Multigene Profiling of CTCs in mCRPC Identifies a Clinically Relevant Prognostic Signature

Prostate cancer is the most common cancer diagnosis and the second leading cause of cancer-related deaths among men in the United States (1). While most patients harbor indolent versions of the disease, a large subset experience metastatic progression, recurrence, and resistance to therapy. Although the androgen signaling pathway is the most common driver of prostate cancer disease progression, a number of other genes and pathways have been recognized as potential prognostic markers and therapeutic targets, and after an initial response to androgen deprivation, metastatic prostate cancer will often progress to castrate-resistant disease (mCRPC; refs. 2–8).

Multiple adaptive mechanisms have been delineated that combat pharmacologic androgen starvation in men with mCRPC, including androgen receptor (AR) amplification, splicing, and mutation, or initiation of parallel androgen synthesis and proliferative pathways (9–11). Wnt signaling is one such pathway that has been implicated as a downstream mediator of AR signaling in mCRPC, and it has been suggested that AR-regulated noncanonical Wnt signals promote prostatic tumor growth, the development of osteoblastic metastasis, prostate cancer stem cell self-renewal and proliferation, and resistance to antiandrogen therapy (12–16).

Through extensive tissue-based sequencing, it is now clear that prostate cancers harbor diverse, clinically actionable drivers of disease progression that may inform both prognosis and potential response to therapy (4, 17–19). While these findings highlight the promise of precision oncology testing in prostate cancer, the utility of these techniques may be limited by both the challenges of extracting viable tissue from bone metastases and the need for serial sampling as new resistance mechanisms arise (18, 20).

Emerging comprehension in the realm of circulating tumor cell (CTC) biology has unveiled the potential for noninvasive, continuous monitoring of cancer, and previous studies have established a correlation between CTC, primary tumor, and metastatic tissue gene expression profiles (21, 22). Previous efforts to inform disease severity in prostate cancer have relied simply on the presence or absence of CTCs, but these studies have not yielded adequate clinical utility to guide therapeutic decision-making for individual patients (23–26). Additionally, genomic profiling of cell-free DNA (cfDNA) represents another feasible and clinically informative liquid biopsy avenue in mCRPC (27, 28). While newly developed approaches show promise in CTC detection, including CTC-iChip and the AdnaTest, the overall landscape of liquid approaches to assess changes in tumor behavior in men with mCRPC represents a major unmet clinical need as robust, readily reproducible methods for isolation and subsequent characterization of CTCs in prostate cancer remain elusive (9, 29, 30).

To facilitate noninvasive detection of genomic alterations and molecular expression in these patients, we developed and implemented a qPCR-based platform for analysis of CTCs from men with mCRPC. This approach was modeled after commercially available, tissue-based qPCR platforms for prostate cancer profiling, with the primary goal of eventual integration into clinical workflows (31, 32). Given that CTCs likely represent cells from disparate metastatic sites, we hypothesized that understanding the expression of key genes of interest within CTCs could have unique utility in precision medicine approaches to predict clinical outcomes in patients with advanced prostate cancer (33).

CTCs were isolated from 5-mL whole blood using anti-EpCAM antibody–conjugated microbeads (Thermo Fisher Scientific). DAPI, CD45, FITC, and PCa-CT-PE (antibody cocktail: PSMA, EGFR, and pan-cytokeratin) were used to confirm the identity of CTCs. Following washing and cell lysis, mRNA from CTCs was captured using Oligo(dT)25 mRNA Dynabeads (Thermo Fisher Scientific). Reverse transcription was performed to obtain cDNA, followed by preamplification to generate target gene amplified libraries, and multiplex qPCR was then performed to evaluate a panel of 96 genes, including controls (Supplementary Table S1). Genes were selected based on relevance to prostate cancer progression and metastasis, as well as EMT and cancer stem cell biology. To account for leukocyte contamination, blood processed from 27 healthy controls was used as a baseline referent. To evaluate gene amplification efficiency for both preamplification and real-time PCR, prostate cancer cells (1.65 × 10⁴ cells, including equal amount of PC3, VCaP, and LNCaP) were lysed with 1 mL lysis buffer. Ten-fold serial dilution was applied with lysis buffer to obtain samples equal to 3,300, 330, 33, 3.3, and 0.33 cells per 200 uL to evaluate the linear relationship between cell number and cycle threshold (Ct) value of individual genes (Supplementary Fig. S1). To further validate the RT and amplification steps, we included Array control RNA Spikes (Thermo Fisher AM1780), using primers from the Fluidigm RNA Spikes Assay kit (100-5582) to validate the reverse transcription, preamplification, and qPCR steps. The kit contains RNA from E. coli with no known homology to mammalian sequences. Spike mix was prepared as described in the Fluidigm delta gene assay protocol, PN 100-4904 J1. Three of the RNA spikes were used, Spike-1 (750 nucleotides), Spike-4 (1000 nucleotides), and Spike-7 (1474 nucleotides) and mixed. Ct value of each RNA spike was acquired with qPCR. The average value plus standard deviation of Ct value from different spikes was calculated and plotted against the copy number of different spikes, which showed a linear correlation between Ct value and copy number (R² = 0.95; Supplementary Fig. S1).

For platform validation, 180 GFP-labeled prostate cancer cells (PC3, LNCaP, and VCaP) were sorted into 5-mL whole blood from a normal donor. Cells were captured using immunomagnetic bead enrichment and visualized by immunofluorescence. Cells were then counted, rather than lysed, to measure the recovery rate. Additionally, 300 PC3 and LNCaP cells were spiked into 5 mL of whole blood from healthy controls and captured by immunomagnetic bead enrichment, followed by staining with DAPI, CD45, and PCa-CT (antibody cocktail: PSMA, EGFR, and pan-cytokeratin) to visualize CTCs. To validate gene expression after CTC isolation, 10 prostate cancer cells (PC3, LNCaP, or VCaP) were spiked into 5 mL of whole blood from a normal control. CTCs were captured by immunomagnetic bead enrichment, lysed, and gene expression was then assessed by qPCR. qPCR gene expression was compared with known microarray profiles for each of these lines in order to evaluate whether the expression profiles matched that of what was expected for PC3, LNCaP, and VCaP cells.

For each sample, gene expression of 96 genes was determined from qPCR in a 384-well format as described above. Ct values were normalized using the ΔΔCt method as previously described (34). Briefly, genes within each sample were first normalized against endogenous controls by subtracting the sample-specific average Ct value for ACTB, HMBS, and TUBA-1B. For each gene, the resulting ΔCt values were corrected for background from contaminating blood cells by subtracting the average ΔCt value for that gene across the 27 healthy donor reference controls to obtain ΔΔCt values. Undetermined Ct values were imputed as 40 (Supplementary Table S2). Finally, gene expression was quantified as 2^−ΔΔCt, which we refer to as “normalized expression.” All analyses were performed with log₂(normalized expression + 1), with the logarithm serving as a variance-stabilizing transformation and the addition of the constant serving to compress expression values comparable to background. In addition, we set gene expression values to zero when the underlying Ct value was greater than 35 as Ct values in this range are generally not quantifiable.

Given that this protocol involves cell lysis rather than visualization and enumeration, we established a model for identifying CTCs using the gene expression of eight epithelial markers: CD326, CDH1, CDH2, DSG2, EGFR, KRT8, KRT18, and KRT19. CTC-positive controls were created using serial dilutions of prostate cancer cell lines, PC3 and LNCaP, spiked into whole blood from healthy donors. Data from 27 healthy donor controls were used as negative controls. We first performed a principal components analysis to verify the accuracy of these genes for distinguishing positive from negative controls. The first principal component clearly separates positive from negative controls, with the second component primarily separating cell lines (Supplementary Fig. S2). These two components account for 92% and 4% of variance for these samples, respectively. The weights of each gene on this component are shown as a bar graph in Supplementary Fig. S2. We define an epithelial score (ES) as the weighted sum of the eight epithelial marker genes using these weights. We then built a classification model for identifying new samples with CTCs using this epithelial score in a logistic regression. The fitted model is:

Patient samples were considered CTC positive if they are above the conservative threshold of 75% on the estimated probability using the model above. In both the control and clinical settings, the majority of samples were toward the very high or very low end of the CTC probability model (Supplementary Fig. S2).

The primary endpoint for this study was overall survival, and in order to identify CTC-based genes associated with mortality, we performed univariate Cox regression analysis in CTC-positive patients for each of the measured genes (35). The Halabi nomogram was utilized to account for baseline clinicopathologic variables, including ECOG performance status, disease site, opioid analgesic use, albumin, hemoglobin, alkaline phosphatase, and PSA in a multivariate Cox model. We obtained 18-month survival probabilities from https://www.cancer.duke.edu/Nomogram/firstlinechemotherapy.html. We subtract this probability from one to define a “Halabi” score so that hazard ratios above one indicate increased risk of death. Cox bivariate regression analyses were then performed for each gene, adjusting for Halabi score, to identify genes associated with mortality independent of measureable clinical factors.

Cox regression was performed for 78 genes, after removing endogenous controls (ACTB, HMBS, and TUBA-1B), epithelial markers (CD326, CDH1, CDH2, DSG2, EGFR, KRT8, KRT18, and KRT19), blood-cell markers (CD20 and CD45), artificial RNAs used in assay development (Spike1, Spike4, and Spike7), and alternate primers for SChLAP1 and SPANXB2. Log-transformed normalized expression values for each gene were rescaled by the interquartile range in order to make hazard ratios easier to interpret. Models were fit using the R function coxph in the survival package (36). Genes were ranked using P values from bivariate proportional hazards models with a single gene and the Halabi score as predictors. These P values were transformed to FDR q-values using the Benjamini–Hochberg method (37). The three genes significant at <10% FDR in the CTC-positive cohort—AURKA, BMP7, and WNT5A—were selected for further analysis. Cox regression and Kaplan–Meier analysis was performed to assess whether CTC-positive or -negative status was associated with survival within the cohort. To aid in the interpretation of hazard ratios, we repeated the analysis using CTC probability based on our epithelial score as a continuous variable. This probability “P” was scaled using the transform: 2 × (P − 0.5), so that one unit change can be thought of as going from P = 0.25 to P = 0.75.

We constructed waterfall plots based on expression of WNT5a, AURKA, and BMP7 for the 39 patients (26 CTC positive) with at least 6-month follow-up with bars colored by vital status at 6 months. In these plots, patients in the CTC-positive subgroup were ordered from lowest to highest normalized expression and patients without CTCs ordered below the CTC-positive group. To assess significance in these plots, we used the hypergeometric distribution to test whether deaths were overrepresented in the top 25% of patients.

We next examined composite scores using these three genes. First, in the CTC-positive group we assigned patients points for each gene, with normalized expression values in the 50th to 75th percentiles given one point, and those above the 75th percentile given two points. Then, we defined prognostic scores by adding point values across either all three genes or one of the three two-gene pairs. For these four scores, we constructed receiver operating characteristic (ROC) curves for predicting vital status at 6 months using the survivalROC function with smoothing parameter lambda = 0.002 from the survivalROC package in R (38, 39).

From the ROC analysis, we identified the best performing genes (WNT5a and AURKA) and added those to the Halabi score to generate a combined molecular and clinical signature. The resulting miCTC score ranges continuously from 0 to 5, with integer values of 0, 1, 2, 3 or 4 coming from AURKA and WNT5A gene expression. We use 1 minus the Halabi 18-month survival probability as the “Halabi component” of the score. As before, we use the survivalROC function in R to construct ROC curves for predicting survival at 6 months (180 days) using the miCTC and Halabi scores. Confidence intervals for the difference in AUCs between these two curves were computed using the appropriate quantiles from 10,000 bootstrap samples. Finally, we define a “high-risk” subgroup as those patients having a miCTC score of two or more, with others forming a “low-risk” subgroup. We compare these risk groups using Kaplan–Meier survival curves and assess significance using the log-rank test. Survival curves and log-rank P values are computed using the survfit and survdiff functions from the survival package (40).

Formalin-fixed paraffin-embedded prostatectomy samples were from two published retrospective patient cohorts from the Mayo Clinic (n = 780; refs. 41–43). As previously described, RNA extraction and hybridization to Affymetrix Human Exon 1.0 ST Arrays (Affymetrix) were performed in a CLIA-certified, clinical operations laboratory (GenomeDx Biosciences, Inc.). Microarrays were normalized using Single Channel Array Normalization (44). Microarray data are available with NCBI GEO accession numbers GSE46691 and GSE62116. The miCTC score was applied to the retrospective samples as described above. Kaplan–Meier curves were generated for the pooled Mayo Clinic cohorts, and hazard ratio estimates were generated using Cox proportional hazards modeling.

All cell lines were obtained from the ATCC, and cell lines were maintained using standard conditions. To ensure identity, all cell lines were genotyped at the University of Michigan Sequencing Core using Profiler Plus (Applied Biosystems) and compared with profiles of respective cell lines available from ATCC. Cell authentication and mycoplasma testing was performed for all cell lines. This study was approved by the Institutional Review Board (IRB# HUM00052405), and written informed consent was obtained from all patients for collection and analysis of samples in accordance with recognized Declaration of Helsinki guidelines.

CTC isolation and subsequent gene expression analysis was conducted using an antiepithelial cell adhesion molecule (EpCAM)-coated microbead protocol, followed by multiplex qPCR of 78 prostate cancer–related genes plus internal controls (Supplementary Fig. S3). Our panel contained genes previously shown to be involved in prostate cancer aggressiveness, metastasis, or progression, including TMPRSS2:ERG, PSA, AR, WNT5a, NKX3.1, EGFR, and SChLAP1 (16, 45–49). Recovered cells were identified by cell surface staining for EpCAM and absent staining for CD45, a common leukocyte marker (50). In order to assess the performance of this platform, small numbers of cells (n = 10) were spiked into 5 mL of whole blood. Gene expression analysis of these cells showed expression profiles that were congruent with previous microarray analyses of these cell lines, supporting the validity of our isolation and expression analytic platform (Supplementary Fig. S4).

To characterize the CTC gene expression pattern in patients, we evaluated 41 patients with mCRPC (Table 1). The median age was 68 (IQR 61–75) and median PSA was 19.7 ng/mL (IQR 3.6–70.6 ng/mL). At the time of enrollment, all patients were on luteinizing hormone releasing hormone (LHRH) agonist therapy (leuprolide and goserelin), 5 patients were on additional treatment with abiraterone, 11 on enzalutamide, 8 on docetaxel, 4 on cabazitaxel, 1 on radium 223, and 1 on dexamethasone. Patients had received a median of two prior therapies (IQR 1–3). All patients had a minimum of 90 days follow-up and there were 21 (51.2%) deaths. The median time to next therapy was 251 days (IQR 119–329) and patients received a median of one subsequent therapy (IQR 0–1).

CTC positivity was determined based on a model using the gene expression of eight epithelial markers (Supplementary Fig. S2) and we identified 27 (66%) patient samples with detectable CTCs. In our protocol, CTC-positive samples are identified after expression profiling, and each profile is potentially a mixture of CTC targets and contaminating blood cells, as this is an enrichment rather than pure isolation approach. For the healthy controls, due to the absence of epithelial CTCs, all RNA in the sequence library is from contaminating blood cells. As described in Materials and Methods, gene expression in patient samples is normalized to both endogenous controls and then to expression in healthy controls to account for background contamination. In order to define potential thresholds regarding positive or negative expression of a given gene in CTCs, we used k-means clustering for each gene to find subgroups among 27 healthy controls and 27 CTC-positive samples. When the composition of these two groups was such that one had significantly more CTC-positive samples than healthy controls (as determined by a χ² test), we used the midpoint between the cluster means to define a threshold at which to call specific samples “positive” for a given gene. Using this approach, 19 samples (70%) showed increased expression of AR and 15 (55%) showed significant expression of AR-V7, with 14 (52%) shared between these two groups. Increased downstream AR signaling was seen in 14 (52%) samples based on joint expression of KLK2 (PSA) and NKX3-1, and 18 (67%) samples showed significant upregulation of Wnt signaling based on expression of WNT5a, WNT5b, and BMP7. Among epithelial–mesenchymal transition (EMT) genes, 9 samples (33%) expressed SOX2 and 21 (67%) had elevated levels of TWIST1, with 7 samples (26%) expressing both genes. Similarly, clustering jointly on MDK and CCNA2, 15 samples (56%) showed high levels of cell cycle activity. Among long noncoding RNAs (lncRNA) associated with prostate cancer, 19 samples (70%) showed elevated levels of the lncRNA SChLAP1, but only 8 (30%) showed increased PCA3 levels. The TMPRSS2:ERG gene fusion was significantly expressed in 11 samples (41%) and based on concordant expression across two distinct amplicons, evidence for the TM2:ETV1 fusion was seen in 4 samples (15%; Fig. 1A).

To acquire a clinically relevant gene signature, we sought to identify genes expressed in CTCs independently associated with survival. To do this, we performed univariable and multivariable Cox regression analyses of individual genes and used a previously validated prognostic model to account for baseline clinicopathologic variables in mCRPC (51). Increased expression of WNT5a (HR 3.62, 95% CI, 1.63–8.05, P = 0.002), AURKA (HR 5.56, 95% CI, 1.79–17.20, P = 0.003), and BMP7 (HR 3.86, 95% CI, 1.60–9.32, P = 0.003) were strongly associated with earlier mortality (FDR < 10%) after adjusting for the Halabi nomogram score. Other genes implicated in prostate cancer biology, including WNT5b, KLK2, and PSA, were significantly associated with overall survival on univariate, but not multivariate, analysis (Fig. 1B). CTC positive or negative status alone was not significantly associated with survival in a Kaplan–Meir analysis (P = 0.143) or univariate Cox model (HR 1.9, 95% CI, 0.6–5.7, P = 0.28) comparing the two groups. CTC status was also not significantly associated with survival in a bivariate model controlling for Halabi 18-month survival probability (HR 1.2, 95% CI, 0.3–4.0, P = 0.81). We repeated this analysis using the CTC epithelial score as a continuous variable, and it was marginally significant in the univariate analysis (HR 3.4, 95% CI, 1.0–11.9, P = 0.06) but not the bivariate model adjusted for Halabi 18-month survival probability (HR 2.1 95% CI, 0.5–8.5, P = 0.28; Supplementary Fig. S5).

Given the ability of WNT5a, AURKA, and BMP7 expression to independently predict survival in this cohort and their previously elucidated roles in metastatic, variant forms of prostate cancer and treatment resistance, we hypothesized that increased expression of these genes within CTCs would correlate with worse prognosis (16, 52–54). Focusing on these three genes, we confirmed that early mortality (≤6 months) occurred more frequently in patients with higher expression of WNT5a, AURKA, and/or BMP7 individually (Fig. 2A). We next developed prognostic models in the 27 CTC-positive patients using composite scores for all three genes and evaluated individual combinations of genes to optimize the model. Scores were determined by normalizing the expression values for each gene. One point was given for expression of a given gene in the 50th to 75th percentiles of the CTC cohort and two points for expression above the 75th percentile. While all three genes were individually prognostic, a composite score comprising WNT5a and AURKA expression outperformed all other combinations (AUC = 0.91; Fig. 2B).

These data were then used to derive a CTC-based prognostic score [hereafter referred to the Michigan CTC (miCTC) score]. This score ranges from 0 to 5, with values of 0, 1, 2, 3, or 4 derived from AURKA and WNT5A gene expression (0–2 points for expression of each gene as described above). In order to account for baseline clinical variables, the Halabi nomogram score was incorporated as the fifth component into the overall score using the 18-month survival probabilities calculated from the Halabi nomogram (see Materials and Methods; Supplementary Fig. S6). For CTC-negative patients, this meant the miCTC score was based solely on the Halabi survival probability and the absence of WNT5A or AURKA gene expression. The miCTC score outperformed the Halabi nomogram alone (AUC = 0.89 vs. AUC = 0.70) across the entire 41 patient cohort (Fig. 2C). When categorizing the present cohort into high-risk (miCTC ≥ 2) versus low-risk (miCTC <2) subgroups, Kaplan–Meier analysis demonstrated significant differences in overall survival over time (Fig. 2D).

To further assess the performance of the miCTC prognostic signature, we first evaluated the level of concordance of gene expression between our qPCR-based platform and tissue-based RNA-sequencing data for a subset of patients enrolled in an ongoing mCRPC profiling study (55). In the 7 patients with both CTC-based qPCR and tissue-based RNA-seq data, a significant level of correlation existed among CTC-to-tissue gene expression (Spearman correlation = 0.38, P = 0.0002), suggesting the genomic data obtained from CTCs accurately reflected tissue-based disease (Fig. 3A and B). We then queried the tissue-based expression of WNT5a and AURKA across a larger database containing samples from normal prostate tissue, localized prostate cancer, and metastatic samples to globally understand the landscape of WNT5a and AURKA expression (18, 56). As expected, WNT5a and AURKA expression was found to be significantly upregulated in metastatic samples compared with normal prostate or localized prostate cancer tissue, further supporting the relevancy of these genes in mCRPC (Fig. 3C). Calculation of the miCTC score (using gene expression only) in this cohort revealed a significantly increased proportion of samples with high-risk scores (miCTC ≥2) in the metastatic group (92%) compared with the primary cancer group (52%; P < 0.001; Fig. 3D).

Lastly, we applied the gene expression component of the miCTC score to a previously published radical prostatectomy cohort in order to assess its ability to predict long-term, objective outcomes in patients with high-risk, clinically localized prostate cancer (41–43). This cohort included 780 patients with a median follow-up of 11.2 years after prostatectomy. A total of 513 patients (65.8%) suffered biochemical recurrence following prostatectomy, 288 (36.9%) developed metastatic disease, 166 (21.3%) died of their disease, and 336 (43.1%) died of any cause. WNT5a and AURKA expression were determined from the genomic data and used to calculate the miCTC score (on a scale of 0–4 without the additional factor of the Halabi score). These scores were then matched with clinical outcomes. As shown in the Kaplan–Meir analyses, high-risk and increasing miCTC scores were associated with a greater risk of biochemical recurrence, metastasis, and death, and decreased prostate cancer-free survival when compared with those with lower-risk miCTC scores (Fig. 4A–D).

Given the relative simplicity of obtaining single blood draws, CTCs obtained from whole blood may provide a compelling source of molecular information, and previous studies have shown that CTCs closely reflect the clinical and functional characteristics of solid tumors (57, 58). While the clinical significance of the presence of CTCs has long been theorized, improvements in the downstream manipulation and comprehensive molecular characterization of liquid biopsy tools are more recent. For example, a plasma-based real-time PCR test for the qualitative detection of defined mutations of the epidermal growth factor receptor (EGFR) gene was recently developed to aid in the selection of non–small cell lung carcinoma (NSCLC) patients for treatment with erlotinib (59). The cobas EGFR Mutation Test, along with the CellSearch platform (50, 60, 61), currently represent the only FDA-approved liquid biopsy tests that may help guide first-line treatment decisions in patients, further underscoring the significant impact prospective cfDNA and CTC-based diagnostic and prognostic tools may have.

While AR-V7 expression in CTCs has received a great deal of attention, we sought to develop a reproducible platform allowing CTC-based expression analysis across a broader panel of prostate cancer–related genes. This targeted approach was selected to improve the efficiency of the assay and potential clinical applicability, given the challenges inherent to CTC isolation and subsequent gene expression profiling. Although an unbiased approach would capture increased genetic heterogeneity and support the discovery of novel gene targets, it would be increasingly difficult to differentiate possible prognostic targets from genetic noise, especially using a qPCR-based approach (62). Additionally, the likelihood of type 1 errors due to multiple testing is substantially greater with a broad, unbiased approach. Our primary goal was to utilize this platform to identify clinically relevant genes able to predict key oncologic outcomes in mCRPC.

In this cohort of mCRPC patients on a variety of therapies, AR-V7 expression was not significantly associated with survival. However, a number of findings appeared to support the clinical validity of the approach utilized here. First, the expression data demonstrated a high degree of AR signaling in a majority of samples, with over half of samples showing AR splice variants. Second, we detected a global increase in prometastatic factors within CTCs, including genes involved in EMT, cell cycle regulation, and lncRNA-mediated tumor suppression (49, 52, 53, 63). Third, in patients with matched tissue-based gene expression available, there was a close correlation between the tissue and CTC-based results. Lastly, the tissue-based analysis in the localized disease setting provides evidence that the markers identified in our CTC-based approach also have applicability in earlier stage disease. While a “proof of concept” rather than a validation, the crude miCTC score based solely on expression was applied to clinical outcomes in prostatectomy patients as a preliminary analysis to further corroborate the prognostic relevance of the genes identified here.

Strikingly, our results add further support to the importance of noncanonical Wnt signaling as an alternative resistance pathway contributing to prostate cancer progression and worse clinical outcomes. While Wnt signaling has been previously linked to prostate cancer aggressiveness through various cellular functions (13, 14, 18, 64, 65), our data strongly corroborate another recent study linking Wnt signaling to antiandrogen resistance in CTCs (16). Additionally, Beltran and colleagues established the role of AURKA in prostate cancer proliferation and, in particular, in driving the development of neuroendocrine prostate cancer (NEPC). This has led to an ongoing phase II trial of aurora kinase A inhibition in NEPC (NCT01799278; refs. 4, 53). The miCTC score, driven by WNT5a and AURKA expression, reliably predicts overall survival in this development cohort, suggesting CTC genomics may mirror tissue-based drivers of variant or lethal prostate cancer.

The present study is limited by its relatively small sample size, and further prospective validation in larger cohorts will be required in order to confirm the clinical validity of these findings. Additionally, patients in this study were on a variety of therapies at the time of analysis. Evaluation of more homogenous cohorts is needed in order to ascertain the impact of specific therapeutic modalities on changes in gene expression over time and to develop predictive signatures of therapeutic response. However, the high event rate allowed us to identify several potential prognostic markers after adjusting for baseline clinical characteristics and determining false discovery rates.

While our CTC isolation protocol allows for simple, rapid processing of samples with high sensitivity, the approach itself has inherent limitations. Specifically, this protocol does not allow for CTC enumeration and no cell morphological assessment can be conducted. As a result, gene expression is not adjusted for number of CTCs and may be impacted by CTC quantity within a sample. However, the lack of correlation between epithelial score and survival indicates that CTC quantity itself is not a useful marker with this platform. Additionally, given that the platform is based on enrichment rather than isolation, gene expression analysis requires adjustment for background leukocyte contamination. Finally, as with other epithelial-based CTC detection approaches, the enrichment platform utilized here could miss CTCs that have undergone EMT and no longer express EpCaM on their cell surface (66).

These limitations notwithstanding, we developed a clinically relevant, multigene signature from CTCs using a scalable qPCR-based platform. Though external validation is needed, this molecular signature appears to strongly inform prognosis in patients with advanced prostate cancer. The strong prognostic performance of this signature in a tissue-based analysis of a large prostatectomy cohort further supports the potential clinical applicability of these data. We anticipate that this approach will facilitate serial patient assessments over time and identification of molecular changes as they occur in response to therapy. With continued improvement in the ability to determine key drivers of tumor progression using CTCs, we anticipate that this noninvasive, liquid-based approach will circumvent the need for invasive extraction of genomic material from bone metastases and guide precision medicine therapeutics in prostate cancer.

R.J. Karnes reports receiving a commercial research grant from GenomeDx, has ownership interest (including patents) in GenomeDx, is a consultant/advisory board member for Genomic Health. S.L. Chang is a bioinformatician at PFS Genomics and has ownership interest in patents pending with GenomeDX Biosciences on genomic signatures unrelated to the work in this article. F.Y. Feng is a consultant/advisory board member for Janssen, Astellas/Medivation, Sanofi, Ferring, and Dendreon. S.A. Tomlins is Co-founder of, consultant for, and Laboratory Director for Strata Oncology, reports receiving a commercial research grants from Astellas/Medivation and GenomeDX, is a consultant/advisory board member for Astellas/Medivation, Janssen, Sanofi, and Almac Diagnostics and has received Travel Support from Thermo Fisher Scientific. No potential conflicts of interest were disclosed by the other authors.

Conception and design: U. Singhal, Y. Wang, Y.S. Niknafs, Y. Qiao, R.S. Taichman, S.A. Tomlins, T.M. Morgan

Development of methodology: U. Singhal, Y. Wang, T.M. Morgan

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): U. Singhal, Y. Wang, Y. Qiao, A. Gursky, D.C. Smith, R.J. Karnes, F.Y. Feng, G.S. Palapattu, T.M. Morgan

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): U. Singhal, Y. Wang, J. Henderson, Y.S. Niknafs, Y. Qiao, S.L.Chang, F.Y. Feng, G.S. Palapattu, A.M. Chinnaiyan, T.M. Morgan

Writing, review, and/or revision of the manuscript: U. Singhal, Y. Wang, J. Henderson, Y. Qiao, A. Zaslavsky, R.J. Karnes, S.L.Chang, F.Y. Feng, R.S. Taichman, A.M. Chinnaiyan, S.A. Tomlins, T.M. Morgan

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): U. Singhal, Y. Wang, Y. Qiao, J.S. Chung, T.M. Morgan

Study supervision: U. Singhal, Y. Wang, G.S. Palapattu, T.M. Morgan

This study was supported by a Department of Defense Physician Research Training Award (W81XWH-14-1-0287) and the Prostate Cancer Foundation (T.M. Morgan, R.S. Taichman, and S.A. Tomlins). T.M. Morgan and S.A. Tomlins are also supported by the A. Alfred Taubman Medical Research Institute. This work was also supported by the NIH: National Cancer Institute PO1CA093900-10 and the University of Michigan Comprehensive Cancer Center Prostate SPORE (P50CA 069568).

The authors would like to thank Jun Luo, PhD, and Ken Pienta, MD, for assistance and mentorship in conducting this work. We also thank Robin Kunkel for help with figure preparation.

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.