The heterogeneity of an individual patient's tumor has been linked to treatment resistance, but quantitative biomarkers to rapidly and reproducibly evaluate heterogeneity in a clinical setting are currently lacking. Using established tools available in a College of American Pathologists–accredited and Clinical Laboratory Improvement Amendments–certified clinical laboratory, we quantified digital pathology features on 9,225 individual circulating tumor cells (CTC) from 179 unique metastatic castration-resistant prostate cancer (mCRPC) patients to define phenotypically distinct cell types. Heterogeneity was quantified on the basis of the diversity of cell types in individual patient samples using the Shannon index and associated with overall survival (OS) in the 145 specimens collected prior to initiation of the second or later lines of therapy. Low CTC phenotypic heterogeneity was associated with better OS in patients treated with androgen receptor signaling inhibitors (ARSI), whereas high heterogeneity was associated with better OS in patients treated with taxane chemotherapy. Overall, the results show that quantifying CTC phenotypic heterogeneity can help inform the choice between ARSI and taxanes in mCRPC patients. Cancer Res; 77(20); 5687–98. ©2017 AACR.

The heterogeneity of cancer has long been recognized by phenotypic differences in cell morphology within a tumor specimen, and the observation that disease is rarely eliminated with any single systemic therapy. Recent multi-gene sequencing efforts have extended the concept of intrapatient cancer heterogeneity to the single cell (1, 2), further revealing immense diversity in tumors between patients (3, 4), individual lesions within the same patient (5, 6), and different sites within each lesion (7, 8). It is therefore unsurprising that reliance on solid tissue biopsies to guide “precision medicine” is not so precise (9, 10). Recently reported profiling studies continue to elucidate the genomic complexity of cancer. Most are observational and are not reported as biomarkers that can be associated with clinical outcomes or used to impact medical decision making in patients (11, 12). Here, we report the development of a quantitative biomarker of CTC heterogeneity fit-for-purpose (13) of informing treatment selection at the time a change in therapy is needed, and explore the relationship between CTC phenotypic heterogeneity and patient survival following treatment with androgen receptor signaling inhibitors (ARSI) or taxane chemotherapies.

It has been hypothesized that the efficacy of pathway-specific targeted agents would be inversely related to the degree of intrapatient heterogeneity present when a therapy is administered (Supplementary Fig. S1A–S1D; refs. 8, 14, 15). The currently approved standard-of-care life-prolonging systemic therapies for progressive metastatic castration-resistant prostate cancer (mCRPC) include two therapeutic drug classes: pathway-specific ARSI and non-pathway–specific taxane-based chemotherapy. ARSI, such as abiraterone acetate, enzalutamide, and apalutamide, target the androgen receptor (AR) signaling pathway, and inhibit the growth of cells dependent on AR signaling for survival, while taxane-based chemotherapies, docetaxel and cabazitaxel, stabilize microtubules, broadly inhibiting cell division in all cells, including those that harbor diverse drivers of resistance to targeted therapy (16).

In current clinical practice, the choice between one therapy over another at a decision point in mCRPC management is largely empiric; the biology of an individual patient's cancer is typically not considered. In the first-line setting following progression on androgen deprivation therapy (ADT), most patients are treated with an ARSI over a taxane based on the high response rates, more favorable safety profile, and long survival times at this decision point (17, 18). This contrasts with the second- or later line setting where there is no consensus on management because the response to sequential ARSI is lower and of shorter duration, and response to the first does not predict response to the other (19). Every treatment to which a tumor has been exposed can alter the biology of the cancer (20), necessitating the serial profiling of disease. Validated therapy-guiding biomarkers to better inform each treatment decision is a significant unmet medical need (21).

High intratumor and intrapatient heterogeneity has been associated with poor prognosis in small cohorts of patients with breast and head and neck cancer (22, 23) and more recently in lung cancer (24). Expanded explorations of the relationship between heterogeneity and other clinical outcomes have been hindered by the range of sampling methods to obtain tumor, the range of devices and analytic methods, and the lack of analytically and clinically validated assays. CTCs are a source of tumor material for biologic profiling that can be obtained from routine, repeated phlebotomy samples with minimal patient discomfort. CTCs can represent multiple metastatic lesions at once and are more likely to provide information on the diversity of a patient's disease than a single site biopsy. However, existing single-cell genomic techniques (25) have yet to meet the regulatory requirements of analytic and clinical validity, turnaround time, and cost for more widespread clinical utility.

In contrast, automated phenotypic (protein and morphologic) assessment of single cells has established clinical utility in the arena of cytologic assessment of Pap smears for cervical cancer detection. In this report, we captured CTCs utilizing an analytically validated microscopy-based CTC detection technology (26) that allows for quantification of protein expression, and further characterized individual cells by digital pathology features previously validated and FDA approved for clinical use to identify and characterize malignant cells in Pap smears (27–29).

Our central hypothesis is that the degree of pretherapy phenotypic heterogeneity of CTCs inversely associates with patient survival following treatment with ARSI but not taxane therapy. To test this, individual features in each cell were used to define phenotypically distinct cell types. Heterogeneity was then quantified on the basis of the diversity of individual cell types in each patient sample using the Shannon index. The Shannon index is widely utilized in ecology and biodiversity research to measure the diversity or entropy of species, or in this case, CTCs. Previous applications of the Shannon index include showing differences in the genomic heterogeneity of separate and distinct metastatic lesions in an individual patient (30), and along with other measures, to predict the transition from Barrett esophagus to esophageal adenocarcinoma (31). Here, we associated this quantitative heterogeneity biomarker with overall survival (OS) following treatment with standard-of-care drug classes to demonstrate clinical utility: that patient outcomes can be improved by using the heterogeneity biomarker test result to inform the choice of treatment relative to nonuse of the test result. We also evaluated whether observed phenotypic heterogeneity was associated with the presence of multiple genomic clones, elucidated by single-cell sequencing of CTCs in a subset of patient samples.

Study design

This is a retrospective analysis of a prospectively accrued cross-sectional cohort of patients treated in a clinical practice setting at Memorial Sloan Kettering Cancer Center (MSKCC, New York, NY). All patients provided signed informed consent for participation on a MSKCC Institutional Review Board/Privacy Board–approved protocol prior to blood sampling and studies were conducted in accordance with the Declaration of Helsinki, Belmont Report, and U.S. Common Rule. The choice of therapy was at the discretion of the treating physician. All patients underwent a history that included details of the stage of disease at diagnosis, initial management, subsequent systemic therapies, physical examination, and laboratory studies including complete blood count, chemistry panel [albumin (ALB), alkaline phosphatase (ALK), lactate dehydrogenases (LDH), hemoglobin (Hgb), prostate specific antigen (PSA)], and serum testosterone to confirm castration status (<50 ng/dL; Table 1).

Table 1.

Demographics and sample characteristics of the CTC contributing cohort and clinical association cohort

CTC contributing cohortClinical association cohort
Patient characteristics 
 Unique patients 179 116 
 Age, years: median (range) 68 (45–91) 69 (48–91) 
Primary treatment 
 Prostatectomy 84 (47%) 55 (48%) 
 Radiation 34 (19%) 20 (17%) 
 Brachytherapy 7 (4%) 7 (6%) 
 None 54 (30%) 34 (29%) 
Sample characteristics 
 Total baseline (pretherapy) samples 221 145 
 Follow-up (on-therapy) samples 98 
 Total samples 319 145 
 Samples with CTCs 264 131 
Prior hormone therapies at baselinea 
 1–2 lines 82 (37%) 25 (17%) 
 3 lines 50 (23%) 38 (26%) 
 4 or more lines 89 (40%) 82 (57%) 
Chemotherapy status at baseline 
 Chemo-naïve 136 (62%) 63 (43%) 
 Chemo-exposed 85 (38%) 82 (57%) 
Metastatic therapy initiated after baseline 
 ARS Inhibitor 150 86 (59%) 
 Taxane 71 59 (41%) 
Line of metastatic therapy at baseline 
 1st line 76 
 2nd line 56 56 
 3rd line 40 40 
 ≥4th line 49 49 
Metastatic sites of disease at baseline 
 Bone onlyb 66 (30%) 46 (32%) 
 Lymph node onlyb 24 (11%) 10 (7%) 
 Bone and lymph nodeb 93 (42%) 63 (43%) 
 Bone and visceral +/− LNb 35 (16%) 25 (17%) 
 Soft tissue only 3 (1%) 1 (1%) 
Laboratory measures at baseline 
 PSA, ng/mL: median (range) 37.7 (0.1–3,728.2) 62.7 (0.1–3,728.2) 
 Hgb, g/dl: median (range) 12 (7–15) 11.6 (7–15) 
 ALK, unit/L: median (range) 110 (25–2,170) 121 (42–1,816) 
 LDH, unit/L: median (range)c 222.5 (123–1,293) 238.5 (123–1,004) 
 ALB, g/dl: median (range) 4.2 (31.–4.9) 4.2 (3.1–4.9) 
AR N-term CTC test at baseline 
 Total CTC/mL: median (range) 6.3 (0–991.3) 6.25 (0–991.3) 
CTC contributing cohortClinical association cohort
Patient characteristics 
 Unique patients 179 116 
 Age, years: median (range) 68 (45–91) 69 (48–91) 
Primary treatment 
 Prostatectomy 84 (47%) 55 (48%) 
 Radiation 34 (19%) 20 (17%) 
 Brachytherapy 7 (4%) 7 (6%) 
 None 54 (30%) 34 (29%) 
Sample characteristics 
 Total baseline (pretherapy) samples 221 145 
 Follow-up (on-therapy) samples 98 
 Total samples 319 145 
 Samples with CTCs 264 131 
Prior hormone therapies at baselinea 
 1–2 lines 82 (37%) 25 (17%) 
 3 lines 50 (23%) 38 (26%) 
 4 or more lines 89 (40%) 82 (57%) 
Chemotherapy status at baseline 
 Chemo-naïve 136 (62%) 63 (43%) 
 Chemo-exposed 85 (38%) 82 (57%) 
Metastatic therapy initiated after baseline 
 ARS Inhibitor 150 86 (59%) 
 Taxane 71 59 (41%) 
Line of metastatic therapy at baseline 
 1st line 76 
 2nd line 56 56 
 3rd line 40 40 
 ≥4th line 49 49 
Metastatic sites of disease at baseline 
 Bone onlyb 66 (30%) 46 (32%) 
 Lymph node onlyb 24 (11%) 10 (7%) 
 Bone and lymph nodeb 93 (42%) 63 (43%) 
 Bone and visceral +/− LNb 35 (16%) 25 (17%) 
 Soft tissue only 3 (1%) 1 (1%) 
Laboratory measures at baseline 
 PSA, ng/mL: median (range) 37.7 (0.1–3,728.2) 62.7 (0.1–3,728.2) 
 Hgb, g/dl: median (range) 12 (7–15) 11.6 (7–15) 
 ALK, unit/L: median (range) 110 (25–2,170) 121 (42–1,816) 
 LDH, unit/L: median (range)c 222.5 (123–1,293) 238.5 (123–1,004) 
 ALB, g/dl: median (range) 4.2 (31.–4.9) 4.2 (3.1–4.9) 
AR N-term CTC test at baseline 
 Total CTC/mL: median (range) 6.3 (0–991.3) 6.25 (0–991.3) 

aIncludes GnRH agonists and antagonists, antiandrogens, and next-generation hormonal therapies (abiraterone acetate, enzalutamide, and apalutamide).

bIncludes patients with other soft tissue disease.

cThree samples did not have LDH available.

Patient selection

Between December 2012 and March 2015, 265 patients with histologically confirmed mCRPC undergoing a change in systemic therapy for progressive disease were treated at MSKCC. Of these, 179 unique patients starting an ARSI (abiraterone, enzalutamide, or apalutamide) or taxane (docetaxel, cabazitaxel, or paclitaxel) provided 319 samples (221 pretherapy and 98 on-therapy samples), which formed two cohorts of samples (Table 1; Fig. 1). The first, the CTC contributing cohort, included all 319 samples and was used for automated feature extraction of protein biomarker expression and digital pathology features (Fig. 2A). The second, the clinical association cohort, is a subset of the CTC contributing cohort that consists only of samples obtained prior to the start of the second- or later line therapy (86 ARSI and 59 taxane, or 145 total pretherapy samples). First-line samples were excluded because the majority of mCRPC patients progressing on standard hormones are treated with an ARSI due to the high response rate. Few receive taxanes in the first line, limiting the ability to explore the association of heterogeneity with outcome to both drug classes. There were no additional exclusions; all available samples were included in the analysis.

Figure 1.

The distribution of patient samples in the CTC contributing cohort and clinical association cohort. CONSORT diagram showing the breakdown of patient samples analyzed for this study. Displayed from top to bottom is the total number of samples collected, samples included for CTC digital pathology for unsupervised clustering and phenotypic assessment (CTC Contributing Cohort), and the subset of samples used for the clinical association cohort by therapy class administered and line of therapy.

Figure 1.

The distribution of patient samples in the CTC contributing cohort and clinical association cohort. CONSORT diagram showing the breakdown of patient samples analyzed for this study. Displayed from top to bottom is the total number of samples collected, samples included for CTC digital pathology for unsupervised clustering and phenotypic assessment (CTC Contributing Cohort), and the subset of samples used for the clinical association cohort by therapy class administered and line of therapy.

Close modal
Figure 2.

CTC and clinical association analysis overview. Shown are schematics for CTC detection and digital pathology analysis on single cells (A), as well as generation of patient-level quantification of phenotypic heterogeneity by Shannon index (B).

Figure 2.

CTC and clinical association analysis overview. Shown are schematics for CTC detection and digital pathology analysis on single cells (A), as well as generation of patient-level quantification of phenotypic heterogeneity by Shannon index (B).

Close modal

CTC collection

Blood samples (7.5 mL) were collected in Streck tubes and processed at MSKCC or shipped to Epic Sciences and processed within 48 hours. Red blood cells were lysed, and approximately 3 × 106 nucleated cells were dispensed onto 10–16 glass microscope slides and placed at −80°C for long-term storage as described previously (26, 32, 33). Sample processing and testing were conducted in laboratories following both Clinical Laboratory Improvement Amendments (CLIA) and College of American Pathologists (CAP) regulations.

CTC immunofluorescent staining and detection

CTCs were characterized by automated immunofluorescent staining for DAPI (a DNA stain), cytokeratins, CD45 (hematopoietic lineage marker), and AR N-terminal domain (AR; Cell Signaling Technology; D6F11) as described previously (26, 32, 33). Up to two slides were evaluated per sample tested. CTCs were identified using fluorescent scanners that imaged every nucleated object on the slides. All CTCs had an intact nucleus (DAPI), no CD45 signal, and morphologic features consistent with malignancy, as published previously (26). Clinical laboratory scientists (licensed in California) conducted final quality control of CTC identification.

Individual CTC biomarker and digital pathology

We built upon digital pathology software from FDA-cleared devices used to characterize Pap smears for the context of use of diagnosing cervical cancer (27–29) to characterize CTCs. A systematic analysis performed by key leaders in the field identified the features that for decades were repeatedly found to be the most diagnostically meaningful in cytopathology, histopathology, and hematopathology applications (34). A subset of these features was found to be suitable for the magnification of the microscope objectives utilized by the Epic Sciences platform. All features considered were measured, and part of the raw data used for analyses, listed in Fig. 2A.

To allow for quantification of digital pathology features on a cell-by-cell basis, individual CTC images of cytokeratin (CK), nuclear, and AR N-terminal staining were segmented from each fluorescence channel image independently utilizing a library of single-cell features optimized for fluorescence microscopy (Fig. 2A; ref. 34). Features utilized for this analysis included morphometric and texture patterns of nuclear and CK staining, as well as the densitometry (intensity) of the CK and AR signal, which were quantified as a ratio of signal intensity on the CTC relative to surrounding white blood cells (WBC; Fig. 2A). The identification of a cluster of two or more CTCs was recorded as a categorical dichotomous variable, as was whether the level of CK and AR expression was above or below an analytic threshold of intensity (Fig. 2A) based on cultured cell line control cells spiked into healthy donor blood as described previously (26). Apoptotic CTCs were not included or reported, in part because nuclear fragmentation and membrane blebbing (35) confound the digital pathology features.

CTC mathematical phenotype identification and Shannon index

Use of the Shannon index necessitated defining unique “species” within a given space. These “species” were defined using an unsupervised clustering analysis of the CTC protein, digital pathology, and categorical features (Fig. 2B) into phenotypically similar subtypes of CTCs from all samples in the CTC contributing cohort. To enable cross-feature comparisons, minimize digital pathology feature redundancy, and distill the distinct sources of variance in the data, quantified biomarkers and features were z-score transformed prior to dimensionality reduction via principal component analysis, with the number of principal components selected via standard 85% variance inclusion (Supplementary Fig. S2A).

K-means clustering was then used to group similar CTCs into discrete phenotypic cell subtypes based on the principal components of digital pathology features while minimizing the amount of variance of biomarker and feature ranges within each “cluster,” or group. The number of clusters (‘k’) was selected by the "elbow method" by which the within cluster sum of squares for k = 1 to k = 40 is plotted, and a human analyst looks for a bend in the curve, signifying a point of diminishing returns where partitioning the data into additional clusters (more ‘k’) only marginally adds to the similarity within each cluster (Supplementary Fig. S2B). All CTCs in the cohort were then assigned a phenotypic cell subtype (“A”–“O”; Fig. 3A and B) reflecting the mathematical unsupervised cluster (not to be confused with a histologic cluster, or clump of CTCs traveling together within the bloodstream) and then grouped by their individual protein, digital pathology, and categorical features (Fig. 4A).

Figure 3.

Phenotypic cell types resulting from unsupervised clustering. A, Shorthand description of each cell type's distinguishing features. B, Example images of each cell type. Blue, DAPI (DNA); red, cytokeratins; green, CD45; white, AR. Note that AR signal indicates AR protein overexpression.

Figure 3.

Phenotypic cell types resulting from unsupervised clustering. A, Shorthand description of each cell type's distinguishing features. B, Example images of each cell type. Blue, DAPI (DNA); red, cytokeratins; green, CD45; white, AR. Note that AR signal indicates AR protein overexpression.

Close modal
Figure 4.

CTC phenotypic features, cell subtype classifications, and Shannon index of CTC phenotypic entropy in patient samples. A, CTC phenotypic features included in this analysis and unsupervised clustering of the CTC phenotypic features identified across all CTCs in the CTC Contributing Cohort (n = 9,225) was used to categorize CTCs into 15 phenotypic subtypes (“A”–“O”). Shown is a heatmap of mean individual cell features per phenotypic subtype. High, red; low, blue. B, Heatmap of CTC phenotype densities detected per patient sample, organized by line of therapy. The bar plot above shows the resulting Shannon index by samples from the observed intra-sample diversity of CTC phenotypes. C, An example of CTCs from a low Shannon index sample. D, An example of CTCs from a high Shannon index sample.

Figure 4.

CTC phenotypic features, cell subtype classifications, and Shannon index of CTC phenotypic entropy in patient samples. A, CTC phenotypic features included in this analysis and unsupervised clustering of the CTC phenotypic features identified across all CTCs in the CTC Contributing Cohort (n = 9,225) was used to categorize CTCs into 15 phenotypic subtypes (“A”–“O”). Shown is a heatmap of mean individual cell features per phenotypic subtype. High, red; low, blue. B, Heatmap of CTC phenotype densities detected per patient sample, organized by line of therapy. The bar plot above shows the resulting Shannon index by samples from the observed intra-sample diversity of CTC phenotypes. C, An example of CTCs from a low Shannon index sample. D, An example of CTCs from a high Shannon index sample.

Close modal

Next, the patient-level frequency of the defined CTC phenotypic subtypes (categorized as “A”–“O”/mL) were determined on a per sample basis (Fig. 4B). These “species densities” were then utilized to calculate a Shannon index for each sample in the clinical association cohort (R package “vegan”) that quantified the entropy of the individual CTC phenotypic subtypes present at the time of the blood draw (Fig. 4B). A higher Shannon index indicates both a higher number of different “species” (CTC phenotypic subtypes) present, as well as a more random distribution of those phenotypic subtypes, meaning that it would be more difficult, compared with lower Shannon index samples, to predict which phenotypic subtypes would be seen next given additional observation. If 0 or 1 CTCs were present, a value of zero is generated due to no sample present or the presence of only a single-cell subtype.

Posttreatment outcomes

Overall survival (OS) was calculated from initiation of therapy to death from any cause. Patients still alive at the time of last follow-up were right-censored.

Statistical analyses

Patient demographics and clinical characteristics at the time of blood draw were evaluated by descriptive statistics. χ2 and Wilcoxon rank sum tests were used to compare treatment groups for categorical and continuous characteristics, respectively. Median survival time as a function of a patient's continuous heterogeneity measure was evaluated using nonparametric kernel estimates (R package “sm,”; ref. 36). Time-to-event outcomes, categorized by group, were evaluated with the Kaplan–Meier method. Differences in time-to-event outcomes between categorized groups were evaluated using the maximum log-rank test, where the grouping was determined adaptively from the data (37).

The association of heterogeneity with time-to-event outcomes was additionally evaluated with HRs estimated from univariable and multivariable Cox proportional hazards (PH) regression methods. Continuous variables with right-skewed distributions were transformed by log2(x + 1) to establish a normal distribution. The pretherapy features evaluated for the multivariable Cox PH models included line of therapy (2, 3, or 4+, as factors), visceral metastases (present vs. not), pretherapy PSA (continuous, log2 +1), LDH (continuous, log2 +1), ALK (continuous, log2 +1), Hgb (continuous), ALB (continuous), patient age (continuous), CTC/mL (continuous, log2 +1), Shannon index (continuous), and therapy class (taxane vs. ARSI). These pretherapy features were individually assessed for correlation with Shannon index (Supplementary Fig. S3).

Visceral metastases, patient age, and ALB were excluded from the final model by using a stepwise selection method based on univariate proportional hazards (Schoenfeld residuals P > 0.05) and univariate significance with outcome (P < 0.05). Included in the models is the interaction between Shannon index and therapy administered. CTC enumeration was also tested for an interaction with therapy and heterogeneity index because of its correlation to Shannon index. All earlier line samples from patients who contributed multiple pretherapy samples but ultimately expired were right censored. In addition, the robust sandwich estimate for the covariance matrix was implemented for all Cox PH models to correct for possible underestimation of variance (38). All statistical tests were two-sided and performed at the 5% significance level. Data consolidation was conducted using KNIME. Statistical analyses utilized the R packages: “survival,” “stats,” “sm,” “vegan,” and “maxstat.” Graphical representations were generated with R packages: “ggplot2,” “gridExtra,” “scales,” “survminer,” and “ggthemes.”

CTC isolation, genome amplification, and next-generation sequencing

Methods for CTC relocation, picking, and sequencing have been described previously (39, 40). In brief, individual CTCs were relocated and recovered from assayed slides using the Eppendorf TransferMan NK4 micromanipulator. Single-cell whole genome amplification (WGA) was performed using the SeqPlex enhanced DNA amplification kit (Sigma). Shotgun libraries were constructed from 100 ng of WGA material using the NEBNext Ultra DNA Library Prep Kit and sequenced to approximately 0.3× depth by 2 × 150 bp PE sequencing.

Genome wide copy number variation (CNV) analysis was performed using the Epic Sciences single-cell CNV analysis pipeline. FASTQ files were aligned to hg38 human reference genome from UCSC Genome database. BAM files were filtered for MAPQ 30 reads, followed by two separate analyses for genome-wide profiling (pipeline 1) and individual gene copy number changes determination (pipeline 2). Pipeline 1: Hg38 human genome was divided into approximately 3,000 1M bp bins and counted across bins for each cell. Read counts per bin were normalized against WBC controls, and the circular binary segmentation algorithm (R Bioconductor package “DNAcopy”) was used to segment DNA copy number data (log2-normalized ratio, sample/reference) and identify abnormal copy number. Pipeline 2: reads were counted for each gene and for each sample, and normalized against the total sequencing reads for the particular sample. Normalized reads were compared with reference WBCs and z-scores were calculated for each gene. Z-score of > 3 and < −3 are used as significant cutoff for calling gene gain or loss.

Clinical characteristics of the patient population

A total of 179 unique patients contributed 319 samples prior to starting (221 pretherapy) or while receiving (98 on-therapy) treatment with an ARSI or taxane therapy. Two cohorts of patient samples were studied: a CTC phenotypic analysis cohort (CTC contributing cohort), and a clinical association cohort that is a subset of the former (see Patients and Methods, Table 1; Fig. 1). The clinical association cohort included only the pretherapy samples from patients about to start a second- or later line of systemic therapy for mCRPC.

Mathematical phenotype identification in pretherapy and on-therapy samples

CTC phenotypic heterogeneity was evaluated using the Shannon index, which measures the entropy of species or, in this case, the phenotypically defined CTC subtypes (Fig. 2B)

Unsupervised clustering of all of the single-cell features (Fig. 4A) from the 9,225 CTCs identified from the 319 samples in the CTC contributing cohort revealed a k of 15 (Supplementary Fig. S2B), which was used to “define” or classify individual cells into 15 phenotypic subtypes (“A”–“O”) where the cells were more similar to members of the subtype to which they were categorized relative to the other 14 subtypes (Fig. 3). Each phenotypic subtype represents a unique signature of the single-cell features (Fig. 4A). Each patient sample was then evaluated for the diversity of the defined CTC subtypes present in a blood draw, and quantified using the Shannon index (see Patients and Methods, Fig. 4B). Examples of CTCs from low Shannon index (Fig. 4C) and high Shannon index (Fig. 4D) samples are shown.

Intersample Shannon index is related to overall survival of ARSI, but not taxanes

The relationship between Shannon index (phenotypic entropy) and OS following treatment with a pathway-specific targeted ARSI or taxane chemotherapy was analyzed on a continuous basis with estimates of median survival by Shannon index (see Patients and Methods) separated by therapy administered (ARSI or taxanes; Fig. 5A). While median survival estimates remained constant for patients on taxanes with respect to Shannon index, the median survival of patients on ARSI was longer than on taxanes with comparable Shannon index at the low end. In contrast, patients with high Shannon index on ARSI had shorter median survival than those on taxanes with comparable Shannon index on the high end. This relationship was further visualized with Kaplan–Meier plots for each therapy class with patient populations dichotomized by the crossover point (Fig. 5A, arrow). Patients going onto ARSI (Fig. 5B) had a large difference in survival between “high” and “low” Shannon index samples (median OS: 8.8 months vs. 28.1 months, P = 0.0015), whereas patients going onto taxanes (Fig. 5C) did not (11.4 months vs. 12.9 months, P = 1).

Figure 5.

The degree of inter-sample Shannon index is related to overall survival of ARSI, but not taxanes. A, The relationship between degree of heterogeneity (Shannon index, x-axis) and overall survival (y-axis) is shown, along with nonparametric kernel estimates of median survival. Colors represent treatment received after pretherapy draw. O, patient alive at last observation; X, patient died at time indicated. Overall survival is alternately visualized with Kaplan–Meier plots from patients starting ARSI (B) and taxanes (C), with survival curves dichotomized with the survival crossover point from A, indicated with an arrow. Individual covariates were tested for additive power to predict overall survival using a Cox proportional hazards (PH) model. D, The resulting P values, HRs, and 95% confidence intervals. E, The interaction of therapy and heterogeneity integrated into the multivariate Cox PH model. The forest plot shows HRs and 95% confidence intervals.

Figure 5.

The degree of inter-sample Shannon index is related to overall survival of ARSI, but not taxanes. A, The relationship between degree of heterogeneity (Shannon index, x-axis) and overall survival (y-axis) is shown, along with nonparametric kernel estimates of median survival. Colors represent treatment received after pretherapy draw. O, patient alive at last observation; X, patient died at time indicated. Overall survival is alternately visualized with Kaplan–Meier plots from patients starting ARSI (B) and taxanes (C), with survival curves dichotomized with the survival crossover point from A, indicated with an arrow. Individual covariates were tested for additive power to predict overall survival using a Cox proportional hazards (PH) model. D, The resulting P values, HRs, and 95% confidence intervals. E, The interaction of therapy and heterogeneity integrated into the multivariate Cox PH model. The forest plot shows HRs and 95% confidence intervals.

Close modal

Multivariate analyses

To correct for potential imbalances in the demographics of the pre-ARSI and pretaxane patient samples due to greater use of the former in the second line and the latter in the third- or later line setting, we evaluated survival in the context of a multivariate Cox PH model utilizing known prognostic factors (see Patients and Methods). Integrated into the Cox PH model was the interaction between CTC phenotypic entropy (Shannon index) and therapy class (taxanes vs. ARSI). CTC enumeration was also considered as an interaction term with therapy class (see Materials and Methods), and while there was an interaction (data not shown), the interaction between Shannon index and therapy class was stronger. Of note, there remained a significant interaction between increasing Shannon index and higher risk of death on ARSI relative to taxanes (HR, 2.48; 95% CI, 1.22–5.03; P = 0.0119; Fig. 5D and E), even when adjusting for pretherapy prognostic factors. To assess the possibility that the observed treatment interaction effects are being driven by low Shannon index due to inclusion of samples with 0 or 1 CTCs, we created a separate model in which we excluded these samples and kept those from the clinical association cohort with a Shannon index greater than 0. The same trends as previously described were upheld (Supplementary Fig. S4A and S4B).

The Pleomorphism index

The relationship between phenotypic heterogeneity, drug class, and OS was additionally evaluated with an alternate measure of intrapatient phenotypic variance: the Pleomorphism index (see Supplementary Methods). Similar to the Shannon index, patients with low Pleomorphism index had longer survival on ARSI, and patients with high Pleomorphism index had longer survival on taxanes, observed as a continuous marker in univariate and multivariate settings (Supplementary Figs. S5A and S5B and S6A–S6E).

Heterogeneous genomic profiles identify frequent subclonal drivers of ARSI resistance observed in high phenotypic heterogeneity samples

Randomly selected CTCs from 17 patient samples, each with 10 or more CTCs, were single-cell sequenced to assess intrasample genomic heterogeneity with CNVs. From a time and resource perspective, it is currently impractical to individually sequence all CTCs in each sample to inform clinical decisions. However, we still sought to explore genomic heterogeneity at the single-cell level in a subset of samples using these research tools.

Shown are two examples of patient samples with multiple distinct genomic profiles with unique CNV patterns (Fig. 6A and B). One patient (Fig. 6A) had 22 CTCs sequenced and appears to have two major genomic patterns (I and II). Three sub-patterns were observed in pattern II with additional chromosome Y loss or a chr5q deletion, indicated by the red circles in Fig. 6A. The second patient (Fig. 6B) had 62 CTCs sequenced and four major genomic patterns identified (I–IV). Pattern II had many chromosomal breakpoints and alterations, including chr8p loss and 8q gain, which commonly occur in prostate cancer, and III and IV had single chromosome loss or gain. This patient also had nine additional CTCs that did not fit any of the four major genomic patterns identified in Fig. 6B, each instead representing its own unique pattern. These samples also had high Shannon index scores of phenotypic heterogeneity, and are provided as examples of the diverse genomic profiles observed in patient samples. They are not intended to offer correlations between phenotypic and genotypic heterogeneity. Properly powered comparisons between phenotypic and genotypic heterogeneity are planned for future work.

Figure 6.

Heterogeneous genomic profiles observed in high phenotypic heterogeneity samples. Copy number variation plots are displayed in A and B, where chromosomes are shown left to right, 1 to 22, X and Y, odd as red, even as blue. Log2-normalized copy number ratio (sample/reference) is indicated on the y-axis for two patient samples that exhibited high CTC phenotypic heterogeneity (by both Shannon index and Pleomorphism index). A, Red circles indicate chromosome Y loss and a chr5q deletion that were observed in subpatterns of genomic pattern II. All cells identified for this patient are visualized here. B, The four dominant profiles found across 53 of the 62 individual cells sequenced are illustrated. C, Heatmap of degree of subclonality of genomic alterations present in high phenotypic heterogeneity samples, with tile darkness indicating the proportion of single-cell whole genome sequenced CTCs with the indicated genomic alteration per sample. Black tiles, complete clonality of a given alteration; white tiles, complete absence; gray tiles, degrees of subclonality.

Figure 6.

Heterogeneous genomic profiles observed in high phenotypic heterogeneity samples. Copy number variation plots are displayed in A and B, where chromosomes are shown left to right, 1 to 22, X and Y, odd as red, even as blue. Log2-normalized copy number ratio (sample/reference) is indicated on the y-axis for two patient samples that exhibited high CTC phenotypic heterogeneity (by both Shannon index and Pleomorphism index). A, Red circles indicate chromosome Y loss and a chr5q deletion that were observed in subpatterns of genomic pattern II. All cells identified for this patient are visualized here. B, The four dominant profiles found across 53 of the 62 individual cells sequenced are illustrated. C, Heatmap of degree of subclonality of genomic alterations present in high phenotypic heterogeneity samples, with tile darkness indicating the proportion of single-cell whole genome sequenced CTCs with the indicated genomic alteration per sample. Black tiles, complete clonality of a given alteration; white tiles, complete absence; gray tiles, degrees of subclonality.

Close modal

A limitation of this analysis is the reliance on CNVs, as the single-cell sequencing technique utilized does not measure mutations or translocations that could reveal additional intercellular diversity. For example, common driver mutations in prostate cancer include alterations like SPOP mutations, AR mutations, and ERG rearrangements, which were not seen with these CNV analyses but very well could exist in the flat genomes shown in Pattern I of Fig. 6A and B.

A broader view of CTCs sequenced from high phenotypic heterogeneity samples showed varying degrees of subclonal gain or loss of key alterations known to drive resistance to ARSI (Fig. 6C). No patients were dominated by what is a more-or-less clonal genotype (i.e., white or black, but not gray tiles). Instead, most patients had many "gray" areas representing subclonal alterations of other drivers, suggesting that a single dominant clone was not identified or not present across all CTCs.

Certain alterations appeared to be more clonal than others, such as MYC, which when amplified, appeared in high proportions of CTCs within a patient sample. Other alterations, such as AR amplification, appeared frequently as subclones. These observations are consistent with intrapatient tumor clonality analyses (41), where MYC and AR amplifications identified in tissue biopsies were primarily truncal and subclonal, respectively, in a cohort of 10 autopsies. In contrast, the analyses reported here were conducted on cells from living patients, with the potential to use results to inform clinical decision making.

Developing biomarkers of heterogeneity to guide treatment selection and improve patient outcomes is an unmet medical need. Our objective was to develop a quantitative CTC heterogeneity biomarker assay that has achieved the level of method validation to be fit-for-purpose (13) of exploring its relationship to survival in mCRPC patients treated with an ARSI or taxane therapy. To do so, we retrospectively applied the Shannon index to prospectively collected and clinically annotated patient samples about to start a second or later line of treatment. The analysis showed that patient survival on taxanes was unrelated to the degree of heterogeneity, while lower heterogeneity scores were associated with a longer median survival on ARSI relative to taxanes, and higher heterogeneity with shorter median survival on ARSI relative to taxanes. A significant interaction was also observed between the Shannon index (Fig. 5D and E) and therapy class in multivariate models correcting for potential intra-cohort imbalances. The results support the hypothesis that CTC phenotypic heterogeneity measured prior to starting systemic therapy associates with differential outcomes on pathway-specific hormonal agents but not on non-pathway–specific chemotherapy.

As noted by a recent heterogeneity focus group, there are no standard methods to quantify heterogeneity, or consensus on whether the method to evaluate heterogeneity should be genetic, transcriptomic, phenotypic, epigenetic, or a combination thereof (42). All may prove relevant. It is critical to develop and test the association of a reproducibly measurable heterogeneity biomarker(s) with clinical outcomes: the context of use for which the biomarker result will be used to inform treatment decisions in the clinic. The outcome explored here was OS.

In this analysis, we expanded upon previously validated digital pathology features that were components of FDA-cleared devices for single-cell characterization with proven clinical utility for detecting cervical cancers from cytology Pap smears and applied these same tools to characterize CTCs phenotypically. Relative to single-site biopsies, CTCs can originate from multiple tumor sites within a patient, and provide a more global picture of an individual patient's cancer. In contrast to single-site biopsies, the acquisition of CTCs through phlebotomy poses minimal risk to the patient, and is amenable to repeated sampling to monitor disease evolution over time.

Critical to the assessment of intrapatient heterogeneity in CTCs is a CTC detection method that does not rely on pre-enrichment by epitope, size, or shape. These positive selection methods often miss critical CTC subtypes and do not consistently assess cell features essential to the analysis of heterogeneity. The CAP-accredited and CLIA-certified Epic Sciences platform used here has undergone the necessary analytic validation (accuracy, linearity, specificity, and intra/interassay precision) for CTC detection and enumeration (26) and has been previously demonstrated to detect a wide range of CTC phenotypes (32, 33, 40, 43). These include rare CTCs that are negative for epithelial markers with malignant genomics (40, 43) and CTCs that are smaller than WBCs (32, 43).

The approach used to analyze the digital pathology features represents a novel confluence of techniques adapted from diverse fields. The z-score transformation and principal component analysis is used in gene array applications to avoid redundancy and distill distinct sources of variance within the data (44). K-means clustering (45) is a machine learning technique with applications including, but not limited to, e-commerce, defense, ecology, and astronomy (46). Shannon index is a measure of species diversity frequently utilized in ecology research (47), where quantification and monitoring of species biodiversity and evolutionary events is common. Shannon index increases with both the number of unique species present and the evenness in species distribution. When applied to CTCs, it indicates the level of uncertainty to predict what phenotypic subtypes of CTCs are going to be detected next: the higher the uncertainty (entropy), the higher the heterogeneity. We additionally developed a simpler, alternate measure of CTC phenotypic heterogeneity within patient samples that does not require unsupervised clustering, the Pleomorphism Index (see Supplementary Methods; Supplementary Fig. S5A and S5B), performed identical survival analyses, and observed similar results with respect to phenotypic heterogeneity, drug class administered, and OS (Supplementary Fig. S6A–S6E). These results suggest that there are likely multiple ways to model heterogeneity, but the underlying phenomena observed is consistent with our central hypothesis.

A limitation of this approach is the potential lack of utility in the setting of low or no CTC counts. Recognizing this, as stated in Patients and Methods, our focus was the second or later line of therapy setting where the frequency of detection and individual CTC counts are higher, and the decision to select one form of therapy versus another is more critical. In our analyses, samples with 1 or 0 CTCs were assigned a heterogeneity score of 0 and included in our OS association analyses. However, despite these potential shortcomings, Shannon index does demonstrate the ability to differentially predict OS by therapy class, even when total CTC count is included as a covariate in multivariate models. A supplemental sub-cohort analysis that excluded samples with heterogeneity scores of 0 yielded similar results as well (Supplementary Fig. S4A and S4B). Ultimately, it may be concluded that the determination of a heterogeneity score (for which there is currently no unified definition, units, or reference range determined by the field; ref. 42) might have a requisite minimal number of CTCs for optimal quantification.

Another emerging technique that can infer heterogeneity on the genomic level is cell-free DNA (cfDNA; ref. 48). cfDNA is several steps removed from the cells of origin and represents a homogenized mixture, or pool, of tumor DNA and nontumor DNA from multiple cells, which poses additional technical hurdles when trying to distill features of entropy or variance within a patient sample. cfDNA can also include fragments of apoptotic or dead cells that could represent parts of tumor that were sensitive to the treatment and consequently eliminated (49). In part due to these reasons, we excluded the analysis of apoptotic CTCs and only scored intact, whole CTCs.

In addition to clinical utility, the abilities to provide results in time to inform clinical decisions and at a reasonable cost are essential for the use of a biomarker in clinical practice. Image-based phenotypic measures of tumor heterogeneity meet this requirement. Previously, our group analytically validated a single-CTC whole genome sequencing technology (39) to potentially assess intrapatient genomic heterogeneity. While promising, this approach is not scalable with available sequencing technologies when considering reproducibility, turnaround time, and cost. Acknowledging these caveats, we performed whole-genome sequencing on CTCs from a subset of samples with at least 10 evaluable CTCs. The results showed frequent subclonal alterations associated with ARSI resistance (Fig. 6C), consistent with the poor outcome of patients on ARSI who had high phenotypic heterogeneity determined with the Shannon index reported herein (Fig. 5). A greatly expanded cohort of single-cell sequencing results would be required for meaningful statistical analyses of genomic heterogeneity. However, these data suggest that samples with high phenotypic heterogeneity have heterogeneous genomic profiles as well.

In previous work, we studied an analytically validated IHC assay for AR-V7 as a treatment-selection biomarker for the same context (50), and found that 100% of patients with nuclear-localized AR-V7 protein in their CTCs had poor responses on ARSI and lived longer on taxanes (50). The assay, however, did not identify all patients who would have poor response on ARSI, as the frequency of AR-V7 detection was only 20% among patients who had less than 50% PSA decline by 12 weeks. Thus, most of the patients who did not respond to ARSI were AR-V7–negative. In the current study, patients with an increasing Shannon index had an increasing risk of death on ARSI relative to taxanes (Fig. 5). This suggests that patients with low heterogeneity scores would be more likely to survive longer on an ARSI over a taxane, and patients with high heterogeneity scores would likely survive longer on a taxane than an ARSI. Specifically in the context of AR-V7–negative patients, this measure could have the potential to more reliably inform treatment selection at this clinical decision point. Comparisons of AR-V7 to heterogeneity indices are planned for future studies.

CTC heterogeneity biomarkers may also be useful in the clinical development of novel therapies, as many trials evaluate cohorts of heavily pretreated patients who might exhibit high levels of heterogeneity that could portend resistance to targeted therapeutics. Such patients may be more suitably offered participation in studies of novel drug combinations, or combinations that include a cytotoxic drug more likely to affect diverse cell populations.

While the results reported here have the potential for broad implications, this study was designed to test a hypothesis that represents a first step in the clinical evaluation of an analytically valid, quantitative measurement of heterogeneity that could serve as a therapy-guiding biomarker to inform the choice between an ARSI or taxane for mCRPC patients starting a second- or later line therapy. The gold standard for evaluating the clinical utility of a predictive biomarker is through randomized, interventional trials. A limitation of observational studies like this one is the possibility of unobserved confounding factors. An additional limitation of our study is the lack of an external validation cohort. Definitive clinical utility of these heterogeneity biomarkers will require a sequence of trials, analogous to the development of a drug focused on a context of use. Further clinical validation in separate, larger cohorts is planned and assessment of longitudinal and kinetic changes over time in response to therapy are ongoing, both within the context of mCRPC clinical decisions and outside of prostate cancer.

H.I. Scher reports receiving a commercial research grant from Janssen and is a consultant/advisory board member for Astellas, Janssen Research and Development, LLC, and Medivation. Y. Wang has ownership interest (including patents) in Epic Sciences. D. Lu is a senior scientist at Epic Sciences. G. Heller reports receiving a commercial research grant from Janssen Diagnostics. No potential conflicts of interest were disclosed by the other authors.

Conception and design: H.I. Scher, R.P. Graf, M. Landers, R.V. Dittamore

Development of methodology: H.I. Scher, R.P. Graf, A. Jendrisak, Y. Wang, S. Greene, M. Fleisher, M. Landers, R.V. Dittamore

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): N. Schreiber, B. McLaughlin, S. Greene, R. Krupa, D. Lu, H.A. Vargas, M. Fleisher, M. Landers

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): H.I. Scher, R.P. Graf, A. Jendrisak, Y. Wang, S. Greene, H.A. Vargas, M. Fleisher, M. Landers

Writing, review, and/or revision of the manuscript: H.I. Scher, R.P. Graf, N. Schreiber, B. McLaughlin, Y. Wang, P. Bamford, H.A. Vargas, R.V. Dittamore

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): R.P. Graf, N. Schreiber, B. McLaughlin, Y. Wang, J. Lee, R. Krupa, D. Lu, J. Louw, L. Dugan

Study supervision: H.I. Scher, R.P. Graf, J. Louw, L. Dugan, M. Landers, G. Heller, R.V. Dittamore

We would like to thank the patients and their families for taking part in this study, and the clinical and laboratory staff at MSKCC and Epic Sciences.

This work was supported by the NIH/NCI P50-CA92629 SPORE in Prostate Cancer, NIH/NCI Cancer Center Support Grant P30-CA008748, and Department of Defense Prostate Cancer Research Program (PC121111 and PC131984). Prostate Cancer Foundation Challenge Award and David H. Koch Fund for Prostate Cancer Research were used to support the design and conduct of the study.

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.

1.
Pestrin
M
,
Salvianti
F
,
Galardi
F
,
De Luca
F
,
Turner
N
,
Malorni
L
, et al
Heterogeneity of PIK3CA mutational status at the single cell level in circulating tumor cells from metastatic breast cancer patients
.
Mol Oncol
2015
;
9
:
749
57
.
2.
Paguirigan
AL
,
Smith
J
,
Meshinchi
S
,
Carroll
M
,
Maley
C
,
Radich
JP
. 
Single-cell genotyping demonstrates complex clonal diversity in acute myeloid leukemia
.
Sci Transl Med
2015
;
7
:
281re2
.
3.
Roberts
NJ
,
Norris
AL
,
Petersen
GM
,
Bondy
ML
,
Brand
R
,
Gallinger
S
, et al
Whole genome sequencing defines the genetic heterogeneity of familial pancreatic cancer
.
Cancer Discov
2016
;
6
:
166
75
.
4.
Skoulidis
F
,
Byers
LA
,
Diao
L
,
Papadimitrakopoulou
VA
,
Tong
P
,
Izzo
J
, et al
Co-occurring genomic alterations define major subsets of KRAS-mutant lung adenocarcinoma with distinct biology, immune profiles, and therapeutic vulnerabilities
.
Cancer Discov
2015
;
5
:
860
77
.
5.
Brastianos
PK
,
Carter
SL
,
Santagata
S
,
Cahill
DP
,
Taylor-Weiner
A
,
Jones
RT
, et al
Genomic characterization of brain metastases reveals branched evolution and potential therapeutic targets
.
Cancer Discov
2015
;
5
:
1164
77
.
6.
Russo
M
,
Siravegna
G
,
Blaszkowsky
LS
,
Corti
G
,
Crisafulli
G
,
Ahronian
LG
, et al
Tumor heterogeneity and lesion-specific response to targeted therapy in colorectal cancer
.
Cancer Discov
2016
;
6
:
147
53
.
7.
Gerlinger
M
,
Rowan
AJ
,
Horswell
S
,
Larkin
J
,
Endesfelder
D
,
Gronroos
E
, et al
Intratumor heterogeneity and branched evolution revealed by multiregion sequencing
.
N Engl J Med
2012
;
366
:
883
92
.
8.
Burrell
RA
,
McGranahan
N
,
Bartek
J
,
Swanton
C
. 
The causes and consequences of genetic heterogeneity in cancer evolution
.
Nature
2013
;
501
:
338
45
.
9.
Hiley
C
,
de Bruin
EC
,
McGranahan
N
,
Swanton
C
. 
Deciphering intratumor heterogeneity and temporal acquisition of driver events to refine precision medicine
.
Genome Biol
2014
;
15
:
453
.
10.
McGranahan
N
,
Favero
F
,
de Bruin
EC
,
Birkbak
NJ
,
Szallasi
Z
,
Swanton
C
. 
Clonal status of actionable driver events and the timing of mutational processes in cancer evolution
.
Sci Transl Med
2015
;
7
:
283ra54
.
11.
Alberter
B
,
Klein
CA
,
Polzer
B
. 
Single-cell analysis of CTCs with diagnostic precision: opportunities and challenges for personalized medicine
.
Expert Rev Mol Diagn
2016
;
16
:
25
38
.
12.
Gorges
TM
,
Kuske
A
,
Rock
K
,
Mauermann
O
,
Muller
V
,
Peine
S
, et al
Accession of tumor heterogeneity by multiplex transcriptome profiling of single circulating tumor cells
.
Clin Chem
2016
;
62
:
1504
15
.
13.
Cummings
J
,
Raynaud
F
,
Jones
L
,
Sugar
R
,
Dive
C
. 
Fit-for-purpose biomarker method validation for application in clinical trials of anticancer drugs
.
Br J Cancer
2010
;
103
:
1313
7
.
14.
Swanton
C
. 
Intratumor heterogeneity: evolution through space and time
.
Cancer Res
2012
;
72
:
4875
82
.
15.
Crockford
A
,
Jamal-Hanjani
M
,
Hicks
J
,
Swanton
C
. 
Implications of intratumour heterogeneity for treatment stratification
.
J Pathol
2014
;
232
:
264
73
.
16.
Mohler
JL
Armstrong
AJ
Bahnson
RR
D'Amico
AV
Davis
BJ
Eastham
JA
et al 
Prostate Cancer, Version 1.2016
.
J Natl Compr Canc Netw
2016
;
14
:
19
30
.
17.
Ryan
CJ
,
Smith
MR
,
de Bono
JS
,
Molina
A
,
Logothetis
CJ
,
de Souza
P
, et al
Abiraterone in metastatic prostate cancer without previous chemotherapy
.
N Engl J Med
2013
;
368
:
138
48
.
18.
Beer
TM
,
Armstrong
AJ
,
Rathkopf
DE
,
Loriot
Y
,
Sternberg
CN
,
Higano
CS
, et al
Enzalutamide in metastatic prostate cancer before chemotherapy
.
N Engl J Med
2014
;
371
:
424
33
.
19.
Schrader
AJ
,
Boegemann
M
,
Ohlmann
CH
,
Schnoeller
TJ
,
Krabbe
LM
,
Hajili
T
, et al
Enzalutamide in castration-resistant prostate cancer patients progressing after docetaxel and abiraterone
.
Eur Urol
2014
;
65
:
30
6
.
20.
Gillessen
S
,
Omlin
A
,
Attard
G
,
de Bono
JS
,
Efstathiou
E
,
Fizazi
K
, et al
Management of patients with advanced prostate cancer: recommendations of the St Gallen Advanced Prostate Cancer Consensus Conference (APCCC) 2015
.
Ann Oncol
2015
;
26
:
1589
604
.
21.
Scher
HI
,
Morris
MJ
,
Stadler
WM
,
Higano
C
,
Basch
E
,
Fizazi
K
, et al
Trial design and objectives for castration-resistant prostate cancer: updated recommendations from the prostate cancer clinical trials working group 3
.
J Clin Oncol
2016
;
34
:
1402
18
.
22.
Seol
H
,
Lee
HJ
,
Choi
Y
,
Lee
HE
,
Kim
YJ
,
Kim
JH
, et al
Intratumoral heterogeneity of HER2 gene amplification in breast cancer: its clinicopathological significance
.
Mod Pathol
2012
;
25
:
938
48
.
23.
Mroz
EA
,
Tward
AD
,
Pickering
CR
,
Myers
JN
,
Ferris
RL
,
Rocco
JW
. 
High intratumor genetic heterogeneity is related to worse outcome in patients with head and neck squamous cell carcinoma
.
Cancer
2013
;
119
:
3034
42
.
24.
Jamal-Hanjani
M
,
Wilson
GA
,
McGranahan
N
,
Birkbak
NJ
,
Watkins
TBK
,
Veeriah
S
, et al
Tracking the evolution of non–small-cell lung cancer
.
N Engl J Med
2017
;
376
:
2109
21
.
25.
Navin
NE
. 
The first five years of single-cell cancer genomics and beyond
.
Genome Res
2015
;
25
:
1499
507
.
26.
Werner
SL
,
Graf
RP
,
Landers
ML
,
Valenta
DT
,
Schroeder
M
,
Greene
SB
, et al
Analytical validation and capabilities of the epic CTC platform: enrichment-free circulating tumour cell detection and characterization
.
J Circulating Biomarkers
2015
;
4
:
3
.
doi: 10.5772/60725
.
27.
Wied
GL
,
Bahr
GF
,
Oldfield
DG
,
Bartels
PH
. 
Computer-assisted identification of cells from uterine adenocarcinoma. A clinical feasibility study with TICAS. I. Measurements at wavelength 530 nm
.
Acta Cytol
1968
;
12
:
357
70
.
28.
Zahniser
DJ
,
Oud
PS
,
Raaijmakers
MC
,
Vooys
GP
,
Van de Walle
RT
. 
BioPEPR: a system for the automatic prescreening of cervical smears
.
J Histochem Cytochem
1979
;
27
:
635
41
.
29.
Tucker
JH
,
Shippey
G
. 
Basic performance tests on the CERVIFIP linear array prescreener
.
Anal Quant Cytol
1983
;
5
:
129
37
.
30.
Almendro
V
,
Kim
HJ
,
Cheng
YK
,
Gonen
M
,
Itzkovitz
S
,
Argani
P
, et al
Genetic and phenotypic diversity in breast tumor metastases
.
Cancer Res
2014
;
74
:
1338
48
.
31.
Maley
CC
,
Galipeau
PC
,
Finley
JC
,
Wongsurawat
VJ
,
Li
X
,
Sanchez
CA
, et al
Genetic clonal diversity predicts progression to esophageal adenocarcinoma
.
Nat Genet
2006
;
38
:
468
73
.
32.
Beltran
H
,
Jendrisak
A
,
Landers
M
,
Mosquera
JM
,
Kossai
M
,
Louw
J
, et al
The initial detection and partial characterization of circulating tumor cells in neuroendocrine prostate cancer
.
Clin Cancer Res
2016
;
22
:
1510
9
.
33.
Punnoose
EA
,
Ferraldeschi
R
,
Szafer-Glusman
E
,
Tucker
EK
,
Mohan
S
,
Flohr
P
, et al
PTEN loss in circulating tumour cells correlates with PTEN loss in fresh tumour tissue from castration-resistant prostate cancer patients
.
Br J Cancer
2015
;
113
:
1225
33
.
34.
Rodenacker
K
,
Bengtsson
E
. 
A feature set for cytometry on digitized microscopic images
.
Anal Cell Pathol
2003
;
25
:
1
36
.
35.
Wickman
G
,
Julian
L
,
Olson
MF
. 
How apoptotic cells aid in the removal of their own cold dead bodies
.
Cell Death Differ
2012
;
19
:
735
42
.
36.
Scher
HI
,
Jia
X
,
de Bono
JS
,
Fleisher
M
,
Pienta
KJ
,
Raghavan
D
, et al
Circulating tumour cells as prognostic markers in progressive, castration-resistant prostate cancer: a reanalysis of IMMC38 trial data
.
Lancet Oncol
2009
;
10
:
233
9
.
37.
MIller
R
,
Siegmund
D
. 
Maximally selected chi square statistics
.
Biometrics
1982
;
38
:
1011
6
.
38.
Lin
DY
,
Wei
LJ
. 
The robust inference for the proportional hazards model
.
J Am Stat Assoc
1989
;
84
:
1074
8
.
39.
Greene
SB
,
Dago
AE
,
Leitz
LJ
,
Wang
Y
,
Lee
J
,
Werner
SL
, et al
Chromosomal instability estimation based on next generation sequencing and single cell genome wide copy number variation analysis
.
PLoS One
2016
;
11
:
e0165089
.
40.
Anantharaman
A
,
Friedlander
TW
,
Lu
D
,
Krupa
R
,
Premasekharan
G
,
Hough
J
, et al
Programmed death-ligand 1 (PD-L1) characterization of circulating tumor cells (CTCs) and white blood cells (WBCs) in muscle invasive and metastatic bladder cancer patients
.
J Clin Oncol
34
;
2s
, 
2016
(
suppl; abstr 446
).
41.
Gundem
G
,
Van Loo
P
,
Kremeyer
B
,
Alexandrov
LB
,
Tubio
JM
,
Papaemmanuil
E
, et al
The evolutionary history of lethal metastatic prostate cancer
.
Nature
2015
;
520
:
353
7
.
42.
Alizadeh
AA
,
Aranda
V
,
Bardelli
A
,
Blanpain
C
,
Bock
C
,
Borowski
C
, et al
Toward understanding and exploiting tumor heterogeneity
.
Nat Med
2015
;
21
:
846
53
.
43.
McDaniel
AS
,
Ferraldeschi
R
,
Krupa
R
,
Landers
M
,
Graf
R
,
Louw
J
, et al
Phenotypic diversity of circulating tumour cells in patients with metastatic castration-resistant prostate cancer
.
BJU Int
2016
Aug 18.
[Epub ahead of print]. doi: 10.1111/bju.13631
.
44.
Cheadle
C
,
Vawter
MP
,
Freed
WJ
,
Becker
KG
. 
Analysis of microarray data using Z score transformation
.
J Mol Diagn
2003
;
5
:
73
81
.
45.
Jain
AK
. 
Data clustering: 50 years beyond K-means
.
Pattern Recognit Lett
2010
;
31
:
651
66
.
46.
Honarkhah
M
,
Caers
J
. 
Stochastic simulation of patterns using distance-based pattern modeling
.
Math Geosci
2010
;
42
:
487
517
.
47.
Brocchieri
L
. 
Phylogenetic diversity and the evolution of molecular sequences
.
J Phylogenet Evol Biol
2015
;
3
:
e109
.
doi: 10.4172/2329-9002.1000e109
.
48.
Wyatt
AW
,
Azad
AA
,
Volik
SV
,
Annala
M
,
Beja
K
,
McConeghy
B
, et al
Genomic alterations in cell-free DNA and enzalutamide resistance in castration-resistant prostate cancer
.
JAMA Oncol
2016
;
2
:
1598
606
.
49.
Tannock
IF
,
Hickman
JA
. 
Limits to personalized cancer medicine
.
N Engl J Med
2016
;
375
:
1289
94
.
50.
Scher
HI
,
Lu
D
,
Schreiber
NA
,
Louw
J
,
Graf
RP
,
Vargas
HA
, et al
Association of AR-V7 on circulating tumor cells as a treatment-specific biomarker with outcomes and survival in castration-resistant prostate cancer
.
JAMA Oncol
2016
;
2
:
1441
9
.