Between January 2015 and January 2020, 18 mCRPC and 43 mHSPC patients were recruited with informed consented at St Bartholomew’s Hospital, Barts Health NHS, London, UK. Clinical characteristics for individual patients are shown in Supplementary Table 1 . Peripheral blood samples were collected into EDTA tubes ≤2 months before commencing 6 cycles of docetaxel. Patients with mHSPC had started initial hormone-therapy <3 months before blood collection. Patients received CT and bone scans before and after treatment. Serum PSA, ALP, and LDH were measured together with CTC sampling. Radiological response assessment was based on RECIST criteria (25): (1) complete response (CR): disappearance of all target lesions; (2) partial response (PR): at least 30% decrease in the sum of the longest diameter of target lesions, taking as reference the baseline since treatment started; (3) progressive disease (PD): at least 20% increase in the sum of the longest diameter of target lesions, taking as reference the baseline since treatment started; (4) stable disease (SD): neither sufficient shrinkage to qualify for PR nor sufficient increase to qualify for PD. Assessments of response by bone scan were classified as follows: (1) CR: disappearance of all bone metastasis; (2) PR: a decrease in number, extent or intensity of bone lesions was detected; (3) PD: appearance of new bone lesion(s) and/or apparent enlargement of the bone metastases; (4) SD: little or no change in the number, extent or intensity of bone metastases was observed. PSA progression was defined as two consecutive rises above PSA nadir at least two weeks apart. CTC measurements were also investigated for their ability to prognose OS and PFS outcomes.

The docetaxel-resistant human PCa cell line PC3-D12 and the sensitive counterpart PC3-Ag, were gifted by A.J. O’Neill (10), University College Dublin. PC3-D12 cells were treated every 4 weeks with 12 nM docetaxel in order to maintain resistance. The cells were maintained in RPMI-1640 medium supplemented with 10% foetal bovine serum and 2 nM L-glutamine (Invitrogen, Waltham, Massachusetts, Unites States).

CTCs were isolated from 7.5 mL of whole blood using the Parsortix^® (Angle Plc, Guildford, UK) isolation system and identified for CTC enumeration using four-colour immunofluorescence as previously described (32, 37). Briefly, 7 mL of blood was transferred to 50 mL LeucoSep tubes (Greiner Bio-One, Frickenhausen, Germany) with 15.3 mL of Ficoll-Paque Plus (GE Healthcare, Chicago, Illinois, Unites States) and centrifuged at 1000 g for 15 min with the break off at room temperature to recover the peripheral blood mononuclear cell (PBMC) fraction. The PBMC fraction along with the plasma above the fit of the LeucoSep tube was removed into a new 50 mL falcon tube and pelleted at 200 g for 8 min at room temperature. The pellet was then re-suspended in 4.5 mL of isolation buffer (PBS containing 1% BSA and 2 nM EDTA) and added back to the remaining 0.5 mL of whole blood and loaded onto the Parsortix^® for CTC isolation. Once samples are loaded, cells are separated based on cell size and deformability according to a pre-set programme PX-S99F that uses 6.5 µm-gap cassette and 99 mbar pressure for isolation. Cells were then harvested using a pre-set programme and transferred onto glass slides for downstream analysis. All blood samples were processed within 4 hrs of collection. Slides were stained using mouse monoclonal PE-conjugated anti-CD45 (Miltenyi Biotec, Bergisch Gladbach, Germany), mouse monoclonal FITC-conjugated anti-Cytokeratin (Miltenyi Biotec, Bergisch Gladbach, Germany), Alexa Fluor 647-conjugated anti-Vimentin (Abcam, Cambridge, UK), and counterstained using SlowFade gold antifade mountant with DAPI (Life Technologies, Carlsbad, California, United States). CTCs were identified as Cytokeratin (CK)+/Vimentin (VIM)−/CD45- (epithelial-CTCs), CK+/VIM+/CD45- (EMTing-CTCs) and CK-/VIM+/CD45- (mesenchymal-CTCs). Patients with ≥1 epithelial-CTC and/or ≥1 EMTing-CTC and/or ≥4 mesenchymal-CTCs were defined as CTC-score ‘positive’ using our previously established definition based on the analysis of healthy control blood samples (32)

CTCs were isolated from a separate 7.5 mL of whole blood using the Parsortix^® and collected into a 1.5 mL low-retention eppendorf. Total RNA was extracted using miRNeasy micro kit (Qiagen, Hilden, Germany) following manufacturer’s instructions but eluted with a final volume of 11.5 μL. The total 11.5 µL of RNA extracted from CTCs was mixed with 0.5 µL of random primers and denatured at 65 °C for 5 min. After incubation for 5 min on ice, 4 µL of first strand buffer, 2 µL of 0.1 M DTT, 1 µL of 10 mM dNTPs (Roche, Basel, Switzerland), 0.5 µL of water and 0.5 µL of Superscript II (Thermo Fisher Scientific, Waltham, Massachusetts, Unites States) were added and cDNA synthesis was performed at 42 °C for 2 hrs, followed by enzyme inactivation by heating at 70 °C for 15 mins. Multiplex RT-qPCR was performed by Barts and the London Genome Centre using BioMark HD system (Fluidigm Corporation, South San Francisco, California, United States). 96.96 Dynamic Array Integrated fluidic circuit (IFC) was used to test expression levels of 32 assays in triplicates within one reaction plate. A list of TaqMan probes (Applied Biosystems, Massachusetts, Unites States) used are shown in Supplementary Table 2 . The brief workflows were as follows: (1) pooling the TaqMan assays. Combine equal volumes of each 20X TaqMan Gene Expression assays in a 0.5 mL microcentrifuge tube, up to 100 μL in total. Dilute the pooled assays using DNA Suspension Buffer (10mM Tris, pH 8.0, 0.1 mM EDTA) so that each assay is at a final concentration of 0.2X. (2) Combine 2.5 μL of TaqMan^® PreAmp MasterMix (Life Technologies, Carlsbad, California, Unites States), 1.25 μL of pooled assay mix and 1.25 μL of cDNA to make the final sample mixture in each aliquot. (3) Place reaction tubes in the thermal cycler and cycle as (95 °C for 10 minutes followed by 14 cycles of 95 °C for 15 secs and then 60 °C for 4 mins). Only the targets of interests are amplified and this results in small amount of cDNA being amplified equally without introducing bias. Following pre-amplification, the samples were diluted 1:5 (v/v) in DNA suspension buffer. Reactions were then assayed using Dynamic Arrays prepared as instructed by the manufacturer. PCR was performed with 40 cycles of reactions.

Two microarray expression profile data sets (GSE36135 (39), GSE33455 (11)) were downloaded from the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo), which are based on the GPL571 Affymetrix Human Genome U133A 2.0 Array [HG_U133A_2] and GPL570 Affymetrix Human Genome U133 Plus 2.0 Array [HG-U133_Plus_2], respectively. The original Series Matrix data files were analysed with GEO2R (using the GEOquery and limma R packages from Bioconductor (http://www.bioconductor.org/) to identify differentially expressed genes (DEGs) in each of the paired docetaxel-resistant and docetaxel-sensitive cell lines in the datasets. A DEG was considered to be significant according to the following criteria: Fold-change (FC) >2 and false discovery rate (FDR) <0.05. Genes that were not upregulated in ≥2 DOC resistant cell lines were excluded to control for random variance in gene expression. Additionally, we identified reported PCa-specific and/or docetaxel-resistance related genes by literature search, and the two gene lists were combined to form a test gene panel. To select for genes suitable for CTC analysis, the genes were searched in The Genotype-Tissue Expression Portal V7 database for their expression in prostate and whole blood. Genes were selected based on their relative high expression in the prostate and low/no expression in whole blood. Candidate gene expression was subsequently validated in a panel of PCa cell lines and PBMC samples from five PCa biopsy negative males.

Mann-Whitney U test was applied to assess differences in clinical characteristics between patient groups. Data were shown as median (interquartile range [IQR]). Spearman’s rank correlation was used to assess associations of CTC counts and gene expression with concurrent PSA, ALP, LDH levels and OS/PFS. A combined risk score (CRS) was computed as the linear predictor of the fitted bivariate logistic model with PSA, ALP, CTC-score, total-CTC number and KLK2 count as only predictors (as CRS = a * Y + b * X…, where the values of ‘a’ and ‘b’ are the estimated log odds ratios). Survival curves were generated using the Kaplan-Meier method and compared using the logrank test. Bonferroni correction method was applied to adjust p values (p_adj) for multiple testing. RT-qPCR 2^-ΔΔCt was used to compare mRNA expression levels in patient samples. Receiver operating characteristic (ROC) curve analysis was used to evaluate prediction values. Follow-up time started on the date of administration of the first docetaxel dose. Observations were censored on the date of last follow-up.

Statistical analyses were performed using GraphPad Prism 9. All tests were two-sided with p values of <0.05 considered statistically significant.

Clinical characteristics and CTC counts are summarised in Table 1 . At least one CTC was detected in 12/18 (67%) of mCRPC patients, 50% of patients had a positive CTC-score (≥1 epithelial-CTC and/or ≥1 EMTing-CTC and/or ≥4 mesenchymal-CTCs) and the median total CTC count was 1.5 (interquartile range= 0-5.8). In mHSPC patients, ≥1 CTC was detected in 26/43 (60%), 51% of patients had a positive CTC-score and the median total CTC count was 1 (interquartile range = 0-4.3). Figure 1A shows an example of immunofluorescence staining for three CTC subtypes. Figure 1B shows individual CTC subtype counts in mHSPC and mCRPC patients. No significant differences were found between CTC subtype numbers.

We subsequently investigated the relationship between CTCs and clinicopathological data. Spearman’s correlation was performed between CTCs and serum PSA and ALP, results are shown in Table 2 . In mCRPC patients, serum PSA was significantly correlated with total-CTC (ρ= 0.51, p= 0.032), epithelial-CTC (ρ= 0.51, p= 0.030), EMTing-CTC numbers (ρ= 0.53, p= 0.024) and positive CTC-score (ρ= 0.68, p= 0.0021). Serum ALP was significantly correlated with total-CTC (ρ= 0.51, p= 0.046) and epithelial-CTC numbers (ρ= 0.62, p= 0.012). In mHSPC patients, serum ALP was significantly correlated with mesenchymal-CTC numbers (ρ= 0.34, p= 0.044), however no other significant correlations were observed.

To assess if CTCs could predict radiological response to docetaxel following treatment cycles, we compared CTC numbers between partial response (PR), stable disease (SD) or progressive disease (PD) groups ( Supplementary Table 1 ), and performed ROC analysis. Due to the limited sample size, we combined patients who had SD or PD at the end of docetaxel treatment into one group. While there were no significant differences in CTC numbers between patients with PR and SD/PD in this small cohort, trends were observed. In mCRPC patients with PR, limited mesenchymal-CTCs and no epithelial- and EMTing-CTCs were detected. Total-CTCs trended towards a significantly lower number in patients with a PR (p= 0.073) compared to those with SD/PD ( Figure 2A ) with an AUC of 0.80, p= 0.071 in predicting SD/PD ( Figure 2C ). In comparison, serum PSA had an AUC of 0.78, p= 0.089, and ALP had an AUC of 0.74, p= 0.20 ( Figure 2C ). In order to improve our ability to predict radiological response to docetaxel using blood-based biomarkers, we generated a combined risk score (CRS) combing the total number of CTCs, serum PSA and ALP levels for SD/PD prediction as CRS-TPA= 0.7414 * Total-CTC number + 0.02909 * PSA + 0.01423 * ALP, which resulted in an AUC of 0.90, p= 0.037, with a sensitivity of 84.62% and specificity of 100% when the cut-off was set at >3.45 ( Figure 2C ). The CRS-TPA performed better than serum PSA and ALP alone, although the difference was not statistically significant (p= 0.66, and 0.23 respectively), likely due to the limited sample size.

In mHSPC, epithelial- and EMTing-CTC numbers did not differentiate patients based on radiological response to docetaxel. Conversely, high mesenchymal-CTC numbers trended towards (p= 0.079) higher chance of SD/PD ( Figure 2B ) with an AUC of 0.65, p= 0.099 ( Figure 2D ). Serum ALP levels were best able to predict SD/PD, with an AUC of 0.68, p= 0.077, compared to that of serum PSA which had an AUC of 0.52, p= 0.82 ( Figure 2D ). A CRS comprised of mesenchymal-CTC number and serum ALP levels as CRS-MA= 0.3599 * mesenchymal-CTC number + 0.002037 * ALP, increased the AUC to 0.69, however with only a trend towards significance (p= 0.064) ( Figure 2D ).

To assess the prognostic value of CTCs, we correlated CTC measurements with patient OS and PFS. Long term follow-up data was available for 18 mCRPC and 42 mHSPC patients. The median follow-up time for mCRPC patients was 22.7 months (range 8.0-53.1 months), during which time 13/18 (72%) patients progressed and/or died. Spearman’s correlation ( Table 3 ) showed that OS significantly inversely correlated with total- (ρ= -0.66, p= 0.0027), epithelial- (ρ= -0.62, p= 0.0057) and EMTing-CTC (ρ= -0.65, p= 0.0034) numbers and a positive CTC-score (ρ= -0.80, p< 0.0001) in the mCRPC cohort. Additionally, PFS was significantly inversely correlated with epithelial-CTC numbers (ρ= -0.63, p= 0.0049) and a positive CTC-score (ρ= -0.65, p= 0.0033). We also evaluated the performance of serum biomarkers PSA and ALP. We found that serum PSA levels significantly inversely correlated with OS (ρ= -0.72, p= 0.0008) and PFS (ρ= -0.50, p= 0.035). With a view to improve the sensitivity and specificity of pre-docetaxel biomarkers for the prediction of OS, we generated a combined risk score using both PSA and CTC-score data. The AUC of a CRS comprised of PSA (AUC= 0.93) and CTC-score (AUC= 0.89) (CRS-PS= 0.08127 * PSA + 4.159 * CTC-score) to discriminate mCRPC patients with <24 months OS from those with ≥24 months OS reached 0.96, p= 0.0009, with a sensitivity of 88.9% and a specificity of 100% when the cut-off was set to <5.96 ( Figure 3 ). This made an improvement on the AUC of PSA alone but without significance, p= 0.60.

We then performed Kaplan Meier survival analysis using total and subtype CTC numbers to predict OS and PFS with optimal CTC number cut-offs evaluated ( Table 4 ). In mCRPC patients, each total-CTC number cut-off that was explored (<2 vs. ≥2 to <6 vs. ≥6 CTCs) was significantly associated with patients with short median OS. The detection of a positive CTC-score (12.80 vs. 37.33 months, HR= 5.08, p= 0.0005) ( Figure 4A ), ≥3 total-CTCs (12.80 vs. 37.33 months, HR= 3.84, p= 0.0053) ( Figure 4B ), ≥1 epithelial-CTC (14.30 vs. 37.33 months, HR= 3.89, p= 0.0041) ( Figure 4C ) and ≥1 EMTing-CTC (11.32 vs. 32.37 months, HR= 6.73, p= 0.0001) ( Figure 4D ) were most significantly associated with shorter median OS. Importantly, when the Bonferroni correction method was applied to adjust p values for multiple testing, a positive CTC-score (p_adj= 0.0055), ≥1 epithelial-CTC (p_adj= 0.045) and ≥1 EMTing-CTC (p_adj= 0.0011) remained significantly associated with shorter median OS. The detection of ≥2 epithelial-CTCs was most significantly associated with shorter median PFS (7.52 vs. 18.83 months, HR= 3.93, p= 0.0058) ( Figure 4E ).

The median follow-up time for mHSPC patients was 29.5 months (range 5.9-48.8 months), during which time 25/42 (60%) patients progressed and 11/42 (26%) died. Spearman’s correlation did not show associations of CTC measurements with PFS and OS, however, ALP was significantly inversely correlated with OS (ρ= -0.45, p= 0.0068) and PFS (ρ= -0.62, p< 0.0001) ( Table 3 ). Kaplan Meier analysis revealed that patients with ≥5 CTCs experienced the most significantly shorter median OS (24.57 vs undefined months, HR= 4.14, p= 0.0097) ( Table 4 , Figure 4F ).

Subsequently, we interrogated CTC mRNA expression to enhance the efficiency of CTCs as predictive biomarkers beyond CTC enumeration in the mHSPC cohort. Up-regulated differentially expressed genes in docetaxel-resistant cell lines from two microarray datasets were considered for docetaxel-resistant CTC detection. There were 162 genes commonly up-regulated in docetaxel-resistant cell lines and considered for further validation. Additionally, we identified 75 reported docetaxel-resistance related genes by literature search of relevant publications regarding the mechanisms of docetaxel resistance. These panels were combined to form a 237-candidate gene panel ( Supplementary Table 3 ). As the enriched CTC fraction that is harvested from the Parsortix® is not pure, it was necessary to account for white blood cell contamination in the sample. Therefore, we performed two steps of gene expression analysis to exclude any of the 237 candidate genes that were expressed in leucocytes, 1. in silico analysis and 2. in vitro experiments using PCa cell lines and patient derived leucocyte samples. Firstly, the 237 genes were searched in The Genotype-Tissue Expression Portal V7 database for their expression in prostate and whole blood. Genes were selected based on their relatively high expression in prostate tissue and low/no expression in whole blood. This resulted in a candidate panel of 39 genes. Secondly, using RT-qPCR in paired docetaxel-sensitive (PC3-AG) and docetaxel-resistant (PC3-D12) PCa cell lines, and five PBMC samples from biopsy-negative patients, we experimentally validated 23/39 genes for lack of expression in leucocytes. A further seven genes, including EMT and stem cell markers (PTPRC (CD45), SNAI1, ZEB1, NANOG, POU5F1, PROM1 and SOX2) were included due to their potential prognostic value in clinical samples and lack of expression in leucocytes, along with housekeeping genes GAPDH and MRFAP1. Unfortunately, the mCRPC sample size available at this time point was limited and gene expression analysis was not performed. However, the relationship between candidate gene expression and OS/PFS was investigated in 33 mHSPC patient samples.

Survival analysis was performed by separating patients into the following groups: 1) Expression vs. no expression 2) 50% highest expression vs. 50% lowest expression. Kaplan Meier analysis revealed that KLK2 expression was significantly associated with shorter median OS (27.17 vs. undefined months, HR= 3.87, p= 0.037) ( Figure 5A ) and PFS (8.13 vs. 26.57 months, HR= 5.15, p= 0.0002) ( Figure 5B ). KLK4 expression was significantly associated with shorter median PFS (10.17 vs. 26.57, HR= 3.01, p= 0.034) ( Figure 5C ), while KLK3 (PSA) expression had only a trend towards shorter median OS (30.13 vs. undefined months, HR= 3.18, p= 0.068) ( Figure 5D ) and did not separate patients based on PFS. Patients with 50% highest SNAI1 and/or ADAMTS1 expression experienced significantly shorter median OS (31.07 vs. undefined months, HR= 9.51, p= 0.0090 ( Figure 5E ); 31.07 vs undefined months, HR= 4.30, p= 0.047 ( Figure 5F ), respectively) and patients with the 50% highest ZEB1 expression experienced significantly shorter median PFS (11.50 vs. 24.40 months, HR= 2.5, p= 0.036) ( Figure 5G ). Hazard ratios for each gene for OS and PFS are presented in Figure 5H .

We then investigated the clinical outcomes of patients who had CTCs expressing more than one high-risk gene. The median PFS for patients with KLK2+ZEB1 ^hi (8.13 months Logrank p= 0.0004) ( Figure 6A ) and/or KLK2+SNAI1 ^hi (5.9 months, Logrank p= 0.0019) ( Figure 6B ) and/or KLK2+ADAMTS1 ^hi (5.9 months, Logrank p= 0.0004) ( Figure 6C ) was significantly shorter than for patients with no KLK2 expression and low expression of each gene. Patients expressing both genes also had high total CTC numbers detected in paired samples (KLK2+ZEB1 ^hi: ≥8 CTCs, KLK2+SNAI1 ^hi: ≥5 CTCs, KLK2+ADAMTS1 ^hi: 0, 5 and 9 CTCs).

Using receiver operating characteristic curve analysis, high expression of ADAMTS1 was significantly predictive of shorter OS with an AUC of 0.73, p= 0.043 ( Figure 7A ). Neither serum PSA or ALP levels significantly differentiated patients based on OS, although ALP showed a trend towards significance (AUC= 0.55, p= 0.66; AUC= 0.73, p= 0.072, respectively) ( Figure 7A ). Combining ADAMTS1, ALP and ≥5 total CTCs to form a CRS as CRS-AA5o= 0.06386 * ADAMTS1 + 0.001465 * ALP + 2.169 * ≥5 total CTCs, which increased the AUC to 0.83, p= 0.0070 (65% sensitivity and 87.5% specificity at a cut off of <0.48) vs AUC of 0.73 for ALP alone, ( Figure 7A ) however without a significant difference (p= 0.38).

We then evaluated the candidate genes for the prediction of PFS and found that high expression of ZEB1, SNAI1 and ADAMTS1 was significantly predictive of progression within 24 months (AUC= 0.77, p= 0.0092; AUC= 0.71, p= 0.039; AUC= 0.71, p= 0.043, respectively) ( Figure 7B ). Expression of KLK2 had a trend of correlated with <24 months PFS, but the AUC was not significant (AUC= 0.64, p= 0.18) ( Figure 7B ). ALP but not serum PSA levels significantly discriminated patients with <24 months PFS from those with ≥24 months PFS (AUC= 0.86, p= 0.0015; AUC= 0.52, p= 0.86, respectively) ( Figure 7B ).