Multiplex protein detection on circulating tumor cells from liquid biopsies using imaging mass cytometry

Whole blood samples from primary and metastatic cancer patients were collected in cell-free DNA blood collection tubes (Streck, Omaha, USA) and shipped to the Kuhn laboratory at USC in temperature stabilized shippers (Standard 71, LLC, Los Angeles). Samples were processed within 24 h of collection using the HD-SCA protocol [6] and archived at  −80 °C. The sample MDA-42109 was collected from a patient with metastatic prostate cancer in 2013 with informed consent and under IRB approval as previously described [8]. Normal control donor blood samples were obtained from The Scripps Research Institute, La Jolla, California, USA. All specimen collections were performed under IRB approval.

Cell lines were obtained from the American Type of Culture Collection, (ATCC, Manassas, VA, USA). Prostate cancer line LNCaP was cultured in RPMI-1640 medium and breast cancer line MDA-MB-231 was cultured in DMEM/F12 medium (Corning, Manassas, VA, USA). Cells were harvested using Versene solution (Gibco) and washed in PBS. For validation of the metal-labeled antibodies in the combined workflow, cultured cells were added at a ratio 1:10 000 to whole blood from normal blood donors and processed according to the standard HD-SCA protocol [6], including storage at  −80 °C for a minimum of 24 h.

To complete the HD-SCA procedure, slides were thawed, stained with DAPI (nuclear stain) and fluorescent antibodies against CD45 (leukocyte marker) and pan-cytokeratin (CK, epithelial marker), and imaged in three colors using a high-speed automated microscope scanner. As part of the normal HD-SCA image analysis, the candidate CK positive cells were identified and coordinates registered. High resolution 40×  images of candidate cells were captured using a Nikon 80i microscope. Prior to staining with the metal-labeled antibody cocktail for IMC analysis, the slides were stored at 4 °C for several days to parallel normal operation of the HD-SCA protocol. Finally, cover slips were removed in PBS with minimal disruption of the cell layer. Slides were then washed with PBS to dissolve the imaging media in preparation for the metal-labeled antibody incubation. For the HD-SCA/Genomics workflow results presented in figure 3, samples were stained with 4-colors, adding AR to the DAPI, CD45 and CK panel described above.

Metal-labeled antibodies were provided by Fluidigm, either from the standard CyTOF catalog (http://maxpar.fluidigm.com/product-catalog-metal.php) or as custom conjugates (supplementary table 1 (stacks.iop.org/CSPO/2/015002/mmedia)). Metal-labeled antibody cocktails were prepared in 0.1% Tween-20, 1% BSA in PBS as per the dilution scheme presented in supplementary table 1. All samples, prepared as described above, were first blocked with 1% BSA and 0.2 mg ml⁻¹ mouse IgG Fc fragment (Thermo Scientific) in PBS for 30 min and then incubated with antibody cocktail for 1.5 h at RT, followed by washing with PBS and staining with DNA intercalator Ir-193 (Fluidigm) and cell membrane counter-stain In-115 (Fluidigm) for 30 min. Slides were again washed with PBS and rinsed with ddH20 for 5 s and dried overnight at room temperature prior to IMC analysis.

Cell boundaries were determined by the segmentation method described in the results section, using the Bioconductor package EBImage in R [11]. The standard deviation over the mean (SDOM) per cell for each channel was defined as the difference between the average ion count for the cell and the mean signal for all the cells divided by the standard deviation of all cells within each ROI (400  ×  400 mm). The noise level, used as a denominator for the S/N ratio was approximated as 1.5  ×  IQR (inter-quartile range) outside the upper quartile of all cells. When applicable, the limit of detection (LOD) for each marker was set as to mean  +  3.3  ×  standard deviation (one sided 95% * 2) or S/N  =  3 [12]. For figure 5, the dataset was exported in a FCS format using the FlowCore package in R and the gating viSNE plots were generated with the CYT graphical software package [13] in MATLAB.

Both the HD-SCA and the IMC were designed to operate on standard glass microscope slides in keeping with general procedures in the clinical pathology laboratory setting. The common slide format facilitated the integration of the two technologies into a single workflow viable for both basic and clinical research applications. The workflow, shown in figure 1(a), begins with processing the blood sample and spreading all nucleated cells on slides, followed by 3- or 4-color immunofluorescent staining and high-speed digital scanning to identify and image cells of interest. The coordinates of each candidate cell are recorded and re-imaged at 40×  for morphometric measurements, as previously reported [6, 8]. In preparation for the IMC (figure 1(b)), slides are re-incubated with MaxPar^™ metal-labeled antibodies. The slide is placed in the IMC instrument and candidate cells are relocated using a previously determined offset between coordinates from the 40×  imaging microscope and the IMC stage. A 400  ×  400 µm ROI around the cell of interest undergoes laser ablation aerosolizing a 1 µm² area/pulse (200 Hz), followed by ionization and quantification in the CyTOF Helios instrument. Ion mass data is collected for each pulse and processed to render images for each individual channel at 1 µm resolution, where the intensity of each pixel corresponds to the ion count value. The images are analyzed using an in-house developed software application in R. Raw data and processed results are stored in a PostgreSQL database enabling export of datasets and integration with HD-SCA assay parameters (figure 1(c)).

Zoom In Zoom Out Reset image size

Although the basic formats of the two assays are similar, adapting the IMC technology to the HD-SCA workflow required the development and validation of staining and analysis methods consistent with the previously validated HD-SCA protocols. The antibody preparation and staining steps were initially developed and optimized using cell lines spiked into normal donor blood samples (Methods).

For this study, a panel of metal-labeled antibodies was designed to include targets relevant to prostate cancer. The panel included PSA, PSMA and two epitopes for AR, along with epithelial markers EpCAM, E-cadherin, cancer stem cell marker CD44, EMT marker B-catenin, as well as a subset of markers for leukocyte characterization [14–16]. HLA-DR and CD14 were selected for monocytes identification along with CD3, CD8, CD4, CD45RA and CD38 to further characterize the T cell population [17]. CD66 was selected to identify mainly the neutrophil population, however, CD66 or CEA (carcinoembryonic antigen) is also expressed in in tumor cells and is known to have a tumor suppressive role in prostate cancer [18]. All catalog antibodies had been previously validated by Fluidigm for use in the CyTOF/Helios systems, but had not been validated in the IMC modality nor tested in the context of the HD-SCA workflow.

Each metal-labeled antibody was first tested for specificity and sensitivity against the prostate cancer line LNCaP, and metastatic breast cancer line MDA-MB231 (Materials and Methods and supplemental figure S1), using the leukocyte populations surrounding the candidate cells as an additional control. The unique challenge of determining the presence of the various antigens on the single candidate cell within each ROI in comparison to flow-based methods required the development of routines for quantification and scoring as described below.

As the pulsed laser sequentially ablates subcellular areas across the candidate cell along with all of the ~300 leukocytes present in the defined ROI, an ion count value is reported for each isotope at each of the 160 000 acquired pulses. The resulting data matrix is stored in a text file from which false color 1 µm resolution images are generated from each array of ion count values. The range of ion count values is internally normalized within each image, usually so that the brightest pixel represents the 98th percentile of the cumulative ion count signal.

Figure 2 shows an example of the images obtained from the IMC analysis of the two spiked cell lines. In the top row, an aggregate of LNCaP cells is seen in the center of each frame. A single LNCaP cell is also captured within the ROI. The different localized signal for PSA (membrane, figure 2(e)), E-cadherin (cell junctions, figure 2(f)), and androgen receptor (AR) (translocated to nuclei upon activation, figure 2(g)), can clearly be seen. For segmentation, the DNA intercalator data (figure 2(d)) is used to identify the objects in each frame. Adaptive thresholding [11] is utilized to generate a binary image, and then applying a watershed algorithm [19] to separate objects that are touching. Finally, the membrane counterstain is used for Voronoi propagation [20]to define the cell boundaries of each object with the primary mask acting as seeds (figure 2(c)). The mask is used to quantify the signals associated with each cell for each channel, by applying the mask on the raw data array of ion count values, rather than the normalized image. The ion count signals within the boundaries of each segmented cell in the ablation field are used to generate biaxial plots showing the average signal per cell for each marker relative to a chosen leukocyte marker, typically CD45 (figures 2(m)–(o)). This quantitation leads to a heuristic scoring system for each marker as described below.

Zoom In Zoom Out Reset image size

In practice, the quantitation of different markers is hampered by (i) subjectivity in adjusting the gain for normalization of images, (ii) a wide dynamic range of protein expression, (iii) inherent differences in detector sensitivity for the metal ions [21], (iv) experimental variation due to instrument fluctuation [22] and differences between staining batches etc, and (v) the difficulty of separating low abundant protein signals from background noise. As alluded to above, these challenges exacerbated in the setting of evaluating rare single cells, as compared to the common application of cytometry where larger populations are characterized.

Classification based on ion count alone was deemed unfeasible due to vast differences in dynamic range for each antibody-analyte pair, impeding normalization of the signals and their interpretation in heat maps. Thus, to establish an objective, semi-automated means for evaluating signals, a four level scoring system was created, with 0 being below the LOD, 1 exceeding LOD, and levels 2 to 3 set to discriminate the highest signal levels. The scoring system is based both on SDOM, and signal to noise (S/N) of the ion count value for the candidate cell to the surrounding leukocyte population, providing a relative measurement that is less sensitive to experimental variation than ion count alone. The LOD for each marker was set as equal to S/N  =  3 or SDOM  =  3.3 [17]. Cells exceeding the LOD were scored as 1+, signals with S/N of 7-20 or SDOM  >  6 were scored 2+, and finally score 3+  was assigned to signals of S/N  >  20 or SDOM  >  12. In figures 2(m)–(p), the scores are marked in the plots along with the estimated LOD based on either S/N  =  3 (dashed lines) or SDOM  =  3.3.

Some of the markers tested, including CD44, CD66, and vimentin, are also present on subpopulations of leukocytes. As seen in figures 2(i)–(k), most of the cells in the ROI express either CD44 or CD66, comprising two mutually exclusive subpopulations. In such cases, manual or automated gating can be readily applied to score the cell of interest, as demonstrated for a spiked MDA-MB-231 cell shown in figure 2(p). In this case, the CD44 signal was above the gated CD44+  leukocyte population, and was given a score of 3. Additional scoring of the different markers for LNCaP and MDA-MB-231 are presented in supplemental figure 1.

To test the integration of IMC with the HD-SCA workflow and to demonstrate the added information provided by IMC over 4-color immunofluorescence, the combined HD-SCA/IMC workflow was performed on matched blood and bone marrow aspirate samples from a patient with metastatic castrate resistant prostate cancer (MDA-42109) collected as part of a larger study [8]. In the initial HD-SCA assay of MDA-42109 conducted in 2013, 9 CTCs ml⁻¹ from blood and 207 DTCs ml⁻¹ in bone marrow were identified, with 48% of the CTCs and 78% of the DTCs also scoring positive for AR expression [23]. High resolution (40×) images had been collected for all the CTCs and a subset of DTCs to verify the AR positivity and subcellular localization prior to isolating the cells for whole genome amplification using the HD-SCA genomics workflow [7]. The results showed that 41/42 of the successfully amplified cells yielded highly rearranged and closely related copy number profiles with common elements typical of metastatic prostate cancer, indicating a single cancer lineage with four distinct sub clones as shown in the heat map and representative profiles (figure 3(a)). The high degree of clonality in the CNV profiled cells provided assurance that the candidate CTCs and DTCs selected for subsequent IMC analysis were highly likely to be bona fide cancer cells from the same lineage. However, we note that despite sharing the same genotype, individual cells exhibit significant phenotypic variability in AR expression and subcellular location, consistent with previous results [7]. In most cases, amplification of the AR gene on chromosome X corresponds to a strong nuclear AR signal in the corresponding fluorescent image (figure 3(b), panels 3, 4, 5 and 7). In other cases, the gene encoding AR was amplified even though the protein could not be detected by fluorescence (figure 3(b), panels 1, 2 and 6). Notably, there is also a genomic subclone comprising cases where AR amplification is not observed in the CNV profile (figure 3(b), panels 9–12) but AR is still detected in some of the fluorescent images (figure 3(b), panels 11 and 12). Morphological traits, such as cell and nuclear size and CK staining also varied among individual cells as is evident from the fluorescent images.

Zoom In Zoom Out Reset image size

To evaluate the performance of the combined HD-SCA/IMC analysis in relation to the results from the HD-SCA fluorescent assay alone, we retrieved one unstained slide each from blood and bone marrow of MD42109 from the  −80 °C archive where they had been stored for three years. The slides were stained and scanned using the standard HD-SCA 3-color protocol (DAPI, CD45 and pan-CK) to identify candidate cancer cells (figure 1). A total of 21 candidate DTCs and three CTCs were identified and their positions located on the respective slides. The stained slides were then stained with metal-labeled antibodies in preparation for imaging on the Hyperion instrument. The combined results of the HD-SCA immunofluorescence and IMC analysis of these cells are presented in figures 4 and 5.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Figures 4(a) and (b) show a selection of images generated in the IMC analysis of two representative cells, DTC #1271 from the bone marrow aspirate and CTC #1290 from blood, along with separate images of the markers being tested in the two modalities. The images from the fluorescence analysis of the representative cells (top panel) show that both cells were strongly CK positive and CD45 negative. The CD45 expression in leukocytes seen in the full ROI composite image generated by the IMC, show a high degree of correlation of CD45 expression ranging from low to high with the same cells in the corresponding ROI captured in the HD-SCA fluorescence assay image (figure 4(a), supplementary data, figure 1). The composite fluorescence images (figure 4(c)), demonstrate a phenotypic heterogeneity of both the morphology (size) and the variation in the CK signal among the cancer cells.

The heat map (figure 4(d)) provides a summary of the IMC signals in each channel for all 24 cells analyzed. The signal intensity for CK 8/18 varies even among the positive cells, and some of the cells that were positive for CK in the HD-SCA analysis were below the LOD in the IMC for CK 8/18, likely due to the fact that more CK epitopes are measured with the pan-CK antibody cocktail used in HD-SCA staining. Fifty percent (12/24) of the cells showed AR positivity in the IMC, in agreement with the parallel HD-SCA assay. Of the AR N-terminal positive cells, the majority were also positive for the C-terminal specific antibody, indicating that the full-length receptor is present in these cells. The rendered images from the IMC show a clear nuclear localization of AR, which follows upon activation of the receptor [24] (figures 4(a) and (b)). A majority of the cells (18/24) were positive for the epithelial marker EpCAM and 6 cells were also positive for E-cadherin, including the cell clusters 1277 and 1285 (figures 4(c) and (d)). Importantly, all of the cells were PSMA positive and a majority were also PSA positive, providing independent evidence for the prostate origin of the cells (figures 4(a), (b) and (d)). Thus, the IMC analysis confirmed and complimented the fluorescent HD-SCA and genomic signatures by providing additional markers for phenotypic profiling. Further, these results demonstrate the feasibility of using archived samples for IMC analysis, allowing us to revisit and further characterize long stored and previously analyzed samples in the HD-SCA biobank.

The HD-SCA assay is unique among the current CTC isolation methods in that it also assays and displays the full complement of leukocytes along with candidate CTC, enabling the analysis of a significant number of immune cells for multiple markers while analyzing CTCs. To demonstrate the feasibility of such an application, the surrounding leukocytes from the prostate cancer samples presented above were characterized by combining data aggregated from multiple ROIs on the slide into one dataset for each of the blood and bone marrow samples. To establish a reproducible analytical workflow, the first step was to remove all poorly segmented cells by size selection. Next, by gating the data (2089 cells) based on the average ion counts for each leukocyte on biaxial plots, the ratio compared to all nucleated cells of neutrophils (CD66⁺/CD44^low, 67%) monocytes (HLA-DR⁺/CD14⁺, 7%) and T-cells (CD3⁺/HLA-DR^–, 9%) could be determined (figure 5(a)). The ratio correlated well with the manual differential count performed by a trained clinical lab technician (69% neutrophils, 14% lymphocytes, 8% eosinophils, 8% monocytes, 1% basophils). Furthermore, the T-helper CD4⁺ cells (2%) and cytotoxic CD8⁺ (5%) were also identified within the T-cell population. Amongst each of the T-helper and cytotoxic T-cell populations, we also detected small populations of activated (CD38⁺, <1%) and naïve/effector cells (CD45RA⁺/CD4⁺, 1% and CD45RA⁺/CD8⁺, 3.8%). In figure 5(b), the signal intensity for different populations are shown overlaid in a 2D-viSNE map [13]. Similar data analysis for the bone marrow sample (2267 cells, figure 5(c)) demonstrated slightly lower ratio of neutrophils (55%), monocytes (4%) and lymphocytes (10%) compared to the blood. Thus, despite lacking markers for precursor cells in this particular analysis, the current results still indicate higher levels of immature cells present in the bone marrow compared to the blood, as expected [25].

In summary, we have developed an integrated approach of CTC and DTC characterization based on the HD-SCA workflow with the subsequent downstream multiplex proteomic analysis using IMC. The method was demonstrated using CTCs, DTCs as well as surrounding leukocytes in samples from a bone metastatic prostate cancer patient, showing added value to the data generated by the HD-SCA assay alone. Results were compared to genomic profiling of matching cells from the same samples, showing a higher degree of heterogeneity on a protein level than on the genomic level.