A multiplex implantable microdevice assay identifies synergistic combinations of cancer immunotherapies and conventional drugs

A multiplex implantable microdevice assay identifies synergistic combinations of cancer immunotherapies and conventional drugs


Murine models

Mice were purchased from Jackson Laboratory. All animal studies were conducted in accordance with protocols approved by the Institutional Animal Care and Use Committee at Oregon Health & Science University (protocol no. IP00000956). All mice were bred and housed under specific pathogen-free conditions under a standard 12-hour light/dark cycle. C57LB/6, BALB/c and FVB/N mice were purchased from Jackson Laboratory. MMTV-PyMT mice were from Lisa Coussens and purchased from Jackson Laboratory. Virgin female mice 8–24 weeks of age were used for all experiments.

Cell lines

EMT6 (mouse breast cancer) cells were purchased from the American Type Culture Collection and were maintained in Waymouth’s medium with 10% FBS and 2 mM L-glutamine. E0771 (mouse breast cancer) cells were purchased from CH3 BioSystems and were cultured in RPMI-1640 with 10% FBS and 10 mM HEPES. Both cell lines were pathogen tested and were grown at 5% CO2 and 37 °C.

Experimental design

The objective of the studies in the figures is to show how intact TME responds to local stimulus of drug release and to test whether this response was significantly different from the baseline TME state in tumor regions distant from the drug site. The number of independent biological replicates of each experiment (n) performed is given in the figure legend. Spatial systems analyses were designed to quantitatively define directional spatial cell dependencies and cause consequence cell association with distance from the reservoir. These ultimately translated to models of drug response. Within these models, we aimed to identify therapeutic vulnerabilities to predict rational immune or TME-modulating treatment combinations and their optimal schedule and sequencing, which we then validated in traditional whole animal studies.

Microdevice implantation studies and sample collection

Nanodose drug delivery devices were manufactured and implanted as described previously in ref. 24. In brief, cylindrical microdevices 5.5 mm in length and 750 μm in diameter were manufactured from medical-grade Delrin acetyl resin blocks (DuPont) by micromachining (CNC Micro Machining Center) with 18 reservoirs of 200 μm (diameter) × 250 μm (depth) on the outer surface. Reservoirs were packed with drugs mixed with PEG (MW 1450, Polysciences) polymer at the concentrations indicated in Supplementary Table 1. Recommended systemic dose in patients with cancer was derived from the web page to June 2017. Systemic doses ranging among 0–1 mg kg−1, 1–2 mg kg−1, 2–4 mg kg−1 and >4 mg kg−1 translate to 20%, 25%, 30% and 40% of drug concentration in PEG, respectively, when released from the nanowell. The calibration was determined previously using mass spectrometry measurements24. Pure PEG was used in control conditions. Implanting multiple devices per tumor and/or multifocal animal models can increase the throughput up to 50–70 times as compared to conventional systemic treatment studies. Microdevices were implanted for 3 days in MMTV-PyMT with late-stage spontaneously growing tumors in all experiments. Tumor size was between 1.2 cm and 1.5 cm in the longest dimension at the time of implant. Tumors were excised at 3 days after device implantation unless otherwise stated, fixed for 48 hours in 10% formalin or 4% paraformaldehyde and then perfused with paraffin. Specimens were sectioned using a standard microtome, and 5-μm tissue sections were collected from each reservoir. Dry FFPE tissues were baked in a 65 °C oven for 30 minutes. After de-paraffinization with xylene and rehydration in serially graded alcohol to distilled water, slides were subjected to endogenous peroxidase blocking in fresh 3% H2O2 for 10 minutes at room temperature. Sections were then stained by mIHC and/or cycIF (Extended Data Fig. 1b,c).


Before iterative cycles of (1) staining, (2) whole slide scanning and (3) fluorophore bleaching, the slides were subjected to heat-mediated antigen retrieval by being immersed in citrate buffer (pH 5.5, HK0809K, BioGenex Laboratories, Citra Plus Antigen Retrieval) for 25 minutes and then briefly rinsed in a hot bath and then immersed in Tris/EDTA buffer (pH 9.0, S2368, Dako Target Retrieval Solution) for 15 minutes, all using a Cuisinart Electric Pressure Cooker (CPC-600N1). Protein blocking was performed for 30 minutes at room temperature with 10% normal goat serum (S-1000, Vector Laboratories) and 1% BSA (BP1600-100) in 1×PBS. (1) Slides were incubated with primary antibody (concentrations defined in Supplementary Table 3) for 2 hours at room temperature while being protected from light in a dark humid chamber. All washing steps were performed for 3 × 2–5 min in 1×PBS while agitating. Slides were mounted with SlowFade Gold antifade mountant with DAPI (S36938) using a Corning Cover Glass (2980-245). (2) Images were acquired using Zeiss Axio Scan.Z1 Digital Slide Scanner (Carl Zeiss Microscopy) at ×20 magnification, after which the coverslips were gently removed in 1×PBS while agitating. (3) Fluorophores were chemically inactivated using 3% H2O2 and 20 mM NaOH in 1×PBS for 30 minutes at room temperature while being continuously illuminated. The fluorophore inactivation was repeated twice with a short, 10-minute, 1×PBS wash in between. Efficacy of bleaching was imaged before antibody incubation (baseline autofluorescence) and every third to fourth cycle on average. After protein blocking, samples were subjected to the next round of staining. Single-cell feature extraction was not applied to evaluate sections stained by cycIF.


Before iterative cycles of (1) staining, (2) whole slide scanning and (3) heat and chemical stripping of antibodies and chromogen, the slides were subjected to staining with F4/80 and CSF1R antibodies (cycle zero, no antigen retrieval; Supplementary Table 2) and hematoxylin staining (S3301, Dako) for 1–5 minutes, followed by whole slide scanning. Slides were then subjected to the first heat-mediated antigen retrieval in 1× pH 5.5–6 citrate buffer (BioGenex Laboratories, HK0809K) for 90 seconds in a low-power microwave and 16 minutes in a steamer, followed by protein blocking with 10% normal goat serum (S-1000, Vector Laboratories) and 1% BSA (BP1600-100) in 1×PBS for 30 minutes at room temperature. (1) Slides were incubated with primary antibodies (concentrations defined in Supplementary Table 2) for 1 hour at room temperature or 16–17 hours at 4 °C while being protected from light in a dark humid chamber. Signal was visualized with either anti-rabbit or anti-rat Histofine Simple Stain MAX PO horseradish peroxidase (HRP)-conjugated polymer (Nichirei Biosciences), followed by peroxidase detection with 3-amino-9-ethylcarbazole (AEC). Two or three drops of HRP polymer were used for up to nickel-size or whole slide tissue sample, respectively. Timing of AEC development was determined by visual inspection of positive control tissue (Extended Data Fig. 1d–f) for each antibody. All washing steps were performed for 3 × 5–10 minutes in 1×PBS while agitating. Slides were mounted with a filtered 1×PBS with 0.075% Tween 20 (BP337100) using a Signature Series Cover Glass (Thermo Fisher Scientific, 12460S). (2) Images were acquired using the Aperio ImageScope AT (Leica Biosystems) at ×20 magnification, after which the coverslips were gently removed in 1×PBS while agitating. (3) Within one cycle, removal of AEC and HRP inactivation was accomplished by incubating the slides in 0.6% fresh H2O2 in methanol for 15 minutes. AEC removal and stripping of antibodies was accomplished by ethanol gradient incubation and heat-mediated antigen retrieval such as described above between cycles. After washing and protein blocking, samples were subjected to the next round of staining.

The readout antibody panel was carefully designed so that it broadly captures all major TME subtypes and allows to find synergy with the most established and/or emerging immunotherapies (Supplementary Table 4). Based on this, we defined a minimal essential set of 13 markers that classifies distinct myeloid and lymphoid lineages as well as components of non-immune stroma (for non-immune TME modulation). Staining the baseline discovery set of 13 markers can be completed in 4–7 days considering that 3–4 markers and two markers are currently detected in one cycle (1 day) in the mIHC and cycIF procedures, respectively. Before that, an additional 3 days are required for sample fixation, paraffin embedding and FFPE block cutting, resulting in total turnaround time of 7–10 days from sample collection to data acquisition and interpretation. However, the method is flexible such that markers can be subtracted or added to allow for deeper cell characterization of identified phenotypes based on investigator interest. We also envision that, by accommodating an increased number of markers per cycle (for example, by using spectral deconvolution techniques), we can further reduce the turnaround times.

The cost of the MIMA workflow has two major components: one, the cost of the drug-loaded microdevices, which is ~$600–800 per device for a typical study, depending on the number and cost of individual drugs loaded into the device reservoirs; and two, the cost of the cycIF/mIHC, which is ~$50 per slide per cycle with basic (single-stain) immunohistochemistry infrastructure in place. It should be noted, however, that up to six tumor/device specimens are embedded in a single paraffin block so as to reduce the total number of slides required.

Image processing and feature extraction of mIHC images

The iteratively digitized images were co-registered using MATLAB (MathWorks, version 2019b) using the detectSURFFeatures algorithm from the Computer Vision Toolbox. The imperfectly registered images were additionally processed using the Linear Stack Alignment with SIFT plugin (Fiji) so that cell features overlap down to a single-pixel level. Hematoxylin-stained images were color deconvoluted for single-cell nuclear segmentation to generate a binary mask using the watershed function and standard image processing steps (noise removal, erosion and dilation; Fiji)72. AEC chromogenic signal was extracted using the NIH plugin RGB_to_CMYK to separate AEC signal into the yellow channel for improved sensitivity of immunohistochemistry evaluation73,74. Grayscale images of all proteins and the binary mask were imported to CellProfiler (version 3.1.8, Broad Institute)75 to quantify single-cell signal mean intensity as defined by mask, which was scaled to a range of 0–1. The IdentifyPrimaryObjects module was used to identify nuclei from mask; the MeasureObjectIntensity module measured mean intensity for each object for each protein. The mean signal intensity per cell output was imported to FCS Express 6 and 7 Image Cytometry Software (De Novo Software) to perform multidimensionality reduction to classify ‘standard cell types’. Gating strategies and hierarchical cell classification are presented in Fig. 1e and Extended Data Fig. 2e. Polygonal gates moving around the central vertex without changing the polygon shapes were used to obtain quantitatively reproducible multiplex data, batch to batch, independent of the condition measured. Positive control tissues were used to help define the single-parameter threshold for positivity by manual gating. A total of 3,000–5,000 cells were analyzed for feature extraction in the assay area located above the drug-releasing site with ±300 total cells for paired, experimental versus control, region. Minimum population proportion within 5% margin of error and 95% confidence level was set to 0.75% (represents 12 cells) to discriminate noise from specific cell enrichment induced by, for example, increased protein expression or cell recruitment into the assay region. Experimental condition of the assay area was compared to random control intratumoral region located perpendicular and/or far from the drug-releasing reservoir. To obtain greater control over confounding variables, paired sample one-tailed t-tests were used to determine enrichment of induced TME states. Percentage of positivity and significance were presented in form of a heat map or bar graphs. Quality of the single-cell data was ensured by excluding deformed (folded), lost or unevenly stained tissue (border effects). The assay area was determined by the first 3,000–5,000 cells above the well excluding these deformed regions. Single-cell data from FCS Express were extracted in a data grid to MATLAB for downstream spatial systems analyses. In computed images, neutrophils are presented independent of the Epcam+/− status.

Spatial systems analyses

The distance-based cluster function finds clusters in a set of spatial points expressed in xy space (adapted and modified from Yann Marcon; MATLAB October 2019). The clustering is based on Euclidean distance between the points (cells). The function does not require the number of clusters to be known beforehand. Each cell clusters with the closest neighboring cell if the distance between the two cells is shorter than the defined threshold. The minimal number of cells per cluster is defined by the user. The function outputs non-clustering cells in gray color, and each cluster meeting the defined parameters (minimal number of cells within maximum distance range) is presented in randomized colors. Clusters within the maximum defined distance merge and share one color. Number of clusters and total coverage in the assay area were calculated using distinct cluster sizes (defined by minimal number of cells within maximum distance range) for control PEG and palbociclib, which identified that cells cluster in response to treatment if a minimum of ten cells are present within a maximum distance range of 30–75 μm (systematic comparison not shown in this study). Cluster parametrization using as few as five cells and distances as large as 100 μm resulted in treatment non-specific cluster formation in PEG negative control. Treatment-specific cluster formation with cluster definition of a minimum of ten cells within 50-μm distance was generalizable to all marker and standard cell types, which was confirmed in panobinostat condition by comparing assay area and distal region side by side in one field of view (Extended Data Fig. 6e). This treatment-specific cluster parametrization was applied in downstream analytics to identify hotspots/zones of interest (for example, proximal, border, distal, network adjacent, CD11c+ DC clusters) in an objective, biology-driven manner.

For the relative abundance profile plot, marker-positive cells and the standard cell types were extracted to xy coordinate space; signal was blurred using Gaussian blur filter; and relative abundance of positive cells was displayed with distance from the well in a profile plot as outlines in corresponding Extended Data figures. A moving average filter with 50-μm and 100-μm window size (movmean function, MATLAB) was additionally applied to smoothen the feature signal for palbociclib and panobinostat condition, respectively. Signal in the profile plots was not scaled.

Inside the hotspot, spatial (geographical) interactions between marker-positive cells were determined by proximity measurements in local microculture by using the pdist2 function in MATLAB (version 2019b), which returns the distance of each pair of observations (positive cells) in x and y using metric specified by Euclidean distance. Random circular regions of 175-μm diameter (defined by Extended Data Fig. 6f) were selected in the border, CSC-rich zone of the panobinostat assay area, and Euclidean distance was measured between Sox9+ and other marker-positive cells. The number of distances was presented in the form of a histogram. To quantify spatially interrelated phenomenon, proportions of distances lower than 50 μm (as defined by distance-based cluster analyses) were compared between different cell pairs (for example, Sox9+Ly6G+ versus Sox9+CD11c+).

Extended hierarchical cell classification was applied to characterize the significantly enriched cell phenotypes forming zones of interest that were outside the standard cell type classification (for example, less-differentiated macrophages or phagocytic DCs). Probe combination and number of cells analyzed within number of clusters are defined in the figures and figure legends.

Two-dimensional composite images were presented by using Fiji72.

The spatial systems analyses were used to identify drug models of response (presented as line diagrams), and the identified therapeutic vulnerabilities were tested in whole animal studies.

Whole animal treatment studies

Although the high-throughput IMD experiments were performed in the MMTV-PyMT model30,31,76,77 with spontaneously growing tumors, the whole animal validation studies of predicted immune-modulating combinations were performed using transplantable breast cancer cell lines in syngeneic mice to avoid extensive breeding and colony maintenance necessary to test synergy of multiple predicted combinations. E0771 and EMT6 models, which are typically used in breast cancer research involving immunotherapy testing32,78,79, were selected randomly for validation of different combinations. The combination of panobinostat and anti-PD-1 was tested in both transplantable models. The most potent triple combination of panobinostat, venetoclax and anti-CD40 was additionally tested in the MMTV-PyMT model with spontaneously growing tumors.

The MMTV-PyMT model has a 100% penetrance and shows good consistency in latency times and similar tumor characteristics76. The model was developed in 1992 in the Muller laboratory30, and, despite the PyMT not being a human oncogene, it mimics the signaling of RTKs, which are often activated in human malignancies, including breast cancer. PyMT expression under the MMTV promoter results in rapid transformation and generation of multifocal tumors that metastasize to lungs. Tumors arising in luminal cells progress through distinct histological stages that mimic human ductal breast cancer progression (hyperplasia, adenoma, MIN and early and late carcinoma)31. Loss of ER and PR expression is observed as the disease progresses31. By gene expression profiling, this model clusters with luminal B subtype32,80,81.

Transcriptionally, the orthotopic syngeneic models fall into luminal A (EMT6) and luminal B (E0771) intrinsic subtype despite being aggressive with poorly differentiated or spindle-shaped histopathology. Both models showed transcriptomic characteristics of ‘claudin-low’ human subtype with a high score for EMT, low differentiation and low proliferation82.

MMTV-PyMT transgenic mice that were 80 days of age were randomized and included in the study when their total tumor burden was 150–550 mm3 (treatment initiation). For the orthotopically induced tumor models of mammary carcinoma, EMT6 (0.5 × 106 in 1×PBS per site) and E0771 (0.5 × 106 in Corning matrigel per site) cells were injected into the #4 mammary fat pad of female virgin BALB/c and C57LB/6, respectively. One tumor was induced in the E0771 and two tumors were induced in the EMT6 model. Caliper measurements were used to calculate the tumor volumes using the formula length × width2 / 2. Treatments were initiated when total tumor burden was 60–150 mm3. For all models, the endpoint was determined by tumor volume above 2,000 mm3 in two consecutive measurements or one measurement above 2,200 mm3. Treatments were administered by intraperitoneal injection. Dose, schedule and duration are indicated in the respective figures and figure legends. We note that the doses for panobinostat and venetoclax were decreased from 15 mg kg−1 to 11.5 mg kg−1 and from 22 mg kg−1 to 18 mg kg−1, respectively, when the two drugs were combined (Fig. 6e). Treatment schedule was estimated depending on the location of the targetable cell phenotype in proximity to the well or more distal from the drug source. For example, cells in the immediate proximity of the drug well at 3 days of exposure were likely recruited first to the drug assay area; thus, early targeting (pre-treatment) of these cells is preferred. Inversely, cells located in distal regions should be targeted by post-treatment approach. Diluent and IgG2a isotype control (Bio X Cell) concentrations were equivalent to the highest dose of the respective drug used in each experiment. Mice that survived the first treatment cycle were allotted an 8-day break before the start of one additional treatment cycle with the same administration of drug doses and duration.

The mice were monitored daily to determine any possible effects on the general condition of the animals using parameters as established by Morton and Griffiths (1985). The guidelines for pain, discomfort and distress recognition were used to evaluate weight loss, appearance, spontaneous behavior, behavior in response to manipulation and vital signs. Specifically, general appearance (dehydration, missing anatomy, abnormal posture, swelling, tissue masses and prolapse), skin and fur appearance (discoloration, urine stain, pallor, redness, cyanosis, icterus, wound, sore, abscess, ulcer, alopecia and ruffled fur), eyes (exophthalmos, microphthalmia, ptosis, reddened eye, lacrimation, discharge and opacity), feces (discoloration, blood in the feces and softness/diarrhea) and locomotor (hyperactivity, coma, ataxia and circling) were monitored to determine loss of body condition (BC) score, namely: BC 1 (emaciated) score was applied when skeletal structure was extremely prominent with little or no flesh/muscle mass and vertebrae was distinctly segmented; BC 2 (under-conditioned) score was applied when segmentation of vertebrate column was evident, dorsal pelvic bones were readily palpable and muscle mass was reduced; and BC 3 (well-conditioned) was applied when vertebrae and dorsal pelvis were not prominent/visible and were palpable with slight pressure. Loss of BC was also considered when anorexia (lack or loss of appetite) or failure to drink; debilitating diarrhea and dehydration/reduced skin turgor; edema, sizable abdominal enlargement or ascites, progressive dermatitis, rough hair coat/unkempt appearance, hunched posture, lethargy, loss of righting reflex, neurological signs or bleeding from any orifice appeared in treated mice. Most treated groups were well-conditioned (BC score 3); less than 20% of mice in each group experienced mild diarrhea for up to 2 days once during the course of treatment (typically after the first or second therapy administration). Mice receiving palbociclib monotherapy were under-conditioned (BC score 2) starting from day three until the end of the treatment. Two out of eight mice in the MMTV-PyMT model died within 1–3 days after the first injection of αCD40 immunotherapy when administered as a single agent. Lethal toxicity of anti-CD40 used as a single agent was previously reported due to a shock-like syndrome58, and our data also strongly suggest that this immunotherapy is tolerable only with prior administration of anti-cancer agent(s). Surviving mice receiving venetoclax/anti-CD40 combination experienced fur graying to different degrees starting approximately 4 weeks after treatment. No signs of pain, discomfort or distress were observed in the surviving mice. Neither emaciated (BC score 1), over-conditioned (BC score 4) nor obese (BC score 5) were observed in our studies.

To show CD8+ T cell infiltration inside the tumor bed, ErbB2ΔEx16 mice83 with spontaneously growing late-stage tumors were intraperitoneally injected with panobinostat (15 mg/mg) on day zero, day two and day four. Tumors were extracted at day seven, were FFPE processed and were stained for CD8 to compare the rate of intratumoral CD8+ T cells in panobinostat-treated versus control (diluent)-treated mice.

Vaccination study

EMT6 and E0771 cells in tissue culture were treated with a soluble drug panobinostat at 5 μM concentration when they would reach 60–70% confluency. After 2 days, the cells were harvested and were injected subcutaneously (total 2–3 × 106 cells) into the lower left flank of BALB/c and C57Bl6 mice, respectively. Cells freeze–thawed three times served as negative control for non-immunogenic form of cell death. After 7–8 days, the mice were re-inoculated by injecting living cells orthotopically into one #4 mammary fat pad (total 0.5 × 106 cells), and tumor appearance was monitored by minimal tumor size approximately 5 mm and 3.5 mm in the longest dimension for E0771 and EMT6 models, respectively (palpable tumors). We note that the E0771 tumors after re-challenge appeared at the primary subcutaneous site, and no tumors were developed in the orthotopic site.

Statistics and reproducibility

All data are combined from 2–3 independent experiments, unless specifically noted. To accomplish randomization for systemic mouse experiments, animals were sorted by a blinded investigator, and then groups were assigned. Each group was checked post hoc to verify no statistical significance in average starting tumor size. There was no sample size estimation in standard drug treatment experiments. Data are shown as mean ± s.e.m., unless otherwise noted. For tumor growth rate, significance was calculated by unpaired two-tailed t-test with equal variance. For survival and tumor-free analyses, Kaplan–Meier curves were generated to demonstrate time to event, and the log-rank (Mantel–Cox) test was used to evaluate statistical significance. For representative micrographs, each experiment was repeated at least three times with similar results, unless stated otherwise.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Share this post

Similar Posts