Cancer-Associated Fibroblasts Promote Epithelial–Mesenchymal Transition and Classical to Basal Subtype Shift in Pancreatic Cancer

Multidimensional analyses highlight the diverse role of cancer-associated fibroblasts in influencing cells in the tumor microenvironment and provide a platform for evaluating emerging therapeutic approaches and studying mechanisms dictating tumor behavior.

Abstract

Pancreatic ductal adenocarcinoma (PDAC) carries an extremely poor prognosis, in part resulting from cellular heterogeneity that supports overall tumorigenicity. Cancer-associated fibroblasts (CAF) are key determinants of PDAC biology and response to systemic therapy, and multiple CAF subtypes have been defined. However, defining the effects of patient-specific CAF heterogeneity and plasticity on tumor cell behavior is required to better characterize the role of CAFs in PDAC. In this study, we used multiomic analyses to characterize the tumor microenvironment (TME) in tumors from patients undergoing curative-intent surgery for PDAC. In these same patients, matched tumor organoid and CAF lines were established to functionally validate the impact of CAFs on the tumor cells. CAFs promoted epithelial–mesenchymal transition and a switch in tumor cell classification from classical to basal subtype. Furthermore, CAF-specific interleukin 8 functioned as a modulator of tumor cell subtype. Finally, neighborhood relationships between tumor cells and T cell subsets were defined, demonstrating a distinct spatial coordination among CAF and tumor cell subtypes. Overall, this study provides data supporting CAF signaling as a regulator of the cellular and behavioral heterogeneity in the PDAC TME. These findings can be used to explore rational approaches to improve therapies for this difficult-to-treat disease.

Significance:

Multidimensional analyses highlight the diverse role of cancer-associated fibroblasts in influencing cells in the tumor microenvironment and provide a platform for evaluating emerging therapeutic approaches and studying mechanisms dictating tumor behavior.

Graphical Abstract

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is a debilitating disease with a dismal 5-year survival rate of 13% (1). These poor outcomes are due to numerous factors, including late-stage diagnosis, resistance to current therapeutic regimens, and, importantly, a complex, heterogeneous tumor microenvironment (TME; refs. 2–4). The TME consists of a diverse and evolving ecosystem of epithelial, endothelial, immune cells, and cancer-associated fibroblasts (CAF; refs. 5–7). CAFs are present in abundance in the PDAC TME and have been implicated as drivers of an immunosuppressive milieu that provides a niche for tumor cells, enhances tumor growth, and ultimately supports metastasis (5, 8–13).

Molecular and cellular classification, including both CAF and tumor cell subtypes, may influence clinical decision-making with both current and emerging therapeutic approaches (14). Two main subtypes of tumor cells have been described: a classical subtype that retains epithelial markers and is often well-differentiated and a basal subtype that is quasi-mesenchymal. Patients with tumors predominantly comprising PDAC cells with a basal subtype have poorer responses to chemotherapy and worse overall survival (15). Additionally, CAF subtyping has associated unique gene expression programs with specific roles within the TME. The most common CAF classifications include inflammatory (iCAF), with the capacity to secrete IL6; myofibroblastic (myCAF) defined by high levels of alpha smooth muscle actin (αSMA) expression; and antigen-presenting (apCAF) defined by the expression of HLA-DR with antigen processing and presentation capacity (13, 14, 16–19). Although these distinctions underscore TME complexity in PDAC, the dynamic and nuanced interactions between CAFs and epithelial tumor cells remain incompletely understood.

To untangle these relationships, we generated multiomic (transcriptomics and spatial proteomics) data to examine how CAFs alter tumor biology through signaling mechanisms within the PDAC TME. We developed an approach to integrate multiplex imaging data from patient tissue slides with functional in vitro cocultures of autologous patient-derived organoids (PDO) and CAFs. This approach demonstrated a spatial relationship associating iCAFs with classical tumor cells and myCAFs with basal tumor cells. These data suggest that the CAF proximity to tumor cells can promote a classical to basal phenotype switch and drive epithelial–mesenchymal transition (EMT) signaling. This switch was associated with decreased CAF-derived IL8 secretion and impaired maintenance of a classical tumor cell phenotype among the tumor epithelial cells. Additionally, tumor regions with increased basal gene expression were associated with increased infiltration of activated T cells, suggesting that CAF–epithelial cross-talk serves to define the immune landscape in this disease. Thus, this study provides new data supporting CAF signaling as a regulator of the cellular and behavioral heterogeneity in the PDAC TME. These findings can be used to explore rational approaches to improve therapies for this difficult-to-treat disease.

Materials and Methods

Imaging mass cytometry

Resected pancreas slides were baked at 60°C for 2 hours, dewaxed in histologic-grade xylene, and then rehydrated in a descending alcohol gradient. Slides were incubated in Antigen Retrieval Agent pH 9 (Agilent, S2367) at 96°C for 1 hour and blocked with 3% BSA in Maxpar PBS at room temperature for 45 minutes. Immunohistochemical staining was done using individually conjugated mass cytometry antibodies. An antibody cocktail was prepared, detailed in Supplementary Table S1, and used to stain the slides at 4°C overnight. Custom antibodies were conjugated in-house, diluted to a concentration of 0.25 mg/mL to 0.5 mg/mL, and then titrated empirically. Immune, stromal, and architectural antibodies were verified for staining quality using control tonsil tissue (Supplementary Fig. 1). Cell-ID Intercalator-Ir (Standard BioTools, 201192A) was diluted at 1:400 in Maxpar PBS and used for DNA labeling (20). Ruthenium tetroxide 0.5% Aqueous Solution (Electron Microscopy Sciences, 20700-05) was diluted at 1:2,000 in Maxpar PBS and used as a counterstain. Images were acquired through the Hyperion Imaging System (Standard BioTools; RRID:SCR_023195) at the Johns Hopkins Mass Cytometry Facility.

Images were prepared for analysis similarly to prior descriptions (21). In brief, images were segmented using nuclear (Ir191 and Ir193) and plasma membrane staining (IMC Segmentation Kit, Standard BioTools, TIS-00001). Sixty images were evaluated to assign pixel classifications and establish probability maps using Ilastik (RRID:SCR_015246; ref. 22). CellProfiler (23, 24) version 4.2.4 (RRID:SCR_007358) was then used to generate segmentation masks for these images based on the resulting probability maps. The quality of segmentation was explored visually, and per-cell data were exported using histoCAT (RRID:SCR_026499; ref. 25). As stromal cells can vary in morphology and obscure segmentation quality, we validated the fidelity of segmentation output with an orthogonal segmentation method based on pixels instead of individual cells, Pixie (26). Clustering of individual cells was achieved using the relative expression of cell subtyping and functional markers using Phenograph (RRID:SCR_016919; ref. 21). The density of cell types was determined by dividing the number of cells detected per cluster by the area of tissue analyzed. Top neighbor analysis was performed by compiling the top three neighbors for each cell. Heatmaps were generated to display aggregated data and clearly label defined clusters. Representative images were prepared using MCD Viewer (Standard BioTools; RRID:SCR_023007), overlaying multiple stains and adjusting the threshold to minimize background. These were then exported as 16-bit images. Box plots were generated in R version 3.6.3 using ggplot2 (RRID:SCR_014601).

Patient sample acquisition and generation of matched organoid and CAF cell lines

Patients with PDAC undergoing surgical resection were enrolled in tissue acquisition protocols at Johns Hopkins Hospital (IRB: NA_00001584, 00138903; Table 1). This study was conducted in accordance with the U.S. Common Rule, approved by the Johns Hopkins Institutional Review Board (IRB), and written informed consent was obtained from all prospectively enrolled patients. PDOs were generated from patient surgical specimens following a combination of mechanical and enzymatic dissociation, as described previously (27, 28). Organoid lines were maintained in Matrigel (Corning, 356234) with Human Complete Feeding Media (HCPLT media), detailed in Supplementary Table S2. For organoid passaging, media was aspirated, and then Matrigel domes were resuspended in Cell Recovery Solution (Corning, 354253) and incubated on ice at 4°C for 45 minutes to allow Matrigel depolymerization. Cells were then pelleted and washed in human organoid wash media (Advanced DMEM/F-12, 10 mmol/L HEPES, 1× GlutaMAX, 100 μg/mL Primocin, 0.1% BSA) prior to pelleting again. Cell pellets were passaged at a ratio of 1:2 and replated in Matrigel in new 24-well plates and placed in the incubator for 10 minutes to allow Matrigel to harden. A volume of 500 μL of Human Complete Feeding Media was added on top of the Matrigel domes, and the plates were returned to the incubator for further expansion.

Table 1.

Patient demographics.

Patienta	Age	Sex	Time to recurrence (months)	KRAS NGS	Neoadjuvant treatment	Adjuvant treatment	Surgical procedureb	Pathologic stage
JHH317	72	F	20.2	G12D	Gemcitabine nab–paclitaxel and immunotherapy	Gemcitabine nab–paclitaxel and immunotherapy	DPS	T3N2
JHH348	84	F	9.9	G12R	Untreated	None	Whipple	T3N2
JHH352	57	F	14.7	G12V	FOLFIRINOX	FOLFIRINOX	Whipple	yT1N2
JHH357	62	M	22.9	G12V	FOLFIRINOX	Gemcitabine–capecitabine	Whipple	yT2N2
JHH361	63	M	20.7	G12D	FOLFIRINOX	Gemcitabine–capecitabine	DPS	yT2N0
JHH362	83	F	—	—	Untreated	—	DPS	T2N1
JHH368	79	F	23.3	G12R	Gemcitabine nab–paclitaxel	Gemcitabine	DPS	yT2N2
JHH369	57	M	4.7	G12D	FOLFIRINOX	Gemcitabine nab–paclitaxel	Whipple	yT2N1
JHH372	57	M	7	Q61H	FOLFIRINOX	Gemcitabine nab–paclitaxel	Whipple	yT2N0
JHH380	85	F	6.2	G12V	Untreated	None	DPS	T3N1
JHH383	57	F	12.2	G12R	Immunotherapy	FOLFIRINOX and gemcitabine nab–paclitaxel	Total pancreatectomy and splenectomy	yT3N2
JHH387	55	M	8.2	G12D	FOLFIRINOX	—	DPS	yT4N2
JHH388	52	M	5.9	G12D	FOLFIRINOX	Gemcitabine nab–paclitaxel	DPS	yT3N0
JHH390	54	M	—	G12D	FOLFIRINOX	FOLFIRINOX olaparibc	Whipple	T2N0
JHH417	72	F	9.8	G12V	Untreated	FOLFIRINOX	Whipple	T2N2

^aJHH348, JHH362, and JHH380 were included in pilot bulk RNA-seq; JHH388 and JHH417 were added to the IMC cohort and are not represented in the RNA-seq data; the JHH390 cocultured organoid sample did not meet quality parameters for analysis and thus is omitted from direct comparisons between monoculture and cocultured organoids.

^bDPS, distal pancreatectomy and splenectomy.

^cPatient has a germline BRCA2 mutation.

CAFs were extracted from surgical resection specimens after remnant tissue was washed twice with human organoid wash media and strained through a 70 μm cell strainer. Cells were plated into one well of a six-well plate and allowed to expand before further cell passaging and expansion. CAFs were expanded by trypsinization and expanded at a ratio of 1:2 in CAF media [RPMI (Fisher 11-875-085) + 10% FBS + 1% penicillin/streptomycin (Pen/Strep; Gibco, 5140122) + 1% L-glutamine (Gibco, 25030081) + 0.1% amphotericin B (Sigma, A2942)]. Both cell types were tested for Mycoplasma using Invivogen MycoStrips (rep-mys-100) upon establishment and at 6-month intervals thereafter.

To achieve adequate biomass, both cell types were passaged and expanded following individual cell type establishment and identity verification matching the patient of origin. Primary CAFs were not immortalized and were used for experiments at passage numbers below 10. PDOs were similarly expanded and used for experiments at passage numbers between 10 and 20.

Coculture setup and cell acquisition

Prior to coculture, PDOs and CAFs were lineage-verified using short tandem repeat evaluation completed at the Johns Hopkins Genetics Resource Core Facility (Supplementary Table S3). CAFs were grown in monolayer and characterized with flow cytometry using common CAF markers, including VIM, PDPN, PDGFRα, FAP, and αSMA, to demonstrate preserved interpatient heterogeneity (Supplementary Fig. S2A–S2C). Epithelial PDOs were established and expanded from primary PDAC tumors, as described previously (27). PDOs were confirmed to be viable and EpCAM-positive (Supplementary Fig. S2D–S2F). Further CAF characterization was performed, in line with prior studies, as cell type markers can display high levels of variability (10, 17, 29, 30). To better resolve CAF heterogeneity and confirm marker specificity, we performed qPCR profiling of seven CAF and five tumor markers across eight CAF lines and six PDOs (Supplementary Fig. S2G). This allowed us to better define the CAF subtypes prior to inclusion in a three-dimensional organotypic coculture system.

We further optimized our coculture method by taking a reductionist approach and considering the effects of media composition on resulting cell counts (Supplementary Fig. S3A–S3C; Supplementary Table S2) and Matrigel dissociation methods on cell yield (Supplementary Fig. S3D) prior to proceeding with further coculture studies. We also compared the effects of different PDO-CAF ratios on CAF expression of αSMA (Supplementary Fig. S3E and S3F); PDOs and CAFs were combined at ratios of 1:1, 1:2, 1:3, 1:5, and 1:10 organoids to CAFs. A ratio of 1:3 PDO/CAFs was selected for experiments to match the expected overall CAF abundance and proportion of αSMA-positive CAFs present within patient tumors as assessed by histopathology. Cocultured cells were resuspended in titrated concentrations of Matrigel, from 100% to 25% Matrigel (Corning, 356234), to assess the effect on the cell interactions in the model TME. To generate lower Matrigel content conditions, Matrigel was supplemented with 5% CAF media (RPMI + 5% FBS + Pen/Strep, 0.1% amphotericin B) and plated in triplicate in 24-well tissue culture dishes. Matrigel domes were allowed to harden at 37°C for 1 hour before adding 500 μL of CAF media containing 5% FBS to overlay the dome. To extract cells from coculture, the supernatant was aspirated, and each dome was digested with either 1 mg/mL dispase solution or 2 mg/mL collagenase IV for 45 minutes at 37°C. Wash media was added to each well to quench the digest, and each well was transferred to a 96-well deep well plate to collect. The plate was centrifuged for 5 minutes at 1,500 RPM, the supernatant was aspirated, and the cells were resuspended in wash media. Centrifugation and supernatant aspiration were then repeated. Cell pellets were resuspended in TrypLE Express (Thermo Fisher Scientific, 12604013) following the manufacturer’s instructions to dissociate organoids into a single-cell suspension for use in downstream assays. As CAFs were initially the more challenging cell type to extract from Matrigel, we verified CAF viability following the dispase extraction method prior to setting up subsequent experiments (Supplementary Fig. S3G). Cocultures were not passaged after setup.

RNA sequencing

Sample preparation and alignment

RNA sequencing (RNA-seq) sample preparation, library construction, quality control, sequencing, and alignment were done as previously described in detail by Guinn and colleagues (31). Briefly, cocultured cells were isolated, sorted by FACS, and immediately underwent RNA extraction (QIAGEN RNeasy Kit, 74004). Quality was assessed by Nanodrop1000 (Thermo Fisher Scientific; RRID:SCR_016517) and via an external Novogene assessment using an Agilent 2100 Bioanalyzer (Agilent; RRID:SCR_018043). mRNA was purified using a magnetic bead approach, and cDNA was constructed using random hexamer primers. Libraries were checked with Qubit (RRID:SCR_020553) and pooled for sequencing on Illumina platforms to a depth of 40 million reads per sample. The cocultured organoid sample for JHH 390 did not meet quality metrics upon sequencing, so it was withheld from analysis. All downstream analyses that include a monoculture to coculture comparison exclude JHH 390. Alignment was performed utilizing Salmon version 1.9.0 (RRID:SCR_017036; ref. 32) on the Joint High-Performance Computing Exchange. “salmon_partial_sa_index__default.tgz” was used as the index for alignment for the HG38 genome, which was premade and available on refgenie (RRID:SCR_017574; ref. 33).

Differential gene expression and pathway analysis

We evaluated sample quality from the distribution of reads as visualized in a boxplot of log counts. We observed no samples with zero median expression, reflective of a low read count, so all samples were of good quality. We used principal component analysis (PCA) of the variance stabilization transformed RNA-seq data to evaluate sample clustering. Two samples were identified as not expressing canonical markers of their respective cell types, and as these were from the same patient, a renaming was completed to correct for this. Differential expression analysis was completed in multiple ways for this study using DESeq2 version 1.32.0 (RRID:SCR_015687; ref. 34): coculture organoids compared with monoculture organoids and coculture CAFs compared with monoculture CAFs.

Estimated fold changes were shrunk with apeglm (35) using lfcShrink to account for the variation in the samples in this dataset. Genes were statistically significant if the absolute log₂ fold changes after shrinkage were greater than 0.5 and the FDR-adjusted P values were below 0.05. Gene set statistics were run with fgsea (bioRxiv 10.1101/060012; RRID:SCR_020938) using MSigDb (36) version 7.4.1 (RRID:SCR_016863) pathways annotated in the HALLMARK, Kyoto Encyclopedia of Genes and Genomes, Reactome, ONCOGENIC, and GO databases (RRID:SCR_002811). Gene sets were considered significantly enriched with FDR-adjusted P values below 0.05. The results were visualized with ggplot2 (RRID:SCR_014601; ref. 37). Projection of the CoGAPS (RRID:SCR_001479; ref. 38) single-cell RNA-seq (scRNA-seq) pattern associated with the co-occurrence of EMT and inflammatory signaling in tumor-adjacent normal and neoplastic cells (CoGAPS pattern 7) was performed with ProjectR (39) as described previously (31).

CIBERSORTx

To better understand the cellular heterogeneity in the PDOs and CAFs, raw counts were uploaded to CIBERSORTx (RRID:SCR_016955; ref. 40) for further deconvolution of cell subtypes. Imputation of cell fractions was completed using a signature matrix generated from reference scRNA-seq data of PDAC tumors, which we previously collated and standardized from a range of public domain datasets (31). This allowed for the estimation of cell fractions based on the bulk gene expression using the default parameters and 1,000 permutations for statistical analysis. Results were plotted on a heatmap utilizing ComplexHeatmap (41) version 2.8.0 (RRID:SCR_017270) and comparison boxplots using ggplot2 (RRID:SCR_014601).

Pixel-based clustering validation

To validate accurate cell segmentation and clustering, we applied a pixel-level clustering algorithm with the Pixie pipeline (ark-analysis version 0.7.2) as described previously by the Angelo Lab (https://github.com/angelolab/ark-analysis, accessed June 2025; ref. 26). Briefly, single-page 32-bit tiff files were exported from MCD files using MCD Viewer (Standard BioTools; RRID:SCR_023007), which then underwent median filtering using the median_filter function from the scipy.ndimage module (version 1.15.2) in Python (version 3.10.16) to remove background signal. Next, filtered tiffs underwent pixel-level clustering using the Pixie pipeline from the ark-analysis toolkit, implemented in a Python (RRID:SCR_008394) environment (ark_env). Tiff images, in which each channel corresponds to a specific marker, were normalized across samples using the normalize_rows function. Per-pixel intensities were then clustered using self-organizing maps (SOM), followed by hierarchical consensus clustering to aggregate SOM clusters into meta-clusters, which were annotated into a final nine clusters representing major cell types present in the data. The total pixel area was quantified for each annotated cluster across all samples. To evaluate the relationship between Pixie-derived cluster abundance and cell-level cluster abundance obtained from PhenoGraph (RRID:SCR_016919), we performed a correlation analysis between the total pixel area per Pixie cluster and the corresponding cell counts per PhenoGraph (RRID:SCR_016919) cluster, which were annotated to represent the same major cell types as annotated in Pixie. A Pearson correlation test was conducted using the stats package in R (version 4.4.3; RRID:SCR_001905) to assess the linear relationship between the two clustering approaches.

Quantitative PCR

To evaluate gene expression, total RNA was extracted from PDO and CAF lines using the RNeasy Mini Kit (QIAGEN, 74104) according to the manufacturer’s specifications. cDNA synthesis was performed using TaqMan Reverse Transcription Reagents (Invitrogen, N8080234), following the manufacturer’s instructions. Real-time quantitative PCR was completed using the Thermo Fisher Scientific TaqMan Gene Expression Assays according to the manufacturer’s protocol in the QuantStudio 6 Flex System (Applied Biosystems; RRID:SCR_020239). mRNA targets are listed in Supplementary Table S4. Relative gene expression was quantified using the 2^−ΔΔCt method as previously described (42), and GAPDH was used as the endogenous control. Data were analyzed using Applied Biosystems QuantStudio Real Time PCR System Software (version 1.7.1).

Flow staining and cell sorting

Organoids were extracted using Cell Recovery Solution (Corning, 354253) and incubated on ice at 4°C for 45 minutes for Matrigel depolymerization. Cells were then pelleted and washed in human organoid wash media. Organoids were dissociated into single cells using TrypLE Express and washed in MACS buffer (PBS + 5 mmol/L EDTA + 1% FBS). Cells were then resuspended in PBS + Zombie NIR (BioLegend 423106; dilution 1:1,000) + Human TruStain FcX (BioLegend, 422302; RRID:AB_2818986; dilution 1:100) for 10 minutes at room temperature in the dark. Cells were then quenched with MACS buffer and spun down. Cells were resuspended in surface stain for 20 minutes on ice at 4°C. Cells were washed twice in MACS buffer. Flow cytometry analysis was performed on the Beckman Coulter CytoFLEX (RRID:SCR_019627) and analyzed via FlowJo version 10.8.0 (RRID:SCR_008520).

For coculture cell sorting, 1 mL of 1 mg/mL Dispase II (Thermo Fisher Scientific, 17105041) in wash media was added to each coculture and monoculture dome to depolymerize Matrigel for 1 hour at 37°C. The digest was quenched with 1 mL of wash media, and cells were spun down. Cells were resuspended in PBS + Zombie NIR (BioLegend, 423106; dilution 1:1,000) + Human TruStain FcX (BioLegend, 422302; RRID:AB_2818986; dilution 1:100) for 10 minutes at room temperature in the dark. Cells were quenched with MACS buffer, centrifuged, and then resuspended in surface stain for 20 minutes on ice at 4°C in the dark: PE FAP (R&D Systems, FAB3715P; dilution 1:75) and APC EpCAM (BioLegend, 324208; RRID:AB_756082; dilution 1:200). Cells were washed twice in MACS buffer. Flow cytometry was performed on the Beckman Coulter CytoFLEX (RRID:SCR_019627) and analyzed by FlowJo version 10.8.0 (RRID:SCR_008520).

For EMT, organoids were extracted using Cell Recovery Solution and incubated on ice at 4°C for 45 minutes for Matrigel depolymerization. Cells were then pelleted and washed in human organoid wash media. PDOs were taken down to single cells prior to plating; therefore, no TrypLE was used in the takedown process to maximize cell recovery. Cells were then resuspended in PBS + Zombie NIR (BioLegend, 423106; dilution 1:1,000) + Human TruStain FcX (BioLegend, 422302; RRID:AB_2818986; dilution 1:100) for 10 minutes at room temperature in the dark. Cells were then quenched with MACS buffer and spun down. Cells were resuspended in surface stain for 20 minutes on ice at 4°C: PE-Cy7 CD325 (N-Cadherin; BioLegend, 350812; RRID:AB_2562684; dilution 1:100), CoraLite Plus 647 Cytokeratin 17 (Thermo Fisher Scientific, CL647-17516; RRID:AB_2934927; dilution 1:100), Alexa Fluor 488 PanCK (BioLegend, 914214; RRID:AB_3675198; dilution 1:100), and APC vimentin (Invitrogen, MA5-28601; dilution 1:100). Cells were washed twice in MACS buffer. Flow cytometry analysis was performed on the Cytek Aurora (RRID:SCR_019826) and analyzed via FlowJo version 10.8.0 (RRID:SCR_008520).

Automated immunoblotting

As previously described (43), immunoblotting for GATA6, TFF1, VIM, and ECAD was carried out using the Wes Simple Western system (ProteinSimple, Bio-Techne), a capillary-based platform for automated size separation and protein detection. The 12 to 230 kDa separation module (SM-W004) was used according to the manufacturer’s instructions. Lysates prepared from PDO cultures were diluted using 0.1× sample buffer. A 5× fluorescent master mix was then added to the diluted lysates in a 1:4 ratio, resulting in a final protein concentration of 0.33 mg/mL. Samples were denatured at 95°C for 5 minutes prior to loading. Primary antibodies were used at the following dilutions: anti–GATA6 (Cell Signaling Technology, 5851S; RRID:AB_10705521) at 1:25 and anti-TFF1 (Cell Signaling Technology, 15571S; RRID:AB_2798747) at 1:25. A universal anti-rabbit secondary antibody (ProteinSimple, Anti-Rabbit Detection Module, DM-001) was used for detection. Alongside samples, the subsequent required reagents were loaded into the assay microplate and centrifuged at 2,500 RPM for 5 minutes: the biotinylated molecular weight ladder, antibody diluents, streptavidin–horseradish peroxidase (HRP), and chemiluminescent substrate (luminol–peroxide mix). The prepared microplate and capillary cartridge were then inserted into the Wes instrument, in which protein separation, capture, and detection occurred within individual capillaries. Separated proteins were immobilized using proprietary photoactivated chemistry and visualized via chemiluminescence. Quantitative analysis was performed using Compass for Simple Western software version 6.1.0 (RRID:SCR_022930), which automatically calculated peak height, peak area, and signal-to-noise ratios. To normalize protein levels, the total protein detection module (ProteinSimple, DM-TP01) was used as per the manufacturer’s guidelines. Protein expression was quantified by calculating the ratio of each target protein’s area under the curve to the total protein signal.

Slide stains

CD3/CD68 dual immunohistochemistry staining

Immunostaining was performed at the Sidney Kimmel Comprehensive Cancer Center Oncology Tissue Services Core. Dual chromogenic immune-labeling for CD3 and CD68 was performed on formalin‐fixed, paraffin-embedded (FFPE) sections on a Ventana Discovery Ultra autostainer (Roche Diagnostics). Briefly, following dewaxing and rehydration, epitope retrieval was performed using Ventana Ultra CC1 buffer (6414575001, Roche Diagnostics) at 96°C for 64 minutes. The primary antibody, anti‐CD3 (Abcam, ab16669; RRID:AB_443425; dilution 1:200), was applied at 36°C for 60 minutes and detected using an anti-rabbit HQ detection system (Roche Diagnostics, 7017936001 and 7017812001), followed by the ChromoMap DAB immunohistochemistry (IHC) detection kit (Roche Diagnostics, 5266645001). Primary antibodies were stripped at 95°C for 12 minutes, and residual HRP was neutralized (Roche Diagnostics, 7017944001). Next, primary antibody CD68 (Dako, M0814; RRID:AB_2314148; dilution 1:5,000) was applied at 36°C for 60 minutes and detected using an anti-mouse HQ detection system (Roche Diagnostics, 7017936001 and 7017782001), followed by the Discovery Teal Detection kit (Roche Diagnostics, 8254338001), counterstaining with hematoxylin, bluing, dehydration, and mounting. Slides were imaged using a Hamamatsu NanoZoomer (RRID:SCR_023762) digital slide scanner at 20× magnification, using the associated NDP.toolkit slide processing software. Slides were quantified by tiling the tumor area in a grid, with individual regions set to 1 μm² to account for the size of imaging mass cytometry regions of interest (ROI) and analyzed for total cell count utilizing HALO Multiplex IHC version 3.4.9 software (Indica Laboratories; RRID:SCR_018350).

Single-target IHC staining

Paraffin-embedded patient tumors were sectioned at 5 μm thickness, and tissue was rehydrated through a series of xylene and alcohol washes, followed by a rinse in water and a PBS wash. Slides were placed in an antigen retrieval buffer (Vector Labs, 3300) and steamed for 20 minutes. Slides were then blocked in a peroxide blocking buffer (Abcam, ab64218) for 15 minutes, followed by a protein block (Abcam, ab64226) for 5 minutes. They were then incubated with anti-vimentin primary antibody (Cell Signaling Technology, 5741S; RRID:AB_10695459; dilution 1:200), which was prepared in antibody diluent (Dako, S0809). Slides were placed in a humidified chamber at 4°C overnight. The next day, slides were washed with PBS and incubated in biotinylated anti-rabbit antibody (Abcam, ab64256; RRID:AB_2661852), followed by streptavidin–HRP solution (Abcam, ab64269) at room temperature for 20 minutes. Samples were then washed with PBS and incubated with 3-amino-9-ethylcarbazole chromogen (Vector Laboratories, SK-4200). Slides were then washed with water and incubated in Mayer’s hematoxylin (Sigma, MHS1) for 1 minute, rinsed with water, and mounted in VECTASHIELD PLUS (Vector Laboratories, H-1900). Slides were imaged using a Hamamatsu NanoZoomer (RRID:SCR_023762) digital slide scanner at 20× magnification, using the associated NDP.toolkit slide processing software.

Slides were quantified by tiling the tumor area in a grid with individual regions set to 1 μm² to account for the size of imaging mass cytometry ROIs and analyzed for total cell count utilizing HALO Multiplex IHC version 3.4.9 software (Indica Laboratories; RRID:SCR_018350).

Trichrome staining

Staining was performed by the Johns Hopkins Reference Histology Core. Analysis was performed using the colour_deconvolution2 plugin for ImageJ (RRID:SCR_003070). Images were threshold adjusted on the blue image to cover all regions of collagen that were seen and analyzed through ImageJ (RRID:SCR_003070). Defined regions of tumor-rich tissue were quantified and analyzed in GraphPad Prism (RRID:SCR_002798).

Hematoxylin and eosin staining

Hematoxylin and eosin (H&E) staining was performed by the Johns Hopkins Reference Histology Core.

Immunofluorescence

CAF cells were grown to 80% confluence in 96-well flat-bottom tissue culture-treated plates. When the desired confluence was reached, media was aspirated, and cells were washed twice with PBS. PBS was aspirated, and cells were fixed using 10% formalin for 10 minutes. Cells were washed twice more with PBS after fixation was complete. Surface markers were blocked using 5% BSA in PBS for 1 hour at room temperature. Cells were then permeabilized with 0.1% Tween for 10 minutes. Primary antibody (VIM, BioLegend, 677809; RRID:AB_2650955) was added, and cells were incubated at 4°C overnight. After incubation, cells were washed twice with PBS and counterstained with DAPI for 30 seconds. DAPI was pipetted off, and a small amount of PBS was added to keep the cells wet. Cells were imaged using an inverted microscope with a filter for GFP.

Wound healing assay

Pancreatic cancer cells (PANC10.05, ATCC, CRL-2547; RRID:CVCL_1639) were plated at a density of 5 × 10⁴ cells/well in a 24-well plate and cultured in the indicated media for 48 hours. A 200 μL pipette tip was used to introduce a wound into the cell monolayer. Phase-contrast images were taken after the media was replaced at the time of wound creation and at the indicated time points. Images were analyzed with Fiji ImageJ analysis software (RRID:SCR_002285) to define the area of the wound at time = 0 and again at the end of the experiment. The fraction of wound closure was calculated by dividing the final wound area by the initial wound area and subtracting the resulting value from 1.

Proteome profiler

Two percent low serum conditioned media was collected after a 3-day incubation with CAFs that were previously determined via bulk RNA-seq to either cause a classical to basal shift in PDOs upon coculture (plastic) or maintain the original classical/basal phenotype (n = 2 per condition). Media was filtered through a 0.45 μm filter to remove cell debris and frozen at −80°C. Frozen conditioned media was then thawed and used for cytokine dot blots (R&D Systems, ARY022B) according to the manufacturer’s instructions. Dot blots were analyzed using Ideal Eyes Systems QuickSpots densitometry software as per the manufacturer’s instructions. Briefly, the chemiluminescence image of the spot arrays was uploaded to QuickSpots software. The appropriate manufacturer and array template were selected and positioned on each spot membrane to generate results. Normalized Z scores were tabulated and analyzed with GraphPad Prism version 9.5.0 (RRID:SCR_002798). A heatmap rank ordered by plastic CAF line Z scores was generated in R version 4.4.0 using ComplexHeatmap (RRID:SCR_017270).

ISH RNAscope

The FFPE-embedded pancreas tumor tissue sections (5 μm thick) were cut onto SuperFrost Plus slides. Staining was completed with an RNAscope Multiplex Fluorescent Reagent kit version 2 assay (Advanced Cell Diagnostics, Bio-Techne, 323135). Briefly, target RNAs were hybridized to ssDNA “z-probes” complementary to the RNA of interest. Oligos were bound to the tail region of the z-probe, which was then bound to amplifiers labeled with HRP and fluorophores. Probes generated by Advanced Cell Diagnostics (520721-c1, KRT6A; 463661-c2, KRT17; 603131-c3, GATA6; 417281-c4, TFF1) were uniquely amplified with Opal dyes (Akoya Biosciences, FP1487001KT, FP1488001KT, FP1497001KT, FP14695001KT) and then counterstained with DAPI (1 μg/mL). Slides were mounted using Pro-Long Gold mounting medium (Invitrogen, p10144). Slides were imaged on a Vectra Polaris scanning microscope (RRID:SCR_025508). Quantification was performed using inForm imaging software (version 2.5.1) to demultiplex the sample images and create individual tiles. Tiles were reconfigured and analyzed in HALO (RRID:SCR_018350) using HighPlex FL module software version 4.2.14. The following channels were used: DAPI (excitation, 375 nm; emission, 435–480 nm), Opal 520 (excitation, 488 nm; emission, 500–550 nm), Opal 570 (excitation, 561 nm; emission, 570–630 nm), Opal 620 (excitation, 588 nm; emission, 650–760 nm), and Opal 690 (excitation; 676 nm, emission, 694 nm).

Results

Spatial proteomics of patient tumors define distinct CAF and tumor cell neighborhoods

Two subtypes of tumor cells (classical and basal) and three subtypes of CAFs (iCAF, myCAF, and apCAF) are accepted as important contributors to PDAC heterogeneity (15, 17). To further delineate these subtypes in a patient-specific manner, we comprehensively analyzed tumor specimens from 15 patients with PDAC who underwent surgical resection at our institution (Table 1). Imaging mass cytometry (IMC) was used to label human tissue with a panel of 43 antibodies (Supplementary Table S1) designed to capture spatially resolved profiles of tumor regions, including CAFs, epithelial cells, and immune cells. Specific ROIs were selected for annotation by an experienced pathologist (J.W. Lee; Supplementary Fig. S4A). To ensure similar sampling of the TME from every patient, ROIs were selected for malignant epithelial-rich or stroma-rich areas and then further divided as immune-rich or immune-poor (Fig. 1A). To annotate immune-rich or immune-poor regions, dual IHC staining for CD3 (T cells) and CD68 (myeloid cells) was performed on sequential slides, and a grid system was applied to account for the size of ROIs to be ablated for IMC (Supplementary Fig. S4B). We intentionally avoided tertiary lymphoid structures when selecting ROIs, as these regions have been shown to be restricted to tumor edges (44). The IMC panel was curated to ensure comprehensive analysis of CAF subtypes as well as endothelial, epithelial, and immune cell subtypes (NK, B, myeloid, and multiple subsets of T cells; Fig. 1B). All cell types were analyzed by relative expression of each marker included in the panel (Fig. 1C). Phenograph clustering analysis identified groups of epithelial cells (CK⁺), CAFs (αSMA⁺, VIM⁺, PDGFR⁺, and/or FAP⁺ and CK^low), and immune cells [both myeloid (CD68⁺) and lymphoid (CD3⁺ and CD20⁺)] as the most highly identified clusters (Fig. 1C). Cells belonging to a cluster characterized by low expression of collagen and vimentin were localized in the stromal regions only and were labeled as “CAF undefined” (Supplementary Fig. S4C). To validate the accuracy of our cell segmentation and clustering workflow, we applied an orthogonal method based on pixel-level clustering (Pixie; ref. 26). Using matching annotations for major cell types for both pixel-based (Supplementary Fig. S4D) and cell-based (Supplementary Fig. S4E) workflows, we verified that cell type proportions from each of the methods strongly correlated with one another (r = 0.802, P = 3.316e−100) without apparent patient-driven effects (Supplementary Fig. S4F). The total number of cells and distribution of cell types acquired for each patient reflected patient-specific heterogeneity, and importantly, all broad cell types were represented in every patient sample (Fig. 1D and E). Proportionally, CAFs and tumor cells [basal (KRT17⁺) or classical (TFF1⁺)] were the most prevalent cell types across our heterogeneous patient cohort (Supplementary Fig. S5A).

Figure 1.

Spatial proteomic profiling of patient tumors identifies defined clusters and regions of cells in PDAC. A, Representative visualization of distinct regions acquired for IMC analysis—stromal high and tumor high regions were annotated by an experienced pathologist (J.W. Lee). Immune high and immune low regions were chosen when possible. B, IMC panel markers grouped by cell type. C, Cell type heatmap annotations demonstrating the relative expression of relevant markers included in the IMC panel. D, Total proportion of broad cell types visualized as a distribution by patient. E, Representative images from four patients demonstrating classical (TFF1: green), basal (KRT17: red), myCAF (αSMA: blue), and activated CAF (VIM: pink) markers. Merged and matching H&E images are in the two right columns. Images observed at 10× magnification (scale bar, 100 μm; N = 15).

With the primary goal of directly evaluating the relationship between CAFs and tumor cells, we clustered our data to look exclusively at CAF and tumor cell populations, both across the cohort and at the level of individual patients (Fig. 2A and B). We performed nearest neighbor analysis to understand interactions between CAF and tumor cell subsets (basal, classical, or mixed). Mixed tumor cell designation was assigned to tumor cells expressing both classical and basal markers, with these data suggesting plasticity of the epithelial cohort between subsets and a transitional state. Upon enumerating the top three nearest cells within a 4 μm distance radius from each of the basal, classical, or mixed tumor cells and filtering the counts by CAF subtypes only, we found that iCAFs were the most frequent cell type nearest in distance to classical tumor cells, whereas myCAFs were most frequently nearest to basal cells (Fig. 2C). To investigate neoadjuvant treatment effects, the patient cohort was stratified into untreated (n = 4) and FOLFIRINOX (5-fluorouracil, irinotecan, oxaliplatin, leucovorin) treated (n = 7) patients. The limited size of the untreated patient cohort reflects current clinical practice, with most receiving neoadjuvant chemotherapy prior to surgical resection. Similar trends were observed between subgroups, with the exception of iCAF localization shifting toward mixed tumor cells in the FOLFIRINOX-treated specimens (Supplementary Fig. S5B and S5C). Similarly, counts were filtered at 10 and 30 μm radii to examine the effect of larger spatial niches, and localization relationships remained consistent (Supplementary Fig. S5D and S5E). ApCAFs, which represented a minority of CAFs (Fig. 2C and D), were also observed as a nearest neighbor of basal cells (Fig. 2C and D). These data suggest that CAF and epithelial subtypes are paired together uniquely in spatial neighborhoods (iCAF/classical and myCAF/basal) across a more broadly heterogeneous TME (Fig. 2D). These associations were also visually confirmed within ROIs displaying both classical-rich and basal-rich areas of tumor cells (Fig. 2E). Taken together, these data demonstrate a distinct spatial coordination among CAF and tumor cell subtypes.

Figure 2.

Spatial proteomics analysis reveals that myCAFs are spatially located nearest to basal PDAC tumor cells. A, Cell type heatmap for the data subset for CAFs and tumor cells. Cell type annotations were determined by the relative expression of relevant markers included in the IMC panel. B, Distribution of cell types analyzed and visualized by patient. C, Heatmap showing the nearest neighbor analysis identifying distinct CAF subtypes that are closest to basal, classical, and mixed tumor cells. D, Stacked bar plot showing cell type frequencies of CAF subtypes that are present nearest to basal, classical, and mixed tumor cells, dictating increased myCAF frequency near basal cells and increased iCAF frequency near classical cells. E, Representative images of IMC and H&E for two patients that have basal-rich and classical-rich regions displayed in one ROI. Images depict more myCAF presence (αSMA⁺) and intensity in basal-rich (KRT17⁺) regions (top row for each patient); H&E scale bar, 50 μm.

Bulk RNA-seq of patient-matched PDO and CAF cocultures defines CAFs as drivers of EMT in PDOs

To more comprehensively evaluate CAF-tumor cell cross-talk within the IMC-defined neighborhoods, we completed bulk RNA-seq of flow-sorted CAF and PDO cocultures from 13 patients with PDAC. The generalizability of these data to diverse cohorts of patients with PDAC was preserved, given that the current set of samples represents three untreated patients, seven previously treated with neoadjuvant FOLFIRINOX, and three who received gemcitabine and other neoadjuvant treatments (Fig. 3A and B; Table 1). Data were first evaluated by computing principal components, which demonstrated that cell type was the predominant source of variation in the data and distinguished CAFs from organoids (Fig. 3C). Further visualization of covariates on the PCA did not find significant variation in the transcriptional data related either to culture conditions (coculture or monoculture as control, Fig. 3D) or individual patients (Fig. 3E). The coculture organoid condition for JHH 390 did not meet quality metrics, so a 12-patient cohort was carried forward for coculture comparisons. Although not a dominant source of variation at a global transcriptional level in the PCA, we still hypothesized that the coculture would affect transcriptional profiles relative to monoculture. Differential analysis of gene expression compared cocultured PDOs to monoculture PDOs to identify CAF-mediated effects (Fig. 3F). Following PDO-CAF coculture, there was a significant increase in the expression of genes associated with fibrogenesis and malignant progression, including POSTN, DCN, COL1A2, and MMP2 and INHBA, respectively (Fig. 3F). We further identified enrichment of gene sets representing EMT, TNFα response, E2F targets, KRAS, and MYC targets in coculture PDOs relative to monoculture using the Hallmark gene sets from the Molecular Signatures Database (Fig. 3G; ref. 45).

Figure 3.

RNA-seq of patient-matched PDO and CAF coculture reveals that CAF presence drives heightened EMT in PDOs. A, Graphical schema depicting the creation of the biobank from patient tumors and the workflow for performing RNA-seq. B, Representative image of PDO–CAF coculture prior to harvest for RNA-seq. The large arrow points toward CAF, and the smaller arrow points toward PDO (scale bar, 250 μm). C, Unsupervised clustering of viable cells from 13 patient-derived CAF–PDO cocultures represented as a PCA plot. D, Resolution of coculture and monoculture cells represented by a PCA plot. E, Resolution of each patient represented by a PCA plot. JHH 390 co-org did not meet quality parameters upon sequencing, so it was excluded from the analysis, resulting in a final N = 12 used in coculture comparisons. F, Differential gene expression of PDO samples that are monocultured (left) compared with cocultured (right). G, Overrepresented MSigDB hallmark gene sets in PDO samples from monoculture (left) compared with coculture (right). H, ProjectR transfer learning of gene signatures inferred in scRNA-seq data of PDAC tumors (31) onto the bulk RNA-seq data demonstrates that the pattern associated with inflammatory signaling and EMT (CoGAPS Pattern 7) inferred in our prior study is enhanced in only PDO cells from coculture relative to PDO cells from monoculture, P = 5.6e–5 by two-tailed paired Student t test. I, Representative time-course images of PDO monoculture and PDO–CAF coculture demonstrating the morphologic changes that occur in PDO over the course of 2 weeks that are not seen in the monoculture PDOs, images taken at 10× on an Echo Rebel microscope (scale bar, 250 μm). J and K, Quantification of N-cadherin (J) and KRT17 (K) surface expression on PDOs via flow cytometry, statistically supported by a one-way ANOVA with Dunnett test. L, Representative images from IMC showing traditional EMT markers, E-cadherin, N-cadherin, and vimentin (scale bar, 250 μm). Quantification of E-cadherin (M) and vimentin (N) expressed in basal tumor cells near a CAF or not near a CAF. Five patients were profiled with four technical replicates per patient. Significance is denoted as follows: ns, not significant; **, P < 0.01; ***, P < 0.001. FC, fold change. A, Created in BioRender. Guinn, S. (2024) https://BioRender.com/3vb522k.

These findings of enhanced EMT in coculture support our prior work examining the global transcriptional impact of CAFs on tumor cells (31). In our prior study, analysis of a comprehensive atlas of public domain single-cell RNA-seq data of PDAC tumors identified a gene expression pattern in tumor cells associated with increased expression of inflammatory, fibrogenic, and EMT pathway genes (31). We referred to this as the “Inflammatory/EMT” pattern, and the enrichment of this pattern seemed to be driven by the presence of, and regulation by, CAF signaling that was observed in coculture data of PDOs from three treatment-naïve patients. In the full 12-patient cohort explored here, we identified CAF-PDO coculture as inducing increased expression of inflammatory/EMT gene pathways (also referred to as CoGAPS Pattern 7), as compared with those from monoculture (Fig. 3H). This suggests that CAF signaling to epithelial tumor cells plays a critical role in the promotion of tumor inflammation, fibrogenesis, and EMT. In addition to the broader transcriptomic findings, we observed morphologic changes consistent with EMT in cocultured PDOs compared with monoculture counterparts (Fig. 3I). To quantify these phenotypic changes, we cocultured five different PDOs with CAFs (at a ratio of 1:3 PDOs to CAFs) or in CAF-conditioned media for 10 days and characterized the PDOs via flow cytometry. PDOs cocultured with CAFs showed greater N-cadherin surface expression compared with PDO monoculture (Fig. 3J). CAF-conditioned media alone did not induce changes in cell morphology or increase the N-cadherin surface expression we used as a marker of EMT in the PDOs, suggesting that cell–cell contact may be necessary for this shift in our system (Fig. 3J). We also examined the surface expression of the basal marker KRT17⁺ on day 3 of coculture and found that both PDOs in coculture with CAFs and CAF-conditioned media displayed higher KRT17 surface expression than monoculture PDOs (Fig. 3K). This is consistent with previous work demonstrating that basal tumor cells exhibit mesenchymal-like gene expression (15). The same trends were observed across individual patients (Supplementary Fig. S6A and S6B). Further evidence to support CAF-directed EMT was found by directly evaluating patient tissues using IMC staining for the expression of E-cadherin, N-cadherin, and vimentin (Fig. 3L). We hypothesized that tumor cells in close proximity to CAFs would demonstrate increased evidence of EMT through the loss of E-cadherin and increased vimentin. Based on IMC analysis, E-cadherin expression was, in fact, greater in tumor cells distant from a nearby CAF (Fig. 3M), whereas vimentin expression was greater in tumor cells near a CAF (Fig. 3N; refs. 31, 46). Taken together, these data support the hypothesis that CAFs are critical to and may be controlling drivers of EMT in PDAC tumor cells at a population level.

CAFs promote changes in PDO gene expression from classical to basal tumor cell classification

We next revisited the classical/basal tumor classification described by Moffitt and colleagues to assess gene signatures in our PDOs and determine how CAFs influence tumor cell subtype (15). Consistent with a prior report using scRNA-seq, PDO monoculture demonstrated heterogeneity in the relative expression of key classical and basal classifier genes (Fig. 4A; ref. 47). As CAFs were introduced into cultures, PDO gene expression was often altered, and profiling revealed plasticity of gene expression in the epithelial compartment in a manner dependent on the presence of CAFs (Fig. 4B). In contrast to the increase of EMT in the inflammatory/EMT pattern defined from our atlas (31), the transition to a more basal-like phenotype was only observed in a subset of the organoids. In the five patients for whom the gene expression was most dynamic, there was a significant shift toward increased basal gene expression (Fig. 4C). These five patients did not have clear common clinical similarities, as they had different KRAS mutations, neoadjuvant treatments, and times to recurrence (Table 1). There were no patient samples in which tumor cell exposure to CAFs through coculture induced a more classical gene expression pattern. Moreover, this state transition seemed independent of the initial subtype classification of the organoid monoculture, although the four lines with the highest basal scores in monoculture did not significantly change subtype.

Figure 4.

CAFs drive changes in PDO gene expression from classical to basal tumor cell classification. A, Heatmap displaying basal and classical gene expression in PDOs between co- and monoculture samples. B, Basal gene expression on a per-patient basis. Red hashed box demonstrates the PDOs that upregulate basal gene expression in coculture PDOs (n = 5). C, Top five PDO lines demonstrating classical to basal transcriptional plasticity when cocultured with CAFs. Comparisons of conditions are statistically supported using the two-tailed Student t test with equal variance, P = 0.0022. D, Differential gene expression showing the genes upregulated in n = 7 lines not displaying transcriptional plasticity (left) and genes upregulated in PDOs that become more basal (n = 5; right), visualized by volcano plot. E, Overrepresented MSigDB Hallmark gene sets in PDO samples that do not shift (left) compared with PDO samples displaying transcriptional plasticity toward a basal gene signature (right). FC, fold change.

We termed these five PDOs “plastic,” indicating the transcriptional changes resulting from coculture. A quantitative assessment of gene expression signatures identified a strong basal shift in these five patients, with increased relative basal expression of global gene programs by at least 14% and up to 28% (Fig. 4C). Assessment of the expression of specific basal genes such as KRT6A, COL1A2, and cadherins, as well as cancer stem cell genes such as CD44, demonstrated increased pathway signaling with CAF coculture suggesting CAFs are driving a more basal and more aggressive PDO phenotype in part through these critical genes (Fig. 4D). In these samples, and in keeping with a more basal phenotype, increased PDO tumorigenicity was inferred by Hallmark pathway analysis, showing upregulation of EMT and proliferation by E2F targets, as well as upregulation of angiogenesis and hypoxia signaling pathways (Fig. 4E). To validate our coculture findings as relevant to patient disease, we returned to primary patient tissues and used multiplex RNAscope to probe for basal (KRT17 and KRT6a) and classical (TFF1 and GATA6) RNA transcripts (Supplementary Fig. S7A) on tissue sections. Similar to our coculture findings, there was a range of expression of representative basal and classical genes across patients at baseline. Additionally, there was a direct correlation between the selected basal RNAs and classical RNAs, reinforcing their coexpression and validity as markers to designate tumor cell subtype (Supplementary Fig. S7B).

To determine if the basal phenotype was associated with an abundance of CAFs, as seen in our coculture, we used trichrome staining to infer CAF density on whole tissue sections, with concurrent H&E staining to confirm CAF morphology. To further demonstrate that collagen detected by trichrome staining is primarily derived from CAFs rather than from tumor cells, we analyzed average collagen expression per ROI in our IMC cohort. This analysis demonstrated significantly greater expression of collagen in CAFs compared with tumor cells (Supplementary Fig. S8A). In samples with pronounced trichrome staining indicative of a high CAF presence, corresponding PDOs were more likely to be either strongly basal at baseline or inducible to a basal phenotype upon CAF coculture (Supplementary Fig. S8B and S8C). These findings further support the hypothesis that dynamic CAF-mediated signaling modulates epithelial gene expression in a manner that influences clinically relevant tumor classification.

We then sought to determine whether coculture with PDOs also led to gene expression changes within the CAFs (mean myCAF signature of approximately 0.985 vs. 0.975, respectively; Supplementary Fig. S9A–S9C). We hypothesized that the limited changes observed at the whole transcriptome level could be due to the limitations of bulk RNA-seq and, therefore, queried specific CAF markers with qPCR to examine gene expression from cocultures relative to monocultures (Supplementary Table S4). Using this more granular approach, we identified significant changes in the expression of ACTA2 (αSMA), VIM (vimentin), CTGF (CCN2, connective tissue growth factor), and COL1A1, suggesting a limited but present myCAF gene reprogramming in coculture (Supplementary Fig. S9D).

CAFs drive basal gene expression in PDOs through secreted proteins

Although our data suggest that proximity between CAFs and tumor cells alters tumor cell classification and drives tumor EMT, it remained unclear if this tumor cell plasticity was due to direct cell–cell interaction or a result of signaling mediated by the CAF secretome. To evaluate the role of CAF-secreted factors on tumor cell molecular states, we treated PDOs with CAF-conditioned media and queried classical and basal gene expression changes using qPCR (Supplementary Table S4). Overall, culturing in CAF-conditioned media induced altered PDO gene expression from classical toward increased basal programming (Fig. 5A). In exploring the functional implications of CAF-conditioned media, and to examine the capacity of the secretome to alter cell migration, we performed a wound healing assay using Panc 10.05 cells (48). Cells treated with CAF-conditioned media from basal patient lines demonstrated an increased rate of wound closure compared with those cultured with conditioned media from CAFs matched to patients with classical PDO lines (Supplementary Fig. S10A and S10B). This suggests that the CAFs are contributing to both enhanced expression of basal-associated genes as well as a basal-like phenotype.

Figure 5.

CAFs that drive basal gene expression in PDOs secrete distinct proteins. A, PDOs treated with CAF-conditioned media demonstrate decreased classical gene expression (GATA6, TFF1, CLDN18, and LGALS4; left) and increased basal gene expression (KRT17, KRT6a, KRT7, S100a2, and BCAR1; right) following qPCR assessment. JHH352 did not amplify for KRT17. B, STRING database protein interaction network. C, Classical gene expression qPCR readout from PDOs treated with rIL8 compared with untreated (left), basal gene expression qPCR readout from PDOs treated with rIL8 compared with untreated (right). Comparisons of rIL8 and untreated conditions are statistically supported using one-way ANOVA with Tukey test in PRISM [version 9.2.0 (283)]. D, IL8 secretion from proteome screen (Supplementary Fig. S11A) analyzed by normalized Z score. E and F, Automated immunoblotting protein bands showing GATA6 (E) or TFF1 (F) protein expression in PDOs treated with CAF-conditioned media or 50 ng/mL of IL8, with total protein visualized via heatmap underneath. Bar graphs show GATA6 (E) or TFF1 (F) density normalized to total protein and analyzed by a paired two-tailed Student t test. G, Resolved heatmap of CAF-defined markers demonstrating IL8⁺ CAF are nearest to classical or mixed tumor cells. H, Representative IMC images from three patients showing colocalization of IL8 (red) and classical tumor marker TFF1 (green; scale bar, 200 μm). Significance is denoted as follows: ns, not significant; *, P < 0.05; **, P < 0.01; ****, P < 0.0001.

To identify secreted proteins that may be responsible for these classical versus basal molecular and functional differences, we analyzed the secretome of conditioned media from patient-derived CAFs, comparing media from six CAF lines: CAFs that maintain a classical PDO (n = 2), CAFs that maintain a basal PDO (n = 2), and CAFs that drive a classical-to-basal shift in their matched corresponding PDOs (n = 2). Upon profiling the CAF secretomes, we found 43 secreted proteins unique to patients with PDOs that demonstrate classical-to-basal gene expression plasticity (Supplementary Fig. S11A). We analyzed these data with the STRING database to infer specific networks of cellular cross-talk driven by protein-specific interactions and selected three dominant clusters (49). These included immune-related proteins, growth factor–related proteins, and cell mobility and migration (Fig. 5B). Further pathway-specific analysis suggested that 14 proteins were significantly enriched in the PDOs demonstrating classical-basal plasticity. Although these results infer proteins distinctly enriched in the group with PDO plasticity, they do not indicate whether the proteins favor classical or basal gene expression patterns. To determine the functional roles of these enriched proteins, we treated PDOs with recombinant proteins for these 14 hits to evaluate their capacity to alter PDO gene programming from a classical to basal phenotype (Supplementary Fig. S11B and S11C). Of the 14 proteins screened, we identified IL8 as a key protein in the secretome that maintains gene expression corresponding to a classical phenotype in tumor cells (Fig. 5C). However, recombinant IL8 treatment resulted in no meaningful change to the PDO gene expression of basal targets KRT17 and S100A2 (Fig. 5C). Thus, it seems that CAF-derived IL8 signaling in classical phenotype tumors, when lost, can induce basal gene expression. In returning to our initial CAF secretome data, IL8 secretion was nearly undetectable in the basal CAF lines but highly detectable in the CAF lines associated with PDO subtype plasticity (Fig. 5D). This suggests that CAF-secreted IL8 facilitates the maintenance of a classical tumor gene expression profile when evaluated within the context of a competent 3D model of the PDAC TME. Using the same PDO model, we examined whether this classical shift is induced exclusively by the presence of IL8. PDOs from five patients were cultured for 4 days in 2% FBS CAF media with recombinant IL8 added or CAF-conditioned media from a single CAF line that demonstrated induction of tumor cell plasticity. PDOs were then analyzed via automated immunoblotting to assess the extent of classical gene expression in response to IL8 treatment. Across the five patients, there was variability in classical marker protein response to IL8 treatment (Fig. 5E and F; Supplementary Fig. S11D and S11E). Expression of either the common classical marker TFF1 or GATA6 was preferentially increased in four of five patients, though not statistically significant (Fig. 5E and F). This lack of statistical significance may be accounted for by sample size, interpatient variability, and the initial transcriptional definition of classical and basal tumor cells. Upon transition from our coculture model to patient tissue sections, we found that FAP⁺, CXCL12⁺, IL6⁺, and IL8⁺ CAFs were nearest to classical tumor and mixed tumor cells (Fig. 5G and H). These data suggest that CAF-derived IL8 contributes to maintaining the classical tumor subtype, revealing an additional regulatory role for IL8 within the TME that warrants further investigation.

Activated T cells are associated with basal tumors

Finally, as the CAF secretome analysis identified differences in the expression of immunomodulatory ligands, we broadly examined the relationship between immune cell type in the TME and the heterogeneity in myCAF/basal and iCAF/classical neighborhoods. There has been limited clinical success employing immune checkpoint blockade to treat PDAC (50). The ineffectiveness of immune checkpoint blockade has been attributed to immune evasion, low neoantigen burden, and immunoediting; recently, CAFs have been hypothesized to affect the immune landscape of the disease (51, 52). IMC analysis demonstrated that across our selected ROIs and among broad immune populations, T cells were collectively the immune cells most often the nearest neighbors to all tumor cells and CAF subtypes (Fig. 6A–D). Based on our findings that IL8, a cytokine involved in T cell suppression and resistance to checkpoint immunotherapy (53), is an important determinant of classical phenotype, we hypothesized that the T cell frequency may be lower near classical cells. Using a nearest neighbor analysis examining the three nearest immune cells, we quantified the frequency of T cells adjacent to tumor subtypes. Basal tumor cells displayed a higher frequency of T cell adjacency events than mixed or classical tumor cells (Fig. 6E). Effector CD4 and CD8 T cells comprised the largest fractions of T cells, most noticeably in the basal and mixed neighborhoods (Fig. 6E). This observation remained consistent regardless of patient treatment status (Supplementary Fig. S12A and S12B). Moreover, T cells were adjacent at the highest frequency near myCAFs (Fig. 6F), irrespective of patient treatment status (Supplementary Fig. S12C and S12D). Collectively, these data demonstrate that the character of the adaptive immune infiltration in the TME is heterogeneous and associated with distinct tumor cell and CAF subtypes. Future studies should examine how CAF-derived secreted factors, including IL8, affect immune recognition and function near basal tumor cells and the spatial coordination of immune cells near classical tumor cells, to better understand immune infiltration, tumor recognition, and immunotherapy resistance in PDAC.

Figure 6.

T cells are located in the closest proximity to tumor cells and CAFs. A, Representative images of patient IMC showing tumor cells (KRT17⁺, TFF1⁺) and immune cells (CD3⁺, CD163⁺) colocalizing (scale bar, 200 μm). B and C, Heatmap of the nearest neighbor analysis of immune and tumor cells (B) and CAFs (C), in which T cells are the predominant immune cell near all classifications of tumor cells and CAFs when evaluating nearest neighbors. D, Cell type heatmap for the data subset for T cells. Cell type annotations were determined by the relative expression of relevant markers included in the IMC panel. E and F, Stacked bar plot of the frequency of T-cell subtypes in association with tumor cell (E) and CAF (F) classifications, in which T cells are found at a higher frequency in association with basal tumor cells (E) and myCAFs (F).

Discussion

Here, we report a multidisciplinary study using comprehensive IMC to complement novel TME culture methods. We employed a living biobank of patient-matched CAFs and PDOs, enabling us to broadly recapitulate the functional biology of CAFs in vitro and explore CAFs as critical regulators of the PDAC TME. By leveraging a combination of high-dimensional data and patient-derived coculture, we identified CAF-driven EMT, a neighborhood-specific association of IL8-secreting iCAF cells with classical tumor cell designation and loss of IL8 expression associated with basal tumor designation. Finally, we identified T cells as being associated with tumor-CAF neighborhoods, with increased abundance in myCAF neighborhoods.

CAF-promoted EMT is a complex phenomenon that can occur across tumor types. Prior work has shown that epithelial cells can lose expression of E-cadherin and transform into mesenchymal cells with a greater ability to migrate through the basement membrane, promote extracellular matrix deposition by stromal cells, and increase resistance to chemotherapeutics (54–56). These malignant epithelial cells can then reach distant organs such as the liver or lung and establish metastatic deposits (57, 58). These data further support a therapeutic emphasis on reprogramming CAFs to limit EMT and, therefore, prevent metastatic spread. In this current study, we found that our previously identified transcriptional signature of CAF-induced EMT observed in human tumors (31) is also associated with EMT transitions in PDO-CAF coculture, independent of prior treatment or subtype. Additional studies are needed to identify specific mechanisms responsible for these protumorigenic changes and to distinguish EMT induction from subtype switching in PDAC. Here, in our PDO model, we identified one mechanism utilized specifically by CAFs to promote a transition from classical to basal tumor cell phenotype in a heterogeneous manner, with the capacity to influence patient-specific outcomes. We identified IL8 as a CAF-secreted factor enriched in CAFs derived from classical tumors that helps maintain the classical tumor cell phenotype. In contrast, the loss of CAF-derived IL8 within the TME was associated with a shift toward the basal phenotype, a PDAC subtype linked to particularly poor prognosis. These findings support further investigation of IL8 preservation or modulation in next-generation therapeutic strategies.

The role of IL8 in tumor behavior is complex and incompletely understood, particularly in PDAC. IL8 has been implicated as an important chemokine in PDAC aggressiveness through in vitro work, with studies suggesting tumor cell expression of IL8 as a hallmark of cells undergoing EMT (59). However, there are fewer studies examining the impact of IL8 that is secreted from stromal cells in the TME. Our findings suggest that IL8 is central to maintaining a classical tumor cell phenotype, implying that (i) IL8 in the PDAC TME may have more of a functional duality than previously appreciated and (ii) the source of IL8 in the TME is not exclusively epithelial cancer cells. Consistent with our findings, Carpenter and colleagues previously described KRT17^highCXCL8⁺ (an alternative name for IL8) tumor cells as an intermediary phenotype between classical and basal and correlated with intratumoral myeloid abundance (60). Unlike the work from Carpenter and colleagues, we exclusively examined the impact of IL8 on tumor cell subtype. This is a critical caveat, as extensive literature supports an immunosuppressive role for IL8 through the recruitment of neutrophils and other myeloid-derived suppressor cells (61, 62).

Characterization of CAF heterogeneity remains challenging, given their diverse molecular functions and the limitations of available model systems (8, 17, 63–65). For example, CAFs have been implicated as having both tumor-promoting and tumor-restraining properties, furthering the challenges associated with classification (66, 67). These collective works demonstrate that changes in stromal composition can have many effects on tumor development and phenotype. Additionally, PDAC tumor subtypes can be classified transcriptionally and can correlate with different types of stroma (15, 68, 69). Our work, along with others, further reveals the heterogeneous landscape and spatial relevance of both tumor and nonepithelial cells in the environment, providing an additional source of gene expression, secreted proteins, and other drivers of the dynamic behavior in the PDAC TME (18, 70–74).

Understanding CAF–tumor cell relationships at a single-cell level is critical to more reliably representing the inherent intratumoral heterogeneity. This is particularly important when considering that many of the genes we have focused on are shared between CAFs and tumor cells, such as EMT markers like vimentin. Although the results from this study further resolve the spatial relationships in the TME, there are also shortcomings to this study. Our IMC data are restricted to a panel that estimates cellular colocalization in cell state transitions through static snapshots, using space as a surrogate for time, rather than explicitly inferring cell state transitions through methods such as trajectory inference (75) or in dynamic models (76). This presents a challenge when trying to relate cellular dynamics to cellular function more deeply; importantly, the number of markers on the panel further limits the cell types we could comprehensively analyze. Myeloid cells are significant contributors to the TME, yet we were unable to fully characterize these populations due to technical and antibody panel limitations; nevertheless, myeloid biology remains an important area for future investigation. Our CAF–PDO coculture model represents an exemplary system to further study the TME using dynamic approaches. However, one of the caveats is that CAF heterogeneity is not fully preserved in vitro, thereby limiting the granularity with which our cocultures can evaluate nascent CAF subtypes. Furthermore, our CAF–PDO model has limitations, given that the technical complexity we did not explicitly explore immune interactions in these cocultures. Inclusion of immune cell populations will be essential for future studies, providing opportunities for further immune phenotyping and therapeutic intervention to overcome the immunosuppressive barriers. These shortcomings demonstrate the importance of employing complementary systems of assessment for studies evaluating the PDAC TME.

Altogether, these findings demonstrate the importance of understanding the complex interplay between CAFs and tumor cells in the PDAC TME. We have defined a relationship whereby CAFs drive EMT and basal gene expression and form distinct cellular neighborhoods of iCAFs/classical and myCAFs/basal cells. We also identified differential T cell presence within tumor neighborhoods, a finding that warrants additional exploration as we aim to improve immunotherapeutic strategies for PDAC. Future work will prioritize understanding the immunofunctional implications of these tumor cell and CAF relationships.

Supplementary Material

Supplemental Table 1

IMC Markers

Supplemental Table 2

HCPLT Media

Supplemental Table 3

Lineage Verification

Supplemental Table 4

mRNA Target Assay ID

Supplemental Figure 1

Representative images of immune, stromal, and architectural markers in the IMC panel

Supplemental Figure 2

Organoid and CAF cell type identification

Supplemental Figure 3

Coculture setup and deconstruction.

Supplemental Figure 4

IMC workflow and quality control

Supplemental Figure 5

Cells analyzed per patient and subset nearest neighbor analysis by treatment status and at increased radii

Supplemental Figure 6

Quantification of N-cadherin and Krt17 surface expression on organoids with and without coculture

Supplemental Figure 7

RNAscope of classical and basal tumor cell markers

Supplemental Figure 8

Collagen presence in IMC cohort

Supplemental Figure 9

Comparison of CAF gene expression in monoculture and coculture

Supplemental Figure 10

Wound healing assay with CAF conditioned media

Supplemental Figure 11

Proteome profiler from CAF conditioned media

Supplemental Figure 12

IMC subset analysis of T cell relationship to tumor cells and CAFs

Data Availability

There are restrictions on the availability of sequencing data. For the protocol under which patients JHH317 and JHH361 were consented, data sharing was not explicitly defined, and it is not possible to reconsent these patients with highly aggressive and rapidly lethal disease. As such, the IRB has requested that we do not publicly share individual-level sequencing data from these two patients. They are, however, securely stored within a Johns Hopkins University patient data system and will be shared by the lead contact (J.W. Zimmerman) upon request. For the remaining patients who received neoadjuvant therapy, raw reads for the RNA-seq data are publicly available in the Database of Genotypes and Phenotypes (dbGaP; RRID:SCR_002709) at phs004499.v1.p1, and processed read counts are publicly available in the Gene Expression Omnibus (GEO; RRID:SCR_005012) at GSE313826. The bulk RNA-seq data generated from patients who did not receive neoadjuvant treatment (JHH348, JHH362, JHH380) are publicly available in dbGaP (RRID:SCR_002709) at phs003563.v1.p1, with processed read counts publicly available in GEO (RRID:SCR_005012) at GSE245319. The code for RNA-seq analysis is available on GitHub: https://github.com/jzimme27/PDAC-PDO-CAF-coculture. Patient IMC data are available at Zenodo [RRID:SCR_004129; https://doi.org/10.5281/zenodo.14871170 (raw MCD); https://doi.org/10.5281/zenodo.17887333 (raw TIFFs)].

Code Availability

The code for IMC analysis is available on GitHub: https://github.com/wjhlab/CAF-neighborhoods-PDAC.

Authors’ Disclosures

D.J. Zabransky reports grants from Roche/Genentech and personal fees from Omni Health Media, Sermo, Atheneum, eScientiq, Deerfield Institute, ZoomRx, Neucore Bio, HAB Central, M3 Global Research, GLG, and Tegus Inc. outside the submitted work. S.M. Shin reports other support from Standard BioTools outside the submitted work. L. Danilova reports grants from the NIH during the conduct of the study. L. Zheng reports grant support from Bristol Myers Squibb, Merck, AbbVie, AstraZeneca, and Ipsen, reports other from Biosion, Pfizer, Amberstone, Fortress, Tallac, and Fortvita (consultant/advisory board member); that L. Zheng is in the Speaker Bureau of Hengrui Pharma and Servier Pharmaceuticals; and owns shares in Amberstone, Alphamab, Cellaration, and Mingruizhiyao. E.J. Fertig reports grants from the NIH and Lustgarten during the conduct of the study, as well as grants from AbbVie Inc., Kura Oncology Inc., Break Through Cancer, NFCR, and Roche/Genentech and personal fees from Mestag Therapeutics and Resistance Bio/Viosera Therapeutics outside the submitted work. L.T. Kagohara reports grants from the Lustgarten Foundation and the NIH during the conduct of the study. E.M. Jaffee reports other support from AB Meta, Adventris Pharmaceuticals, and Bristol Myers Squibb; personal fees from HDT Bio, Mestag, Surge, NeoTX, NEUVOGEN, and CAndel; and grants from Lustgarten, Genentech, and Break Through Cancer outside the submitted work. W.J. Ho reports grants from NIH S10OD034407 during the conduct of the study, as well as other support from Rodeo/Amgen, Standard BioTools, and Exelixis and grants from NeoTX, Riboscience, and Sanofi outside the submitted work. J.W. Zimmerman reports other support from the Charles and Margaret Levin Family Foundation and the Dana and Albert Broccoli Charitable Foundation during the conduct of the study, as well as grants from imCORE (Genentech) and personal fees from Sermo and ZoomRx outside the submitted work. No disclosures were reported by the other authors.

Authors’ Contributions

K. Belanger: Data curation, validation, investigation, methodology, writing–review and editing. S. Guinn: Conceptualization, data curation, formal analysis, validation, investigation, visualization, methodology, writing–original draft. B. Perez: Validation, investigation, methodology, writing–review and editing. Y. Cho: Data curation, formal analysis, validation, writing–review and editing. J.A. Tandurella: Data curation, formal analysis, investigation, methodology. M. Ramani: Validation, investigation. J.W. Lee: Supervision, investigation, visualization, methodology. D.J. Zabransky: Supervision, visualization, writing–review and editing. E. Kartalia: Investigation, visualization. J. Patel: Investigation, visualization, project administration. H. Zlomke: Investigation, visualization, project administration. N.G. Nicolson: Investigation, writing–review and editing. N.E. Gross: Data curation, formal analysis, visualization, methodology, writing–review and editing. S.M. Shin: Formal analysis, validation, investigation, visualization. B.C. Barrett: Validation, investigation, methodology. S. Sun: Investigation. N. Sun: Investigation. E. Firth: Investigation. A.G. Hernandez: Investigation, visualization. E.M. Coyne: Investigation, visualization. C.D. Cannon: Investigation, visualization, methodology. E. Moon: Validation, investigation. S. Charmsaz: Validation, investigation, methodology. L. Danilova: Data curation, formal analysis, supervision, writing–review and editing. J.M. Leatherman: Supervision, project administration. M.R. Lyman: Investigation, methodology. J.T. Mitchell: Formal analysis, investigation, methodology. L. Zheng: Supervision, writing–review and editing. M.G. Goggins: Resources, project administration, writing–review and editing. K.J. Lafaro: Resources. J. He: Resources. C.R. Shubert: Resources. W.R. Burns: Resources. E.J. Fertig: Supervision, investigation, writing–review and editing. L.T. Kagohara: Supervision, investigation. E.M. Jaffee: Conceptualization, resources, supervision, funding acquisition, writing–review and editing. R.A. Burkhart: Conceptualization, resources, supervision, funding acquisition, investigation, writing–review and editing. W.J. Ho: Conceptualization, resources, supervision, validation, writing–original draft, project administration, writing–review and editing. J.W. Zimmerman: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, methodology, writing–original draft, writing–review and editing.