Single-Cell Circulating Tumor Cell Analysis Reveals Genomic Instability as a Distinctive Feature of Aggressive Prostate Cancer

Prostate cancer is associated with considerable clinical heterogeneity and molecular diversity between and within patients (1–8). Tumor burden, anatomical distribution, and line of treatment constitute the foundation of the clinical classification of advanced disease, but these factors do not fully capture the heterogeneity observed. Developing a more precise stratification based on biological characteristics has been hindered by the difficulty in reliably accessing metastatic tumor tissue, as the dominant site of prostate cancer metastasis is the bone.

Although metastatic castration-resistant prostate cancer (mCRPC) responds in a majority of cases to novel androgen receptor (AR) signaling inhibitors (such as abiraterone and enzalutamide), resistance almost invariably develops over months to years (1, 9, 10). In 20% to 30% of the patients, the disease directly fails to respond. AR inhibition–refractory and a number of resistant forms tend to present with atypical clinical characteristics, occasionally with neuroendocrine features upon pathologic assessment, and to behave with increased aggressiveness, resulting in poor prognosis and increased morbidity (11, 12). Because those patients with aggressive variant prostate cancer (AVPC) may benefit from intensified treatments (13), a number of clinical criteria were proposed to facilitate AVPC recognition (11). However, the clinical presentations may be difficult to distinguish, and treatment options remain limited and suboptimal, together making molecular characterization and more precise identification priorities for the field (11, 12, 14–16).

Preclinical and clinical studies have established candidate molecular markers for the classification of AVPC using prostate cancer cell lines, patient tumor-derived xenografts, and biopsy tissue specimens from primary and metastatic sites (12, 14, 15). Genomic aberrations affecting the tumor suppressor genes (TSG) PTEN, RB1, and TP53 are among the most frequently enriched in mCRPC relative to earlier prostate cancer presentations (2–4, 17). TSG loss-of-function has been linked to an adverse prognosis (3, 17–22). In particular, two-hit RB1 loss is frequently found in neuroendocrine prostate cancer, often together with concurrent alterations in PTEN and/or TP53 (12, 14, 15). Indeed, RB1-loss promotes lineage plasticity, anti-androgen blockade indifference, and a switch to a neuroendocrine transcriptional program in PTEN-deficient preclinical models in cooperation with TP53 loss (23–26). Moreover, the presence of at least two of those cooperative TSG alterations may indicate benefit from carboplatin in AVPC (13).

The collective knowledge of cancer genomics has been primarily built upon tissue sampling of the solid phase of disease: primary tumors and metastatic tissue. Comparatively, knowledge about the genomics of the seeds of metastases in transit, circulating tumor cells (CTC), is slim. Even slimmer is genomic data about this fluid phase at clinically relevant points in disease evolution. Compared with metastatic tissue sampling, CTC provides a minimally-invasive opportunity to serially characterize tumor heterogeneity at single-cell resolution (8). The Epic Sciences CTC platform (EsCTC) is based on direct and unbiased identification of all nucleated cells on a slide, inclusive of heterogeneous CTC populations (27, 28). EsCTC has been used to develop CTC phenotypic heterogeneity indices associated with increased overall survival (OS) benefit from taxane-based chemotherapy relative to AR-targeted treatments in mCRPC (27), and to portray CTC characteristics in patients with neuroendocrine prostate cancer (28, 29). In addition to describing morphometric and organizational features, EsCTC is amenable to genomic characterization of CTC at the single-cell level via low-pass whole-genome sequencing analysis of copy-number alterations (CNA). This enables assessment of focal gains or losses of chromosomal regions as well as genome-wide estimations of genomic instability such as large-scale transitions (LST; ref. 30). LST is defined as chromosomal breakpoints between adjacent DNA segments >10 MB in size, and its measurement has been used as a functional surrogate to identify tumors with homologous recombination deficiency (HRD), which may be associated with increased clinical efficacy from platinum-based chemotherapy and/or PARP inhibitors (30, 31).

Here we report on the application of the EsCTC workflow for the analysis of single CTC genomics in blood samples collected from patients with mCRPC with clinically defined AVPC as well as non-AVPC “conventional” mCRPC. Broadly, we sought to evaluate if the TSG losses previously linked to AVPC can be detected in CTC, with clinical outcomes consistent with previous reports for tissue assessments, and to identify relationships between established biology of tumor suppressor losses and emergent phenomena of genomic instability resolvable via clinically friendly blood sampling. Under the premise that CTC in this clinical setting provides a faithful representation of the tumor clones driving progression, we next used single-CTC resolution to refine the set of genomic biomarkers associated with aggressive prostate cancer behavior and poor prognosis.

Blood samples were collected immediately before starting treatment from participants on trial NCT01505868, “Study of cabazitaxel with or without carboplatin in patients with metastatic castration-resistant prostate cancer.” This was a prospective, randomized, two-site (MD Anderson Cancer Center and Barbara Ann Karmanos Cancer Institute), phase I/II study evaluating the efficacy of cabazitaxel in combination with carboplatin relative to cabazitaxel alone in men with mCRPC (13), including patients with prospectively stratified AVPC (features described in ref. 11). The study was approved by the corresponding institutional review boards and was conducted in accordance to ethical principles founded in the Declaration of Helsinki. All patients gave written informed consent. No tumor biopsies were required prior to trial enrollment. Three subcohorts were evaluated on the basis of the availability of samples for analysis (Supplementary Fig. S1): (i) CellSearch: samples with matched EsCTC and CellSearch (clinically measured CTC number) draws (n = 47); (ii) CTC-sequenced: with at least one whole genome-sequenced CTC (n = 47); and (iii) circulating cell-free tumor DNA (ctDNA)-matched (n = 57). We defined high disease volume as >10 focal bone metastases and/or tumor mass >4 cm at any site, and/or extension to at least three organ sites with one lesion at least 2 cm in diameter. Low disease volume included ≤ 4 bone metastases with or without extension to lymph nodes up to 2 cm in diameter. The accelerated rate of progression applied to patients with worsening performance status, pain, or other symptoms related to tumor growth in the 6 weeks prior to the blood specimen collection, and/or development of >2 new metastatic lesions in a single site or new nonnodal organ site extension occurring in the previous 3 months.

Blood (7.5 mL) was collected in Streck tubes at the MD Anderson Cancer Center from each subject before the start of treatment and sent to the Kuhn-Hicks Laboratory at the University of Southern California, for slide creation and banking within 24 hours, as described previously (27, 32). Sample processing for CTC detection and characterization was conducted at Epic Sciences. Two slides per patient were processed, corresponding to a median of 0.9 mL blood (Table 1). CTC was identified as cells containing an intact nucleus, without CD45 expression or leukocyte nuclear morphology, and with cytokeratin or AR N-terminus expression, under Clinical Laboratory Improvement Amendments (CLIA)-like conditions (Fig. 1A and B). CTC clusters were defined as >1 CTC sharing a boundary. AR expression was calculated from AR-fluorescence intensity compared with neighboring white blood cells. California-licensed clinical laboratory scientists conducted final QC of CTC classification.

We followed previously described methods for CTC relocation, isolation, and genomic sequencing (27, 30). If <12 CTC were present, all were picked and processed for sequencing. If ≥12 were present, CTC was picked at random within a sample based on sequencing plate availability. Genome-wide CNA analysis was performed using the Epic Sciences single-cell CNA analysis pipeline. FASTQ files were aligned to hg38 human reference genome from the University of California Santa Cruz Genome database. BAM files were filtered for MAPQ 30 reads followed by discrete analyses for genome-wide profiling, individual gene copy-number changes, and LST.

For genome-wide profiling, the hg38 human genome was divided into ∼3,000 1Mbp bins and counted across bins for each cell. Read counts were normalized against white blood cell controls, and the circular binary segmentation (CBS) algorithm “DNAcopy” was used to segment DNA copy-number data (log2 normalized ratio, sample/reference) and identify abnormal copy number. For individual gene copy-number changes, reads were counted for each gene and each sample and normalized against the total sequencing reads for the particular sample. Normalized reads were compared with reference leukocytes, and Z scores were calculated for each gene. The significant cutoff for calling gene gain or loss was Z-score > 3 or Z-score < −3, respectively.

LST was determined as contiguous chromosomal breakpoints that were 10Mb in size (30). Each CNA profile was given an LST score based on the number of chromosomal breaks between adjacent regions of at least 10Mb across the entire genomic landscape (Fig. 1C). A CNA neutral profile carries an LST score close to zero, and CTC with a low LST score carries few chromosomal breaks. Gene-based copy-number analysis was performed as described previously, focusing on 578 cancer genes from the Roche Comprehensive Cancer Design panel (Roche Sequencing; ref. 30).

DNA was extracted from plasma (median 3 mL, range 1–4.5 mL) obtained before starting chemotherapy. The peripheral blood mononuclear cell fraction was used to extract genomic DNA as a matched normal sample. The plasma DNA was then size-selected (<1,000 bp) to enrich for tumor DNA. QC for ctDNA at 160 bp was then performed, quantified on TapeStation (Agilent). Samples with ctDNA >2 ng were used for low-input barcoded next-generation sequencing library construction using highly efficient DNA ligases (KAPA HyperPlus). The barcoded libraries were enriched by PCR and used for both whole-genome sparse sequencing (0.1×) for copy-number profiling at high (220 kb) resolution.

Multiplexed reads sequenced using Illumina's HiSeq4000 (76 bp paired-end) were demultiplexed using Illumina's bcltofastq software and stored in individual FASTQ files while allowing one barcode mismatch. The demultiplexed sequencing data were processed using the “variable binning” pipeline. The individual FASTQ files were aligned to human genome assembly NCBI Build 37 (hg19/NCBI37) using Bowtie2 (2.2.6) alignment software. The aligned reads stored as SAM files were converted to BAM files and sorted using SAMtools (1.2). PCR duplicates were marked and removed using SAMtools. The reads were counted using variable bin sizes at an average genomic resolution of 220 kb. Unique normalized read counts were segmented using the CBS method from R Bioconductor “DNAcopy” package followed by MergeLevels to join adjacent segments with nonsignificant differences in segmented ratios. The parameters used for CBS segmentation were alpha = 0.0001 and undo.prune = 0.05. Default parameters were used for MergeLevels, which removed erroneous chromosome breakpoints.

Statistical analysis was done in R v3.4.1. Descriptive statistics were used to evaluate demographic and clinical characteristics at the time of blood draw. Wilcoxon rank-sum tests were utilized to compare differences between patient groups with continuous measures. The coefficient of determination was utilized to compare the fit of linear regressions. The probability of survival over time was assessed using Kaplan–Meier (KM) estimation. Differences in time-to-event outcomes between groups were measured with the log-rank method, hazard ratios (HR) were obtained from a univariable Cox model, and the P-value from the log-rank test within the coxph function, all in the R package “survival.” For comparisons of a continuous variable with a dichotomized time-to-event outcome, a time interval cut point was chosen at a point where censored patients were only in the group after the cut point, and all patients before the cut point had an event (i.e., had progressed or deceased). Time-dependent ROC analyses were conducted with R package “survivalROC.” To normalize for potential differences in scale, features were normalized to a scale of 0 to 1 prior to adding together for ROC analyses. Dimensionality reduction was conducted with the R package “umap” without parameter alteration (33). Power estimations for predictive biomarker associations made use of the method by Peterson and George 1993 for estimating the power of a biomarker-treatment interaction with a time-to-failure outcome (34). With the exception of KM curves (R package “survminer”), data visualizations were created using R package “ggplot2.” For all analyses, P < 0.05 was considered “significant,” without correction for multiple comparisons unless noted. All statistical tests were two-sided unless noted. All statistical analyses were exploratory and not prospectively declared under NCT01505868.

The datasets generated and analyzed during this study are available from the corresponding author on reasonable request.

Peripheral blood samples were collected from 62 patients with mCRPC prospectively stratified as AVPC (29 patients; ref. 11) or conventional mCRPC (33 patients) before starting chemotherapy on trial NCT01505868 (Supplementary Fig. S1). Clinical characteristics are described in Table 1. Using EsCTC, CTC were detectable in 49 patients (79.0%). Of those, enumeration, AR expression, and genomic sequencing data were available from 47 patients. No association was observed between CTC number (as single cells or CTC clusters) and site of metastasis, performance status, tumor load, or clinically defined AVPC characteristics (Supplementary Figs. S2A and S2B). The patients with shorter PFS and OS did however have higher CTC counts (Table 2). CellSearch comparison showing a positive correlation between the two CTC enumeration methods and higher detection sensitivity for EsCTC is available in Table 1 and Supplementary Figs. S3A and S3B.

Genome-wide CNA profiles were used to calculate copy number for individual gene regions. A candidate molecular signature for AVPC defined as loss of at least two of the three (2+)TSG PTEN, RB1, and TP53 in single CTC was evaluated first. A total of 257 CTC were individually sequenced across the 47 patients (1–22 CTC/patient; ref. Fig. 2A). A patient qualified as TSG-signature positive if at least one CTC had a concurrent loss of 2+TSG. PTEN, RB1, or TP53 loss individually was observed in at least one CTC in 21 of 47 patients (44.7%), concomitant 2+TSG loss in 20 of 47 patients (42.6%), and all-3TSG loss in 10 of 47 patients (21.3%; Fig. 2A; Supplementary Table S1 shows single-cell data per patient). Loss of 2+TSG on the same CTC was associated with shorter median progression-free survival (PFS) and overall survival (OS; Fig. 2B and C; Supplementary Figs. S4A and S4B shows all-3TSG loss). Patients with at least 3 CTC sequenced demonstrated similar results (Supplementary Figs. S4C and S4D).

We sought to compare TSG loss detection in DNA from CTC versus ctDNA. In 57 patients with matched samples available for CTC and ctDNA analysis, 24 (42%) had sufficient ctDNA for CNA assessment, and 44 (77%) had CTC detected and sequenced, passing quality control (Fig. 3A). CTC/mL was higher in patients with sufficient versus insufficient ctDNA (median 8.2/mL vs. 1.7/mL; Supplementary Table S2). The 2+TSG signature was positive in 12 of 57 patients (21%) by ctDNA and in 20 of 57 patients (35%) by CTC. Concordance between methods was most common in patients with relatively high CTC burden (Fig. 3B). The median blood volume assessed for CTC in this analysis was 0.85 mL, and median plasma assessed for ctDNA was 3 mL.

High-resolution CTC imaging before genomic analysis allowed for comparison of AR protein expression intensity to CNA profiles. We found that the relationship between AR gene Z-score and AR protein expression could be patient specific (Fig. 4A), suggesting that factors other than the AR gene could contribute to AR protein expression. Across all patient CTC in the cohort, there was a relationship between AR gene gain status and AR protein expression (Fig. 4B). However, the degree of AR protein expression on individual CTC became more resolved when gains in additional genes, such as MYC and NCOA2 (Fig. 4C), as well as gains in the AR gene enhancer region (Fig. 4D), were factored in, suggesting an additive, multifactorial effect of other genes on AR protein expression in individual CTC. We found AR protein expression to be not meaningfully different in the CTC of patients with clinically defined AVPC relative to conventional mCRPC or in CTC with concurrent loss of RB1 and TP53 relative to all other CTC (Supplementary Figs. S5A and S5B).

For every patient, the median LST across all CTC was determined and used as a genomic instability score (Fig. 2A; Supplementary Table S1). Using the CNA profiles of white blood cells as a reference, we estimated that approximately 20% of the CTC sequenced in this cohort had flat genomic traces. CTC/mL and genomic instability score per patient did not show a discernable correlation (Supplementary Figs. S6A and S6B). However, loss of any TSG and, to a lesser degree, median number of AR overexpressing CTC, were associated with higher genomic instability in single CTC (Table 2). Higher genomic instability across patient CTC was also observed in those with clinical AVPC, accelerated progression pretherapy, PFS <4 months, and OS <12 months. Conversely, those with merely higher tumor load did not see an association of similar magnitude (Table 2). Time-dependent ROC analyses for 12-month survival showed a trend for greater AUC when CTC/mL, genomic instability score, and TSG presence in CTC were assessed in combination compared with each individually (Supplementary Fig. S7).

For each gene region altered, incidence among all 257 CTC sequenced and genomic instability score for the CTC carrying the alteration were calculated (Supplementary Table S3). Two regions on chromosomes 8q (gain) and 13q (loss) had ≥3 gene alterations associated with high genomic instability. Particularly in 13q, we sought to evaluate the single-CTC dependency of losses of two closely located genes with independent, previously established causal relationships with genomic instability: BRCA2 (13q13.1) and RB1 (13q14.1; refs. 35, 36; Supplementary Table S4). As expected, BRCA2 loss was strongly associated with higher genomic instability as per LST on a single-cell level (median LST in CTC with vs. without BRCA2 loss: 34 vs. 13.5, respectively; P < 0.0001, Wilcoxon rank sum test). We also observed that relatively more patients with clinically defined AVPC had at least one CTC with BRCA2 and RB1 loss (9/21 patients, 43%) or RB1 loss without BRCA2 loss (10/21 patients, 48%), as compared with those with conventional mCRPC [5/26 (19%) and 7/26 (27%) patients, respectively]. The most common alteration observed overall (44%) was a gain corresponding to the 100kb region spanning PTK2, which encodes focal adhesion kinase (FAK).

To more broadly assess the relationship between common CNA, genomic instability, prior therapy, and survival, we performed dimensionality reduction on the Z-scores of the >20% incidence genes in all sequenced CTC, projecting these gene–gene relationships into two-dimensional space. These projections of gene–gene relationships did not identify trends between CTC from patients previously treated with abiraterone/enzalutamide versus not (Supplementary Fig. S8A). However, the CTC from patients with >2 years survival clustered in a dense region associated with lower genomic instability and lower incidence of TSG loss, broadly suggesting relationships between these genomic alterations and prognosis (Supplementary Figs. S8B and S8D).

In this study, we extended the classification of patients prospectively identified as either clinical AVPC or conventional mCRPC to liquid biopsies through the application of the EsCTC technology for single-CTC genomic analysis. We identified a compound TSG loss-based molecular signature previously related to AVPC (15, 18) and linked it to other specific CNA observed in association with increased genomic instability in single CTC relative to patients' outcomes, thus effectively expanding the set of chromosomal alterations associated with aggressive prostate cancer behavior. Our findings support the feasibility of single CTC genomics to contribute information of biological and clinical relevance beyond CTC enumeration, at a resolution that bulk tissue or ctDNA analyses may not attain.

In line with detection rates reported in tissue-based studies (12, 14, 15, 18, 19), we identified the candidate TSG-based molecular AVPC signature, defined as concomitant loss of at least two of the three TSG in at least 1 CTC, in 32.2% of the 62 initial patients. We next compared TSG status in the CTC versus ctDNA in matched blood samples and found a general degree of concordance that lends credence to both techniques for gene CNA assessment in liquid biopsy. Still, only 42% of the patients in the ctDNA cohort had sufficient ctDNA for CNA analyses, whereas 77% had CTC detected and sequenced, demonstrating increased detection sensitivity through CTC (at least with the methods we used). An important caveat is that only a median of 3 mL plasma and 0.85 mL blood were respectively available and used for ctDNA and CTC analyses, likely underestimating CTC/ctDNA detection rates.

As our group has reported that the presence of the candidate AVPC signature in tumor tissue and/or ctDNA may predict improvements in both PFS and OS with the addition of carboplatin to cabazitaxel relative to cabazitaxel alone (13), we tried to evaluate whether the same relationship could be discerned through CTC. Unfortunately, our cohort represented a subset of the entire cohort and did not have the requisite sample size power to detect predictive biomarker effects less than an exceptionally high treatment-specific interaction hazard ratio (Supplementary Table S5). Regardless, we found the presence of 2+TSG on a single CTC associated with directionally shorter survival outcomes and incremental loss of TSG to be directly proportional to a poorer prognosis, consistent with established biology of more aggressive oncogenic behavior and empirical correlations to poor outcome seen by other groups (15, 18).

Our ability to simultaneously evaluate AR protein and gene resulted in the finding of patients with mixed patterns of AR expression and amplification status in CTC. A probable explanation for some of the AR-expressing mCRPC cases with no or low-level AR amplification may have been the presence of amplification in the enhancer region upstream of AR (5, 6). Our single-CTC genomic analysis also uncovered increased AR expression in relation to gains in the chromosomal regions containing Myc and NCOA2, a nuclear coactivator of AR (2, 37). The concordance between AR protein and gene status across CTC may have clinical significance and should be further examined (16).

We sought to expand the landscape of genomic alterations related to aggressive mCRPC behavior that are resolvable through liquid biopsy, with an eye toward disease evolution monitoring. Our analyses led to a broad surrogate of genomic instability: LST, a genome-wide measure of DNA scarring that has been linked to HRD, with HRD linked to the sensitivity of different cancer types to platinum drugs and PARP inhibitors. Clinically, we found higher genomic instability scores as independent of CTC number, more clearly related to measures of aggressiveness than to tumor bulk, and additive to prognostic models. Molecularly, co-occurring 2+TSG losses were all positively correlated with increasing genomic instability but had less prognostic ability than the genomic instability score. The chromosomal regions most frequently altered in CTC with higher genomic instability included those affecting genes linked to HRD [loss in 13q13.1 (BRCA2)], increasing AR activity [gain in 8q12–13 (NCOA2) and loss in 16q23-24 (CBFA2T3)], lineage plasticity [loss in 13q14.2 (RB1)], and cell proliferation and metastasis [gain in 8q24 (Myc, PTK2, EXT1)]. Although genomic instability via LST appears to be a global measure of prostate cancer aggressiveness in CTC, PTK2 gain, MYC gain, and TP53 loss were the specific gene alterations most strongly associated with poor prognosis in this cohort (Supplementary Table S6).

To more broadly explore the interacting relationships between gene alterations, TSG losses, and emergent phenomena of genomic instability in the context of patient outcomes, we applied UMAP dimensionality reduction to the 32 genes most commonly altered in the cohort. This analysis revealed gene-gene associations that suggest a continuum of genomic instability related to increasing TSG losses and worse survival.

A shortcoming of our study is the limited sampling of CTC (median <1 mL). As such, some patients had very few CTC for low-pass whole-genome sequencing, precluding deeper interrogations. It may well be that the number of CTC detected and the number of CTC sequenced could influence the results, such that patients with lower CTC counts and fewer CTC sequenced might have lower chances to have a CTC with compound TSG losses (but it may not affect the analytically more robust determination of LST/genomic instability, which empirically within this cohort did not seem to be associated to the number of CTC detected or sequenced). Although care was taken to accommodate such limitations (we only included patients with CTC detected into our outcome analyses, and provided extra analyses with the subset of patients with at least 3 CTC sequenced in Supplementary Fig. S4C and S4D, showing similar results), studies with larger numbers of patients and sample volumes available will enable greater resolution of genomic heterogeneity and more refined subclonal assessments of biological and disease classification significance. However, even with our limitations in patient number and CTC sequenced, this study represents the most comprehensive single-CTC sequencing dataset completed to date in metastatic prostate cancer.

Broadly, our single-CTC genomic analysis suggests an increasing continuum of gene alterations or altered chromosomal regions associated with genomic instability towards aggressiveness in mCRPC, and representative of still not well characterized, discrete paths to progression that are likely linked to distinct molecularly defined AVPC categories. TSG seem to be critical components of those paths and based on this study, genomic instability is an additional lens through which future interrogations should be viewed. Serial evaluation of liquid biopsies from larger numbers of similarly-treated patients will be needed to untangle clinical and molecular heterogeneity and the biological paths to progression.

R.P Graf is an employee of Epic Sciences. A. Rodriguez holds ownership interest (including patents) in Epic Sciences. R. Sutton is an employee of Epic Sciences. R. Jiles holds ownership interest (including patents) in Epic Sciences. A. Kolatkar holds ownership interest (including patents) in Epic Sciences. R. Dittamore is an employee of Epic Sciences. J. Hicks holds ownership interest (including patents) in and is an unpaid consultant/advisory board member for Epic Sciences. P. Kuhn holds ownership interest (including patents) in and is an unpaid consultant/advisory board member for Epic Sciences. A.J. Zurita reports receiving commercial research grants from Infinity Pharmaceuticals; reports receiving speakers bureau honoraria from Pfizer Oncology; and reports receiving other remuneration from Bayer. No potential conflicts of interest were disclosed by the other authors.

Conception and design: R.P. Graf, C. Logothetis, A.M. Aparicio, R. Dittamore, P. Kuhn, A.J. Zurita

Development of methodology: P.D. Malihi, R.P. Graf, A. Rodriguez, R. Sutton, C. Ruiz Velasco, C. Logothetis, J. Hicks, A.J. Zurita

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): P.D. Malihi, A. Rodriguez, N. Ramesh, R. Sutton, R. Jiles, E. Sei, C. Logothetis, P. Corn, A.M. Aparicio, A.J. Zurita

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): R.P. Graf, A. Rodriguez, N. Ramesh, J. Lee, R. Sutton, A. Kolatkar, C. Logothetis, N.E. Navin, P. Corn, A.M. Aparicio, R. Dittamore, J. Hicks, A.J. Zurita

Writing, review, and/or revision of the manuscript: P.D. Malihi, R.P. Graf, A. Rodriguez, N. Ramesh, R. Sutton, C. Logothetis, P. Corn, A.M. Aparicio, R. Dittamore, J. Hicks, P. Kuhn, A.J. Zurita

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): N. Ramesh, R. Sutton, A. Kolatkar, N.E. Navin, R. Dittamore, A.J. Zurita

Study supervision: P. Kuhn, A.J. Zurita

Other (sample processing and data acquisition): E. Sei

We would like to thank the patients participating on trial NCT01505868 and their families for contributing to this clinical research. We would additionally like to thank the laboratory staffs at the Eckstein Tissue Acquisition Laboratory at MD Anderson and at Epic Sciences. This work is funded in whole or in part by the Prostate Cancer Foundation Award 17CHAL01 (A.M. Aparicio, P. Kuhn, J. Hicks), the Solon Scott III Prostate Cancer Research Fund (P. Corn), David and Janet Polak Foundation Fellowship in Convergent Science (P.D. Malihi), and the CPRIT Research Training Program RP170067 (N. Ramesh).

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.