Transcriptome Profiling of Circulating Tumor Cells to Predict Clinical Outcomes in Metastatic Castration-Resistant Prostate Cancer

The clinical utility of circulating tumor cells (CTC) as a non-invasive multipurpose biomarker is broadly recognized. The earliest methods for enriching CTCs from whole blood rely on antibody-based positive selection. The prognostic utility of CTC enumeration using positive selection with the FDA-approved CellSearchTM system has been demonstrated in numerous studies. The capture of cells with specific protein phenotypes does not fully represent cancer heterogeneity and therefore does not realize the prognostic potential of CTC liquid biopsies. To avoid this selection bias, CTC enrichment based on size and deformability may provide better fidelity, i.e., facilitate the characterization of CTCs with any phenotype. In this study, the recently FDA-approved Parsortix® technology was used to enrich CTCs from prostate cancer (PCa) patients for transcriptome analysis using HyCEADTM technology. A tailored PCa gene panel allowed us to stratify metastatic castration-resistant prostate cancer (mCRPC) patients with clinical outcomes. In addition, our findings suggest that targeted CTC transcriptome profiling may be predictive of therapy response.


Introduction
The discovery of oncogenes in the 1970s led to the development of targeted therapies and initiated new hope for curative cancer treatments [1]. Solid tumor genotyping has revealed novel druggable oncogenic alterations that contribute to therapy resistance in patients with mCRPC. This has led to the development of novel treatment modalities such as poly ADP-ribose polymerase inhibitors (PARPi) for tumors with multiple lossof-function alterations in DNA-repair genes [2]. Unfortunately, response rates have been highly variable and unpredictable [3].
Continued reliance on solid tissue biopsies to identify druggable alterations may be limiting progress. Conventional biopsies do not capture clonal heterogeneity and are limited by their invasive nature and the inaccessibility of certain lesions [4]. Inadequate insight into dynamic spatiotemporal alterations in PCa tumors has restricted the overall clinical benefit of novel therapies. Furthermore, the detection limit of traditional diagnostic imaging techniques is currently about one billion cells; below this threshold, malignant lesions are undetectable [5]. Together, these lesions can present a significant tumor burden, yield a high degree of heterogeneity, and carry underlying therapy resistance-associated genetic aberrations [6].
Though CellSearch was introduced in 2005, CTC enumeration has not been implemented as a prognostic, or an early response biomarker in daily clinical PCa management [7]. Recent developments in molecular profiling and cell sorting technologies have increased sensitivity and decreased costs. This could make CTC genotyping viable for precision diagnostics in the near future [8,9]. The collection of tumor material via liquid biopsies allows oncologists to deliver precision medicine and stratify therapy response by the detection of biomarkers, such as androgen receptor splice variant 7 (ARv7), in CTCs [10,11]. With longitudinal sampling, CTC transcriptomes present temporally accurate tumor phenotypes with clinically actionable potential [12]. While biomarker discovery is a growing field in clinical research, therapeutic innovations have outpaced diagnostic capabilities in oncology. As a result, patient stratification for personalized therapeutics remains an unmet clinical need.
In the present work, we used label-free CTC enrichment coupled with CTC transcriptome profiling to characterize multiple oncogenic pathways in mCRPC patients starting a new line of therapy. The Parsortix PR1 microfluidic system, which utilizes the same technology as the FDA-approved PC1 system, was used to enrich CTCs from whole blood using physical properties such as a larger size and reduced deformability compared to leukocytes [13]. This approach enables the capture of epithelial and mesenchymal CTCs in both single and aggregate forms (CTC clusters) [13]. The resulting tumor-enriched samples contain CTCs in a background of 200-800 leukocytes per mL of processed blood [13]. CTC-enriched samples were analyzed using a patented hybridization technology known as HyCEAD [14,15]. This multiplex gene expression assay provides the required sensitivity and specificity to enable the detection of a single CTC target in a background of other circulating epithelial cells and leukocytes [15]. HyCEAD has previously been used to predict malignancy in women with pelvic masses [16].
The primary aim of this study was to test the feasibility of label-free CTC enrichment and targeted transcriptome profiling in mCRPC to prognosticate patients. A secondary aim was identifying transcriptome profiles predictive of therapy response in CTC-enriched mCRPC samples.

CTC Transcriptome Gene Panel
The CTC transcriptome panel consists of genes with high expression in PCa and low expression in leukocytes. Genes were selected as described in Section 4.1. In total, 64 genes were included; each gene and its relevance to PCa are shown in Table S1.

Clinical Follow-Up
In total, 40 patients provided baseline CTC samples (i.e., before starting a new line of therapy). Patients received various therapies: androgen receptor inhibitors (ARSI) (abiraterone or enzalutamide; n = 22), chemotherapy (docetaxel or cabazitaxel; n = 9), immunotherapy (Ipilimumab and Nivolumab; n = 2), radioligand therapy (PSMA-Lu177 or Radium-223; n = 2), PARPi (Olaparib; n = 2), and two were actively surveilled until signs of disease progression (i.e., they did not receive systemic therapy). The clinical trajectories, time of inclusion (CTC collection), and lines of therapy are shown in Figure 1. In one case, two CTC samples were collected from a patient receiving PARPi, 'RAD30' at baseline followed by 'RAD41' 14 days later ( Figure S1). 'RAD41' was excluded from the primary analysis as it was not collected at baseline. Baseline alkaline phosphatase and PSA levels, age, de novo metastasis, years since CRPC diagnosis, type of therapy received following CTC collection, and 1-year survival per therapy group are shown in Table 1.
Due to overlap in the DNA sequencing and CTC transcriptome panels, AR and BRCA2 genomic alterations in solid tissue biopsies were associated with expression in CTCs. In eight of nine patients with AR amplifications, the expression of AR and/or associated signaling genes (KLK3, ARv7, and FOXA1) was increased in matched CTC samples, suggesting high concordance between solid tissue and liquid CTC biopsies [17]. In the patient with both AR amplification and mutation, expression of AR, KLK3, ARv7, and/or FOXA1 were not increased in CTCs ( Figure S3A). Surprisingly, in two patients with BRCA2 shallow deletions, BRCA2 and associated DDR genes (BRCA1, FANCA, RRM2, and TOP2A) were expressed in matched CTCs in one case ( Figure S3B). This suggests the loss of BRCA2 may not always be concordant between solid tissue and CTCs biopsies [18].
The relative prognostic value of genes from the complete gene panel was determined via multivariate analysis on PFS. Genes were ranked using HR ( Figure S5). Notably, the epithelial marker, EPCAM, was not a high-ranking prognosticator in this cohort, yet CTCs expressing EPCAM are known to be prognostically relevant in various malignancies [7,19]. Instead, genes involved in AR signaling (KLK3, HOXB13, and GRHL2), and metastasis (MYO6 and TRPM8) were more informative [20,21]. Prognostic candidates were cross-referenced with relevant literature and went through several iterations of clustering/survival analyses to form a tailored panel consisting of the following genes: AR, ARv7, FOLH1 (aka PSMA), KLK2, KLK3, and TMPRSS2 (hereafter referred to as the 'agnostic gene panel') [22].

CTC Profiles Predict Therapy Response
To explore whether targeted CTC transcriptome profiling has predictive value regarding therapy responses, patients were assessed based on their subsequent therapy.
Genes with the potential to predict an ARSI response were identified via a multivariate analysis on PFS using the subset of ARSI patients (n = 22) and the complete gene panel. Genes were ranked using HR ( Figure S8). Amongst the top-ranking genes, most were involved in AR signaling such as GRHL2, KLK3, HOXB13, FOXA1, and AGR2. Several iterations of clustering and survival analyses produced a gene panel (aka the 'ARSI gene panel') with improved prognostic value over the complete gene panel (Figures 3A-C and S9A-C).

Longitudinal CTC Profiling-A Case Discussion
Divergent from the main protocol, one patient was sampled twice: a baseline sample, 'RAD30', and a follow-up sample, 'RAD41', 14 days after starting PARPi therapy. Both samples had comparable PTPRC levels and relatively similar CTC transcriptome profiles, with some notable exceptions. A clear upregulation of SOX2, TRPM8, NAALADL2, ARv7, DLX1, TUSC3, GHR, FOXA1, FOLH1, and HOXC6 along with a downregulation of WNT5A expression can be seen in 'RAD41' when compared to baseline. A smaller, but noticeable, increase in EPCAM expression indicates the epithelial cell fraction in 'RAD41' may be higher than in 'RAD30' (Figure S1). Many of the upregulated genes are associated with AR signaling, suggesting an increase in AR pathway activity in 'RAD41', either via changes in gene expression or CTC load.

Discussion
The present work highlights the prognostic utility of CTC transcriptome profiling in advanced PCa. Based on transcriptome analysis of tumor tissue samples, 64 genes were selected for targeted CTC transcriptome profiling. From this gene panel, AR signaling genes were amongst the strongest prognosticators, yielding two transcriptionally distinct groups with differing clinical trajectories. A recent publication from Sperger et al. corroborates these findings [22]. While the present study focused on advanced PCa, the Sperger study included a significant portion of castration-sensitive prostate cancer (CSPC) patients (28%). As highlighted by the considerable overlap between the Sperger-gene panel and our therapy-agnostic gene panel, AR signaling continues to drive both CSPC and mCRPC progression. In recent years, ARv7 has been a primary focus for biomarker research in PCa [11,23]. Its ability to drive AR signaling in the absence of a ligand would suggest ARv7 to be prognostically useful in mCRPC [23]. However, we found CTC ARv7 expression to be an insignificant independent prognosticator for survival in the therapy-agnostic cohort, likely due to the presence of alternative resistance mechanisms [24].
In patients receiving ARSI therapy, CTC ARv7 expression did have prognostic value in terms of OS but not PFS. Not surprisingly, AR signaling genes remained strong prognosticators for survival in the ARSI cohort. Stratifying therapy response in patients receiving ARSIs using a tailored gene panel targeting AR pathway activity significantly improved the prognostic value of CTC transcriptome profiling (Figures 3A and S9A). Scoring AR pathway activity has been successfully used to stratify response to hormone therapy in hormone-sensitive AR-positive salivary duct carcinoma (SDC) [25]. Interestingly, SRD5A1 expression was a stronger prognosticator than AR pathway activity in hormone-sensitive SDC [25]. SRD5A1 increases intracellular dihydrotestosterone (DHT) indicating canonical (androgen-dependent) AR pathway activity [26]. Unlike other AR-associated genes, SRD5A1 expression was unable to stratify the ARSI response in the present study, underscoring the prevalence of non-canonical AR signaling in these patients ( Figure S8) [27]. AR/ARv7 also lacked prognostic utility as independent biomarkers for PFS in the ARSI cohort. The expression of AR/ARv7 alone is not informative of AR nuclear translocation, which is a prerequisite for genomic signaling [28]. Instead, the ability of AR/ARv7 to prognosticate PFS relied on the co-expression of AR co-factors and their transcriptional target genes. These findings highlight the importance of using a multifaceted approach when profiling AR pathway activity for stratifying ARSI response in mCRPC.
CTC transcriptome profiling in patients receiving chemotherapy yields a surprisingly different set of prognostic genes when compared to those receiving ARSIs (Figures S8 and S12). AKR1C3, DLX1, HOXC6, and MYO6 were the strongest prognosticators for survival in this cohort.
DLX1 and HOXC6 are primary biomarkers in the SelectMDx liquid biopsy test for the diagnosis of high-risk localized PCa [29]. HOXC6 expression is upregulated in localized, advanced, and metastatic PCa, promoting proliferation [30]. As an androgen-independent AR co-factor, HOXC6 can drive noncanonical AR signaling under castrate conditions and theoretically promote DLX1 expression in ERG-negative mCRPC [30][31][32]. This could ex-plain their consistent prognostic utility in both CTC and urinary liquid biopsy studies [33]. Combined, HOXC6 and DLX1 could act synergistically with regards to driving progression in mCRPC patients receiving chemotherapy, since upregulation of DLX1 promotes metastasis in mice with advanced PCa [32].
AKR1C3 is an androgenic enzyme that is highly expressed in mCRPC. Through intracellular biosynthesis of DHT, AKR1C3 can drive canonical AR signaling in the absence of testosterone [34]. Surprisingly, AKR1C3 is also highly expressed in AR-negative mCRPC, pointing to an AR-independent oncogenic function [34]. One such moonlighting function was found in esophageal adenocarcinoma where AKR1C3 regulates AKT phosphorylation, leading to chemotherapy resistance [35].
MYO6 encodes an actin motor protein with a key function in cell motility [36]. MYO6 is upregulated in PCa and promotes the proliferation of CRPC cells [37]. MYO6 may be of importance for metastasis, as the knockdown of MYO6 mRNA in PCa cell lines impairs cellular migration [38]. In this light, CTC MYO6 expression could be reflective of metastatic potential. We found MYO6 to be uniquely prognostic in patients receiving chemotherapy, a cohort with a high metastatic burden as reflected by their alkaline phosphatase levels ( Table S4).
As we were unable to determine CTC counts, it is difficult to elucidate its impact on CTC transcriptome profiles. However, EPCAM, the cell adhesion marker for prognostic CTC-enumeration in the CellSearch platform, was not a strong prognosticator in this study ( Figures S5, S8 and S12) [7]. Thus, CTC transcriptome profiling likely carries added benefit above enumeration.
We found AR pathway activity to be a prognosticator across treatment modalities, underscoring its broad scope of oncogenic functions in mCRPC. Recent studies have shown synthetic lethality between AR and PARP inhibition, implying a functional role of AR in DDR [39,40]. PARP inhibition reduces genomic stability and inhibits AR nuclear translocation in PCa [39][40][41]. However, we found the expression of several AR signaling genes upregulated in CTCs of a patient receiving PARPi when compared to baseline ( Figure S1). This patient harbored a BRCA2 mutation, which has been associated with increased Src signaling [42]. Phosphorylation of AR by Src kinase promotes AR nuclear translocation constituting a potential mechanism for PARPi resistance [42]. Monitoring CTC AR pathway activity could serve as an early response biomarker for patients with BRCA2 mutations receiving PARPi. Further research is needed to substantiate this claim.
So far, this discussion has focused on the clinical utility of CTC transcriptome profiling in mCRPC. However, CTC survival, invasiveness, and ability to resist vascular stressors are unlikely to be mCRPC-specific [43,44]. CTC clusters are associated with worse clinical outcomes in several cancers due to their resilience to vascular stressors and high metastatic potential [44]. The formation of CTC clusters is highly dependent on the expression of the basal cell marker, KRT14, in ovarian and breast cancer [44,45]. Basal PCa cells are known for their aggressive phenotype [46]. The cell membrane channel, TRPM8, promotes tumor cell invasiveness in several cancers [47]. MYO6 upregulation drives metastasis and progression in breast, gastric, and prostate cancer [36,37,48]. The prognostic value of KRT14, TRPM8, and MYO6 in the present study suggests they may be indicative of CTC metastatic potential ( Figures S5 and S12).
CTC transcriptomes offer temporal indications for therapy resistance and metastatic potential. Label-free CTC enrichment coupled with tailored transcriptome profiling is a promising avenue for the stratification of therapy response in mCRPC. Our approach did not permit the quantification of CTC counts. Hence, the inability to establish a correlation between CTC burden and transcriptome profiles is a limitation in this study.

Gene Panel Development
An in-house microarray transcriptome dataset containing PCa, and leukocyte samples was used to select a panel of 64 genes [33] (Table S1). The panel consists of genes that are highly expressed in localized PCa and mCRPC, with limited or no expression in peripheral blood cells, and/or have known roles in PCa biology. A leukocyte-specific gene, PTPRC, was included to assess the leukocyte background signal. PTPRC expression was excluded from survival analyses. The 64-gene panel will be referred to as the 'complete gene panel' hereafter. Mock CTC samples containing representative fractions of PCa cells and leukocytes (i.e., 10-1000 LNCaP cells spiked in 8 mL whole blood of young and healthy volunteers) were used to test the feasibility of our approach (ATCC, Manassas, VA, USA). Tumor-specificity of the complete gene panel was tested using healthy volunteer blood from age-matched males, young adult males, and age-agnostic females.

Patient Recruitment
Patients were recruited from a single-center observational study at Radboudumc, Nijmegen, the Netherlands (PROMPT, NCT04746300). Eligible patients were men with confirmed mCRPC. Patients were either therapy naïve or had received a maximum of one line of systemic therapy in the castrate setting. Treatment with up to 6 cycles of docetaxel or ARSI in castrate-sensitive disease was allowed. All patients provided written consent for their participation in the PROMPT study and the biobanking of blood and urine samples for pre-defined biomarker studies.

Healthy Donor Recruitment
EDTA blood samples from healthy volunteers were collected at the Sanquin blood bank in Nijmegen, Netherlands. In total, 29 donors participated, including 9 age-matched males (≥50 years), 10 young adult males (<50 years), and 10 females. Healthy donor demographics are shown in the supplementary tables (Table S5).

Next-Generation DNA Sequencing
Molecular profiling of tumor material via next-generation DNA sequencing using the TruSightOncology 500 (TSO500; Illumina, San Diego, CA, USA) panel was done for all PROMPT participants (Illumina, San Diego, CA, USA). Formalin-fixed paraffin-embedded prostate or metastatic tissue biopsy material was used for sequencing, preferably obtained in castrate state. A virtual PCa-specific diagnostic panel was applied, limiting the analysis to 44 genes with prognostic and/or druggable potential in PCa (Table S6).

Presentation of Genomic Alterations
Genomic alterations from solid tissue biopsies and matched transcriptomic alterations from liquid CTC biopsies were presented in an oncoprint using the online OncoPrinter tool from cBioPortal [49,50].

Sample Processing
Blood was drawn into 8 mL EDTA vacutainers following inclusion and processed within 48 h. When necessary, whole blood was stored at 4 • C until processing. Blood was passed through a 6.5 µm filtration cassette at 50 mbar using the Parsortix PR1 system, according to the manufacturer's instructions (ANGLE plc, Guildford, UK). After filtration, a pulsating backflush using phosphate-buffered saline (PBS) was passed through the cassette to harvest trapped cells in HyCEAD lysis buffer per the manufacturer's protocol. Lysates were stored at −80 • C. All samples were shipped on dry ice to ANGLE Biosciences, Inc. laboratories in Toronto, ON, Canada for transcriptome profiling.

Transcriptome Profiling
Lysed samples were analyzed using sense-strand HyCEAD probes and Ziplex ® Flow-Thru Chip ® (ANGLE plc, Guildford, UK) [14,15,51]. Raw expression values were floored to 1 and log2 transformed before analysis. Transcriptome profiles were generated using unsupervised hierarchical clustering as described in Section 4.9.

Clinical Follow-Up and Data Collection
Patient inclusion started in September 2020 and ended in January 2021. Baseline clinical parameters and genomic alterations in metastatic tissue or prostate biopsies were recorded at the time of CTC collection. Clinical and genomic data were prospectively collected, and clinical outcomes were updated following inclusion until May 2022 or death.

Statistical Analysis
Clinical trajectories were visualized in a swimmer plot created in Rstudio using the following packages: Tidyverse and swimplot (Rstudio version 4.2.2) [52].
Unsupervised hierarchical clustering of log2 transformed expression values was done using Rstudio and the following packages: Tidyverse, pheatmap, pvclust, RColorBrewer, and Grid [52]. Clustering of CTC transcriptome profiles was done using Euclidean distance and complete clustering. Multiscale bootstrap resampling was done to determine the statistical significance for each cluster using 10 4 bootstrap replications. The prognostic value of CTC gene expression, serum PSA levels, and age were determined via survival analyses using the median as a cut-off value.

Conclusions
Label-free CTC enrichment provides a noninvasive and non-biased opportunity for transcriptome profiling of tumor material. CTC transcriptome profiles can stratify mCRPC patient survival and therapy response. Conventional methods for detecting therapy resistance and disease progression in mCRPC are limited. Tailored transcriptome profiling of liquid CTC biopsies is a promising avenue for the development of novel prognostic and predictive clinical tests.