Abstract
The architecture of normal and diseased tissues strongly influences the development and progression of disease as well as responsiveness and resistance to therapy. We describe a tissue-based cyclic immunofluorescence (t-CyCIF) method for highly multiplexed immuno-fluorescence imaging of formalin-fixed, paraffin-embedded (FFPE) specimens mounted on glass slides, the most widely used specimens for histopathological diagnosis of cancer and other diseases. t-CyCIF generates up to 60-plex images using an iterative process (a cycle) in which conventional low-plex fluorescence images are repeatedly collected from the same sample and then assembled into a high dimensional representation. t-CyCIF requires no specialized instruments or reagents and is compatible with super-resolution imaging; we demonstrate its application to quantifying signal transduction cascades, tumor antigens and immune markers in diverse tissues and tumors. The simplicity and adaptability of t-CyCIF makes it an effective method for pre-clinical and clinical research and a natural complement to single-cell genomics.
INTRODUCTION
Histopathology is among the most important and widely used methods for diagnosing human disease and studying the development of multi-cellular organisms. As commonly performed, imaging of formalin-fixed, paraffin-embedded (FFPE) tissue has relatively low dimensionality, primarily comprising Hematoxylin and Eosin (H&E) staining supplemented by immunohistochemistry (IHC). The potential of IHC to aid in diagnosis and prioritization of therapy is well established1 but IHC is primarily a single channel method: imaging multiple antigens usually involves the analysis of sequential tissue slices or harsh stripping protocols (although limited multiplexing is possible using IHC and bright-field imaging2,3). Antibody detection via formation of a brown diamino-benzidine (DAB) or similar precipitate is also less quantitative than fluorescence4. The limitations of IHC are particularly acute when it is necessary to quantify complex cellular states and multiple cell types, such as tumor infiltrating regulatory and cytotoxic T cells,5 in parallel with tissue and pharmaco-dynamic markers.
Advances in DNA and RNA profiling have dramatically improved our understanding of oncogenesis and propelled the development of targeted anti-cancer drugs.6 Sequence data are particularly useful when an oncogenic driver is both a drug target and a biomarker of drug response, such as BRAFV600E in melanoma7 or BCR-ABL8 in chronic myelogenous leukemia. However, in the case of drugs that act through cell non-autonomous mechanisms, such as immune checkpoint inhibitors, tumor-drug interaction must be studied in the context of multi-cellular environments that include both cancer and non-malignant stromal and infiltrating immune cells. Multiple studies have established that these components of the tumor microenvironment strongly influence the initiation, progression and metastasis of cancer9 and the magnitude of responsiveness or resistance to therapy.10
Single-cell transcriptome profiling provides a means to dissect tumor ecosystems at a molecular level and quantify cell types and states.11 However, single-cell sequencing usually requires disaggregation of tissues, resulting in loss of spatial context.11,12 As a consequence, a variety of multiplexed approaches to analyzing tissues have recently been developed with the goal of simultaneously assaying cell identity, state, and morphology.13–17 For example, FISSEQ18 enables genome-scale RNA profiling of tissues at single-cell resolution, and multiplexed ion beam imaging (MIBI) and imaging mass cytometry achieve a high degree of multiplexing using antibodies as reagents, metals as labels and mass spectrometry as a detection modality.13,19 Despite the potential of these new methods, they require specialized instrumentation and consumables, which is one reason that the great majority of basic and clinical studies still rely on H&E or single-channel IHC staining. Moreover, methods that involve laser ablation of samples such as MIBI inherently have a lower resolution than optical imaging.
Thus there remains a need for highly multiplexed tissue analysis methods that (i) minimize the requirement for specialized instruments and costly, proprietary reagents, (ii) work with conventionally prepared FFPE tissue specimens collected in clinical practice and research settings, (iii) enable imaging of ca. 50 antigens at subcellular resolution across a wide range of cell and tumor types, (iv) collect data with sufficient throughput that large specimens (several square centimeters) can be imaged and analyzed, (v) generate high resolution data typical of optical microscopy, and (vi) allow investigators to customize the antibody mix to specific questions or tissue types. Among these requirements the last is particularly critical: at the current early stage of development of high dimensional histology, it is essential that individual research groups be able to test the widest possible range of antibodies and antigens in search of those with the greatest scientific and diagnostic value.
This paper describes a method for highly multiplexed fluorescence imaging of tissues, tissue-based cyclic immunofluorescence (t-CyCIF), inspired by a cyclic method first described by Gerdes et al.14. t-CyCIF also extends a method we previously described for imaging cells grown in culture.20 In its current implementation, t-CyCIF assembles up to 60-plex images of FFPE tissue sections via successive rounds of 4-channel imaging. t-CyCIF uses widely available reagents, conventional slide scanners and microscopes, manual or automated slide processing and simple protocols. It can, therefore, be implemented in most research or clinical laboratories on existing equipment. Our data suggest that high dimensional imaging methods using cyclic immunofluorescence have the potential to become a robust and widely-used complement to single cell genomics, enabling routine analysis of tissue and cancer morphology and phenotypes at single-cell resolution.
RESULTS
t-CyCIF enables multiplexed imaging of FFPE tissue and tumor specimens at sub-cellular resolution
Cyclic immunofluorescence14 creates highly multiplexed images using an iterative process (a cycle) in which conventional low-plex fluorescence images are repeatedly collected from the same sample and then assembled into a high dimensional representation. In the implementation described here, samples ~5 μm thick are cut from FFPE blocks, the standard in most histopathology services, followed be dewaxing and antigen retrieval either manually or on automated slide strainers in the usual manner.21 To reduce auto-fluorescence and non-specific antibody binding, a cycle of “pre-staining” is performed; this involves incubating the sample with secondary antibodies followed by fluorophore oxidation in a high pH hydrogen peroxide solution in the presence of light (“fluorophore bleaching”). Subsequent t-CyCIF cycles each involve four steps (Figure 1A): (i) immuno-staining with antibodies against protein antigens (three antigens per cycle in the implementation described here) (ii) staining with a DNA dye (commonly Hoechst 33342) to mark nuclei and facilitate image registration across cycles (iii) four-channel imaging at low and high magnifications (iv) fluorophore bleaching followed by a wash step and then another round of immuno-staining. In t-CyCIF, the signal-to-noise ratio often increases with cycle number due to progressive reductions in background intensity over the course of multiple rounds of fluorophore bleaching. This effect is visible in Figure 1B as the gradual disappearance of an auto-fluorescent feature (denoted by a dotted white oval and quantified in Figure 1-figure supplement 1; see detailed analysis below). When no more t-CyCIF cycles are to be performed, the specimen is stained with H&E to enable conventional histopathology review. Individual image panels are stitched together and registered across cycles followed by image processing and segmentation to identify cells and other structures.
t-CyCIF allows for one cycle of indirect immunofluorescence using secondary antibodies. In all other cycles antibodies are directly conjugated to fluorophores, typically Alexa 488, 555 or 647 (for a description of different modes of CyCIF see Ref 20). As an alternative to chemical coupling we have tested the Zenon™ antibody labelling method22 from ThermoFisher in which isotype-specific Fab fragments pre-labelled with fluorophores are bound to primary antibodies to create immune complexes; the immune complexes are then incubated with tissue samples (Figure 1-figure supplement 2). This method is effective with 30–40% of the primary antibodies that we have tested and potentially represents a simple way to label a wide range of primary antibodies with different fluorophores.
Imaging of t-CyCIF samples can be performed on a variety of fluorescent microscopes each of which represent a different tradeoff between data acquisition time, image resolution and sensitivity (Table 1). Greater resolution (a higher numerical aperture objective lens) typically corresponds to a smaller field of view and thus, longer acquisition time for large specimens. Imaging of specimens several square centimeters in area at a resolution of ~1 μm is routinely performed on microscopes specialized for scanning slides (slide scanners); we use a CyteFinder system from RareCyte (Seattle WA) configured with 10X 0.3 NA and 40X 0.6 NA objectives but have tested scanners from Leica, Nikon and other manufacturers. Figures 2A-B show an H&E image of a ~10 × 11mm metastatic melanoma specimen and a t-CyCIF image assembled from 165 individual image tiles stitched together.
Stitching is performed using Ashlar, a newly developed software package; the BaSiC algorithm is used to correct for uneven illumination (Figure 2-figure supplement 1)23. In Figure 2B tumor cells staining positive for S100 (a melanoma marker in green24) are surrounded by CD45-positive immune cells (CD45RO+ cells in white) and by stromal cells expressing the alpha isoform of smooth muscle actin (α-SMA in red). By zooming in on one tile, single cells can be identified and characterized (Figure 2C); in this image, CD4+ and CD8+ T-lymphocytes and proliferating pRB+ positive cells are visible. At 60X resolution on a confocal GE INCell Analyzer 6000, kinetochores stain positive for the phosphorylated form of the Aurora A/B/C kinase and can be counted in a mitotic cell (yellow arrowhead in Figure 2D). Nominally super-resolution imaging on a GE OMX Blaze Structured Illumination Microscope25 (using a 60X 1.42 Plan Apo objective) reveals very fine structural details including differential expression of Lamin isotypes (in a melanoma, Figure 2E and Figure2-figure supplement 2) and mitotic spindle fibers (in cells of a xenograft tumor; Figure 2F and Figure2-figure supplement 2). These data show that t-CyCIF images have readily interpretable features at the scale of an entire tumor, individual tumor cells and sub-cellular structures. Little subcellular (or super-resolution) imaging of clinical FFPE specimens has been reported to date (but see Ref. 26), but fine subcellular morphology has the potential to provide dramatically greater information than simple integration of antibodies intensities across whole cells.
To date, we have tested commercial antibodies against ~200 different proteins for their compatibility with t-CyCIF; these include lineage makers, cytoskeletal proteins, cell cycle regulators, the phosphorylated forms of signaling proteins and kinases, transcription factors, markers of cell state including quiescence, senescence, apoptosis, stress, etc. as well as a variety of non-antibody based fluorescent stains (Table 2). Multiplexing antibodies and stains makes it possible to discriminate among proliferating, quiescent and dying cells, identify tumor and stroma, and collect immuno-phenotypes.19,27,28 Use of phospho-specific antibodies and antibodies against proteins that re-localize upon activation (e.g. transcription factors) make it possible to assay the states of signal transduction networks. For example, in a 10-cycle t-CyCIF analysis of human tonsil (Figure 3A) subcellular features such as membrane staining, Ki-67 puncta (Cycle 1), ring-like staining of the nuclear lamina (Cycle 6) and nuclear exclusion of NF-κB (Cycle 6) can easily be demonstrated (Figure 3B). The five-cycle t-CyCIF data on normal skin in Figure 3C shows tight localization of auto-fluorescence (likely melanin) to the epidermis prior to pre-bleaching and images of three non-antibody stains used in the last t-CyCIF cycle: HCS CellMask Red Stain for cytoplasm and nuclei, Actin Red, a Phalloidin-based stain for actin and Mito-tracker Green for mitochondria.
In the current work, we rely exclusively on commercial antibodies that have previously been validated using IHC or conventional immunofluorescence; when feasible we confirm that staining by t-CyCIF resembles what has previously been reported for IHC staining. This does not constitute a sufficient level of testing or validation for discovery science or clinical studies and the patterns of staining described in this paper should therefore be considered illustrative of the t-CyCIF approach rather than definitive descriptions; we are currently developing a database of matched t-CyCIF and IHC images across multiple tissues and knockdown cell lines to address this issue and share validation test data with the wider research community.
Fluorophore inactivation, cycle count and tissue integrity
The efficiency of fluorophore inactivation by hydrogen peroxide, light and high pH varies with fluorophore but only minimally with the antibody to which the fluorophore is coupled (Alexa Fluor 488 is inactivated more slowly than Alexa Fluor 570 or 647; Figures 4B and Figure 4-figure supplement 1). We typically incubate specimens in bleaching conditions for 60 minutes, which is sufficient to reduce fluorescence intensity by 102 to 103-fold (Figure 4C). When testing new antibodies or analyzing new tissues, imaging is performed after each bleaching step and prior to initiation of another t-CyCIF cycle to ensure that fluorophore inactivation is complete. In preliminary studies, we have tested a range of other fluorophores for their compatibility with t-CyCIF including FITC, TRITC, phycoerythrin, Allophycocyanin, eFluor 570 and eFluor 660 (eBioscience). We conclude that it will be feasible to increase the number of t-CyCIF channels per cycle from four to at least six (3 to 5 antibodies plus a DNA stain). However, all of the images in this paper are collected using a four-channel method.
The primary limitation on the number of t-CyCIF cycles that can be performed is the integrity of the tissue: some tissues samples are physically more robust and can withstand more staining and washing procedures than others (Figure 4D). To study the effect of cycle number on tissue integrity, we performed a 10-cycle t-CyCIF experiment on a tissue microarray (TMA) comprising a total of 40 cores from 16 different tissues and tumor types. After each t-CyCIF cycle, the number of nuclei remaining was quantified for each core relative to the initial number. For example, Figure 4D shows breast, bladder, lung and prostate cores in which cell number fell between ~2% and an unusually high 46% (apparent increases in cell number in these data are caused by fluctuation in the performance of cell segmentation routines and are not statistically significant). Cells that were lost appear red in these images. The data show that loss is often uneven across samples, preferentially affecting regions of tissue with low cellularity.
Overall, we found that the extent of cell loss varied with tissue type and, within a single tissue type, from core to core (6 breast cores are shown; Figure 4E). For many tissues we have not yet attempted to optimize cycle number and the experiments performed to date do not fully control for pre-analytical variables29 such as fixation time and the age of tissue blocks. As a rule, we find that normal tonsil, skin, glioblastoma, ovarian cancer, pancreatic cancer and melanoma can be subjected to >15 cycles with less than 25% cell loss. Figure 4F shows a melanoma specimen subjected to 20 t-CyCIF cycles with good preservation of cell and tissue morphology (Figure 4G). We conclude that t-CyCIF is compatible with multiple normal tissues and tumor types but that some tissues and/or specimens can be subjected to more cycles than others. One requirement for high cycle number appears to be cellularity: samples in which cells are very sparse tend to be more fragile. We expect improvements in cycle number with additional experimentation and the use of fluidic devices that deliver staining and wash liquids more gently.
One concern about cyclic immunofluorescence is that the process is relatively slow; each cycle takes 6–8 hours and we typically perform one cycle per day. However, a single operator can easily process 30 slides in parallel, and in the case of TMAs, 30 slides can over 2000 more different samples. Under these conditions, the most time-consuming step in t-CyCIF is collecting 200–400 images for each slides. Time could be saved by imaging fewer cells per sample, but the results described below (demonstrating substantial cellular heterogeneity in a single piece of a tumor resection) strongly argue in favor of analysis of as large a fraction of each tissue specimen as possible. As a practical matter, data analysis and data interpretation remain more time-consuming than data collection. We also note that the throughput of t-CyCIF compares favorably with other tissue-imaging platforms with single-cell transcriptome profiling.
Impact of cycle number on immunogenicity
Because t-CyCIF assembles multiplex images sequentially, it is sensitive to factors that alter immunogenicity as cycle number increases. To investigate such effects, we performed a 16-cycle t-CyCIF experiment in which the order of antibody addition was varied between two immediately adjacent tissue slices cut from the same tissue block (Figure 5A; Slides A and B); the study was repeated three times, once with tonsil and twice with melanoma specimens with similar results ((~1.8 × 105 cells were used for the analysis and overall cell loss was <15%).
This experimental design made it possible to judge: (i) the repeatability of staining a single specimen using the same set of antibodies (Figure 5A, denoted by yellow highlight) (ii) the similarity of staining between slides A and B (blue highlight) and (iii) the effect of swapping the order of antibody addition (cycle number) between slides A and B (blue lines). Comparisons within a single slide were made on a cell-by-cell basis but because slides A and B contain different cells, comparisons between slides were made at the level of intensity distributions (computed on a per-cell basis following segmentation). The repeatability of staining (as measured in cycles 3,7,12 and 16) was performed using anti-PCNA-Alexa 488, anti-Vimentin-Alexa 555 and anti-Tubulin-Alexa 647 which bind abundant proteins with distinctive intracellular distributions (Figure 5B). Repeated staining of the same antigen is expected to saturate epitopes, but we reasoned that this effect would be less pronounced the more abundant the antigen. For PCNA, the correlation in staining intensities across 4 cycles was high (ρ = 0.95 to 0.99) and somewhat lower in the case of Vimentin and Tubulin (ρ = 0.80 to 0.95; Figure 6A; a more extensive comparison is shown in Figure 6-figure supplement 1). When we examined the corresponding images, it was readily apparent that Tubulin, and to a lesser extent Vimentin, stained more intensely in later than in earlier t-CyCIF cycles (see intensity distributions in Figure 6A and images in Figure 6B). When images were scaled to equalize the intensity range (by histogram equalization), staining patterns were indistinguishable across all cycles and loss of cells or specific subcellular structures was not obviously a factor (Figure 6B, left vs right panels and Figure 6C). Thus, for at least a subset of antibodies, staining intensity increases rather than decreases with cycle number whereas background fluorescence falls. As a consequence, dynamic range, defined here as the ratio of the least to the most intense 5% of pixels, frequently increases with cycle number (Figure 6A and Figure 6-figure supplement 1). These effects were reproducible across slides A and B in all three experiments performed.
When we compared staining between slides A and B for the same antibodies and cycle number, the overlap in intensity distributions was high (>0.85), demonstrating good sample to sample reproducibility. The overlap remained high for the majority of antibodies even when they were used in different cycles on slides A and B, but for some antibodies, signal intensity clearly increased or decreased with cycle number (Figure 6D; blue and red outlines). In the case of 8 antibodies (including tubulin, as discussed above) the overlap in intensity distributions was <0.6 as a consequence of both increases and decreases in staining intensity with cycle number (Figure 6E). Overall, we found that the repeatability of staining between two biological samples was highest when the antibodies were used in the same cycle on both samples, lower when the antibodies were used in different cycles on the sample, and lowest when both the order and sample were different (Figure 6F).
The reasons for these differences are not known, but the fact that the same changes in staining intensity were observed across multiple experiments (for any single antibody) suggests that they arise not from irreproducibility of the t-CyCIF procedure but rather from changes in epitope accessibility. Even in these cases, it appears that it is absolute intensity rather than morphology that is variable. Thus, while changes in staining intensity with cycle number are a concern for a subset of t-CyCIF antibodies, it should be possible to minimize the problem by staining all samples in the same order. Other approaches will also be important; for example, using calibration standards and identifying antibodies with the least variation in cycle number.
One way to reduce artefacts generated by differences in the order of antibody addition is to create a single high-plex antibody mixture and then stain all antigens in parallel. This approach is not compatible with t-CyCIF but is feasible using methods such as MIBI or CODEX19,28. However, there is substantial literature showing that the formulation of highly multiplex immuno-assays is complicated by interaction among antibodies30 that has a physicochemical explanation in some cases in weak self-association and viscosity.31 Consistent with these data, we have observed that when 8 or more unlabeled antibodies are added to a t-CyCIF experiment, the intensity of staining can fall, although the effect is smaller than observed with the antibodies most sensitive to order of addition. We conclude that the construction of sequentially-applied t-CyCIF antibody panels and of single high-plex mixtures will both require optimization of specific panels and their method of use.
Analysis of large specimens by t-CyCIF
Review of large histopathology specimens by pathologists involves rapid and seamless switching between low power fields to scan across large regions of tissue and high-power fields to study cellular morphology. To mimic this integration of information at both tissue and cellular scales we performed 8-cycle t-CyCIF on a large 2 × 1.5 cm resection specimen that includes pancreatic ductal adenocarcinoma (PDAC) and adjacent normal pancreatic tissue and small intestine (Figure 7A-C). Nuclei were located in the DAPI channel and cell segmentation performed using a watershed algorithm (Figure 7-figure supplement 1: see Methods section for a discussion of the method and its caveats) yielding ~2 × 105 single cells each associated with a vector comprising 25 whole-cell fluorescence intensities. Differences in subcellular distribution were evident for many proteins, but for simplicity, we only analyzed fluorescence intensity on a per-antigen basis integrated over each whole cell. Results were visualized by plotting intensity value onto the segmentation data (Figure 7D), by computing correlations on a cell-by-cell basis (Figure 7E), or by using t-distributed stochastic neighbor embedding (t-SNE)32, which clusters cells in 2D based on their proximity in the 25-dimensional space of image intensity data (Figure 8A).
The analysis in Figure 7E shows that E-cadherin, keratin and β-catenin levels are highly correlated with each other whereas vimentin and VEGFR2 receptor levels are anti-correlated, recapitulating the known dichotomy between epithelial and mesenchymal cell states in normal and diseased tissues. Many other physiologically relevant correlations are also observed, for example between the levels of pERKT202/Y204 (the phosphorylated, active form of the kinase) and activating phosphorylation of the downstream kinase pS6S235/S236 (r = 0.81). When t-SNE was applied to all cells in the specimen, we found that those identified during histopathology review as being from non-neoplastic pancreas (red) were distinct from PDAC (green) and also from the neighboring non-neoplastic small intestine (blue) (Figure 8 B-D). Vimentin and E-Cadherin had very different levels of expression in PDAC and normal pancreas as a consequence of epithelial-to-mesenchymal transitions (EMT) in malignant tissues as well as the presence of a dense tumor stroma, a desmoplastic reaction that is a hallmark of the PDAC microenvironment.33 The microenvironment of PDAC was more heavily infiltrated with CD45+ immune cells than the normal pancreas, and the intestinal mucosa of the small intestine was also replete with immune cells, consistent with the known architecture and organization of this tissue.
The capacity to image samples that are several square centimeters in area with t-CyCIF can facilitate the detection of signaling biomarker heterogeneity. The WNT pathway is frequently activated in PDAC and is important for oncogenic transformation of gastrointestinal tumours34. Approximately 90% of sporadic PDACs also harbor driver mutations in KRAS, activating the MAPK pathway and promoting tumourigenesis35. Studies comparing these pathways have come to different conclusions with respect to their relationship: some studies show concordant activation of MAPK and WNT signaling and others argue for exclusive activation of one pathway or the other36. In t-SNE plots derived from images of PDAC, multiple sub-populations of cells representing negative, positive or no correlation between pERK and β-catenin levels can be seen (marked with labels “a”, ‘b” or “c”, respectively in Figure 8A). The same three relationships can be found in non-neoplastic pancreas and small intestine (Figure 8A and 7C). In PDAC, malignant cells can be distinguished from stromal cells, to a first approximation, by high proliferative index, which can be measured by staining for Ki-67 and PCNA37. When we gated for cells that were both Ki67high and PCNAhigh, and thus likely to be malignant, the co-occurrence of different relationship between pERK and β-catenin levels on a cellular level was again evident. While we cannot exclude the possibility of phospho-epitope loss during sample preparation, it appears that the full range of possible relationships between the MAPK and WNT signaling pathways described in the literature can be found within a specimen from a single patient, illustrating the impact of tissue context on the activities of key signal transduction pathways.
Multiplex imaging of immune infiltration
Immuno-oncology drugs, including immune checkpoint inhibitors targeting CTLA-4 and the PD-1/PD-L1 axis are rapidly changing the therapeutic possibilities for traditionally difficult-to-treat cancers including melanoma, renal and lung cancers, but responses are variable across and within cancer types. The hope is that tumor immuno-profiling will yield biomarkers predictive of therapeutic response in individual patients. For example, expression of PD-L1 correlates with responsiveness to the ICIs pembrolizumab and nivolumab38 but the negative predictive value of PD-L1 expression alone is insufficient to stratify patient populations39. In contrast, by measuring PD-1, PD-L1, CD4 and CD8 by IHC on sequential tumor slices, it has been possible to identify some immune checkpoint-responsive metastatic melanomas10. To test t-CyCIF in this application, 8-cycle imaging was performed on a 1 × 2 cm specimen of clear-cell renal cell carcinoma using 10 antibodies against multiple immune markers and 12 against other proteins expressed in tumor and stromal cells (Figures 9A-B; supplement file 4). A region of the specimen corresponding to tumor was readily distinguishable from non-malignant stroma based on α-SMA expression (α-SMAhigh regions denote stroma and α-SMAlow regions high density of malignant cells).
In the α-SMAlow domain, CD3+ or CD8+ lymphocytes were 4-fold enriched (Figures 9C) and PD-1 and PD-L1 positive cells were 13 to 20-fold more prevalent as compared to the surrounding tumor stroma (α-SMAhigh domain); CD3+ CD8+ double positive T-cells were found almost exclusively in the tumor. Suppression of immune cells is mediated by binding of PD-L1 ligand, which is commonly expressed by tumor cells, to the PD1 receptor expressed on immune cells10. To begin to estimate the likelihood of ligand-receptor interactions, we quantified the degree of co-localization of cells expressing the two molecules. The centroids of PD-1+ or PD-L1+ cells were determined from images (PD-1, red; PD-L1, green, Figure 9E) and co-localization (highlighted in yellow, Figure 9F) computed by k-nearest neighbor analysis. We found that co-localization of PD-1/PD-L1 was ~2.7 fold more likely (Figure 9-figure supplement 1) in tumor and stroma and was concentrated on the tumor-stroma border consistent with previous reports on melanoma10. These data demonstrate the potential of spatially-resolved immuno-phenotyping to quantify state and location of tumor infiltrating lymphocytes; such data may ultimately yield biomarkers predictive of sensitivity to immune checkpoint inhibitor10.
Analysis of diverse tumor types and grades using t-CyCIF of tissue-microarrays (TMA)
To explore the general utility of t-CyCIF in a range of healthy and cancer tissues we applied 8 cycle t-CyCIF to TMAs containing 39 different biopsies from 13 healthy tissues and 26 biopsies corresponding to low and high-grade cancers from the same tissue types (Figure 10A and Figure 10-figure supplement 1, Supplement file 3 for antibodies used, Supplement file 5 for TMA details and naming conventions) and then performed t-SNE on single cell intensity data (Figure 10B). The great majority of TMA samples mapped to one or a few discrete locations in the t-SNE projection (compare normal kidney tissue - KI1, low grade tumors - KI2, and high-grade tumors -KI3; Figure 10C) although ovarian cancers were scattered across the t-SNE projection (Figure 10D); overall, there was no separation between normal tissue and tumors regardless of grade (Figure 10E). In a number of cases, high grade cancers from multiple different tissues of origin co-clustered, implying that transformed morphologies and cell states were closely related. For example, while healthy and low grade pancreatic and stomach cancer occupied distinct t-SNE domains, high grade pancreatic and stomach cancers were intermingled and could not be readily distinguished (Figure 10F), recapitulating the known difficulty in distinguishing high grade gastrointestinal tumors of diverse origin by histophathology.40 Nonetheless, t-CyCIF might represent a means to identify discriminating biomarkers by efficiently sorting through large numbers of alternative antigens and antigen localizations.
Quantitative analysis reveals global and regional heterogeneity and multiple histologic subtypes within the same tumor in glioblastoma multiforme (GBM)
Data from single-cell genomics reveals extensive heterogeneity in many types of cancer41 but our understanding of this phenomenon requires spatially-resolved data.13 We performed 8-cycle imaging on a 2.5 cm × 1.8 mm resected glioblastoma (GBM) specimen imaging markers of neural development, cell cycle state and signal transduction (Figure 11A-B, Supplement file 6). GBM is a highly aggressive and genetically heterogeneous42 brain cancer commonly classified into four histologic subtypes.43 Following image segmentation, phenotypic heterogeneity was assessed at three spatial scales corresponding to: (i) 1.6 × 1.4 mm fields of view (252 total) each of which comprised 103 to 104 cells (ii) seven macroscopic regions of ~104 to 105 cells each, corresponding roughly to tumor lobes and (iii) the whole tumor comprising ~106 cells. To quantify local heterogeneity, we computed the informational entropy on a-per-channel basis for 103 randomly selected cells in each field (Figure 11C, see online Methods for details). In this setting, informational entropy is a measure of cell-to-cell heterogeneity on a mesoscale corresponding to 10–30 cell diameters. For a marker such as EGFR, which can function as a driving oncogene in GBM, informational entropy was high in some areas (Figure 11C; red dots) and low in others (blue dots). Areas with high entropy in EGFR abundance did not co-correlate with areas that were most variable with respect to a downstream signaling protein such as pERK. Thus, the extent of local heterogeneity varied with the region of the tumor and the marker being assayed.
Semi-supervised clustering using expectation–maximization Gaussian mixture (EMGM) modeling of all cells in the tumor yielded eight distinct clusters, four of which encompassed 85% of all cells (Figure 12A and Figure 12-figure supplement 1). Among these, cluster one had high EGFR levels, cluster two had high NGFR and Ki67 levels and cluster six had high levels of vimentin; cluster five was characterized by high keratin and pERK levels. The presence of four highly populated t-CyCIF clusters is consistent with data from single-cell RNA-sequencing of ~400 cells from five GBMs12. Three of the t-CyCIF clusters have properties reminiscent of established histological subtypes including: classical, cluster 1; pro-neural, cluster 3; and mesenchymal, cluster 6, but additional work will be required to confirm such assignments.
To study the relationship between phenotypic diversity and tumor architecture, we mapped each cell to an EMGM cluster (denoted by color). Extensive intermixing was observed at all spatial scales (Figure 12B). For example, field of view 147 was highly enriched for cells corresponding to cluster 5 (yellow), but a higher-magnification view revealed extensive intermixing of four other cluster types on a scale of ~3–5 cell diameters (Figure 12C). At the level of larger, macroscopic tumor regions, the fraction of cells from each cluster also varied dramatically (Figure 12D). None of these findings was substantially different when the number of clusters was set to 12 (Figure 12-figure supplement 2).
These results have several implications. First, they suggest that GBM is phenotypically heterogeneous on a spatial scale of 5–1000 cell diameters and that cells corresponding to distinct t-CyCIF clusters are often found in the vicinity of each other. Second, sampling a small region of a large tumor has the potential to misrepresent the proportion and distribution of tumor subtypes, with implications for prognosis and therapy. Similar concepts likely apply to other tumor types with high genetic heterogeneity, such as metastatic melanoma11, and are therefore relevant to diagnostic and therapeutic challenges arising from tumor heterogeneity.
DISCUSSION
The complex molecular biology and spatial organization of tissues and solid tumors poses a scientific and diagnostic challenge that is not sufficiently addressed using single-cell genomics, in which morphology is commonly lost, or H&E and single-channel IHC staining, which provide data on only a few proteins or molecular features. At the same time, the vast number of FFPE histological specimens collected in the course of routine clinical care and clinical trials (and in the study of model organisms) represents an underutilized resource with great potential for novel discovery. A variety of methods for performing highly multiplexed immune-based imaging of cells and tissues has recently been described including imaging cytometry27, MIBI19, DNA-exchange imaging (DEI)44 and CODEX28; FISSEQ18 directly images expressed RNAs. Like traditional antibody stripping approaches, the cyclic immunofluorescence approach first described by Gerdes et al14 and further developed here assembles highly multiplexed images by sequential acquisition of lower dimensional immunofluorescence images. We show here that the t-CyCIF implementation of cyclic immunofluorescence is compatible with a wide range of antibodies and tissue types and yields up to 60-plex images with excellent preservation of small intracellular structures.
The requirement in t-CyCIF for multiple rounds of staining and imaging might seem to be a liability but it has several substantial advantages relative to all-in-one methods such as MIBI, DEI and CODEX. First, t-CyCIF can be performed using existing fluorescence microscopes. Not only does this reduce costs and barriers to entry, it allows the unique strengths of slide-scanning, confocal, and structured illumination microscopes to be exploited. Using different instruments, samples several square centimeters in area can be rapidly analyzed at resolutions of ~1μm and selected fields of view studied at super-resolution (~ 110 nm on an OMX Blaze). Multiscale imaging makes it possible to combine tissue-level architecture with subcellular morphology, much like a pathologist switching between low to high power fields, but there is little chance that such capabilities can be combined in a single instrument. Because no spectral deconvolution is required, t-CyCIF can use highly optimized filter sets and fluorophores, resulting in good sensitivity. t-CyCIF antibody panels are also simple to assemble and validate using commercial antibodies, including those that constitute FDA-approved diagnostics. This avoids the limitations of pre-assembled reagent kits provided by manufacturers. Finally, t-CyCIF is compatible with H&E staining, enabling fluorescence imaging to be combined with conventional histopathology review.
Commercial systems for non-optical tissue imaging are only now starting to appear and it is difficult to compare their performance to multiplexed immunofluorescence, particularly because the approach published by Gerdes14 is proprietary and available only as commercial service. In contrast, the t-CyCIF implementation described here can easily be implemented in a conventional research or clinical laboratory without the need for expensive equipment or specialized reagents. As MIBI, DEI and CODEX instruments come on-line, direct comparison with t-CyCIF will be possible. We anticipate that high resolution and good linearity will be areas in which fluorescence imaging is superior to enzymatic amplification, laser ablation or mechanical picking of tissues. t-CyCIF is relatively slow when performed on a single sample, but when many large specimens or TMAs are processed in parallel, throughput is limited primarily by imaging acquisition, which is at least as fast as approaches involving laser ablation. Considerable opportunity exists for further improvement in t-CyCIF by switching from four to six-channels per cycle, optimizing bleach and processing solutions to preserve tissue integrity, using fluidic devices to rapidly process many slides in parallel and developing better software for identifying fields of view that can be skipped in large irregular specimens. Because direct fluorescence will remain challenging in the case of very rare epitopes, we speculate that hybrid approaches involving t-CyCIF and methods such as DEI will ultimately prove to be most effective.
As in all methods involving immune detection, antibodies are the most critical and difficult to validate reagents in t-CyCIF. To date, we have shown that over 200 commercial antibodies are compatible with the method as judged by patterns of staining similar to those previously reported for IHC; this is an insufficient level of validation for most studies and we are therefore working to develop a generally useful antibody validation resource (www.cycif.org). Thus, while this paper describes markers relevant to diagnosis of disease, our results are illustrative of the t-CyCIF approach and specific findings might not prove statistically significant when tested on larger, well-controlled sets of human samples.
There is little or no evidence that antigenicity falls across the board in t-CyCIF as cycle number increases; signal-to-noise ratios can even increase due to falling background auto-fluorescence. When samples are stained with the same antibodies in different t-CyCIF cycles, repeatability is high (as measured by correlation in staining intensity on a cell-by-cell basis) as is reproducibility across two successive slices of tissue (as measured by overlap in intensity distributions). Moreover, for the majority of antibodies tested, order of use is not critical. For some antibodies fluorescence intensity increases with cycle number and for others it decreases; these factors need to be considered when developing a staining strategy. While the precise reasons for variation in staining with cycle number are not clear, they can be overcome by staining all specimens with the same antibodies in the same order (which also represents the most practical approach). However, this solution is likely to be insufficient for creation of large-scale t-CyCIF datasets in which diverse tissues will be compared with each other (e.g. in proposed tissue atlases45) and it will therefore be important to identify antibodies for which cycle number has minimal impact and to create effective methods to correct for those fluctuations that do occur (e.g. inclusion of staining controls).
As an initial application of t-CyCIF, we examined a cancer resection specimen that includes PDAC, healthy pancreas and small intestine. Images were segmented and fluorescence intensities in ~105 whole cells calculated for 24 antibody channels plus a DNA stain. Integrating intensities in this manner does not make use of the many subcellular features visible in t-CyCIF images and therefore represents only a first step in data analysis. We find that expression of vimentin and E-cadherin, classical markers of epithelial and mesenchymal cells, are strongly anti-correlated at a single-cell level and that malignant tissue is skewed toward EMT, consistent with prior knowledge on the biology of pancreatic cancer46. The WNT and ERK/MAPK pathways are known to play important roles in the development of PDAC34, but the relationship between the two pathways remains controversial. t-CyCIF reveals a negative correlation between β-catenin levels (a measured of WNT pathway activity) and pERK (a measure of MAPK activity) in cells found in some regions of PDAC, non-malignant small intestine and pancreas, a positive correlation in other regions and no significant correlation in yet others. Thus, the full range of discordant observations found in the literature can be recapitulated within a single tumor, emphasizing the wide diversity of signaling states observable at a single cell level.
As a second application of t-CyCIF, we studied within-tumor heterogeneity in GBM, a brain cancer with multiple histological subtypes whose differing properties impact prognosis and therapy.43,47 Clustering reveals multiple phenotypic classes intermingled at multiple spatial scales with no evidence of recurrent patterns. In the GBM we have studied in detail, heterogeneity on a scale of 10–100 cell diameters is as great as it is between distinct lobes. The proportion of cells from different clusters also varies dramatically from one tumor lobe to the next. Although it is not yet possible to link t-CyCIF clusters and known histological subtypes, cell-to-cell heterogeneity on these spatial scales are likely to impact the interpretation of small biopsies (e.g. a core needle biopsy) of a large tumor sample; the data also emphasize the inherent limitation in examining only a small part of a large tumor specimen (e.g. to save time on image acquisition). At the same time it is important to note that cell-to-cell heterogeneity is caused by processes operating on a variety of time scales, only some of which are likely to be relevant to therapeutic response and disease progression. For example, some cell-to-cell differences visible in GBM images arise from a cyclic process, such as cell cycle progression, whereas others appear to involve differences in cell lineage or clonality. Methods to correct for the effects of variation in cell cycle state have been worked out for single-cell RNA-sequencing48, but will require further work in imaging space.
In a third application of t-CyCIF we characterized tumor-immune cell interactions in a renal cell tumor. Immune checkpoint inhibitors elicit durable responses in a portion of patients with diverse types of cancer, but identifying potential responders and non-responders remains a challenge. In those cancers in which it has been studied38, quantification of single checkpoint receptors or ligands by IHC lacks sufficient positive and negative predictive value to stratify therapy or justify withholding checkpoint inhibitors in favor of small molecule therapy39. Multivariate predictors based on multiple markers such as CD3, CD4, CD8, PD-1 etc. appear to be more effective, but still underperform in patient stratification10 probably because cells other than CD8+ lymphocytes affect therapeutic responsiveness. In this paper, we perform a simple analysis to show that tumor infiltrating lymphocytes can be subtyped by t-CyCIF and analyzed for the proximity of PD-1 and PD-L1 at a single cell level. Next steps involve thorough interrogation of immuno-phenotypes by multiplex imaging to relate staining patterns in images to immune cell classes previously defined by flow cytometry and to identify immune cell states that fall below the limit of detection for existing analytical methods.
In conclusion, t-CyCIF is a robust, easy to implement approach to multi-parametric tissue imaging applicable to many types of tumors and tissues; it allows investigators to mix and match antibodies depending on the requirements of a specific type of sample. To create a widely available community resource, we have posted antibody lists, protocols and example data at http://www.cycif.org and intend to update this information on a regular basis. Highly multiplexed histology is still in an early stage of development and better methods for segmenting cells, quantifying fluorescence intensities and analyzing the resulting data are still required. For example, quantification of cell-to-cell heterogeneity may enable reconstruction of signaling network topologies in situ13,49 by exploiting the fact that protein abundance and states of activity fluctuate from one cell to the next; when fluctuations are well correlated, they are likely to reflect causal associations50. We expect t-CyCIF to be complementary to, and used in parallel with other protein and RNA imaging methods such as FISSEQ51 or DEI52 that may have higher sensitivity or greater channel capacity. A particularly important task will be cross-referencing tumor cell types identified by single-cell genomics or multi-color flow cytometry with those identified by multiplexed imaging, making it possible to precisely define the genetic geography of human cancer and infiltrating immune cells.
Competing Financial Interests
PKS is a member of the Scientific Advisory Board of RareCyte Inc., which manufactures the CyteFinder slide scanner used in this study; research with RareCyte is funded by NIH grant R41 CA224503 (PI E. Kaldjian). PKS is also co-founder of Glencoe Software, which contributes to and supports the open-source OME/OMERO image informatics software used in this paper. Other authors have no competing financial interests to disclose.
Acknowledgements.
This work was funded by NIG grants P50-GM107618 (PKS), U54-HL127365 (PKS), and R41-CA224503 (PKS) and by a DF/HCC GI SPORE Developmental Research Project Award (BI) and DFCI Claudia Adams Barr Program for Innovative Cancer Research Award (BI). BI was also supported by grant K08CA222663. We thank J. Waters and T. Lambert from the Harvard Cell Biology Microscopy Facility for access to the OMX Blaze, their guidance on SIM acquisition and analysis, and L. Shao for CUDA-accelerated SIM reconstruction code., B. Wolpin and C. Lian for providing specimens, Z. Maliga and J. Muhlich for technical support and L. Garraway, and members of Ludwig Center for Cancer Research at Harvard for many fruitful discussions.