Periductal Fibroblast Density Defines Lymphocyte Exclusion via a CD44-Dependent Stromal Checkpoint in Pancreatic Cancer

Abstract

Pancreatic ductal adenocarcinoma (PDAC) exhibits dense fibrosis and immune exclusion. While fibrosis has been studied globally and at the region-of-interest level, its impact on stromal-ductal architecture and immune cell localization remains unknown. Here, we establish cancer-associated fibroblast (CAF)-stratified ductal spatial architecture as a fundamental determinant of immune exclusion in PDAC. Focusing on malignant PDAC epithelial ductal regions, the critical interface where immune cells must access tumor epithelium, we demonstrate that periductal fibroblast organization dictates leukocyte proximity. Through integrative analysis of treatment-naïve patient samples from three independent cohorts – including imaging mass cytometry, multiplex immunohistochemistry, and single-cell RNA sequencing – we uncover that activated, pro-inflammatory leukocytes preferentially localize near malignant ducts in regions with low fibroblast density. Stratifying epithelial-ductal regions by CAF abundance reveal a graded constraint: increasing fibroblast content corresponds to reduced leukocyte–epithelial proximity and elevated collagen I deposition. Despite their exclusion in high-CAF ducts, leukocytes in low-CAF ducts retain functional competence. Mechanistically, ligand–receptor inference implicates collagen–CD44 signaling as an adhesion axis anchoring immune cells within fibroblast-rich zones. CD44 blockade augments natural killer cell motility in vitro, while enhancing lymphocyte collagenase activity through MMP14 overexpression promotes infiltration by overcoming αSMA⁺ CAF-mediated stromal barriers in vivo. Thus, by establishing ductal regions as critical spatial units of immune exclusion, these findings provide a framework for dissecting stromal–immune interactions, reveal targetable “stromal checkpoints”, and provide complementary strategies to overcome CAF-driven barriers to leukocyte motility and infiltration in PDAC.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) has a 5-year survival rate of 12% (1). PDAC is considered to be an immunologically “cold” tumor with minimal infiltration by activated leukocytes and immunotherapy-based clinical trials using checkpoint inhibitors have been largely unsuccessful (2-4). This resistance can be at least partly attributed to the tumor’s desmoplastic stroma, driven by extracellular matrix (ECM) deposition from cancer-associated fibroblasts (CAFs) (5). The resulting fibrotic barrier increases tissue stiffness while restricting activated leukocyte infiltration and access to malignant ductal epithelial cells (6,7). Therefore, defining how CAF spatial organization within the periductal stroma constrains activated leukocyte-epithelial cell interactions is critical to creating successful immune-based therapeutic approaches.

Recent studies have explored the impact of CAFs on tumor cell progression in PDAC and characterized the landscape of the PDAC TME using multiple spatial molecular profiling technologies (8-18). While these studies have provided insights into immune dysfunction, CAF subtype heterogeneity, and desmoplastic remodeling, they have primarily focused on global tissue-level patterns or pre-selected regions (i.e. juxtatumoral vs. panstromal). Furthermore, the complex cellular interactions between CAFs, immune cells, and tumor cells within PDAC malignant ductal regions remain unknown. These ducts are the sites of tumor origin, progression, and metastasis that shape the global PDAC TME and its consequent biology, making the study of cellular colocalization and signaling in these sites critical to understand PDAC carcinogenesis (19-22). Since these ductal regions represent the critical interface between malignant epithelium and infiltrating immune cells, defining the stromal barriers that restrict immune access in these regions is crucial for understanding and overcoming immune exclusion in PDAC.

Here, we identified periductal CAF density in malignant ducts within invasive PDAC as a key tumor organizational feature that dictates immune access to tumor epithelium. Using complementary spatial approaches, including imaging mass cytometry (IMC) and multiplex immunohistochemistry (mIHC), in separate cohorts of PDAC samples, we mapped periductal cellular interactions and found that periductal αSMA⁺ fibroblast organization is the principal determinant of leukocyte infiltration in malignant ductal regions. Increased periductal fibrosis correlated with increased collagen I deposition and increased proliferation of epithelial cells and corresponded with reduced leukocyte-epithelial proximity, particularly with regards to natural killer (NK) cells. Ligand-receptor inference analysis of a recently published single cell RNA-sequencing (scRNA-seq) dataset (23) of PDAC samples from 16 patients revealed increased collagen signaling from CAFs to leukocytes via CD44, and CD44 neutralization in vitro increased NK cell motility through the ECM. Alternatively, augmenting lymphocyte collagenase activity through MMP14 overexpression enabled infiltration in collagen-rich PDAC models in vivo. These results indicate that (1) periductal αSMA⁺ CAF architecture governs leukocyte access while promoting malignant epithelial cell proliferation and (2) targeting stromal checkpoints, such as CD44, or enhancing lymphocyte’s collagenase activity may improve leukocyte infiltration and overcome ECM-mediated immune exclusion. Collectively, CAF-stratified ductal spatial architecture defines a new framework for analyzing PDAC fibrotic restrictions to leukocyte infiltration that can potentially be overcome through neutralization of stromal checkpoints in PDAC and other highly fibrotic tumors.

Materials and Methods

Tissue samples, antibody validation, metal-conjugation, and imaging mass cytometry (IMC)

The Lombardi Comprehensive Cancer Center (LCCC) Histopathology & Tissue Shared Resource (HTSR) provided formalin-fixed, paraffin-embedded (FFPE) human PDAC and spleen tissue samples. For IMC, we used a LCCC-curated FFPE human pancreas tissue microarray (TMA) slide series containing 2 mm cores from normal pancreas, pancreatitis, pancreatic neuroendocrine tumors, intraductal papillary mucinous neoplasm, pancreatic mucinous adenocarcinoma, PDAC, colon adenocarcinoma, testicle, tonsil, and placenta samples, and PDAC cell lines. Information for all TMA samples are available at doi: 10.5281/zenodo.10582038. The IMC study was restricted to 44 cores from 21 previously untreated PDAC patients who underwent Whipple resections or distal pancreatectomies (Supplementary Table S1).

For IMC, we designed a metal-conjugated antibody panel. Carrier-free antibodies not commercially metal-conjugated were validated by IF in PDAC or spleen tissue (Supplementary Table S2). Antibodies with expected expression patterns were metal-conjugated using MaxPar X8 Multimetal Labeling Kits (Standard Bio Tools (SBI); (Supplementary Table S2). Metal-conjugation was validated by cytometry by time-of-flight (CyTOF) (Hyperion Imaging System; SBI) at the LCCC Flow Cytometry Shared Resource (FCSR). Metal-conjugated antibody dilutions were optimized in the pancreas TMA prior to final IMC staining at the HTSR. Antigen retrieval and staining was performed at the LCCC HTSR, while image acquisition was performed at the LCCC Mass Spectrometry and Analytical Pharmacology Shared Resource. A maximum of 2.25 mm² regions-of-interest (ROIs) on TMA cores were selected, laser-ablated and analyzed by CyTOF to quantitate the metal-conjugated antibodies per ROI, generating multiplex images (24). All data were collected using the Hyperion Imaging System (SBI) and qualitatively validated in MCD Viewer (v1.0.560.6; SBI) and QuPath softwares (RRID:SCR_018257).

IMC image processing, data transformation, and single-cell clustering

Raw IMC data was processed following the widely used ImcSegmentationPipeline (25) (https://github.com/BodenmillerGroup/ImcSegmentationPipeline). In the Ilastik software (RRID:SCR_015246), DNA, E-cadherin, pan-cytokeratin, vimentin, and CD45RO markers were used to pixel-train and generate cell segmentation probability maps for nuclear, cytoplasmic, cell membrane, and background areas in all IMC images, which were imported into CellProfiler to generate cell segmentation masks to then extract single-cell protein expression data using histoCAT (26). 412,862 cells were identified across 44 images from 21 PDAC patient samples. Extracted IMC-derived single cell data were imported into R (v4.2.1) for downstream computational analyses. histoCAT-derived single cell data files generated are available at doi: 10.5281/zenodo.10582038. All scripts adapted from this pipeline for downstream analyses are available at https://github.com/Weiner-Lab/NaturalKillerCells_PDAC. Single cell protein expression data was log-transformed, re-integrated, and clustered using Seurat-based analysis (27). As single cell expression may contain “lateral spillover” into neighboring single cells as a result of cell segmentation (25), we therefore used a supervised clustering approach to phenotype cells based on canonical cellular morphology markers (αSMA, CD3, CD4, CD8α, CD11b, CD16, CD20, CD45RO, CD56, CD68, E-cadherin, FAP, FOXP3, NKG2D, NKp44, pan-cytokeratin, vimentin), we found that a clustering resolution = 0.35 provided distinct immune, fibroblast, and epithelial cell populations. Each cell population was subsetted and re-clustered to identify the functional state of subclusters using the following markers: CD69, CD107α, cleaved-caspase 3, collagen-I, granzyme B, HLA-A/B/C, HLA-A1, IFNγ, Ki67, MICA, PCNA, perforin, PD-1, PD-L1, and TGFβ. All cellular subclusters were reintegrated and annotated based on functional classification: active subpopulations were solely immune cell based and were characterized by high expression of known immune cell activation markers, CD69, CD107α, and/or IFNγ; proliferative cell populations were characterized by high expression of proliferation markers Ki67 and/or PCNA; and inhibitory subpopulations were characterized by high TGFβ expression, which is known to be immune suppressive. Subclustering identified 29 unique cell populations, which were validated visually using the Cytomapper R package and the MCD Viewer software. Seurat objects containing all the PDAC single cell metadata is available at doi: 10.5281/zenodo.10582038 and were converted to SingleCellExperiment (25) objects for downstream spatial analyses.

IMC spatial analysis

We generated knn-based interaction graphs for the 15-nearest neighbors and used k-means clustering to group cells into 8 distinct cellular neighborhoods using the SpatialExperiment and imcRtools R packages (25). Using the imcRtools package, permutation test-based cell-cell interaction analysis identified significant interactions or avoidances between cell populations. Using FDR analysis, cell-cell interactions with p < 0.01 were considered statistically significant. Based on Euclidean distance metadata, distance and network analyses were performed, as previously described (28). Seurat’s CellSelector function (27) was used to select and quantify cells in epithelial-ductal ROIs of spatial images. From CellSelector, percentages of cell populations in each epithelial-ductal ROI were quantified, and epithelial-ductal ROIs were stratified in quartiles based on fibroblast abundance (Q1 = low fibroblast quartile, Q4 = high fibroblast quartile). Mean total expression of all markers was quantified in representative epithelial-ductal ROIs of pseudoimages using MCD Viewer. Cell-based expression in epithelial-ductal ROIs was performed using Seurat.

Multiplex immunohistochemistry (mIHC)

Sections of 5μm thickness were cut from FFPE tissue blocks containing PDAC tumor samples. FFPE samples were provided by the Lombardi Comprehensive Cancer Center (LCCC) Histopathology & Tissue Shared Resource (HTSR). Patient samples were from 10 males and 7 females with age ranges between 45-88, and had tumor diameters ranging from 0.1-8 cm (Supplementary Table S3). Samples were stained for CD56, CD16, CD335, panCK, αSMA, and CD68 (Supplementary Table S4).

The slides were baked at 60°, deparaffinized in xylene, rehydrated, washed in distilled water and incubated with 10% neutral buffered formalin (NBF) for an additional 20 minutes to increase tissue-slide retention. Epitope retrieval/microwave treatment (MWT) for all antibodies was performed by boiling slides in the respective Epitope Retrieval buffer (ER1, pH6 or ER2, pH9; Leica Biosystems AR9961 or AR9640, respectively). Protein blocking was performed using antibody diluent/blocking buffer (Akoya, ARD1001EA) for 10 minutes at room temperature. Primary antibody/OPAL dye pairings, incubation conditions, and catalog numbers are listed in Supplementary Table S4. The staining was performed using a Leica Bond autostainer. Sections were counterstained with spectral DAPI (Akoya FP1490) for 5 minutes and mounted with ProLong Diamond Antifade (ThermoFisher, CAT#: P36961) using StatLab #1 coverslips (CAT#: CV102450). The order of antibody staining and the antibody/OPAL pairing was predetermined using general guidelines and the particular biology of the panel. General guidelines include spectrally separating co-localizing markers and separating spectrally adjacent dyes. Multiplex IHC was optimized by first performing single-plex IHC with the chosen antibody/OPAL dye pair to optimize signal intensity values and proper cellular expression, followed by optimizing the full multiplex assay. Slides were scanned at 10X magnification using the Phenoimager Fusion system Automated Quantitative Pathology Imaging System (Akoya Biosciences). Whole slide scans were viewed with Phenochart (Phenochart 2.2.0, Akoya) which also allows for the selection of high-powered images at 20X (resolution of 0.5mm per pixel) for multispectral image capture. Multispectral images of each tissue specimen were captured in their entirety.

Multiplex staining and imaging were performed as previously described (29), except tissue images were captured using Akoya Biosciences Inc.’s PhenoImager Fusion imaging platform at the LCCC HTSR. Images were exported as .qptiff files and visualized and processed using QuPath software (RRID:SCR_018257).

mIHC image processing and analysis

mIHC images were processed using QuPath software (v0.6.0) (30). Spectrally unmixed images were visualized and a ROI grid, with each ROI measuring 500 μm via 500 μm, was overlaid on each image. Cells within each ROI were then automatically segmented using the StarDist machine-learning based pipeline (31), with the DAPI channel used to delineate nuclear staining, normalized percentile set at (1, 99), threshold set at 0.5, and pixel size set at 0.5 μm/pixel. After cell segmentation, duplicate channel training images were created for each marker to create an object classifier for each cell type. Under the direction of an experienced pathologist, an object classifier was trained for each marker. This was repeated for three separate images. Afterwards, object classifiers for each marker were then applied to each image for cell phenotyping. Percentage of cells was calculated by dividing cells detected for marker of interest by total cells in each ROI. All ROIs were then averaged to get a total value for each cell type of interest for each image. Only ROIs with at least 500 cells were analyzed, with biologically incompatible markers excluded from analysis. Graphs of resulting data and statistical analysis were generated using GraphPad Prism v10 software.

Single-cell RNA-sequencing analysis

To infer ligand-receptor (L-R) interactions in PDAC, we utilized a previously published human PDAC scRNAseq dataset (accession number: GSE155698) (23) and subsetted phenotyped NK cells, fibroblasts, and epithelial cells, using the Seurat R package. Malignant and normal epithelial cells were further differentiated using the CytoTRACE R package, a computational framework that predicts cellular differentiation states by measuring the number of genes expressed by each cell (32). The CytoTRACE scores distribution was observed to be bi-modal and using a mixed-model, an ideal cut-off value separating the two distributions was determined to be 0.72 where scores greater than 0.72 indicated a malignant epithelial cells (33). Using known fibroblasts markers, fibroblasts were subclustered into known subpopulations and we identified: antigen-presenting CAFs (apCAFs), inflammatory CAF (iCAFs), myofibroblasts (myCAFs), normal fibroblasts, and stellate-like fibroblasts (34). We then used CellChat, a manually curated protein database of approximately 2,000 L-R pairs (35) to infer potential L-R interactions between epithelial, fibroblast, and NK cell populations. A Wilcox test with a Bonferroni correction from the FindMarkers Seurat function (v4.3) was used for differential gene expression analysis of CellChat-identified L-R genes. Genes with an adjusted-P < 0.01 were considered significantly differentially expressed. R scripts for applied CytoTRACE, fibroblast subclustering, and CellChat analyses are available at https://github.com/Weiner-Lab/NaturalKillerCells_PDAC.

Cell culture

Human PANC-1 and ASPC-1 PDAC cells were obtained from the Georgetown LCCC Tissue Culture and Biobanking shared resource (obtained in 2022 and 2025, respectively) and cultured in supplemented DMEM (Fisher, CAT#: SH30022LS) (10% FBS, 1% penicillin/streptomycin (ThermoFisher, CAT#: 10378016)). BT-474 cells were kindly provided by Dr. Rebecca Riggins (Georgetown University LCCC, Washington DC, USA, obtained in 2025) and cultured in modified IMEM (Fisher, CAT#: A1048901) (10% FBS, 1% penicillin/streptomycin, 2mM L-glutamine (ThermoFisher, CAT#: 10378016)). All cell lines were tested and determined to be free of mycoplasma (IMPACT II profile); no other authentication assay was performed. Donor NK cells were cultured, as previously described (36,37), except donor NK cells received 0.5% v/v of IL-2 (PeproTech, CAT#: 200-02-50UG). Healthy human donor NK cell lines, Donor NK #1, #2, and #4 were purchased (STEMCELL Technologies, Vancouver, Canada), and Donor NK #3 cells were freshly isolated from a leukoreduction system (LRS) chamber (Inova Health System) and isolated via MojoSortTM Human NK Cell Isolation Kit (BioLegend, CAT#: 480053). Donor NK cells were stimulated with the engineered K562 cell line (RRID:CVCL_0004), K562-mb21-41BBL (courtesy of Dr. Dean Lee at Nationwide Children’s) (38). K562-mb21-41BBL cells were cultured in supplemented RPMI (Fisher, CAT#: SH30027FS) (10% FBS, 1% penicillin/streptomycin). For NK cell stimulation, K562-mb21-41BBL cells(courtesy of Dr. Dean Lee, Nationwide Children’s) were collected (5 minutes, 1,000 rpm, 4°C), resuspended in supplemented RPMI with mitomycin C (Sigma-Aldrich, CAT#: 10107409001) (C_f = 0.01 mg/ml) and incubated for 4 hours (37°C). Pre-treated K562-mb21-41BBL cells were collected (5 minutes, 1,000 rpm, 4°C), resuspended in NK cell medium, then co-cultured with donor NK cells (effector:target (E:T) = 1:1); this process was repeated at least once per week. Donor NK cells aged day 1-50 and stimulated 1-3 days prior to in vitro assays were used. Primary NK cells were enriched from peripheral blood from LRS chambers, both male and female, using RosetteSep (StemCell Technologies, CAT#: 15065) and maintained in culture with IL-15 (Miltenyl Biotec, CAT#: 13030-095-764) (1ng/mL) (39). LRS chambers were obtained by a protocol provided by Dr. Miriam Jacobs, approved by the Institutional Review Board of Georgetown University.

Flow cytometry

Collected cells were centrifuged (5 minutes, 1,000rpm, 4°C) washed twice (1X PBS (Fisher, CAT#: MT21040CV)), then resuspended in FACS staining buffer (1% BSA (Millipore, CAT#: A7906-100G) in 1X PBS). Human Fc block (BD Pharmingen, CAT#: 564219) then α-CD44-BV605 (BioLegend, CAT#: 103047) antibodies were added per manufacturer’s recommendations (30 minutes, 4°C). Cell pellets were washed twice (FACS staining buffer), resuspended in fixative (1% PFA (ThermoFisher, CAT#: J61899) in 1X PBS), then processed by the LCCC FCSR in a BD LSRFortessa Cell Analyzer (BD Biosciences). Analysis was performed using FlowJo (v10.8.1; RRID:SCR_008520).

NK cell 2D invasion assay

For invasion assays, the bottom of 5 μm Transwells (Corning, CAT#: 3421) were coated with Matrigel (Corning, CAT#: CB-40234) diluted 1:3 in NK cell medium, Fibronectin (Corning, CAT#: 354008), or Laminin (Corning, CAT#: 354239). CXCL9 (R&D Systems, CAT#: 392-MG) (C_f =100 ng/ml)-supplemented NK cell medium was added to the bottom of each well. Donor NK cells were added to Transwells with or without an CD44 neutralizing antibody (clone IM7; Bio X Cell; C_f = 10 μg/ml) and incubated for 4 hours (37°C). Transwells were removed, cells were stained with Trypan Blue Stain (Gibco, CAT#: 15250061) and counted. Relative invasion is the percent of cells that invaded through the transwell normalized to untreated NK cell conditions. Graphs of resulting data and statistical analysis were generated using GraphPad Prism v10 software.

NK cell 3D spheroid invasion assay

As previously described (39), PANC-1 cells were seeded in each well of a low-adhesion U-bottom plate, spun (5 minutes, 1,000 rpm, 4°C) and cultured overnight. Spheroids were embedded with donor NK cells in an 80:20 Collagen-I (Invitrogen, CAT#: 354236): Matrigel (Corning) mixture in chamber wells (Lab-Tek). Embedded spheroids were treated with α-CD44 (clone IM7; Bio X Cell; C_f = 10 μg/ml) for 24 hours (37°C) prior to immunocytochemistry (IC). For IC, samples were fixed with 4% paraformaldehyde (ThermoFisher, CAT#: J61899) (PFA; 10 minutes, 25°C), washed with 1X PBS (5 minutes), then permeabilized with 0.5% Triton-X100 (Fisher, CAT#: CAS 9002-93-1) (10 minutes). Cells were washed once, blocked (1% BSA in 1X PBS, 30 minutes), then incubated with human Fc block (BD Pharmigen, CAT#: 564219). Samples were incubated with pan-cytokeratin-eFluor^™ 660 (Invitrogen, CAT#: 50-9003-82) and CD56 (Invitrogen, CAT#: MA1-19129) antibodies overnight (25°C). An α-Mouse IgG2a-AF488 secondary antibody (Invitrogen, CAT#: A-21131; 1-4 hours, 25°C), cells were washed 3 times, incubated with DAPI (ThermoFisher, CAT#: D3571) (50 ng/ml; 10 minutes, 25°C), then washed twice. Coverslips were mounted with anti-fade mountant, then and dried. Z-stacks of spheroids were imaged at 20X using a Leica SP8 AOBS microscope at the LCCC Microscopy and Imaging Shared Resource. Z-stacks were ≤ 1 μm. Z-stack images were processed and using FIJI software (40) (v2.14.0; RRID:SCR_002285) for manual NK cell quantitation into PANC-1 spheroids. Graphs of resulting data and statistical analysis were generated using GraphPad Prism v10 software.

NK cell cytotoxicity and ADCC assays

For NK cell cytotoxicity assays, PANC-1 cells were seeded in a treated 24-well, clear plate (ThermoFisher, CAT#: FB012929) at 50,000 cells per well and then incubated overnight at 37°C in 5% CO₂. After 24 hours, eight groups were added in triplicate: 1) Medium (Target); 2) 400,000 primary NK cells from a healthy donor to PANC-1 for a 4:1 E:T ratio (Effector); 3) Medium (Target) with 10 μg/mL CD44 (clone IM7; Bio X Cell; C_f = 10 μg/ml); 4) 400,000 primary NK cells from a healthy donor to BT474 for a 4:1 E:T ratio (Effector) with 10 μg/mL CD44. Percent viability was determined 4 hour post exposure to primary NK cells in the presence or absence of 10 μg/mL CD44 by measuring live and dead cell counts. Target cells were harvested with 0.05% Trypsin-EDTA (ThermoFisherFisher, CAT#: 25300062) and counted using a LUNA-II Automated Cell Counter (ThermoFisher) with live cell counts, dead cell counts, and cell viability identified by Trypan Blue. The percent cytotoxicity of the target cells was calculated as [(percent viability_0:1 – percent viability_4:1)/percent viability_0:1]*100. Graphs of resulting data and statistical analysis were generated using GraphPad Prism v10 software.

For NK cell ADCC assays, BT474 cells were seeded in a treated 24-well, clear plate (ThermoFisher, CAT#: FB012929) at 50,000 cells per well and then incubated overnight at 37°C in 5% CO₂. After 24 hours treatment eight groups were added in triplicate: 1) Medium (Target); 2) 1 μg/mL trastuzumab (MedChemExpress, CAT#: HY-P9907) for BT474; 3) 400,000 primary NK cells from a healthy donor to BT474 for a 4:1 E:T ratio (Effector); 4) the combination of trastuzumab and NK treatment (ADCC); 5) Medium (Target) with 10 μg/mL CD44; 6) 1 μg/mL trastuzumab for BT474 with 10 μg/mL CD44; 7) 400,000 primary NK cells from a healthy donor to BT474 for a 4:1 E:T ratio (Effector) with 10 μg/mL CD44; 8) the combination of trastuzumab and NK treatment (ADCC) with 10 μg/mL CD44. Percent viability was determined 4-h post exposure to primary NK cells in the presence or absence of 1 μg/mL trastuzumab and in the presence or absence of 10 μg/mL CD44 by measuring live and dead cell counts. Target cells were harvested with 0.05% Trypsin-EDTA and counted using a LUNA-II Automated Cell Counter (ThermoFisher) with live cell counts, dead cell counts, and cell viability identified by Trypan Blue. The percent cytotoxicity of the target cells was calculated as [(percent viability_0:1 – percent viability_4:1)/percent viability_0:1]*100. Graphs of resulting data and statistical analysis were generated using GraphPad Prism v10 software.

MMP gene expression in human NK cells

Donor NK cells were incubated with or without α-CD44 (clone IM7; Bio X Cell; C_f = 10 μg/ml) for 1 or 4 hours. Cells were collected (5 minutes, 1,000 rpm, 4°C) and RNA was isolated using the PureLink^™ RNA Mini Kit (ThermoFisher, CAT#: 12183018A). RNA concentration was measured using NanoDrop 8000 (ThermoFisher), then stored at −80°C. Diluted RNA (20 ng/μl) was treated with 10% v/v 10X RQ1 DNase Reaction Buffer (Promega, CAT#: M198A) and 10% v/v RQ1 RNase-Free DNase (Promega, CAT#: M610A; 15 minutes, 25°C). 10% v/v of RQ1 RNase/DNase Stop Solution (Promega, CAT#: M199A) and 10% v/v Oligo(dT) nucleotides (IDT) were added, then heat inactivated at 65°C (10 minutes) then at 4°C (5 minutes). cDNA was synthesized with the GoScript^™ Reverse Transcriptase Kit (Promega, CAT#: A5001). qPCR assessing genes listed in Supplementary Table S5, along with primer sequences, was performed using GoTaq qPCR Master Mix (Promega, CAT#: A6001) with the StepOnePlus^™ Real-Time PCR System (ThermoFisher). Gene expression was quantified (2^−ΔCт) using HPRT (Forward 3’→5’: CTTTCCTTGGTCAGGCAGTA, reverse 3’→5’: TGGCTTATATCCAACACTTCG) as an internal, endogenous control. Fold-change was calculated using the 2^−ΔΔCт method and graphs of resulting data and statistical analysis were generated using GraphPad Prism v10 software. All samples were tested in duplicate and each experiment was repeated at least once.

MMP protein expression in human NK cells

Primary donor NK cells were incubated with or without α-CD44 (clone IM7; Bio X Cell; C_f = 10 μg/ml) for 1 or 4 hours. Western blots were performed as previously described (39). Briefly, cell lysates were extracted with 1% NP-40 lysis buffer containing a protease and phosphatase inhibitors cocktail (Thermo Scientific, CAT#: 87785). Cell lysates were kept on ice for 30 minutes and centrifuged at 13,200 rpm for 10 minutes at 4 °C. Protein concentrations were measured by the Bio-Rad protein assay (Bio-Rad, CAT#: 5000006). A 4X Laemmli sample buffer (Bio-Rad, CAT#: 1610747) containing 2-mercaptoethanol (Bio-Rad, CAT#: 1610710) was added to the 10ug cell lysates and boiled at 60 °C for 10 min. Protein samples were separated on 4%-12% Bis-Tris gels (GenScript, CAT#: M00652) and transferred to polyvinylidene difluoride (PVDF) membranes (Millipore, CAT#: IPVH00010). The membranes were blocked with 5% skim milk in TBS-T (0.1% Tween-20) for 1 hour at room temperature and incubated overnight at 4 °C with primary antibodies: anti-MMP1 (1:1000, Invitrogen, CAT#: MA5-15872), anti-MMP2 (1:1000, Invitrogen, CAT#: 436000), anti-MMP9 (1:2000, ProteinTech, CAT#: #60600-1-lg), anti-MMP14 (1:1000, Invitrogen, CAT#: MA5-32076), and β-actin (1:1000, Santa Cruz Biotechnology, CAT#: sc-47778). The next day, the membrane was probed for 1 hour at room temperature with an appropriate secondary antibody conjugated to horseradish peroxidase. Clarity Western ECL Substrate (Bio-Rad, CAT#: 170-5061) was applied for membrane development, and images were analyzed using the ChemiDoc MP system (Bio-Rad).

Generation of FAP and MMP14 overexpressing cells

MMP14 and FAP were overexpressed in primary donor NK cells as previously described(39). Briefly, MMP14 or FAP(Supplementary Table S6 and S7) were overexpressed (OE) using a third-generation lentiviral system (packaging plasmids: pMDlg, pRSV-rev, pCMV-VSVG). Lentivirus was produced using HEK293T cells and was concentrated using Amicon Ultra tubes (Millipore, CAT#: UFC910024). NK cells were spinfected with the virus 24 hours after isolation. Cells were used 3 to 10 days following infection.

MMP14 overexpression and in vivo NK cell infiltration

ASPC-1 human tumor cells were implanted directly into the pancreas of NSG (NOD.Cg-Prkdc^scid Il2rg^tm1Wjl/SzJ) mice using ultrasound guided injection. Ultrasound guided injection was performed by the Georgetown LCCC Animal Models Shared Resource. Briefly, mice were anesthetized using isoflurane in oxygen (3% for induction, 1.5-2.5% for maintenance). Mice were then placed on the THM150 system (FujiFilm) mouse pad, which comprises a temperature and ECG control and monitoring. Hair was removed and skin surface sterilized. Sterile ultrasound gel was applied to the area above the injection site. The ultrasound transducer (VEVO 3100 Imaging Station, FujiFilm) was moved into position to clearly visualize the pancreas and needle guide function was then activated. The needle guide system then injected 50 μL of cell solution containing 0.2x10⁶ tumors cells into the pancreas. After injection, gel was cleaned off, mouse was removed from the mouse pad, placed in a clean cage and monitored for recovery.

Mice were monitored regularly by ultrasound and when tumors were 150 mm³, mice were injected, by intravenous (i.v.) injection, with PBS, 1x10⁶ of NK cells, 1x10⁶ FAP OE NK cells, or 1x10⁶ MMP14 OE NK cells, all with 60 ng/mL IL-2 and 10 ng/mL IL-15. Seventy-two hours after injection, the same treatment was repeated. Twenty-four hours after the second treatment, mice were euthanized and the tumors were excised. The tumors were submitted to the Georgetown LCCC Histopathology and Tissue Shared Resource Core. Slides were stained for CD56 (Abcam, CAT#: ab133345) at 1:800, αSMA (Abcam, CAT#: ab124964) at 1:800, and Masson’s Trichrome for collagen. Images were then analyzed using QuPath software.

Ethics statement

The human pancreas TMA and PDAC slides used in this study were approved by the Georgetown University LCCC Biospecimen Use Committee. NSG animal study was reviewed and approved by the Georgetown University Institutional Animal Care and Use Committee.

Code and data availability statement

Code used to perform IMC and scRNAseq analyses and generate the figures is available at https://github.com/Weiner-Lab/NaturalKillerCells_PDAC. Raw .mcd files for all cores of the pancreas tissue microarray imaged by IMC and IMC-derived datasets, including histoCAT-derived single cell files, and Seurat, SingleCellExperiment, Cytomapper and distance data matrix R objects, to reproduce figures are deposited at doi: 10.5281/zenodo.10582038. The scRNAseq data was provided to us by the Pasca Di Magliano lab at the University of Michigan, which was previously published at doi: 10.1038/s43018-020-00121-4 (accession number: GSE155698). The authors declare that all other data supporting the findings of this study are available within the paper or its supplementary information files. All other relevant data are available from the corresponding author upon reasonable request.

Results

Fibroblast composition organizes spatial architectures that determine immune access to malignant epithelial ducts in PDAC

We first performed IMC on a tissue microarray (TMA) of primary tumor samples from 21 treatment-naïve PDAC patients to define the major cellular compartments of the tumor microenvironment (TME). IMC is a high-dimensional spatial proteomics platform that maintains tissue architecture while capturing single-cell protein levels (41). We used a custom 40-marker metal-conjugated antibody panel targeting cells from epithelial, immune and stromal lineages (Supplementary Table S2). Following imaging and cell segmentation, we identified 412,862 cells in 44 IMC pseudoimages, with two or more primary tumor cores being imaged for some patient (Supplementary Table S1). Using a Seurat-based analysis pipeline (26,27,42-44), single-cell protein expression data were log-normalized, and cell populations were identified using clustering approaches (45), which resolved eight principal cell types encompassing malignant epithelial cells, cancer-associated fibroblasts (CAFs), leukocytes, and other stromal populations (Supplementary Fig. S1). Subclustering distinguished functional states: active, proliferative, or inhibitory, within each lineage (Supplementary Fig. S2). This cellular map established the framework for subsequent analyses of how CAF organization around malignant ducts constrains leukocyte access within the PDAC TME.

Having defined major cellular lineages by IMC, we next examined how these populations are spatially organized within the PDAC TME. Using cellular neighborhood analysis (46), we classified distinct multi-cellular structures based on their cellular composition and proximity (46). This analysis identified eight distinct cellular neighborhoods (Figs. 1A-B) each defined by characteristic combinations of epithelial, fibroblast, and immune populations. Fibroblast composition influenced neighborhood architecture: FAP⁺ fibroblast-rich regions (neighborhood 6) were associated with proliferative epithelial areas, while αSMA⁺ fibroblast-rich regions (neighborhood 7) were enriched for CD4⁺/CD8⁺ T cells and macrophages (Fig. 1C). In contrast, immune-enriched neighborhood 4 had the majority of apoptotic epithelial cells and was largely depleted of FAP⁺ or αSMA⁺ fibroblasts, consistent with reduced stromal barriers in immune-accessible regions (Fig. 1C, Supplementary Fig. S3). Notably, neighborhood 2 was enriched for IFNγ⁺ positive epithelial cells, a marker of immune activation (47) (Fig. 1C, Supplementary Fig. S3). These epithelial neighborhoods also contained a higher density of leukocytes relative to fibroblasts (Supplementary Fig. S3), supporting that immune cell accumulation occurs preferentially in epithelial compartments with reduced stromal architecture. Consistent with previous reports in untreated PDAC (11), adaptive lymphoid cells (CD4⁺/CD8⁺ T cells, B cells) were confined to their own distinct neighborhood (3) and were spatially segregated from epithelial rich regions (Fig. 1C). In contrast, activated NK cells, based on Granzyme B expression, were observed in epithelial-dominant neighborhoods (1, 2, 6, and 8), suggesting that innate immune cells mediate the primary immune interactions at malignant ducts.

Figure 1: Fibroblast composition organizes spatial architectures that determine immune access to malignant epithelial ducts in PDAC.

A. Corresponding patient IMC pseudoimage for cellular neighborhood (CN) spatial plot in B (αSMA = blue; E-cadherin, Pan-cytokeratin = pink; NKG2D = yellow, scale bar = 200 μm). B. Representative spatial plot of CNs in the human PDAC TME (CN1 = teal, CN2 = yellow, CN3 = purple, CN4 = red, CN5 = blue, CN6 = orange, CN7 = lime, CN8 = pink). C. CN analysis revealed 8 distinct cellular neighborhoods (y-axis: cellular neighborhoods; red = high proportion of cell subpopulation of interest (x-axis) within cellular neighborhoods, blue = low proportion of cell subpopulation of interest within cellular neighborhoods). D. Cellular interaction analysis (blue = strong avoidance, red = strong interaction; FDR, p < 0.01) of IMC-defined cell subpopulations. E. Network graph of the top 75% of interactions between all IMC-defined cell populations in human PDAC showing epithelial cells and NK cells exist in similar networks.

Although CAF-mediated immune exclusion has been reported in PDAC (9,16), our analysis defines how distinct fibroblast subsets establish reproducible spatial architectures that dictate the degree of immune access to malignant ducts. Epithelial dominant neighborhoods (1, 2, 6, and 8) displayed variable leukocyte abundance depending on surrounding fibroblast density (Fig. 1C). Collectively, these analyses reveal that immune accessibility in PDAC is determined not by total stromal content, but by the local fibroblast organization encasing malignant epithelial ducts, prompting a focused analysis of epithelial-ductal regions to quantify how fibroblast abundance constrains leukocyte access within the TME.

Given that neighborhood analysis captures higher-order organization but not direct cellular interactions, we next applied a permutation-based interaction analysis(26), which identifies significant deviations from a random cellular arrangement, to infer colocalization between PDAC cellular subsets identified via IMC. Specifically, this analysis reveals which cellular subsets are spatially interacting more or less frequently than expected due to random variation. As expected, epithelial cells, fibroblasts, and immune cells tended to localize with cells of the same type (Fig. 1D). Confirming neighborhood analysis, activated NK cells, but not T cells, frequently interacted with apoptotic epithelial cells (Fig. 1D) while activated CD4⁺ and CD8⁺ T cells primarily associated with one another rather than with epithelial or stromal cells (Fig. 1D). Macrophages, on the other hand, most often interacted with αSMA⁺ fibroblasts (Fig. 1D).

To visualize these multicellular relationships within the TME, we performed a cell–cell network analysis defined from the average distance of spatial neighbors(28), which shows that macrophage and αSMA⁺ fibroblast nodes have the closest cross-lineage connectivity, whereas NK cell nodes clustered near epithelial cell nodes, distinct from CD4⁺/CD8⁺ T cell and B cell networks (Fig. 1E). This network structure suggests that fibroblast organization not only structures epithelial–stromal interfaces but also constrains immune cell localization within the PDAC TME.

Validation in a completely independent cohort of PDAC tumors (n = 17) using multiplex immunohistochemistry (mIHC) confirmed the spatial relationships observed by IMC (Supplementary Fig. S4A). Fibroblasts and epithelial cells constituted the majority of the PDAC TME (Supplementary Fig. S4B). Notably, αSMA⁺ fibroblast content was inversely correlated with NK cell, but not macrophage, abundance (Supplementary Fig. S4C). CD16⁺ NK cells and macrophages, consistent with a pro-inflammatory phenotype (48,49), were reduced with higher αSMA⁺ fibroblast content in tumors, supporting cell-cell network analysis and suggesting a selective exclusion of effector leukocytes in fibroblast-rich regions (Fig. 1E). Overall, these analyses demonstrate that fibroblast organization determines immune architecture in PDAC, limiting the spatial access of pro-inflammatory, cytotoxic leukocytes to malignant epithelial ducts.

CAF-stratified ductal spatial architecture governs immune access in PDAC

Given that increased αSMA⁺ fibroblasts are inversely correlated with effector leukocyte infiltration (Supplementary Fig. S4C), we next sought to understand how per-ductal fibroblasts influence leukocyte infiltration in regions near the duct. Using epithelial-ductal regions of interest (ROIs) defined from IMC-based cellular neighborhood analysis, we applied Seurat’s (27) CellSelector function to isolate and quantify immune and stromal populations across 952 epithelial-ductal ROIs. These ROIs were stratified into 4 quartiles (Q1-Q4) according to fibroblast abundance, with Q1 representing fibroblast-low and Q4 representing fibroblast-high ducts. This stratification revealed a consistent CAF-stratified ductal spatial architecture (Figs. 2A-B) where increasing fibroblast content corresponded to progressive restriction of leukocyte–epithelial proximity. Across quartiles, fibroblast abundance increased significantly (Fig. 2C). A majority of these fibroblasts were αSMA⁺ fibroblasts, with few FAP⁺ fibroblasts present (Supplementary Fig. S5), supporting previous reports that αSMA⁺ fibroblasts cluster in periglandular regions (50). As expected, increasing αSMA⁺ fibroblast abundance was associated with a corresponding rise in Collagen I expression (Fig. 2D, Supplementary Fig. S5), consistent with fibroblast-driven matrix deposition (51). While total epithelial cells decreased with increasing fibroblast content (Fig. 2E), the proportion of proliferative epithelial cells increased (Fig. 2E). This observation is in contrast to the separation between proliferating ductal cells and αSMA⁺ fibroblasts observed in our IMC data. This suggests that when analyzed at the ductal level, increased periductal αSMA⁺ fibroblast content supports an increased proportion of proliferating epithelial cells that remain trapped along the duct and are not able to infiltrate into the surrounding regions as would be facilitated with ducts surrounded by a less dense stroma.

Figure 2: CAF-stratified ductal spatial architecture govern immune access in PDAC.

A. Representative IMC pseudoimages showing αSMA (left, light blue) or collagen-I (right, dark blue) with pan-cytokeratin (panCK, pink), E-cadherin (Ecad, pink), and NKG2D (yellow); scale bar = 100 μm with white boxes of each fibroblast abundance quartile with corresponding numbers. B. Insets of representative epithelial-ductal ROIs from each fibroblast abundance quartile (Q1 = 238 ducts, Q2 = 235 ducts, Q3 = 240 ducts, Q4 = 239 ducts) shown in the white insets in A. Bar plot of percent cells/ducts (mean ± SEM) of C. total fibroblasts, and D. total collagen-I expression in epithelial-ductal ROIs ranked by fibroblast abundance quartile (2-way ANOVA; *p < 0.05, ** p < 0.01, ***p < 0.001, ****p < 0.0001; Q1 = 55 ducts, Q2 = 58 ducts, Q3 = 57 ducts, Q4 = 60 ducts). E. Connecting line plot of percent cells/ducts (mean ± SEM) of total epithelial cells and proliferative epithelial cells (2-way ANOVA; *p < 0.05, ****p < 0.0001; all quartiles compared to Q1). F. Connecting line plot of percent cells/ducts (mean ± SEM) of NK cells, CD4⁺ T Cells, and CD8⁺ T Cells (2-way ANOVA; *p < 0.05, ****p < 0.0001; all quartiles compared to Q1). G. Violin plot showing NK cell-based expression in Q1 and Q4 fibroblast abundance quartiles of Granzyme B (Kruskal-Wallis test).

We next assessed the impact of αSMA⁺ fibroblasts on the spatial associations of activated leukocytes with ductal epithelial cells. Expression of NKG2D, an activating receptor expressed on human lymphocytes (52), declined as αSMA⁺ fibroblast and collagen content increased (Fig. 2B). Among immune subsets, NK cells showed the strongest sensitivity to αSMA⁺ fibroblast density; their abundance was highest in fibroblast-low ducts (Q1) and decreased progressively across quartiles (Fig. 2F), indicating fibroblast-dependent exclusion of cytotoxic innate lymphocytes. In contrast, CD4⁺ and CD8⁺ T cells (% cells/duct) remained relatively stable (Fig. 2F), consistent with their preferential localization to stromal rather than malignant ductal epithelium (53). The loss of NK cell infiltration was accompanied by reduced IFNγ⁺ epithelial cells (Supplementary Fig. S6), implicating diminished NK-epithelial cell interactions and associated cytolytic signaling. In fibroblast-sparse ducts (Q1-Q2), NK cells retained high expression of activation markers (CD69, CD107a, IFNγ) (Supplementary Fig. S7) and displayed significant increases in Granzyme B expression (Fig. 2G), confirming preserved effector function where malignant epithelial cell access is maintained.

Collectively, these findings establish CAF-stratified ductal spatial architecture as a defining architecture feature of PDAC, whereby periductal αSMA⁺ fibroblast density promotes malignant epithelial proliferation while restricting NK cell infiltration and activation. This represents a mechanism of immune evasion in PDAC and suggests that periductal stromal remodeling to reduce αSMA⁺ fibroblast density or collagen deposition may be required to enable effective lymphocyte-based immunotherapies.

Cell-cell communication analysis implicates CD44 as a stromal checkpoint mediating NK-ECM interactions

Because NK cell access to malignant ducts was inversely correlated with periductal αSMA⁺ fibroblast density (Figs. 2B-G), we next sought to identify the molecular pathways that could mediate these fibroblast-NK cell interactions within the PDAC tumor microenvironment. We applied ligand-receptor inference of cell to cell communication with CellChat (35) to a reference immune-enriched human PDAC scRNAseq dataset containing matched adjacent-normal (AdjNorm) and PDAC tumor samples (23). To distinguish malignant from normal epithelial cells, we used CytoTRACE to infer differentiation states (32) (54). We also subclustered fibroblasts to resolve canonical CAF subsets, including myofibroblasts (myCAFs; FAP, MMP11, HOPX, POSTN, COL12A1) (9), inflammatory (iCAFs; CFD, DPT, AGTR1, CXCL12, CCL2) (9), antigen-presenting (apCAFs; HLA-DRA, HLA-DPA1, HLA-DQA1) (9), normal fibroblasts, and stellate-like fibroblasts (MYH11, RGS5) (34,55) (Supplementary Fig. S8).

Because NK cells were the immune population most affected by periductal fibroblast density (Figs. 2F-G), we next examined whether fibroblasts engage in molecular signaling that could underlie this selective exclusion. CellChat analysis revealed that CAFs in PDAC, particularly iCAFs and myCAFs, exhibited increased outgoing signaling compared with adjacent-normal fibroblasts, whereas NK cells and malignant epithelial cells served as prominent signal receivers (Figs. 3A-B). This suggests that CAF-derived signaling networks may directly modulate NK cell behavior and epithelial interactions, extending their role beyond structural confinement.

Figure 3: Cell-cell communication analysis implicates CD44 as a stromal checkpoint mediating NK-ECM interactions.

A. Incoming and outgoing interaction strength for each cell population grouped by AdjNorm and PDAC disease states. B. Heatmap of differential interaction strengths of target (y-axis) and source (x-axis) cells with bar plots displaying the absolute values of interactions. C. Communication probabilities of signaling pathways targeting NK cells between AdjNorm and PDAC disease states as measured by information flow (pathway legend: pink = AdjNorm-specific, teal = PDAC-specific, black = equally specific). D. Significant ligand-receptor pairs from the COLLAGEN signaling pathways targeting NK cells upregulated in PDAC compared to AdjNorm (interaction legend (x-axis): pink = AdjNorm, teal = PDAC).

Among enriched signaling pathways, collagen signaling showed the largest PDAC-specific increase, driven by upregulation of multiple collagen isoforms (Fig. 3C). Within this network, NK cells were predicted to receive collagen-derived signals through CD44 (Fig. 3D), a principal receptor mediating cell–matrix adhesion on leukocytes (56). Differential expression analysis confirmed upregulated CD44 expression in NK cells within the PDAC TME, concurrent with transcriptional upregulation of collagen genes in malignant epithelial cells, iCAFs, and myCAFs (Fig. 3D, Supplementary Fig. S9).

Together, these findings implicate the collagen–CD44 axis as a critical CAF-to-NK communication route that reinforces stromal confinement of cytotoxic lymphocytes. This represents a stromal checkpoint mechanism in which NK cells are sequestered within collagen-rich regions. We hypothesize that this CAF-NK cell signaling pathway limits the migration of NK cells toward tumor targets. Hence, disrupting CD44–collagen interactions could restore NK-cell motility and enhance antitumor function in PDAC.

CD44 blockade enhances NK cell invasion in vitro

Given that NK cells in the PDAC TME primarily engage the collagen pathway through CD44 (Figs. 3C-D), we hypothesized that CD44-mediated interactions with ECM components physically restrict NK cell infiltration by anchoring them within the stroma. To test this, we first confirmed CD44 expression on human donor NK cells by flow cytometry (Supplementary Fig. S10). We then evaluated NK cell invasion through a collagen-containing matrix in vitro. In a 2D Transwell assay, membranes were coated with Matrigel (rich in collagen (57)), and NK cells were incubated with or without a CD44-neutralizing antibody (α-CD44) (Fig. 4A). CD44 blockade significantly enhanced NK cell invasion in three of four human donor NK cell samples (Figs. 4B-E). We repeated this experiment, including matrices devoid of collagen, and found that CD44 blockade significantly enhanced NK cell motility through fibronectin as well (Supplementary Fig. S11).

Figure 4: CD44 blockade enhances NK cell invasion in vitro.

A. Schematic for Transwell invasion assays. Created in BioRender. Malchiodi, Z. (2026) https://BioRender.com/iuq0xmu Average relative invasion (± SEM) of B. Donor NK #1 (n = 9), C. Donor NK #2 (n = 12), D. Donor NK #3 (n = 6), and E. Donor NK #4 (n = 10) upon CD44 neutralization. *p < 0.05 as determined by Wilcoxon matched-pairs signed rank test. F. Schematic of spheroid invasion assay. Created in BioRender. Malchiodi, Z. (2026) https://BioRender.com/b9uvsgm G. Representative 20X IF images of outlined (yellow) PANC-1 spheroids (Pan-cytokeratin; red) embedded with NK cells (CD56; green), pointed out with white arrows, treated with or without α-CD44. Scale bar = 250 μm. H. Zoomed inset of NK cells in the α-CD44 treatment group in G (white box). Average number of NK cells per PANC-1 spheroid (± SEM) from I. Donor NK #1 (− α-CD44, n = 53; + α-CD44, n = 42), J. Donor NK #2 (− α-CD44, n = 29; + α-CD44, n = 43), and K. Donor NK #3 (− α-CD44, n = 27; + α-CD44, n = 47); n = # of spheroids analyzed; *p < 0.05 as determined by Wilcoxon matched-pairs signed rank test.

To validate these findings in a more physiological context, we employed a 3D in vitro model using a collagen-rich tumor spheroid assay (Fig. 4F). Human PANC-1 spheroids were embedded with donor NK cells in a Collagen I–Matrigel matrix and cultured for 24 hours with or without α-CD44. Immunofluorescence imaging showed increased NK-cell penetration into PANC-1 spheroids upon CD44 blockade across three donor NK cells (Figs. 4G-K, Supplementary Fig. S12). Since CD44 is known to regulate NK cell activation, we tested whether CD44 blockade affected NK cell cytotoxicity. CD44 blockade did not impact NK cell cytotoxicity or ability to facilitate antibody dependent cell mediated cytotoxicity (ADCC) in vitro (Supplementary Fig. S13).

To determine whether this enhanced invasion involved CD44-dependent modulation of extracellular matrix degradation (58,59), we analyzed matrix metalloproteinase (MMP) gene and protein expression in human donor NK cells treated with or without α-CD44 for 1 or 4 hours (58,59). No significant differences were observed (Supplementary Figs. S14-15, Supplementary Table S5), strongly suggesting that CD44 neutralization facilitates NK cell invasion primarily by disrupting adhesion to collagen, rather than altering proteolytic activity.

MMP14 overexpression in NK cells reduces CAF abundance and collagen deposition in vivo

Since CD44 blockade appeared to enhance NK cell invasion independent of MMP upregulation, we next sought to investigate whether enhancing NK cells’ ability to degrade ECM would enhance NK cell invasion. To test this, we overexpressed two proteases with established collagenase activity, MMP14 and FAP, in primary human donor NK cells using lentiviral transduction. We then intravenously injected NK cells into NSG mice bearing orthotopic human ASPC-1 pancreatic tumors. Once tumors were ~150 mm³, mice received an intravenous injection of 1x10⁶ of NK cells or PBS, which was followed by another injection seventy-two hours later (Fig. 5A). Twenty-four hours after the second injection, tumors were harvested, and αSMA immunostaining revealed a significant reduction in αSMA⁺ CAFs in tumors treated with MMP14 or FAP overexpressing (OE) NK cells compared to parental NK cells or PBS (Figs. 5B-C). OE MMP14 resulted in a significant increase in NK cell infiltration, based on CD56 immunostaining, relative to control (Figs. 5B, D). Consistent with decreased αSMA⁺ CAF content, MMP14 OE treated tumors demonstrated a significant reduction in intra-tumoral fibrosis by Masson’s Trichrome staining (Figs. 5B, E). These data suggest that enhancing collagenase activity of NK cells, through MMP14 overexpression, can enhance adoptive NK cell infiltration and reduce αSMA⁺ CAF abundance, resulting in decreased collagen deposition.

Figure 5. MMP14 overexpression in NK cells reduces CAF abundance and collagen deposition in vivo.

A. Schematic of the experimental design. B. Representative immunohistochemistry images of αSMA and CD56 staining are shown for each treatment group (PBS, Parental, FAP OE, MMP14 OE). Masson’s trichrome staining demonstrates collagen deposition within the tumor microenvironment. Quantification of C. αSMA-positive cells per mm² (PBS, n = 38; Parental, n = 46; FAP OE, n = 46; MMP14 OE, n = 35), D. CD56-positive cells per mm² (PBS, n = 23; Parental, n = 27; FAP OE, n = 37; MMP14 OE, n = 19), and E. collagen staining (PBS, n = 3; Parental, n = 5; FAP OE, n = 7; MMP14 OE, n = 6); n = # of regions of interest analyzed. Each dot represents an individual region of interest. Line represents median. *p < 0.05, **p < 0.01, ***p < 0.0001 as determined by one-way ANOVA with multiple comparison test. ns, not statistically significant.

Together, these results provide a mechanistic rationale for reducing the impact of collagen on lymphocyte infiltration. Blocking CD44–collagen binding restores NK-cell mobility and invasive capacity, while MMP14 OE increases NK cell motility and reduces αSMA⁺ CAF abundance. Thus, targeting stromal checkpoints or enhancing lymphocyte collagenase activity present new therapeutic avenues for immunotherapy in PDAC.

Discussion

Recent reports characterizing immune-stromal interactions in PDAC using single-cell technologies have characterized the stromal heterogeneity and impact of CAFs on the PDAC TME (8-18). However, the impact of stromal architecture on CAF, tumor cell, and leukocyte interactions and leukocyte infiltration in malignant ductal regions, the critically important tumor-immune interface, remains unexplored. Here, we define CAF-stratified ductal spatial architecture as a key organizational feature of the PDAC TME that dictates leukocyte access to malignant epithelium. Utilizing an approach that combines IMC, mIHC, and scRNA-seq to map periductal cellular interactions, we have demonstrated that periductal fibroblast organization is a critical determinant of leukocyte infiltration in malignant ductal regions. IMC and mIHC analyses collectively revealed that fibroblast abundance inversely correlated with immune accessibility, with NK cells being the leukocyte subset most sensitive to fibrous density. These data establish periductal fibroblast composition, not total fibrosis, as the dominant structural determinant of immune exclusion in PDAC.

Cell-to-cell communication analysis revealed that CAFs function as the dominant senders of signals within the PDAC TME, engaging both epithelial and NK cells through collagen-mediated pathways. The collagen-CD44 axis emerged as the principal CAF-to-NK signaling pathway, positioning CD44 as a stromal checkpoint that restricts leukocyte motility and reinforces fibroblast-mediated immune exclusion at malignant ducts. Functionally, CD44 blockade restored NK cell infiltration through collagen-rich matrices without altering MMP expression, indicating that neutralizing CD44–collagen adhesion can overcome such barriers. Moreover, endowing lymphocytes with increase collagenase activity, through MMP14 overexpression, enabled rapid infiltration in collagen-rich PDAC models in vivo, with evidence of tumor cell damage only twenty-four hours after infusion. We have previously shown that such enhanced NK cell infiltration has significant antitumor effects in human PDAC xenograft models(39).

In PDAC, stromal elements, including CAF-derived chemokines (e.g., CXCL12) (60) and ECM-derived ligands, such as fibronectin that engage inhibitory receptors like ILT3 (61), have been shown to modulate immune suppression within the TME. For example, elevated stromal CXCL12 correlates with the exclusion of cytotoxic T cells from tumor nests, while fibronectin engagement of the inhibitory receptor ILT3 on intra-tumoral myeloid cells induces a suppressive phenotype, collectively demonstrating that stromal components can directly regulate immune positioning and activation. Furthermore, higher CAF content has been shown to be associated with increased epithelial cell tumorigenicity (22), consistent with our results. Together, these findings indicate that the desmoplastic stroma functions not only as a physical barrier but also as an active regulatory element, acting as a “stromal checkpoint,” in which CAF-derived and ECM-mediated signals restrict leukocyte access and engagement with malignant epithelial cells. Our data extend this framework by identifying impaired immune trafficking as a key regulatory mechanism mediated through CD44–ECM interactions.

Consistent with our results, previous work has demonstrated that higher CAF content has been associated with increased malignant epithelial cell tumorigenicity. Broad stromal ablation and ECM degradation have proven unsuccessful in restoring intra-tumoral leukocyte functionality or limiting tumor growth preclinically and clinically (62,63). However, our data highlight the potential of instead remodeling lymphocyte-ECM interactions in the PDAC TME by either targeting stromal checkpoints, such as CD44, or by enhancing lymphocyte collagenase activity through MMP overexpression. This builds on our previous work(39), further suggesting that approaches to augment lymphocyte motility and infiltration can enhance NK cell-based immunotherapies in PDAC.

One limitation of our study is that our IMC panel was limited to only 36 markers and thus we were not able to include additional markers to further classify tumor cell heterogeneity apart from looking at functional markers of apoptosis (cleaved caspase-3) and proliferation (Ki-67). Future studies, such as single cell spatial transcriptomics, will be needed to further explore the impact of tumor cell heterogeneity on neighborhood stratification.

In summary, our findings establish fibroblast composition as a central determinant of PDAC tissue architecture and immune accessibility. We define CAF-stratified malignant ductal spatial architecture as the structural basis of immune exclusion and identify CD44 as a molecular mediator of this stromal checkpoint. High CAF-containing ducts promote epithelial cell proliferation while restricting lymphocyte access, particularly NK cells, whereas lower CAF-containing ducts preserve effector function and cytolytic ability. Blocking CD44–ECM binding restores NK-cell mobility and invasive capacity, while MMP14 OE increases NK cell motility and infiltration in vivo. Thus, targeting lymphocyte-stromal interaction presents an opportunity to increase leukocyte infiltration and overcome ECM-mediated barriers in PDAC and other solid tumors.

Synopsis.

In PDAC, targeting stromal checkpoints or enhancing lymphocyte collagenase activity can increase leukocyte infiltration, irrespective of ECM degradation.