CCND1 expression (AUC = 0.72), which was previously reported to be associated with RB1 loss. Genes including CAMK2N2, CDKN2A, and GPR137C were positively correlated with RBS score (i.e., high expression in RB1 loss), whereas genes including MED4 and RB1 were most negatively correlated with RBS score (Fig. 2C).

RBS was highly accurate at identifying RB1 loss due to deep deletion and due to shallow deletion with additional mutation, which comprised the large majority of RB1-loss tumors. However, RBS was less effective at detecting the few RB1-loss tumors with 2+ RB1 mutations, suggesting that these tumors may have a distinct gene expression profile. To investigate this further, we performed a bootstrapped differential expression analysis to identify genes over- and underexpressed in CN-0 tumors with two or more RB1 mutations compared to tumors with no RB1 mutations (Materials and Methods). We identified 448 genes significantly overexpressed and 245 genes significantly underexpressed in the tumors with two or more RB1 mutations (Supplementary Table S3). Of these, 16 overexpressed genes (including CCNE2 and CDKN2A) and 3 underexpressed genes (most notably RB1) were also found in RBS. In addition, several known regulators or effectors of RB1 such as CCNE1, CDK2, EZH2, HOXB7, and select E2F-family genes were not in RBS but were differentially expressed in the tumors with two or more mutations in RB1 (30–34). Altogether, these findings suggested that there are some transcriptomic similarities but also notable differences between RB1 loss due to deletion and due to biallelic RB1 mutations.

To assess the utility of RBS in capturing the effects of epigenetic events on gene expression, we examined the correlation between RBS score and the methylation scores of 39 methylation probes in the Pan-Cancer cohort (Fig. 3). To test whether the pattern of correlation between RBS and methylation probe values was significant in the RB1 promoter and gene body regions, we compared our results with the correlation between RBS score and methylation in the promoter and gene body regions of 20 other random tumor suppressors unrelated to RB1 (Supplementary Table S4). We found that the positive correlation between RBS score and RB1 promoter methylation and negative correlation between RBS score and RB1 gene body methylation were significant (P = 0.0077 and P = 0.0016, respectively). The directionality of correlation was also consistent with existing literature, which suggests that promoter methylation is associated with decreased gene expression and gene body methylation is associated with increased gene expression in tumor suppressors (22). These findings supported the hypothesis that RBS could detect the downstream effects of RB1 loss due to multiple etiologies, including those (such as methylation) that may not be captured using DNA-sequencing techniques.

After assessing the accuracy of RBS for predicting RB1 loss, we sought to use RBS to investigate the prognostic significance of RB1 loss across cancer types. For this analysis, we included patients in the TCGA Pan-Cancer dataset with available clinical follow-up. High RBS was defined as scores above a threshold of 0.6, determined based on the Youden Index approach applied to the CCLE training dataset. Of note, we found that the majority of cancer types had an RB1 2-hit prevalence of greater than 5%, suggesting that RB1 loss was common and potentially important in many cancer types. In our pooled analysis of all patient samples across cancer types, we found that RB1 loss defined using RBS was predictive of short PFS [median PFS, 36 vs. 56 months; HR, 1.3; 95% confidence interval (CI), 1.18–1.44; P < 0.0001; Fig. 4A], short DSS (median DSS, 88 vs. 219 months; HR, 1.34; 95% CI, 1.17–1.55; P < 0.0001; Fig. 4B), and short OS (median OS, 70 vs. 94 months; HR, 1.23; 95%, CI, 1.09–1.38; P = 0.0004; Fig. 4C). In a multivariable survival model including both RBS and cancer type, high RBS was found to be independently prognostic of short PFS (HR, 1.12; 95% CI, 1.02–1.26; P = 0.04). These findings supported the hypothesis that RB1 loss is clinically important across cancer types and may indicate more advanced or aggressive disease in general.

We additionally assessed the prognostic significance of a DNA-sequencing–based definition of RB1 loss, namely, having at least two DNA alterations in RB1. We found that similarly to high RBS, presence of 2+ DNA alterations in RB1 was associated with short OS, PFS, and DFS compared with presence of 0 or 1 DNA alterations in RB1 (Fig. 4D–F). These findings suggested that our definition of “RB1 loss” as predicted biallelic loss of RB1 was clinically meaningful.

In our methylation analysis, we showed that RBS could potentially be used to detect RB1 loss through mechanisms that could not be detected by DNA sequencing. In addition, it is known that not all DNA mutations and copy-number loss events in a gene have the same effect on the affected allele (i.e., resulting protein may still be partly or completely functional). Because RBS measures the downstream effects of DNA and non-DNA alterations at the gene expression level, we hypothesized that RBS may offer information on Rb-pathway disruption that is independent of DNA-sequencing results.

To explore this hypothesis, we assessed the prognostic significance of high RBS for predicting survival in the TCGA Pan-Cancer cohort independently of number of observed DNA alterations. For these analyses, we focused on PFS as our clinical endpoint of interest due to a prior study that found that PFS was generally the most accurate endpoint collected across all cancer types in the TCGA Pan-Cancer dataset (24). On multivariable analysis adjusting for number of DNA alterations in RB1, high RBS was independently predictive of short PFS (HR, 1.14; 95% CI, 1.02–1.29; P = 0.03). This suggested that RBS may help distinguish patients with a more pronounced RB1-impaired clinical phenotype from those with a less-pronounced phenotype independently of the number of DNA alterations observed in the gene. Moreover, using a criterion of high RBS or 2+ DNA alterations in RB1 to select RB1-impaired patients resulted in a 73% increase in group size as compared with using the criteria of just 2+ DNA alterations (Supplementary Fig. S1A). Thus, RBS may be useful for identifying a more comprehensive group of patients with Rb-pathway disruption than can be recovered using DNA sequencing alone.

To explore this concept further, we examined a previously published cohort of patients with mCRPC (25)—the lethal subtype of prostate cancer not represented in the TCGA Pan-Cancer cohort. RB1 loss (as defined based on detected DNA alterations in RB1) has been shown to be associated with short survival in mCRPC (35). Interrogating the mCRPC cohort of 101 patients with both whole-genome sequencing and RNA-seq data available, we aimed to assess whether high RBS might be predictive of short OS independently of the number of DNA alterations present. First, we examined the degree of concordance between RB1 status as defined based on number of DNA alterations observed and as defined based on RBS score. We found that although RBS score was strongly related to the number of DNA alterations observed (AUC = 0.90), not all tumors with high RBS score harbored 2+ DNA alterations and vice versa (Fig. 5A). By expanding the DNA-sequencing–based definition of RB1-loss (2+ RB1 DNA alterations) to include tumors with fewer than 2 DNA alterations in RB1 but with high RBS, one could recover 50% more tumors with RB1-impaired status (Supplementary Fig. S1B). Next, we examined the prognostic significance of high RBS in the mCRPC cohort. We found that RB1 loss as defined by high RBS was predictive of short OS in mCRPC (median OS, 15.0 vs. 42.0 months; HR, 2.93; 95% CI, 1.47–5.83; P = 0.001; Fig. 5B). Finally, to assess whether the RNA-seq (high RBS) and DNA-sequencing (number of DNA alterations in RB1) results were independently predictive of survival outcomes, we performed a multivariable analysis including both the RNA-seq and DNA-sequencing definitions of predicted RB1 loss. We found that both the RNA-seq and DNA-sequencing definitions were independently predictive of short OS (P = 0.0036 and P = 0.046, respectively), suggesting that both RNA-seq and DNA sequencing offered unique information on RB1 status that could be used to detect a clinical phenotype of RB1-impaired, clinically aggressive mCRPC.

In order to assess the clinical implications of RB1 loss across cancer types, we developed a pan-cancer RB1-loss signature (RBS) that predicted biallelic loss of RB1 based on gene expression data. We found that RBS was highly accurate at predicting RB1 loss across cancer types compared with existing RB1 gene signatures. Moreover, RBS was able to capture RB1 inactivation due to both DNA and epigenetic changes. Using pan-cancer (N = 10,486) and metastatic prostate cancer (N = 101) cohorts, we demonstrated that high RBS was predictive of poor clinical outcomes independently of the number of DNA alterations in RB1.

There are several possible explanations as to why RBS was much more accurate than the leading existing RB1 signature at predicting biallelic loss of RB1 (AUC of 0.89 vs. 0.66). For one, RBS was the only RB1-loss signature that was designed to be applied across cancer types. Because RBS was trained on CCLE cell-line data derived from many different primary tissue types, it was well-suited to assess RB1 loss in the TCGA Pan-Cancer validation set, which also included patient samples from many different disease sites. Moreover, in contrast to existing RB1-loss signatures, which included genes largely or exclusively based on prior annotations, the RBS gene set was selected in an unbiased, unsupervised manner. Our approach nominated genes from the set of all existing genes that were most differentially expressed in our pan-cancer, RB1-loss training set samples. A final methodological strength of RBS was that it was trained on a very large dataset (N = 995) including many samples with known RB1 loss (N = 133) that could be collectively used to represent a distinct RB1-loss expression pattern.

It is important to note that that the “accuracy” of our model for AUC analyses was defined as concordance between (RBS-based) RB1-loss calls and DNA-sequencing–based variant calls (i.e., mutation, copy-number, and structural variant data when available). This was because DNA-sequencing results are commonly used to predict gene functional status and were the only data available for comparison. However, DNA-sequencing calls do not capture certain forms of gene inactivation such as DNA methylation of the RB1 promoter. Although RBS demonstrated high concordance with DNA-sequencing calls in our pan-cancer and mCRPC-specific analyses (AUCs of 0.89 and 0.87, respectively), the differences in RB1-loss assignments may not be due to error but rather improved identification of RB1 gene inactivation.

This study is not without limitations. We evaluated RBS as a potential tool to identify RB1 loss due to DNA-sequence alterations and DNA methylation at the RB1 locus. However, still other mechanisms of RB1 inactivation exist, such as CDK phosphorylation of the Rb protein (36, 37). It is unclear whether these mechanisms of RB1 inactivation result in a similar pattern of gene expression and whether RBS can be used to identify these Rb-inactivated tumors. Future work may involve collecting and integrating phosphoproteomic data with DNA-sequencing and RNA-seq data to study these additional cases of tumors with RB1 gene inactivation. In addition, because our analysis was conducted primarily using the CCLE and TCGA Pan-Cancer databases (which focus on primary cancers), an extension to metastatic cancers is needed. In particular, as RB1 loss and RB1 underexpression have been implicated as predictors of more advanced disease in various cancers (38–40), future disease-specific studies with a range of indolent and aggressive tumors may leverage RBS to study RB1 loss in the context of disease progression.

The data presented here offer several novel insights and contributions. First, our study is the first to examine the clinical implications of RB1 loss on a pan-cancer scale. We found that RB1 loss was associated with shorter PFS, OS, and DSS, highlighting the widespread clinical importance of the genomic event. Second, our novel transcriptomic signature (RBS) is highly accurate at predicting RB1 loss and can be used as a tool in future studies to shed new light on the biological and clinical impact of RB1 loss. This is especially relevant in light of recent studies which suggest that RB1 loss may associated with response to various cancer therapies including radiotherapy (3, 41), platinum-based chemotherapy (3, 7), and CDK4/6 inhibitors (13, 15) in breast, prostate, and small-cell lung cancers. RBS may be useful for detecting differential response to specific cancer therapies for an even broader range of therapies and cancer types than has been already studied. Third, RBS is specific to RB1-loss and not strongly correlated with cell proliferation scores (in contrast to existing RB1-loss signatures). Altogether, our study along with others suggest RB1 may have important functions aside from regulating cell proliferation, such as DNA damage repair (41–43). Additional studies are needed to assess this in greater detail. Fourth, our transcriptomic signature may be used to identify RB1-impaired tumors that may not be detected using standard DNA-sequencing–based definitions of predicted RB1 loss. The results of our multivariable analyses on two independent cohorts suggest that both RNA-seq and DNA-sequencing results may be useful to identify a more complete set of RB1-impaired patients.

Our approach to developing an RB1-loss signature is generalizable to studying a wide range of genomic alterations and may serve as a paradigm for generating expression-based gene signatures in an unbiased manner. Because RBS is an expression-based signature, it is complementary to and potentially more holistic than DNA-sequencing–based approaches, which may fail to capture the full spectrum of genomic events that can result in a specific gene expression profile or phenotype. Given the plethora of studies highlighting RB1 loss as a driver event in a number of cancer types, the potential clinical implications, and the increasing availability of gene expression data for both retrospective and prospective cohorts, RBS is an immediately useful tool that can be used to assess RB1 loss in a variety of settings. Our analyses and the findings of others suggest that RB1 loss may be predictive not only of survival but also of response to cytotoxic and targeted therapies. RBS may be invaluable for investigating these relationships further with the broader goal of developing personalized cancer treatment regimens.

S.G. Zhao and E. Davicioni hold ownership interest (including patents) in GenomeDx Biosciences. C.A. Maher holds ownership interest (including patents) in Illumina. K.E. Knudsen reports receiving commercial research grants from Celgene and CellCentric, speakers bureau honoraria from Celgene, and is a consultant/advisory board member for CellCentric. E.J. Small holds ownership interest (including patents) in Harpoon Therapeutics and Fortis Therapeutics, and is a consultant/advisory board member for Janssen, Fortis Therapeutics, and Beigene. P.L. Nguyen reports receiving commercial research grants from Janssen and Astellas, holds ownership interest (including patents) in Augmenix, and is a consultant/advisory board member for Augmenix, Boston Scientific, Ferring, Bayer, Dendreon, Blue Earth Diagnostics, Astellas, GenomeDx Biosciences, Nanobiotix, and Cota. F.Y. Feng reports receiving commercial research grants from Zenith, is a consultant/advisory board member for Bayer, Blue Earth Diagnostics, Celgene, Clovis, Janssen, EMD Serono, Sanofi, Dendreon, Ferring, and Astellas, and reports receiving other remuneration from PFS Genomics. No potential conflicts of interest were disclosed by the other authors.

Conception and design: W.S. Chen, M. Alshalalfa, S.G. Zhao, Y. Liu, B.A. Mahal, P.W. Kantoff, E.J. Small, P.L. Nguyen, F.Y. Feng

Development of methodology: W.S. Chen, M. Alshalalfa, S.G. Zhao, Y. Liu, B.A. Mahal, F.Y. Feng

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): W.S. Chen, M. Alshalalfa, Y. Liu, E. Davicioni, E.J. Small

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): W.S. Chen, M. Alshalalfa, S.G. Zhao, Y. Liu, D.A. Quigley, E. Davicioni, T.R. Rebbeck, C.A. Maher, K.E. Knudsen, P.L. Nguyen, F.Y. Feng

Writing, review, and/or revision of the manuscript: W.S. Chen, M. Alshalalfa, S.G. Zhao, Y. Liu, B.A. Mahal, D.A. Quigley, T. Wei, E. Davicioni, T.R. Rebbeck, P.W. Kantoff, C.A. Maher, K.E. Knudsen, E.J. Small, P.L. Nguyen, F.Y. Feng

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): W.S. Chen, E. Davicioni, T.R. Rebbeck, F.Y. Feng

Study supervision: T.R. Rebbeck, E.J. Small, P.L. Nguyen, F.Y. Feng

S.G. Zhao, B.A. Mahal, D.A. Quigley, E.J. Small, and F.Y. Feng are supported by the Prostate Cancer 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.