A Patient‐Derived Organ‐on‐Chip Platform for Modeling the Tumor Microenvironment and Drug Responses in Pancreatic Cancer

ABSTRACT

Pancreatic ductal adenocarcinoma (PDAC) is a fatal malignancy. Current conventional chemotherapeutics are inadequate in controlling the disease; hence, there is an urgent need for precision medicine. Ex vivo models that replicate the tumor and its microenvironment can advance precision medicine in PDAC. Patient‐derived organoids (PDOs) offer a promising solution by retaining the functional features of the tumor, allowing for individualized study of cancer biology and drug response. However, PDOs fall short in replicating the tumor microenvironment (TME), which includes various stromal and immune cells influencing tumor growth and chemoresistance. We hypothesize that combining PDO technology with organ‐on‐a‐chip (OoC) systems can enhance ex vivo cancer modeling. Here, we develop a patient‐derived platform by incorporating PDOs with key components of the TME (fibroblasts, endothelial cells, and immune cells) within a microfluidic system. This OoC model represents the crosstalk between cancer and stroma observed in PDAC in vivo. Targeting the stroma improves the effectiveness of standard chemotherapy in this OoC. Further, using this platform, we are able to model and assess the efficacy of immune checkpoint blockade for T cell cytotoxicity in PDAC. The OoC provides a pathophysiologically applicable system to support future investigations aimed at utilizing precision medicine and testing therapeutics in PDAC.

Researchers have developed a patient‐derived organ‐on‐a‐chip model for pancreatic cancer by integrating cancer cells with supportive stromal and immune cells inside a microfluidic device. This system mimics the tumor microenvironment, enabling personalized testing of chemotherapy and immunotherapy, and offering new insights into how targeting surrounding tissue may affect treatment responses.

1. Introduction

Pancreatic ductal adenocarcinoma (PDAC) is a fatal malignancy with poor survival outcomes, underscoring the inadequacy of conventional chemotherapy approaches [1]. The molecular heterogeneity of PDAC poses a significant challenge, contributing to inconsistent responses even to targeted therapies and underscoring the need for more personalized treatment approaches [2, 3, 4, 5, 6]. These variations in response present substantial obstacles to trial designs for promising agents that pass animal testing [7, 8, 9, 10, 11]. PDAC is particularly challenging as it exhibits variable responses to therapy [12, 13, 14, 15, 16, 17, 18].

It is therefore important to establish efficient approaches for assessing the therapeutic potential of the current and future novel treatments [4, 19]. While animal models are costly, time‐consuming, and often lack translational relevance to the patient's tumor, ex vivo precision models hold promise for improving outcomes tailored to individual tumor characteristics [20, 21, 22, 23, 24, 25, 26].

To achieve this, ex vivo models need to replicate the tumor by incorporating cancer cells with their microenvironment, which plays a significant role in PDAC pathogenesis and therapeutic responses [27, 28, 29, 30, 31, 32, 33, 34]. Patient‐derived organoids (PDOs) offer a promising solution since they retain the functional features of the tumor in vivo, allowing for an individualized study of tumor biology and drug response [35, 36, 37, 38, 39]. Yet PDOs have functional limitations, including the inability to capture all the stromal and immunological cells that make up the tumor microenvironment (TME), absence of vascularization or perfusion, and static culture conditions that fail to mimic the physiological fluid flow in vivo, thereby limiting nutrient and waste exchange [40]. These may limit the utility of PDOs in modeling dynamic tumor‐stroma‐immune interactions for drug testing [41, 42].

The combination of PDO technology with organ‐on‐a‐chip (OoC) using a microfluidic system is a viable approach to enhance ex vivo modeling of PDAC [43, 44]. This system consists of chambers designed to allow rapid growth of PDOs along with culture of various cell types (i.e., stroma and immune cells) and microchannels to allow fluid passage within its chambers [45, 46]. This integrated approach has the potential to bridge the gap between traditional in vitro models and pre‐clinical animal studies, enhancing personalized cancer modeling and accelerating drug screening [47, 48, 49, 50, 51, 52, 53].

In PDAC, the TME is characterized by dense extracellular matrix and complex stromal interactions that contribute to tumor progression and poor drug response [33, 54]. Macrophages and fibroblasts within the TME play crucial roles in tumor progression and chemoresistance [32, 34, 55, 56, 57, 58]. This has fueled ongoing efforts to exploit the TME for therapy in clinical trials [56]. Dampened anti‐tumor immunity is a hallmark of PDAC, driven by the dense desmoplastic stroma and immunosuppressive molecules such as programmed death ligand 1 (PD‐L1), leading to reduced T cell efficacy and resistance to immunotherapy [59, 60, 61, 62, 63, 64, 65]. We present a novel patient‐derived OoC platform by incorporating PDOs with innate and humoral immune cells (monocytes, T cells), along with fibroblasts within a vascularized microfluidic system, which, to our knowledge, has not yet been used to model PDAC TME.

2. Results

2.1. Ex Vivo Modeling of Pancreatic Ductal Adenocarcinoma on the Chip

The PDAC TME comprises a complex mixture of immune cells, cancer‐associated fibroblasts, and endothelial cells [60, 66, 67]. To better mimic in vivo PDAC, we developed a novel ex vivo model that recapitulates key features of the TME. We utilized a microfluidic device (hereafter, Chip) to leverage the potential of biopsy‐based PDO technology to establish a novel PDO‐based OoC model that includes key elements of the TME known to be involved in determining drug response in PDOs.

PDOs and monocytes: Biopsy samples from naïve PDAC were obtained using ultrasound‐guided fine needle biopsy. The isolated cancer cells were cultured to proliferate, forming new spherical structures within the 3D Matrigel environment as described before (Figure 1A) [68]. Hematoxylin and eosin (H&E) staining revealed that both normal (Figure 1B‐i) and cancer (Figure 1B‐ii) PDOs have tubular architectures, consistent with ductal glands. The normal PDO had a uniform, elongated cystic structure, composed of a single layer of epithelial cells surrounding the central lumen. The nuclei are round and uniform in size and shape. In contrast, the cancer PDO showed regions of irregular gland formation with budding glandular protrusions extending from the organoid. These buds indicate uncontrolled growth due to loss of contact inhibition, consistent with a neoplastic phenotype. The cells vary in size and shape (pleomorphism), and cellular crowding, another characteristic of neoplastic transformation, is also present (arrowheads). Dispersion of the nuclear‐to‐cytoplasmic (N/C) ratio was markedly greater in the Cancer PDO (Fligner–Killeen χ²(1) = 14.19, p = < 0.001), with 12.1‐fold higher variance relative to Normal PDO (Figure 1C), consistent with pronounced cell‐to‐cell heterogeneity (nuclear pleomorphism) in PDAC. Next, PDOs from two lines (Cancer PDO and Normal PDO) were dissociated into single cells and cultured under standard organoid conditions. Brightfield images were acquired over time, and individual organoid diameters were quantified. Using a linear mixed‐effects model, both lines increased in size over time (day effect β = 0.188, p = 0.004). The PDO diameter of the cancer PDOs was larger than normal PDOs (β = 0.551 on the log scale, 73% higher diameter, p = <0.001). Growth over time was steeper in Cancer PDOs, with an additional 7.9% increase per day versus Normal (95% CI 1.0% – 15.3%, p = 0.027), corresponding to 20.7%/day for Normal and 30.3%/day for Cancer PDOs (Figure 1D). The tumor purity of the generated PDOs was confirmed with the overexpression of pancreatic cancer marker, CA‐19‐9 (Figure 1E), while organoids from a non‐neoplastic pancreas showed none to very mild CA19‐9 expression. PDOs were also stained for Cytokeratin‐19 (CK19) [69, 70]; representative images show weak/patchy CK19 in the normal PDOs and intense, widespread staining in the cancer PDOs (Figure 1F‐i,‐ii). Quantification on a log10 scale confirmed higher CK19 signal in the cancer PDOs relative to the normal PDOs (W = 36, p = 0.002, Figure 1G). Western blotting of PDO lysates showed higher phosphorylated ERK (p‐ERK) relative to total ERK (p‐ERK/ERK) in cancer organoids compared with normal controls (Figure 1H). Across three PDAC samples, p‐ERK/ERK averaged 0.69 ± 0.09 (mean ± SD), versus 0.39 ± 0.08 in two normal organoids (Welch's t‐test, t = 3.90, p = 0.037), consistent with elevated ERK pathway activation in PDACs (Figure 1I).

FIGURE 1.

Establishment of pancreatic organoids and immune cell isolation workflow. (A) Schematic diagram of the workflow for developing and maintaining human Pancreatic Ductal Organoids (PDOs) from patient tissue. Organoids are cultured in 3D Matrigel, passaged weekly, and maintained in complete organoid media. Created in BioRender. Adnan, D. (2025) https://BioRender.com/y93dwq2 (B) Representative brightfield images of (i) a normal PDO and (ii) a cancer PDO derived from human pancreas samples. (C) Variability of nuclear‐to‐cytoplasmic (N/C) ratio: cancer vs normal. Each dot = single cell (normal n = 16 cells, cancer n = 17 cells). Variance compared by Fligner–Killeen test (two‐sided): χ ²(1) = 14.19, *p < 0.001. Bars show variance on log10 scale ±SE. (D) PDO growth in culture from single cells: mean diameter over days for normal vs cancer lines. Analysis by linear mixed‐effects model (lme4) with fixed effects for Group, Day, and Group×Day and random intercept for PDO line; day effect β = 0.188, p = 0.004; cancer > normal β = 0.551 (log‐scale), p < 0.001; Group×Day interaction p = 0.027, bars show mean ± SE. (E) Immunofluorescence staining of PDOs from normal and cancer donors showing expression of CA19‐9, a pancreatic cancer marker. (F) IF staining of Cytokeratin‐19 (CK19) in normal (i) vs cancer (ii) PDOs (representative fields). (G) Quantified CK19 signal (log10 scale), n = 6 per group. Comparison by Wilcoxon rank‐sum test (two‐sided), showing CK19 expression in both groups, with higher levels observed in tumor‐derived PDOs consistent with known PDAC features: W = 36, p = 0.002. (H) Representative immunoblots for p‐ERK, total ERK, and β‐actin in PDO lysates. (I) p‐ERK/ERK ratio. Each dot = one PDO sample (cancer n = 3, normal n = 2). Welch's t‐test (two‐sided): t = 3.90, p = 0.037. Bars show mean ± SD. (J) Brightfield/IF images of isolated human monocytes. (K) IF confirming polarization: CD68⁺ M1 and CD163⁺ M2 macrophages (representative). (L–P) Cytokine profiling of polarized macrophages (M1 vs M2) at 1, 6, and 24 h. Each point n = 2 per group per time. Two‐sided Welch t‐tests per timepoint with Benjamini–Hochberg (BH) correction within each cytokine. Significant examples: IFN‐γ (M1>M2 at 1 h p = 0.0217, 6 h p = 0.00571, 24 h p = 0.0128), IL‐6 (all times p < 0.0415), TNF‐α (24 h p = 0.0149), IL‐4 (M2>M1 at 1 h and 24 h p = 0.0382; M1 6 h NA), IL‐10 (ns at all times). Lines show mean ± SD.

On the day of the procedure, we obtained blood from the same patients who supplied biopsies. Fresh peripheral blood mononuclear cells (PBMCs) were isolated and cryopreserved in aliquots of 2 million cells. Monocytes were then extracted from PBMCs using an immunomagnetic negative selection method, which yielded ∼2 00 000 monocytes per 2 million PBMCs following several washing steps. The monocytes are highly purified (94%) with this method [71], and we checked their expression using CD68 IF staining (Figure 1J). Additionally, monocytes were differentiated into macrophage subtypes (M1 and M2) using standard polarization techniques to evaluate their phenotype plasticity, followed by immunostaining for CD68 and CD163 to confirm successful differentiation (Figure 1K).

As shown in Figure 1L–P, cytokine analysis confirmed phenotypic differences between M1 and M2 macrophages. After differentiation (day 6), conditioned media were collected at 1, 6, and 24 h and quantified by MSD. All cytokine levels were shown as pg/mL. IFN‐γ was markedly higher in M1 (1 h: 30.08 ± 1.45 vs 1.07 ± 0.11, p = 0.0217; 6 h: 48.50 ± 0.75 vs 1.22 ± 0.32, p = 0.00571; 24 h: 54.34 ± 1.13 vs 0.85 ± 0.14, p = 0.0128). IL‐6 was also higher in M1 (1 h: 3.66 ± 0.36 vs 0.17 ± 0.05, p = 0.0415; 6 h: 3.59 ± 0.30 vs 0.18 ± 0.09, p = 0.0415; 24 h: 3.94 ± 0.27 vs 0.20 ± 0.03, p = 0.0415). TNF‐α was higher in M1 (1 h: 2.66 ± 0.06 vs 0.31 ± 0.03, p = 0.00662; 6 h: 2.47 ± 0.25 vs 0.11 ± 0.00, p = 0.0476; 24 h: 2.76 ± 0.01 vs 0.35 ± 0.06, p = 0.0149). In contrast, IL‐4 was lower in M1 (1 h: 0.017 ± 0.002 vs 1.119 ± 0.094, p = 0.0382; 24 h: 0.029 ± 0.010 vs 1.636 ± 0.095, p = 0.0382; IL‐4 at 6 h not available). IL‐10 increased over time in both conditions (M1: 0.089 ± 0.012 → 0.230 ± 0.019; M2: 0.060 ± 0.013 → 0.169 ± 0.034) but showed no per‐time differences between M1 and M2 (all p ≥ 0.05). Collectively, these cytokine profiles show a clear separation of M1 and M2 phenotypes, supporting their use as positive controls for downstream analyses.

2.2. Building Patient‐Derived Ex Vivo Model of PDAC

Multicellular Chip: The Chip consists of two chambers separated by a 3 µm porous membrane. The upper chamber was used to simulate a vascular channel by seeding it with human umbilical vein endothelial cells (HUVECs), allowing media to flow across an endothelial‐like barrier. The lower chamber serves as the tumor compartment, where we established a multicellular co‐culture including primary human pancreatic stellate cells (PSCs), macrophages, and tumor organoids (Figure 2A).

FIGURE 2.

Building Patient‐Derived ex vivo model of PDAC. (A) Schematic diagram of the microfluidic chip platform (The Chip). The device consists of two chambers separated by a 3 µm porous membrane. The upper chamber mimics a vascular channel seeded with human umbilical vein endothelial cells (HUVECs), while the lower chamber represents the tumor compartment and contains a co‐culture of patient‐derived pancreatic ductal organoids (PDOs), primary pancreatic stellate cells (PSCs), and monocyte‐derived macrophages. Right panels show representative immunofluorescence staining confirming cellular components: CD31 (HUVECs), EpCAM (epithelial/tumor cells), α‐SMA (PSCs), and CD68 (macrophages). Created in BioRender. Adnan, D. (2025) https://BioRender.com/8x9rr7b (B) Timeline schematic detailing the sequential seeding of each cell type and initiation of flow via microfluidic pump. Created in BioRender. Adnan, D. (2025) https://BioRender.com/x5kxoqg (C) Representative brightfield images showing tumor growth in the Chip at days 0, 4, and 9. (D) Scatter plot quantifying PDO diameter over time. PDOs in the full multicellular model exhibited significantly greater growth than PDOs cultured alone Linear mixed‐effects model (random intercept for Chip): Group×Day interaction p = 0.0016. Follow‐up per‐day Welch tests with BH correction: Day 0 ns (p = 0.896), Day 4 p = 0.0478, Day 9 p < 0.001 (full>PDO only). n = 3 Chips per group. Points show Chip means; lines show mean ± SE. (E) Scatter plot illustrating percent change in surface area (S.A.%) of PDOs from day 0 to day 9. Linear mixed‐effects model: Group×Day p = 0.173; per‐day Welch tests BH‐adjusted: Day 4 ns (p = 0.183), Day 9 ns (p = 0.183); values consistently higher in full model. n = 3 Chips per group. (F‐G) Representative BF (F) and biopsy H&E (G) illustrating desmoplastic morphology (no stats). (H) Immunofluorescence staining of PDOs from normal and cancer donors showing differential expression of phosphorylated ERK (p‐ERK), indicating activation of MAPK signaling (representative). (I) Quantified p‐ERK intensity: n = 4 per group. Wilcoxon rank‐sum (two‐sided): W = 16, p = 0.02. Bars show mean ± SD.

The cell seeding strategy and the establishment of the flow using a pump are illustrated in Figure 2B. To promote cell adhesion, we pre‐treated the upper and lower chambers of the Chip with Animal Component‐Free (ACF) Cell Attachment Substrate and Poly‐L‐Lysine solution, respectively. Following a two‐hour incubation at room temperature, which was sufficient to enhance attachment, we seeded 50 000 HUVECs cells into the upper chamber to establish an endothelial layer. The formation of this layer was confirmed both by light microscopy as well as by IF staining for CD31 (PECAM‐1), an endothelial marker (Figure 2A). To assess endothelial barrier function, we quantified FITC–dextran permeability of the vascular layer on days 1, 2, and 3 after seeding HUVECs to the upper chamber. Fluorescence from the effluent in the lower chamber decreased from 25 452–24 353 units on days 1–2 to 1295 relative fluorescence units (RFU) on day 3, indicating progressive maturation of a HUVEC monolayer.

The lower chamber was seeded with 10 000 primary human PSCs (ScienCell, Catalog #3830) and incubated overnight for cell adhesion. Fully grown PDOs were digested into single cells using TrypLE before loading into the Chip. These cells were combined with monocytes at a 1:1 ratio in a matrix composed of 80/20 Matrigel to complete organoid media. We used a 1:1 PDO‐to‐monocyte ratio to provide equal initial opportunity for tumor–myeloid interactions, which is aligned with the prior PDAC organoid–macrophage co‐culture work that employed a 1:1 ratio to interrogate tumor–macrophage interactions and relate tumor‐associated macrophages (TAM) diversity to cancer cell survival [72]. A total of 10,000 combined cells were then seeded into the lower chamber of the Chip. The next day, perfusion was initiated at 10 µL/hour on day and maintained at this rate for the remainder of the experiment. Multicellular tumor Chips were confirmed by identifying PDOs that expressed epithelial cell adhesion molecule (EpCAM), a tumor epithelial transmembrane protein; PSCs that expressed alpha‐smooth muscle actin (α‐SMA); and macrophages that expressed CD68 (Figure 2A). To maximize the reliability and robustness of the experimental conditions, multiple trials were conducted to determine the optimal culture duration for PDOs in the Chip. These trials consistently demonstrated that the PDOs maintained their viability and structural integrity over a nine‐day culture period, which was then fixed as the experimental duration (Figure 2C).

2.3. Stroma‐Cancer Interaction in the Patient‐Derived PDAC Chip Model

To study the effect of stroma on the PDOs growth, the PDO growth was monitored in the presence or absence of stroma throughout the experiment period. As shown in Figure 2D, the PDO diameter in the full model (PDO + PSC + Monocytes) increased more rapidly than in PDOs cultured alone (group×day interaction p = 0.0016). Mean ± SD (µm): Day 0: PDO 34.9 ± 12.7, full 35.7 ± 12.5; Day 4—PDO 62.6 ± 23.4, full 89.9 ± 34.7; Day 9—PDO 64.7 ± 11.9, full 119.7 ± 31.1. Per‐day Welch tests (BH‐adjusted) showed no difference at Day 0 (p = 0.8956), but larger diameters in the full model at Day 4 (p = 0.0478) and Day 9 (p < 0.001). Similarly, tumor growth measured by the change in percent surface area [S.A%], defined as the proportion of the field of view occupied by the tumor, was higher on average in the full tumor model than in the control (PDO only, group×day interaction p = 0.173, Figure 2E). Mean ± SD (S.A.%): Day 0: PDO 0.0 ± 0.0, full 0.0 ± 0.0; Day 4: PDO 0.056 ± 0.049, full 0.268 ± 0.220; Day 9: PDO 0.115 ± 0.019, full 0.320 ± 0.238. Per‐day Welch tests (BH‐adjusted) did not reach significance at Day 4 (p = 0.183) or Day 9 (p = 0.183), though values were consistently greater in the full model. By day 4, we observe the tumor growth by co‐culture (Figure 2F,G), which is also accelerated compared to PDO only (control). When visually compared to the histopathological analysis of EUS biopsy from the same subject, the H&E image (Figure 2G) revealed infiltrating, irregular glandular structures embedded within a desmoplastic stroma. Strikingly, our Chip recapitulated similar features, with PDOs surrounded by a dense fibrotic stroma composed of PSCs and immune cells. These morphological patterns are characteristic of PDAC, showing both the infiltrative growth of malignant ductal cells and the extensive stromal remodeling. The accelerated growth of PDOs was confirmed by the high expression of phosphorylated extracellular signal‐regulated kinase (pERK), which plays a crucial role in cell proliferation and survival. Elevated levels of pERK are associated with tumor growth and progression in pancreatic cancer [73, 74]. Quantification showed that cancer PDOs exhibited significantly higher pERK intensity compared to normal PDOs (W = 16, p = 0.02, Figure 2H,I). Mean ± SD intensities were 6110 ± 1550 for cancer PDOs versus 2115 ± 410 for normal PDOs.

We then investigated changes that may occur in the stroma in response to cancer cells. As shown in Figure 3A, the tumor Chips containing PDOs, PSCs, and macrophages exhibited a denser stromal network compared to Chips with PDO + PSCs (without macrophages), as indicated by Collagen (red) staining of the stroma (Figure 3B). The presence of the full model appeared to enhance stromal remodeling, contributing to elongated PSCs and higher collagen levels, indicating a more contractile microenvironment, in line with the bidirectional interaction between cancer and stromal cells in vivo [75]. To test whether PSC activation is due to the presence of organoids or is specifically enhanced by cancer PDOs, we co‐cultured PSCs with either a normal PDO or a cancer PDO and assessed activation at 48 h by Collagen 1 staining. PSC activation was quantified by measuring the spread area of individual PSCs (n = 13 cells per condition, from triplicate wells). Analysis revealed that PSCs co‐cultured with the cancer PDO exhibited a statistically significantly larger cell areas compared to those co‐cultured with the normal PDO (W = 24, p = 0.002). Mean ± SD areas were 2663 ± 1436 µm² (n = 13) for PSCs in cancer PDO co‐cultures versus 1074 ± 504 µm² (n = 12) for PSCs in normal PDO co‐cultures, indicating that stromal activation is cancer‐specific rather than a general response to PDO+PSC co‐culture (Figure 3C).

FIGURE 3.

Tumor‐derived signaling modulates stromal activation and fibroblast gene expression. (A) Immunofluorescence staining of tumor models containing PDOs, pancreatic stellate cells (PSCs), and monocytes, showing increased collagen deposition (red) and EpCAM+ PDOs (green), indicating stromal remodeling. (B) Immunofluorescence image of tumor model containing only PDOs and PSCs (without monocytes), showing reduced collagen content and less elongated PSC morphology compared to the full model in (A). (C) PSC activation quantified by spread area after co‐culture with cancer PDOs vs normal PDOs. Each dot = single PSC (cancer n = 13 cells, normal n = 12 cells). Wilcoxon rank‐sum: W = 24, p = 0.002. Bars show mean ± SE, dots = cells. (D) Schematic of the transwell co‐culture experiment in which primary pancreatic fibroblasts were seeded in the lower chamber and co‐cultured with either cancerous PDOs or Matrigel‐only controls in the upper insert (3 µm pore size), allowing for exchange of soluble factors. Created in BioRender. Adnan, D. (2025) https://BioRender.com/zc49f8x (E) PCA plot of RNA‐seq data from co‐cultured fibroblasts (n = 1 per group, fibroblasts monoculture vs fibroblasts + cancer PDOs), demonstrating clustering near cancer‐associated fibroblasts (CAFs) identified in a previously published single‐cell RNA‐seq dataset of PDAC. (F) Volcano plot showing significantly upregulated and downregulated genes in fibroblasts co‐cultured with tumor organoids versus controls. EdgeR with TMM normalization and BH‐FDR; genes significant at FDR<0.05. n = 1 per group (G) Heatmap displaying all 219 differentially expressed genes (FDR < 0.05) from the transwell co‐culture experiment. (H) Focused heatmap of selected genes involved in stellate cell activation and extracellular matrix remodeling. (I) Gene set enrichment analysis showing statistically significant hallmark pathways enriched in fibroblasts co cultured with cancerous PDOs, indicating activation of pro‐tumorigenic pathways. (J–O) Monocyte cytokines in Transwell Monocyte alone vs Monocyte co‐cultured with PDOs. Conditioned media collected at 1, 6, 24 h, n = 2 per group per timepoint. Two‐sided Welch t‐tests with BH within cytokine across times. IL‐6 higher at 1, 6, 24 h (p = 0.0108, 0.0272, 0.0233), TNF‐α higher at 24 h (p = 0.0022), IL‐1β higher at 6 and 24 h (p = 0.0090, 0.0090); IFN‐γ ns; IL‐10 and IL‐4 ns. Lines show mean ± SD. (P) MDS positioning shows a shift of Monocytes when co‐cultured with PDOs toward mixed polarization relative to M1/M2 references.

To investigate the mechanisms underlying stroma remodeling, we analyzed transcriptomic changes in primary fibroblasts following co‐culture with PDOs. Primary pancreatic fibroblasts were isolated from a normal post‐mortem donor pancreas (see methods). Briefly, a small piece of non‐cancerous pancreatic ductal tissue was excised, washed four times, and incubated in culture media. Over approximately one month, fibroblasts spontaneously migrated from the tissue and adhered to the culture plate. After removal of the remaining ductal tissue, adherent fibroblasts were expanded through serial passaging for subsequent experiments.

Fibroblasts were co‐cultured with cancerous PDOs or maintained in monoculture as controls using transwell plates (Figure 3D). After two days of co‐culture, cells were harvested for RNA‐seq analysis. Principal component analysis (PCA) of the RNA‐seq data confirmed the fibroblast identity, with cells from both groups clustering closely (Spearman r = 0.738 each) with cancer‐associated fibroblasts (CAFs) from a single‐cell RNA‐seq (scRNA‐seq) dataset of PDAC (Figure 3E) [76]. Fibroblasts enriched genes significantly overlapped the CAF reference (9/30, Fisher's OR = 21.46, p < 0.001). Next, we quantified a canonical myofibroblastic CAF (myCAF) program: ACTA2, TAGLN, MYL9, COL1A1, COL1A2, ITGA11, THBS2, MMP2, PDGFRB, FAP, COL11A1) in bulk RNA‐seq from fibroblasts co‐cultured with cancer PDOs or controls (monocultured). Primary fibroblasts co‐cultured with cancer PDOs showed a higher myCAF score than the control (1.412 vs 1.118), further indicating a pro‐tumorigenic reprogramming of fibroblasts in the presence of primary cancerous epithelial cells.

Differential abundance analysis to compare gene expression between primary fibroblasts monoculture and primary fibroblasts co‐cultured with PDOs discovered 219 significantly differentially expressed genes, indicating substantial transcriptional changes induced by cancer‐stromal interactions (Figure 3G). Genes known to control stellate cell activation, such as LAMC2, LAMB3, IL1B, and ADAMTS4, were significantly upregulated in primary fibroblasts upon co‐culture with cancerous PDOs (Figure 3H). The functional pathway analysis (Figure 3I) using human hallmark gene sets from MSigDB [77] demonstrated that the co‐culture induced shifts in the transcriptome of fibroblasts, which aligned with a pro‐tumorigenic stromal phenotype. In addition to TNFα signaling, NF‐κB, and IL‐6/JAK/STAT3 signaling pathways, we observed enrichment of inflammatory signaling (interferon‐γ response, MAPK/ERK, PI3K/AKT), hypoxia, and stress (e.g., apoptosis) responses as well as metabolic pathways (glycolysis and xenobiotic), which are implicated in stromal remodeling in PDAC [78, 79]. These findings underscore the involved mechanisms that may explain cancer‐stroma interactions in our model, where cancer cells drive stromal activation, which in turn supports tumor progression.

Next, we explored changes that may occur to monocytes in response to cancer PDOs. We used a Transwell co‐culture system. Primary monocytes were seeded in the lower well and maintained for four days. Cancer PDOs were then added to the upper insert and incubated for 48 h. After replacing it with fresh medium, CM was collected at 1, 6, and 24 h, and cytokines were profiled using the MSD human cytokine panel (IFN‐γ, IL‐6, IL‐1β, TNF‐α, IL‐10, and IL‐4). Compared with monocytes alone, when co‐cultured with cancer PDOs, analysis showed a clear pro‐inflammatory shift (Figures 3J‐O). IL‐6 was markedly higher in M0+PDO (1 h: 0.196 ± 0.028 vs 8.821 ± 0.113 pg/mL, BH p = 0.0108; 6 h: 0.294 ± 0.036 vs 8.374 ± 0.502, p = 0.0272; 24 h: 0.240 ± 0.020 vs 10.252 ± 0.353, p = 0.0233). TNF‐α was also significantly elevated at 24 h (0.242 ± 0.066 vs 3.041 ± 0.048, p = 0.0022), though not at earlier timepoints. IL‐1β showed a trend at 1 h (2.197 ± 0.474 vs 4.512 ± 0.139, p = 0.072) and was significantly higher at 6 h (2.525 ± 0.199 vs 6.416 ± 0.258, p = 0.0090) and 24 h (2.433 ±  0.061 vs 6.839 ± 0.004, p = 0.0090). IFN‐γ increased over time but did not reach significance (e.g., 24 h: 1.303 ± 0.202 vs 2.822 ± 0.336, p = 0.145). IL‐10 and IL‐4 were numerically higher in M0+PDO (e.g., IL‐10 at 24 h: 0.163 ± 0.046 vs 1.589 ± 0.440, p = 0.134; IL‐4 at 24 h: 0.036 ± 0.013 vs 0.324 ± 0.048, p = 0.118) but not statistically significant. Multidimensional scaling (MDS) analysis further demonstrated a clear shift of monocytes co‐cultured with cancer PDOs away from the M0 baseline across all timepoints. Positioning relative to M1 and M2 reference positive controls indicated coordinated upregulation of M1‐associated cytokines together with induction of M2‐associated cytokines, consistent with a mixed, pro‐tumorigenic polarization state (Figure 3P). Together, these data indicate that cancer PDOs reprogram monocytes toward a TAM‐like phenotype in our model. We further attempted to assess changes in monocyte polarization at the transcriptome level and performed single‐cell RNA sequencing; however, despite these efforts, monocytes did not survive the library preparation in sufficient numbers to generate reliable, high‐quality transcriptomic data.

2.4. Drug‐Response Assessment in PDAC Chip Model

To evaluate the utility of our tumor Chip model for drug testing, we assessed the response to gemcitabine, a chemotherapy agent routinely used as the standard of care for PDAC. Tumor Chips containing PDOs, autologous monocytes, allogeneic human PSCs, and allogeneic human endothelial cells (seeded in the upper chamber) were established as described previously (Figure 2B timeline). To evaluate how the presence of tumor microenvironment components modulates PDO response to gemcitabine, we first established the gemcitabine IC50 for the PDO in monoculture (Figure 4A). This concentration was then applied to our multicellular Chip model. While we cannot exclude the possibility that IC50 values obtained in PDO monoculture may differ in the multicellular OoC due to additional cell–cell and cell–matrix interactions, this approach allowed us to specifically assess the impact of TME components on gemcitabine response, independent of potential PDO‐intrinsic heterogeneity to the drug.

FIGURE 4.

Evaluation of the tumor Chip model for drug response and stromal modulation. (A) Left: Representative brightfield image of PDOs cultured in an Elplasia 96‐well round‐bottom microplate prior to chemotherapy testing. Right: Dose‐response curve of gemcitabine treatment showing non‐linear regression used to calculate the IC₅₀ value based on five serial concentrations (100, 20, 4, 0.8, 0.16 µm). (B) Immunofluorescence images of tumor (control) or tumor + all‐trans retinoic acid (ATRA), demonstrating that ATRA reduces stromal density and PSC elongation, as evidenced by decreased collagen staining. (C) PSC extension length with or without ATRA in the full model. Each dot = PSC process length (Tumor n = 9 cells, Tumor+ATRA n = 15 cells). Wilcoxon rank‐sum: W = 110.5, p = 0.011. (D) Representative immunofluorescence images showing cleaved caspase‐3 (c.c.3⁺) and EpCAM⁺ cells in tumor Chips treated with DMSO (control), gemcitabine alone, or gemcitabine in combination with ATRA and Clodrosome (a macrophage‐depleting agent). (E) Quantification of apoptotic tumor cells (c.c.3⁺/EpCAM⁺ double‐positive cells) revealed significantly increased apoptosis in the combination treatment group (gemcitabine + ATRA + Clodrosome) compared to gemcitabine alone or DMSO control (DMSO n = 7, gemcitabine n = 6, gemcitabine+ATRA+Clodrosome n = 8). One‐way ANOVA: p < 0.001; Tukey post‐hoc: Combo > DMSO p < 0.001, Combo > gem p = 0.011, gem vs DMSO p = 0.084 (ns). Box/whiskers show median, IQR, 1.5 × IQR.

To further characterize cancer‐stroma interactions functionally, we investigated whether targeting the stromal compartment could enhance chemotherapy efficacy. First, we investigated if treating the Chip with all‐trans retinoic acid (ATRA), which induces PSC quiescence could affect the stroma phenotypes in our Chip. ATRA decreased stromal density and PSC elongation (Figure 4B). Using collagen staining, quantitative analysis (Figure 4C), Mean ± SD PSC extension was 179 ± 48.5 µm (n = 9) in the PDO + PSC + Monocytes condition which reduced to 109 ± 67.1 µm (n = 15) with + ATRA (W = 110.5, p = 0.011), implying that ATRA significantly reduces stroma in our Chip model. Because both ATRA and Clodrosome primarily target stromal or immune compartments, we examined whether these drugs have any direct cytotoxic effects on PDOs. To this end, PDOs were treated with ATRA, Clodrosome, or control (DMSO), and cleaved caspase‐3 staining followed by Kruskal–Wallis and Dunn's tests revealed no significant differences in apoptosis among groups (χ ² = 1.74, p = 0.42; all pairwise p > 0.40), indicating that neither ATRA nor Clodrosome induced apoptosis in PDOs (Figure S1C). Additionally, no differences in organoid size or morphology were observed across treatment conditions, suggesting neither drug directly alters epithelial architecture in monoculture (Figure S1D). To assess whether stromal modulation could potentiate chemotherapy response, we combined ATRA and Clodrosome (a macrophage‐depleting agent) with gemcitabine treatment. Gemcitabine alone modestly increased cancer cell apoptosis, as determined by the proportion of cleaved caspase‐3 positive (c.c.3⁺) EpCAM⁺ tumor cells (Figure 4D). However, the addition of the anti‐stroma (ATRA + Clodrosome) to gemcitabine, caused a substantial elevation in tumor killing (c.c.3+/EpCAM+ cells) compared to the gemcitabine alone or DMSO control (means ± SD: DMSO 9.6% ± 6.3% [n = 7], gem 29.1% ± 8.1% [n = 6], gem+ATRA+Clod 56.2% ± 22.9% [n = 8]; one‐way ANOVA p < 0.001; Tukey: combo > DMSO p < 0.001, combo > gem p = 0.011, gem vs DMSO p = 0.084; Figure 4E), indicating that incorporation of anti‐stromal agents augmented the anti‐cancer efficacy of the chemotherapy. Together, these findings suggest that our tumor Chip platform can serve as a robust ex vivo system to model drug responses and evaluate combinatorial therapeutic strategies.

2.5. Incorporation of Patient‐Derived T Cells Into the Chip Model

Cytotoxic CD8+ T cells play a crucial role in recognizing and destroying cancer cells. To include this key element of adaptive immunity within TME and its interactions with the tumor in our model, we isolated T cells from cryopreserved PBMCs obtained from the same patients who provided biopsy samples for PDO generation. Isolated T cells were cultured in T cell expansion media and activated with IL‐2 and CD3/CD28/CD2 T cell activators. We then tagged the T cells with a Red Trafficking Dye (Abcam, Catalogue No: AB269446) before infusing a total of fifty thousand cells into the upper (vascular) chamber of our Chip model. This setup allowed us to investigate immune cell recruitment from the vascular chamber in response to the tumor signals (lower chamber).

Analysis demonstrated a successful migration of T cells toward the tumor chamber (Figure 5A). The number of T cells migrating into the lower chamber was substantially higher when more PDOs were present, suggesting that cancer‐derived signaling is a driver of immune cell attraction and that the migration is correlated with tumor burden. Furthermore, the presence of a more complete tumor model with inclusion of PSCs and monocytes further enhanced T cell infiltration, with the highest number of recruited T cells observed in the tumor model (Figure 5A,B). Specifically, T cell infiltration was 8.5 ± 3.2 cells/FOV in 1K PDO, 68.5 ± 26.1 cells/FOV in 10K PDO, and 98.8 ± 33.4 cells/FOV in 10K PDO + stroma, suggesting the role of stroma in the tumor immune infiltration in the Chip. Pairwise comparisons showed 10K PDO > 1K PDO (p = 0.009), 10K PDO + stroma > 1K PDO (p < 0.001), and 10K PDO + stroma vs 10K PDO (p = 0.15, ns).

FIGURE 5.

Incorporation of patient‐derived T cells into the tumor Chip reveals stromal effects on T cell infiltration and function. (A) Brightfield and immunofluorescence images showing T cells successfully migrating from the upper vascular chamber into the tumor compartment of the Chip containing PDO only (1K vs. 10K PDOs) and further in PDO + Stroma. (B) Bar plots showing quantification of T cell infiltration into the lower chamber demonstrating increased migration with higher tumor burden (10K vs. 1K PDOs) and further enhancement when stroma (PSC + Monocytes) were present, indicating that the tumor microenvironment facilitates T cell recruitment. 1K (n = 6), 10K (n = 4), 10K+Stroma (n = 5). One‐way ANOVA with Tukey: 10K>1K p = 0.009, 10K+Stroma>1K p < 0.001, 10K+Stroma vs 10K p = 0.15 (ns). Bars show mean ± SD. (C) Bar plot indicating a significantly higher T cell count at the lower chamber in the full model Chips (PDOs + PSCs + Monocytes) compared to the PDOs only Chips. PDO only (n = 3) vs PDO+Stroma (n = 6). Welch's t‐test: p = 0.028. (D) Despite increased T cell infiltration, quantification of cancer cell apoptosis (e.g., cleaved caspase‐3⁺/EpCAM⁺ cells) revealed reduced tumor cell killing in Chips with stromal components compared to PDO‐only Chips, suggesting that stromal cells contribute to T cell dysfunction in this ex vivo model of pancreatic cancer. PDO only (n = 3) vs PDO+Stroma (n = 6). Welch's t‐test: p = 0.028. Bars show mean ± SD. (E) Schematic of the transwell co‐culture experiment, where tumor models (PDOs + PSCs + monocytes) or Matrigel controls were placed in the upper insert, while activated T cells were seeded in the lower well, allowing for exchange of signaling without direct contact. Created in BioRender. Adnan, D. (2025) https://BioRender.com/zc49f8x (F) PCA plot shows that blood‐derived bulk RNA‐seq data from T cells co‐cultured with tumor models cluster closely with tissue‐derived tumor‐infiltrating T cells from a previously published single‐cell RNA‐seq dataset. (G) Heatmap displaying all significantly up‐ and downregulated genes in T cells co‐cultured with the tumor model versus control, indicating substantial transcriptional changes, (edgeR; BH‐FDR<0.05, n = 1 sample per group). (H) Volcano plot highlighting differentially expressed genes from the transwell co‐culture, (edgeR; BH‐FDR<0.05, n = 1 sample per group). (I) Heatmap of selected exhaustion‐associated genes (e.g., CD38 and NOTCH2NL) in T cells co‐cultured with the tumor model, demonstrating the induction of an exhausted T cell phenotype, (edgeR; BH‐FDR<0.05, n = 1 sample per group).

We next examined whether the patient‐derived T cells are functional in the Chip by studying their cancer killing effects. Tumor Chips were prepared by including PDOs along with their TMEs, including autologous monocytes and allogeneic human stellate cells. On day 4, a total of fifty thousand activated blood‐derived cytotoxic T cells were infused into the vessels (upper chamber) for 24 h. Upon quantitative analysis, we found that although more T cells migrated to the tumor Chip where stroma was present (2.33 ± 1.53 vs 19.5 ± 12.5 cells/FOV for PDO only vs PDO + stroma; p = 0.028, Figure 5C), less cancer cell killing occurred (8.83% ± 5.67% vs 1.78% ± 1.08%; p = 0.028, Figure 5D). This indicates that the stroma may render T cells dysfunctional, reminiscent of in vivo reports on stroma hindering cytotoxic T cell function in PDAC [80, 81].

To investigate the mechanisms that led to dysfunction of T cells, we analyzed the transcriptomic profiles of activated T cells cultured with or without the full tumor model (PDOs + PSCs + monocytes) (Figure 5E). RNA‐seq analysis revealed that co‐cultured T cells exhibited transcriptomic signatures resembling tumor‐infiltrating lymphocytes from a publicly available scRNA‐seq dataset of human PDAC (Figure 5F) [76]. Further differential expression analysis revealed 70 differentially expressed genes in patient‐derived T cells upon co‐culture with the tumor (Figure 5G,H). Notably, key markers associated with T cell exhaustion were upregulated, including CD38 and NOTCH2NL (Figure 5I) [82, 83, 84]. Conversely, CD34, a marker of naïve/progenitor T cells [85], as well as transcription factors ATF3 and TLX2, which are involved in T cell differentiation [86, 87]. Additionally, NRP2, which is a gene implicated in T cell‐mediated anti‐tumor immunity, was also downregulated [88]. These changes further support a phenotype shift toward a less effective phenotype upon tumor co‐culture (Figure 5H). Additional genes of (MRPS22, LBX2, RHOQ‐AS1, HCRTR1, RNUVU1‐14, C1orf162, and PLXNA1) were upregulated in exhausted T cells. Next, we conducted a gene set enrichment analysis utilizing the Gene Ontology (GO) [89, 90]. The analysis of the upregulated gene set from the co‐culture experiment of T cells compared to monoculture indicated an upregulation of the N‐linked glycosylation pathway, a process essential to the performance and signaling activation of T cell surface receptors [91]. The findings indicate that tumor‐stroma interactions in the Chip model induce T lymphocytes to shift into an exhausted phenotype, aligning with the immunosuppressive tumor microenvironment in PDAC. These findings highlight the capability of our Chip model as a framework for examining tumor‐immune interactions and responses to immunotherapy.

To determine whether T cells retained phenotypes before infusion, we performed multiparameter flow cytometry at serial culture time points. Flow cytometry analysis revealed dynamic temporal changes in CD8⁺ T cell memory subset distribution following in vitro stimulation (Figure 6A). Terminally differentiated effector memory cells re‐expressing CD45RA (TEMRA, CD45RA⁺CCR7⁻) were progressively lost during prolonged culture. In contrast, both naïve T cells (CD45RA⁺CCR7⁺) and central memory T cells (TCM, CD45RA⁻CCR7⁺) emerged by day 3, with effector memory T cells (TEM, CD45RA⁻CCR7⁻) and TCM subsets being stably maintained through day 10. These findings indicate that less differentiated memory subsets, particularly TCM and TEM, are preserved during extended in vitro expansion, reflecting retained functional plasticity in the cultured T cell population [92]. Following secondary stimulation on day 12, T cells were cultured for an additional five days, during which both TEM and TCM subsets were maintained (Figure 6B). This sustained presence indicates preservation of these memory phenotypes under the culture conditions. Furthermore, to verify functional activity, conditioned medium from cultured T cells was collected at days 5, 7, and 10 and then profiled using mesoscale (MSD) cytokine panels. Our analysis demonstrates sustained cytokine production across time points through day 5–10. (Figure S3).

FIGURE 6.

(A–C) Representative flow plots showing memory subsets (CD45RA/CCR7) over time and PD‐1 expression after co‐culture. (D) Dose‐dependent response to PD‐1 blockades with pembrolizumab in the PDAC tumor Chip model. Quantification of T cell infiltration into the tumor chamber following treatment with low‐dose (10 µg/mL) or high‐dose (100 µg/mL) pembrolizumab. Control (n = 6), pembrolizumab 10 µg/mL (n = 7), pembrolizumab 100 µg/mL (n = 6). One‐way ANOVA: F(2,16) = 4.64, p = 0.026; Tukey: 100 µg/mL > 10 µg/mL p = 0.028; 100 µg/mL vs control p = 0.110 (ns); 10 µg/mL vs control p = 0.800 (ns). Bars show mean ± SD. (E) Representative immunofluorescence images of tumor Chips showing DAPI (nuclei, blue), cleaved caspase‐3 (C.C.3, red), and EpCAM (tumor marker, green). Merged images illustrate enhanced tumor cell apoptosis in Chips treated with high‐dose pembrolizumab. (F) Quantification of tumor cell apoptosis (percent cleaved caspase‐3⁺ among EpCAM⁺ cells). High‐dose pembrolizumab (100 µg/mL) significantly increased tumor cell killing compared to both the low‐dose group (10 µg/mL) and control. Control (n = 6), 10 µg/mL (n = 7), 100 µg/mL (n = 6). One‐way ANOVA: F(2,16) = 7.03, p = 0.006 (η ² = 0.47). Tukey: 100 µg/mL > control p = 0.007; 100 µg/mL > 10 µg/mL p = 0.028; 10 µg/mL vs control p = 0.709 (ns).

Next, to test whether autologous PDOs modulate T cell phenotype beyond transcript‐level changes we observed earlier, we profiled memory‐marker and PD‐1 expression by flow after 48 h transwell co‐culture versus matched controls. Analysis revealed that both naïve T cells and TCM populations were preferentially expanded, consistent with their established enhanced antitumor activity and self‐renewal capacity [93]. These observations align with prior evidence demonstrating that less differentiated memory T cells exhibit robust recall responses and provide sustained long‐term immune surveillance capabilities [94]. Co‐culture with PDOs induced elevated PD‐1 expression in T cells (Figure 6C), indicative of the onset of T cell exhaustion. This phenotypic change is consistent with the well‐characterized capacity of the tumor to promote T cell dysfunction through antigen exposure and sustained immunosuppressive signaling pathways [95].

Notably, CD4⁺ T cells exhibited a similar temporal redistribution and differentiation pattern (Figure S4), underscoring a conserved activation‐driven phenotypic shift that occurs across both major T cell lineages.

2.6. Immunotherapy in PDAC Chip Model

Tumor inhibitory signals on T cells are typically orchestrated by immune checkpoint signals, which can be targeted by immune checkpoint blockade/immunotherapy [96]. This treatment strategy restores T cells ability to induce tumor killing, and has significantly advanced cancer treatment in recent years [97]. We next evaluated the performance of our tumor Chip model in assessing ICB responses by testing the impact of pembrolizumab, a monoclonal antibody that targets programmed cell death protein 1 (PD‐1) and enhances cytotoxic T cell function [98]. Tumor Chips were prepared as described in Section 2.5. T cells were pre‐incubated with control, or with pembrolizumab at either low or high doses for 30 min before being perfused into the vessel of the Chip. On day 4 50 000 activated, blood‐derived T cells were introduced into the vascular channel, either untreated or pre‐incubated for 30 min with low or high doses of pembrolizumab. This experimental design allowed us to examine the immune‐modulatory effects of PD‐1 inhibition on T cell behavior within the context of a complex, patient‐relevant tumor microenvironment.

Following T cell infusion, the Chips were incubated for 24 h to allow for immune cell migration and interaction with the tumor compartment. The lower chamber (tumor) was then fixed with 4% paraformaldehyde (PFA) for analysis. Quantification of T cell infiltration revealed a trend toward increased recruitment with a higher dose of pembrolizumab (100 µg/mL), as shown in Figure 6D. Means ± SD (cells/FOV): control 19.5 ± 12.5 (n = 6), pembrolizumab 10 µg/mL 14.6 ± 8.7 (n = 7), pembrolizumab 100 µg/mL 36.7 ± 19.1 (n = 6). One‐way ANOVA indicated a group effect (F(2,16) = 4.64, p = 0.026). Tukey post‐hoc: 100 µg/mL > 10 µg/mL (p = 0.028); 100 µg/mL vs control (p = 0.110) and 10 µg/mL vs control (p = 0.800) were not significant. To assess the functional impact of PD‐1 blockade, we measured cancer apoptosis by detecting c.c.3 in EpCAM⁺ tumor (Figure 6E,F). While treatment with 10 µg/mL pembrolizumab, did not significantly enhance tumor cell killing, the higher dose of 100 µg/mL led to a significantly increased percentage of apoptotic tumor cells (c.c.3⁺/EpCAM⁺) compared to both lower dose and control conditions (Figure 6F). Group means ± SD were: control 25.8% ± 9.3% (n = 6), 10 µg/mL pembrolizumab 30.1% ± 11.8% (n = 7), and 100 µg/mL pembrolizumab 45.7% ± 6.9% (n = 6). A one‐way ANOVA showed a treatment effect, F(2,16) = 7.03, p = 0.006, η ² = 0.47. Tukey post‐hoc comparisons revealed that 100 µg/mL > control (p = 0.007) and 100 µg/mL > 10 µg/mL (p = 0.028), whereas 10 µg/mL vs control was not significant (p = 0.709). These results suggest a dose‐dependent effect of PD‐1 inhibition in restoring T cell cytotoxicity within the tumor Chip model.

3. Discussion

In this study, we developed a first‐of‐its‐kind ex vivo tumor model of PDAC by integrating patient‐derived cancer organoids along with the major components of the tumor microenvironment (TME), including endothelial cells, fibroblasts, and immune cells. This model successfully recapitulated the in vivo cancer‐stroma interaction of PDAC. As proof of concept, we verified the anti‐tumor effect of a standard chemotherapy in PDAC, which was enhanced when the stroma was targeted, demonstrating the model's functional capacity in studying cancer‐stroma interactions for patient‐based combination therapies. Additionally, by utilizing patient‐derived T cells, we established a tumor immune model for the first time, allowing us to test immunotherapy effects in our PDAC Chip model ex vivo.

The traditional in vitro models, such as 2D cell lines and spheroids, fail to recapitulate the complex features of PDAC and its TME, hence limiting their translational relevance. In the pursuit of mimicking cancer pathophysiology, Organs‐on‐Chip (OoC) have emerged as promising platforms capable of more accurately modeling the PDAC TME by integrating dynamic fluid flow and precise control of multicellular interactions [99]. The PDAC TME consists of a mixture of immune cells, and fibroblasts [60, 66, 67]. While TME promotes cancer behavior in the cancerous epithelium, cancer cells modulate the TME to create a favorable environment for their growth and spread [54]. This multicellular communication between malignant epithelial cells and the surrounding stroma activates a desmoplastic reaction by TME cells, forming a dense stroma around the malignant cells [33]. By integrating PDOs with stromal components in a microfluidic system, we observed pronounced stromal remodeling, including fibroblast activation and collagen deposition, consistent with in vivo stroma activation by tumor‐derived signals [100].

Besides tumor growth, the PDAC TME can modulate drug response [32]. The use of the anti‐fibroblast agent ATRA attenuated the cancer‐associated stroma phenotypes in our model. Priming the tumor with anti‐stromal agents enhanced the chemotherapy effect, consistent with the reported effects of stroma in attenuating the chemotherapy effects in PDAC [101]. In line with a stromal‐mediated mechanism, ATRA and Clodrosome exhibited no direct impact on PDOs when tested in monoculture, as neither condition changed organoid size, shape, or apoptosis levels, highlighting the significance of the TME in influencing such a response to the addition of anti‐stroma to chemotherapy.

Cytotoxic CD8⁺ T cells play a crucial role in anti‐tumor immunity [102]. Cancer cells attract T cells primarily through the release of chemokines and/or presenting tumor‐related epitopes to T cells, leading to their activation [102]. The activated T cells then attack and destroy cancer cells. We leveraged our platform to advance ex vivo modeling of tumor immune response in PDACs. Patient‐derived T cells migrated more to the tumor as tumors grew, and the number of cancer cells increased. Notably, stroma enhanced the recruitment of cytotoxic T cells, yet these cells were not able to kill cancer cells, likely due to T cell dysfunction, according to our transcriptomic analysis [103, 104].

It is well established that T cell function is affected by cancer cells that express immune checkpoint molecules such as PD‐L1 that allow cancer to avoid destruction by T cells [105, 106] Immunotherapy, which has shown success in treating many cancers, works by re‐activating the T cells’ ability to killing cancer cells by blocking checkpoint inhibitors such as PD‐L1 with drugs such as pembrolizumab [96]. However, PDAC remains largely resistant to immune checkpoint blockade in clinical settings. An ex vivo model capable of recapitulating T cell recruitment and functional response within the PDAC TME is essential for advancing immunotherapy testing. In our PDAC model, we demonstrated that high‐dose (100 µg/mL) pembrolizumab, but not the lower dose (10 µg/mL), significantly enhanced T cell infiltration and restored tumor cell killing. The results from the low‐dose pembrolizumab, which falls in the range of systemic drug levels observed clinically, are consistent with the low efficacy of pembrolizumab immunotherapy reported in PDAC trials and lend promise to our ex vivo platform for modeling immunotherapy in PDACs [107]. The improved tumor killing observed with high‐dose pembrolizumab in our model could be due to more effective PD‐1 blockade in response to higher doses as reported elsewhere [108, 109, 110], and suggests higher concentrations may achieve better functional blockade, especially in dense tumor microenvironments like PDAC. Although simply increasing the pembrolizumab dose is not a viable solution to PDAC's resistance clinically [110, 111], other options for targeting mechanisms that drive T cell exhaustion could be considered [112].

Additionally, these findings are also consistent with PDAC's resistance to pembrolizumab immunotherapy observed in clinical trials and lend promise to our platform for advancing ex vivo modeling of immunotherapy in PDACs. Although clinically increasing the pembrolizumab dose is not a viable solution to PDAC's resistance [110, 111], other options for targeting mechanisms that drive T cell exhaustion could be investigated [112]. Our ex vivo model provides a comprehensive platform for future studies to test immunotherapy by targeting mechanisms that promote T cell exhaustion, including modulation of various components of the TME in a physiologically relevant framework [110, 111].

We are aware that our model is still limited in incorporating various immune cells present in the TME, such as natural killer cells (NKs), neutrophils, dendritic cells (DC), and other types of lymphocytes. Additionally, the use of allogeneic PSCs raises the possibility of MHC‐I mismatch contributing to the observed increase in T cell migration, potentially confounding the interpretation of immune recruitment. The PDOs were established from small endoscopic ultrasound‐guided biopsy (EUS) samples. While it was occasionally possible to grow tumor‐associated fibroblasts from these biopsies, the process is inconsistent. Even in cases where fibroblasts did emerge from the same specimen, their continued expansion and maintenance for reproducible organ‐chip cultures proved technically challenging. As a result, consistent autologous PDO–fibroblast pairing was not feasible at this time. We highlight the need for future models incorporating patient‐derived fibroblasts to better reflect autologous interactions. While our cytokine profiling provides functional evidence supporting TAM‐like polarization in our model, we acknowledge that future studies incorporating single‐cell transcriptomic analysis will be necessary to more precisely characterize CAF and TAM, define distinct cellular subpopulations, and validate the molecular programs underlying the functional phenotypes observed in this model.

We note that variability in PDO size and morphology was observed across experiments, which is expected for cancer PDOs and likely reflects differences in passage number, matrix stiffness, cancer heterogeneity, and co‐culture conditions. Importantly, despite this, PDOs consistently retained histologic features characteristic of PDAC, including epithelial organization with a basement membrane and luminal structures. We acknowledge that IC50 values obtained in PDO‐only cultures may differ from those in multicellular OoC models, where additional cell–cell and cell–matrix interactions can influence drug uptake, distribution, and resistance pathways. Our analysis evaluated ATRA in combination with Clodrosome, which prevented us from assessing the individual effects of each agent. Including additional conditions such as ATRA or Clodrosome alone would have required multiple simultaneous organ‐chip setups with replicates, which is technically challenging given the limited expansion and shorter‐term viability of primary organoids compared to cell lines. Therefore, we are unable to determine whether macrophage depletion is necessary to enhance ATRA's effects, highlighting the need for future studies to assess whether its stromal reprogramming effects are sufficient on their own or are augmented by simultaneous macrophage targeting. Despite these limitations, our model presents significant advantages, such as precise and compositional control of cells, a microscale format enabling low input of patient‐derived samples, and the capacity to model complex multicellular interactions within a tumor‐relevant architecture.

4. Conclusion

We present a platform for evaluating interactions between primary cancer cells and their TMEs. This approach can facilitate the development of more precise, pathophysiologically relevant platforms that better replicate in vivo tumors, thereby advancing personalized therapies for PDAC, which is largely resistant to current treatments.

5. Experimental Section/Methods

5.1. Human Cancerous Pancreatic tissue

We collected tissue biopsies from patients with primary PDAC who were undergoing an ultrasound endoscopy (EUS) to get a biopsy taken from their primary tumors from the pancreas for diagnostic purposes, as clinically indicated and determined by Rush's gastroenterologist for diagnostic purposes, after suspicion of PDAC based on clinical symptoms and radiological findings. During the EUS procedure, tissue biopsies were taken from the core of the mass via a fine needle aspiration method and then immediately preserved in the Advanced DMEM / F12 Media (Catalogue #12634‐010) until transferred to the lab. The diagnosis of PDAC was confirmed from the Pathology reports on EPIC by clinically certified GI Pathologists.

5.2. Human Normal Pancreatic Tissue

Normal post‐mortem pancreatic tissues were obtained from a donor provided by the non‐profit organ donation organization, Gift of Hope. The organization's transplant surgeons performed a pancreatectomy, excising the entire pancreas. The pancreas was preserved in UW solution, maintained on ice, and transported to Rush University Medical Center within 16 h for processing and organoid development. A midline cut was performed on the anterior surface of the pancreas to access the pancreatic ducts and obtain the ducts and adjacent pancreatic tissue.

5.3. Blood Collection to Isolate PBMCs

Fasting peripheral blood was collected in purple‐top tubes (BD Vacutainer, Catalogue #366643) from the same patients with PDAC at Rush University during the morning of the EUS procedure and processed immediately after collection to isolate PBMCs using the Lymphocyte Separation Medium LSM (Corning Cat. No. 25‐072‐Cl). The protocol by Corning was followed; shortly, the blood was aseptically layered on top of the LSM media, and the density gradient centrifugation was used to separate layers of lymphocytes from other blood components. The isolated lymphocytes were then transferred to a new tube and washed three times before resuspending in Cell Culture Freezing Medium (Gibco, Catalogue No: Q32853) and then cryopreserving the tubes in liquid nitrogen for future use.

5.4. Isolation of Monocytes

Primary monocytes were isolated using the EASYSEP isolation kit (Catalogue #19319) from the cryopreserved PBMCs following the protocol provided by the company. Primary monocytes were isolated on the day of the co‐culture with other cells and were not monocultured since they tend to stick to the plate strongly and differentiate into macrophages. We wanted the differentiation to happen in the co‐culture tumor model.

5.5. Isolation and Culture of T Cells

T cells were isolated from PBMCs using the EASYSEP isolation kit (Catalogue #17951) following the protocol provided by the company. Then, the isolated cells were monocultured for expansion using STEMCELL XF‐T cell media supplemented with 10 ng/mL Human Recombinant IL‐2 (STEMCELL; Catalogue #78145). T cells were activated on day 0 with 25 µL/mL of ImmunoCult Human CD3/CD28/CD2 T Cell Activator (Catalog #10970). T cells were cultured for 10 days before being used in co‐culture and cryopreservation. If T cells were cultured for more than 12 days, they were restimulated by adding 25 uL/mL of ImmunoCult Human CD3/CD28/CD2 T Cell Activator.

5.6. Flow Cytometry

At each indicated time point (days 0, 3, 5, 7, and 10 after initial stimulation, and day 17 following a day‐12 re‐stimulation), T cells were collected and washed in cold PBS. After Fc receptor blocking, cells were stained for surface markers in FACS buffer. The panel included anti‐human CD3, CD4, CD8, CD45RA, CCR7, and a fixable viability dye. Following two washes, cells were analyzed using a Sony MA900 flow cytometer. The gating strategy involved sequentially identifying singlets, live cells, and CD3⁺ T cells. The CD3⁺ population was then delineated into CD4⁺ or CD8⁺ subsets, from which memory populations were defined by CD45RA and CCR7 expression: naïve (CD45RA⁺CCR7⁺), central memory (TCM; CD45RA⁻CCR7⁺), effector memory (TEM; CD45RA⁻CCR7⁻), and TEMRA (CD45RA⁺CCR7⁻).

For the co‐culture experiment, a Transwell system was utilized where PDOs were grown in the upper insert, and autologous T cells were placed in the lower well. Co‐cultures were maintained for 48 h. Afterward, T cells were recovered from the lower well, washed, and stained with a surface panel containing CD3, CD4, CD8, CD45RA, CCR7, and PD‐1. Gating and analysis followed the flow‐cytometry pipeline detailed above.

5.7. Pancreatic Stellate Cells and Endothelial Cells

We utilized commercially available pancreatic stellate cells (PSCs, ScienCell, Catalog #3830), which get activated during co‐culture with PDOs, therefore eliciting a stromal response and collagen deposition [113]. PSCs were monocultured in Stellate Cell Media (SteCM, Cat# 5301) and passaged on a weekly basis for maintenance. PSCs less than passage 10 were used for downstream co‐culture experiments. We additionally utilized a commercially available HUVEC cell line (Promocell, Catalogue# 22022), which was isolated from the umbilical vein of pooled donors. These endothelial cells were monocultured in complete EC‐Cult‐XF media (STEMCELL, Catalogue #08003) and passaged regularly for maintenance before downstream co‐culture experiments.

5.8. Development of Patient‐Derived Organoids (PDOs)

The biopsy samples were processed immediately after collection. First, the samples were washed with a wash buffer, digested to isolate cancer cells with 200 U/mL Collagenase Type 2 (Gibco, Catalogue # 17101015) and generate 3D PDOs in Matrigel Matrix (Corning, Catalogue #356231) domes following the well‐recognized previously published protocol [68], as shown in Figure 1A. The PDOs were left to grow for ∼16 days until they were fully grown. Once grown, PDOs were then passaged with mechanical breakage for growing/cryopreserving purposes. PDOs were maintained in monoculture and passaged regularly until other tumor components were ready for the co‐culture experiments (Figure 1B).

5.9. Design and Manufacture of Microfluidic Device (The Chip)

The microfluidic devices were fabricated using polymethyl methacrylate (PMMA) sheets, a material commonly employed in rapid prototyping of microfluidic systems due to its optical clarity, biocompatibility, and ease of machining. The device consisted of four layers: a base, a lower microchannel layer, an upper microchannel layer, and a cover. Each layer was designed using CAD software and precision‐cut with a laser cutter. After cutting, the components were thoroughly cleaned with 70% ethanol and deionized water, followed by air drying. Residual particulates were removed using Scotch tape to ensure optimal surface cleanliness before bonding. A porous membrane (3 µm pore size/polyethylene terephthalate) was carefully aligned and sandwiched between the lower and upper microchannel layers. The layers were assembled and bonded using a solvent‐assisted thermal bonding technique, wherein a mixture of 80% isopropanol and 20% dichloromethane was applied to the contact surfaces. The stacked device was then incubated at 75°C for 15 min to promote complete bonding and sealing without deforming the microstructures. Following assembly, metal connectors were inserted into the inlet and outlet ports, and silicone tubing was affixed to facilitate fluid handling and connection to external flow systems.

5.10. FITC‐Dextran Permeability (Endothelial Barrier Integrity)

To promote adhesion, the upper (vascular) chamber of each Chip was pre‐treated with Animal Component‐Free (ACF) Cell Attachment Substrate (STEMCELL, Catalogue #07130) for 2 h at room temperature. HUVECs (50 000 cells per Chip) were then seeded into the pre‐coated upper chamber and allowed to adhere overnight, after which perfusion was initiated at 10 µL/h. Endothelial barrier function was assessed on days 1, 2, and 3 post‐seeding using a FITC–dextran permeability assay: 4 kDa FITC–dextran (1 mg/mL in assay medium) was introduced into the upper chamber, incubated for 2 min, and effluent from the lower channel was collected. Fluorescence was measured on a plate reader (Ex 485 nm/Em 528 nm); permeability was reported as effluent fluorescence units. A 7‐point 1:2 serial dilution of 4 kDa FITC–dextran (1000–15.6 µg/mL; blank = 0) was run in duplicate; fluorescence values were normalized to the maximal signal, concentrations were log10‐transformed, and a four‐parameter logistic (variable‐slope) model (“log[agonist] vs. normalized response”) was fit in GraphPad Prism 9.5.0. The fit was excellent (R ² = 0.966; n = 21 points, DF = 19) with EC₅₀ = 408.2 µg/mL (95% CI 364.5–455.5) and Hill slope = 2.242 (95% CI 1.800–2.839), confirming an appropriate dynamic range for the assay.

5.11. Transwell Experiments

Transwell permeable plates (Costar, Catalogue #3415) were utilized to study the impact of PDAC tumor co‐culture on primary fibroblasts or T cells. First, the tumor (PDO + Monocytes + PSCs) was seeded in the upper insert of a transwell chamber (3 µm polycarbonate membrane) or control (Matrigel only) while target cells (primary fibroblasts or T cells) were plated in the bottom well plate, allowing for exchange of soluble factors without direct cell contact. Complete culture media were added to both compartments, and the plate was placed at 37°C in a humidified incubator with 5% CO₂ for 2 days. Cells were then harvested for downstream RNA analysis.

5.12. Calculation of IC50 of Gemcitabine

PDOs were dissociated into single cells with TrypLE, and 10,000 cells were seeded per well in 96‐well U bottom Elplasia microplates (Corning, Catalogue #4442) in complete organoid media supplemented with 20% Matrigel. Each Elplasia well contains 79 microwells, enabling the growth of 79 individual organoids (Figure 4A). The cells were incubated in 37 Celsius and 5% CO2 and allowed to form PDOs. On day three, when PDOs were established, triplicate wells were treated with five serial dilutions of gemcitabine (100, 20, 4, 0.8, and 0.16 µm) for 48 h. Cell viability was assessed using the ApoLive‐Glo Multiplex Assay (Promega, Cat# G6410). IC50 values were calculated using a non‐linear regression model of normalized viability data in GraphPad Prism version 9.5.0.

5.13. Clodrosome Dosage and Delivery and Validation

Clodronate liposomes (Clodrosome, Catalogue #CLD‐8909) were administered by local perfusion into the upper (vascular) channel of the OoC at 16.58 µg/mL, using the same syringe‐pump setup and flow rate used for other agents (10 µL/h). Perfusion was continuous for 72 h. This dose was selected based on prior work showing selective macrophage susceptibility without direct cytotoxicity to PDAC cells [45]. Before drug‐response experiments, we verified on‐target activity: primary monocytes were cultured for two days, then treated with Clodrosome (16.58 µg/mL) or vehicle control for 72 h, fixed, and stained for cleaved caspase‐3 (c.c.3). Clodrosome increased c.c.3‐positive cells relative to controls (OR 3.57, 95% CI 2.33–5.47, p < 0.001; Figure S2), confirming effective macrophage depletion.

5.14. Monocyte – M1 – M2 Differentiation and Cytokine Assay

Primary monocytes were isolated from human PBMCs using the Monocyte Isolation Kit (STEMCELL, Catalogue No: 19359), according to the manufacturer's protocol. Purified monocytes were cultured for 4 days in complete ImmunoCult‐SF Macrophage Medium (STEMCELL, Catalogue No: 10961), supplemented with 50 ng/mL of Human Recombinant M‐CSF. On day 4, subsets of monocytes were polarized toward M1 or M2 macrophages following the STEMCELL macrophage differentiation protocol: M1 macrophages were generated by supplementation with 10 ng/mL LPS and 50 ng/mL IFN‐γ, while M2 macrophages were generated by stimulation with 10 ng/mL IL‐4. On day 6 (48 h after incubation), culture medium was replaced with fresh complete macrophage differentiation medium. Conditioned media (CM) were collected at 1, 6, and 24 h following the media change. Cytokine levels were then quantified using Meso Scale Discovery (MSD) V‐PLEX Human Pro‐Inflammatory Panel, according to the manufacturer's protocols. The following cytokines were assessed: interferon‐γ (IFN‐γ), tumor necrosis factor‐α (TNF‐α), interleukin‐6 (IL‐6), interleukin‐4 (IL‐4), and interleukin‐10 (IL‐10).

5.15. Monocyte – PDO Transwell Cytokine Assay

Primary human monocytes were seeded in the lower chamber of a Transwell system and maintained for four days. Cancer PDOs were then added in Matrigel/Media (80/20) to the upper insert and co‐cultured for 48 h. After changing the media with fresh media, CM was collected at 1, 6, 24 h and assayed on the Meso Scale Discovery (MSD) V‐PLEX Proinflammatory Panel 1 Human kit (Catalogue No: K15049D‐1) per manufacturer's instructions. Two technical replicates were used for each timepoint on a single biological sample. For analysis, each cytokine was modeled as Value ∼ Condition + Timepoint to compare monocytes only (control) with monocytes + PDOs with timepoint as a fixed effect; p‐values were adjusted across cytokines using Benjamin‐Hochberg.

5.16. RNA‐Seq and Data Processing

We extracted RNA using the RNeasy Micro kit (Catalogue No: 74004, Qiagen) following the protocol provided by the company. The RNA concentration was quantified utilizing Qubit (Catalogue No: Q10211, Invitrogen). We carried out RNA sequencing using Revvity's kits. We then performed ribodepletion (Catalogue No: NOVA‐512965, Revvity) before generating cDNA (Catalogue No: NOVA‐5198‐53) utilizing an automated workstation (Sciclone G3 NGSx iQ). Next, the libraries were sequenced using an Illumina NovaSeq X sequencer equipped with a 10B flow cell, targeting an approximate read depth of 35 million aligned reads per sample. Paired‐end FASTQ files were aligned to the Homo sapiens reference genome (hg38) using STAR aligner (RRID:SCR_004463). Expression counts were generated using the featureCounts program (RRID:SCR_012919). Differential expression analysis was performed using the edgeR package in R (RRID:SCR_012802). Genes with a p‐value < 0.05 and a false discovery rate (q‐value) < 0.05 were considered statistically significant.

5.17. Immunofluorescence Staining

At the end of the experiments, Chips were perfused with 4% PFA for 30 min, followed by a wash with PBS. Permeabilization was performed using 1% Triton X‐100 (Sigma, Catalogue #9002‐93‐1). Next, we infused a blocking buffer for 1 h, added primary antibodies, and incubated at 4°C overnight. The next day, the Chips were washed three times, then secondary antibodies were added for 1 h at room temperature. DAPI was used to counterstain the nucleus. Images were obtained with an Inverted Axio Observer 7 fluorescent microscope (Zeiss, 491917‐0001‐000Z), which was equipped with Apotome 3 and a long objective. The CZI microscope images were then converted to TIF format and analyzed using ImageJ software (RRID:SCR_003070). Plots were made using ggplot2 (RRID:SCR_014601) and ComplexHeatmap (RRID:SCR_017270). The following primary antibodies were used: Anti CA19‐9 (abcam, Catalogue #AB116024), Anti‐Hu CD31 (PECAM‐1) (Invitrogen, Catalogue #14‐0319‐82), Anti p‐ERK (P‐p44/42 MAPK T202/Y204) (Cell Signaling, Catalogue #9101L), Anti CD68 (D4B9C) XP (R) (Cell Signaling, Catalogue #76437S), Anti CD163 (abcam, Catalogue# ab182422), Anti EpCAM (VU1D9) (Cell Signaling, Catalogue #2929S), Anti Alpha Smooth Muscle Actin (Invitrogen, Catalogue #MA1‐06110), Anti Collagen 1 (Invitrogen, Catalogue #PA5‐95137), and Anti Cleaved Caspase 3 (Asp175) (Cell Signaling, Catalogue #9661S). The following secondary antibodies were used in this study as appropriate: Alexa Flour 555 donkey anti‐mouse (Invitrogen, Catalogue #A31570), Alexa Flour 488 donkey anti‐mouse (Invitrogen, Catalogue #A21202), Alexa Flour 555 donkey anti‐rabbit (Invitrogen, Catalogue #A31572), Alexa Flour 488 donkey anti‐rabbit (Invitrogen, Catalogue #A21206).

5.18. Statistical Analysis

We performed data analysis using R version 4.5.0 (released April 11, 2025), running within RStudio 2025.05.0+496 (RRID:SCR_000432). All statistical analyses, data processing, and visualizations (except the IC50 analysis) were done in R. For calculating the IC50 drug response analysis, we utilized GraphPad Prism (version 9.5.0) to fit dose‐response curves (four‐parameter logistic model).

For the nuclear‐to‐cytoplasm (N/C) ratio (Figure 1C), single cell N/C ratios were compared between groups using a Fligner–Killeen variance test (two‐sided, α = 0.05; normal n = 16 cells, cancer n = 17 cells). CK‐19 immunofluorescence intensities (Figure 1G) were background‐corrected, log10‐transformed, and compared by Wilcoxon rank‐sum (two‐sided; n = 6 fields per group). Western blot p‐ERK/ERK ratios (Figure 1H–I) were obtained by band densitometry from ImageJ, normalized to total ERK, and compared by Welch's t‐test (two‐sided; cancer n = 3 PDO samples, normal n = 2). Macrophage cytokines (Figure 1L–P) were quantified from MSD standard curves; for each cytokine, independent Welch tests at 1, 6, and 24 h compared M1 vs M2 with Benjamini–Hochberg adjustment across timepoints (n = 2 per group per time). Unless noted, data reported in the texts as mean ± SD; normality/variance were inspected (Shapiro–Wilk/Levene) to guide test choice; no outliers were removed a priori.

For PDO diameter over time (Figure 2D) and tumor growth (% surface area) (Figure 2E), repeated measures from PDO‐only vs full model (PDO+PSC+Monocytes) across days were analyzed with a linear mixed‐effects model (lme4) including fixed effects for Group, Day, and Group×Day; per‐day contrasts were assessed with two‐sided Welch t‐tests and Benjamini–Hochberg adjustment within figure/panel. Sample sizes: n = 3 per group for both panels. For p‐ERK immunofluorescence intensity (Figure 2I), results were compared between normal vs cancer using a two‐sided Wilcoxon rank‐sum test (n = 4 FOVs per group). Data were reported as mean ± SD in the text; assumptions (distribution/variance) were inspected to guide test choice; α = 0.05.

For the PSC spread area (Figure 3C), analysis from PSCs co‐cultured with cancer PDO vs normal PDO was compared with a two‐sided Wilcoxon rank‐sum test (cancer n = 13 cells; normal n = 12 cells). For fibroblast bulk RNA‐seq (Figure 3E–I), counts were filtered for low abundance, TMM‐normalized (edgeR), and an assumed common dispersion = 0.20 was used in the classic edgeR pipeline to test differential expression; p‐values were BH‐adjusted and significance defined at FDR<0.05; volcano plots display log₂ fold‐change vs –log₁₀(FDR) and heatmaps show z‐scored log₂‐CPM of DE genes. Pathway enrichment used MSigDB Hallmark gene sets with FDR control. For alignment with scRNA‐seq (GSE205013), raw single‐cell data were reprocessed to derive cluster centroids (epithelial, endothelial, CAF, T cells); bulk RNA‐seq samples were combined with these centroids, log‐normalized, visualized by PCA (prcomp), and compared by Spearman correlation to CAF centroids; a myCAF program score was computed on bulk RNA as the mean z‐score of the canonical set (ACTA2, TAGLN, MYL9, COL1As1, COL1A2, ITGA11, THBS2, MMP2, PDGFRB, FAP, COL11A1). For monocyte cytokines (Figure 3J–O), MSD concentrations (pg/mL) were analyzed per cytokine at 1, 6, and 24 h with two‐sided Welch t‐tests comparing M0 vs M0+PDO, applying BH adjustment across times within each cytokine (n = 2 per group per time). Mean ± SD was reported in the text. For the multidimensional scaling plot (MDS), cytokine concentrations (pg/mL) for IFN‐γ, TNF‐α, IL‐4, and IL‐10 were log10‐transformed and z‐scored within cytokine across all samples (M0, M0+PDO, M1, M2; 1, 6, 24 h). For each condition/time, we computed composite signatures: M1 score = mean(z [IFN‐γ], z_[TNF‐α]) and M2 score = mean(z_[IL‐4], z_[IL‐10]). Points were plotted in a 2D space with x = M1 score and y = M2 score, using M1 and M2 as positive‐control anchors; lines connect the trajectory from M0 to M0+PDO over time. n = 2 per group per time were averaged for display.

IC₅₀ for gemcitabine (Figure 4A) was obtained by fitting a four‐parameter logistic (4PL) dose–response model to log₁₀‐transformed concentrations in GraphPad Prism (Version 9.5.0); the IC₅₀ and 95% CI were taken from the nonlinear regression (single PDO line). PSC extension (Figure 4C) in the full tumor model vs tumor+ATRA was quantified and compared with a two‐sided Wilcoxon rank‐sum test (tumor n = 9, tumor+ATRA n = 15). Apoptosis (% c.c.3⁺ among EpCAM⁺) across DMSO, gemcitabine, and gemcitabine+ATRA (Figure 4E) was analyzed at the FOV level using a one‐way ANOVA with Tukey post‐hoc tests (DMSO n = 6, Gemcitabine n = 6, gemcitabine+ATRA n = 8); a Kruskal–Wallis test with Dunn's adjustment was performed. Data were reported as mean ± SD, tests are two‐sided (α = 0.05).

For T cell infiltration (Figure 5B), the number of cells/FOV was analyzed across 1K (n = 6), 10K (n = 4), and 10K+Stroma (n = 5) using a one‐way ANOVA with Tukey post‐hoc comparison. For T cell counts in the lower chamber (Figure 5C), PDO‐only (n = 3) vs PDO+Stroma (n = 6) were compared with a two‐sided Welch's t‐test. For tumor apoptosis (Figure 5D; outcome = % c.c.3⁺ among EpCAM⁺ cells), PDO‐only (n = 3) vs PDO+Stroma (n = 6) were compared using a two‐sided Welch's t‐test. For T cell RNA‐seq (Figure 5F–I), raw counts were filtered for low abundance, TMM‐normalized (edgeR), and an assumed common dispersion of 0.20 was used in the classic edgeR pipeline; p‐values were BH‐adjusted with significance at FDR<0.05. Volcano plots display log₂ fold‐change vs –log₁₀(FDR); heatmaps show z‐scored log₂‐CPM; GO enrichment reports FDR‐controlled pathways. Data were reported in the results section as mean ± SD, tests were two‐sided (α = 0.05). For the PCA (Figure 5F), bulk T cell RNA‐seq (± co‐culture) was log‐normalized and projected with prcomp alongside PDAC scRNA‐seq cluster centroids (GSE205013; T cells, CAF, epithelial, endothelial).

For flow cytometry (Figure 6A–C), events were gated as singlets → live → CD3⁺, then split into CD4⁺/CD8⁺ and memory subsets by CD45RA/CCR7 (naïve, T_CM, T_EM, T_EMRA); outcomes are reported as percent of parent (technical replicates averaged), and given a single donor longitudinal series and the exploratory co‐culture, results are descriptive without formal hypothesis testing. For T cell infiltration (Figure 6D; cells/FOV), chip‐level means were analyzed across Control (n = 6), pembrolizumab 10 µg/mL (n = 7), and 100 µg/mL (n = 6) using one‐way ANOVA with Tukey post‐hoc comparisons (assumptions checked; Kruskal–Wallis/Dunn's prespecified if violated). For tumor apoptosis (Figure 6F; % c.c.3⁺ among EpCAM⁺), chip‐level means were compared across three groups with one‐way ANOVA and Tukey post‐hoc (effect sizes reported as η ² when applicable). Tests were two‐sided (α = 0.05), and data were reported as mean ± SD in the results section.

For the statistical analysis of the effects of ATRA and Clodrosome on PDOs, percent apoptosis was quantified as % c.c.3⁺ among EpCAM⁺ PDO cells at the FOV level; groups were Control (n = 4 FOVs), ATRA (n = 6), Clodrosome (n = 4). No outliers were removed. Distributional assumptions were inspected; because data were non‐Gaussian across groups, we used a Kruskal–Wallis test (two‐sided, α = 0.05) with Dunn's post‐hoc pairwise comparisons and multiplicity adjustment.

For testing the impact of Clodrosome on monocytes (Figure S2), the outcome was the c.c.3⁺/DAPI ratio for monocytes ± Clodrosome (16.58 µg/mL), analyzed at the measurement level (Control n = 6, Clodrosome n = 10) with a binomial GLMM (logit link) including a random intercept for Subject (two biological donors) to account for between‐donor variability; two‐sided α = 0.05. The model showed higher odds of c.c.3 positivity with Clodrosome (OR = 3.57, 95% CI 2.33–5.47, p < 0.001); estimated marginal means were 29% (Control) vs 59% (Clodrosome), with adequate fit (Pearson χ²/df = 1.12). Statistical significance was determined based on a p value less than 0.05. Asterisks were used in the figures to mark degree of significance: * p < 0.05, ** * p < 0.01, *** p < 0.001.

5.19. Alignment With the ScRNA‐Seq

We ran an analysis to examine the alignment of our transcriptomic data from RNA‐seq analysis with publicly available scRNA‐seq data of PDAC tissue [76]. First, raw scRNA‐seq data were downloaded from GEO with accession number GSE205013 from the following subject IDs (P04, P07, P15, P19, and P23). All these subjects were treatment naïve, and the single cells were obtained from a pancreas resection. We followed similar methods provided within the manuscript to re‐analyze the data to identify various clusters in the pancreatic cancer tissue, such as epithelial, endothelial, CAFs, and T cells. Then, we extracted gene expressions from the clusters and combined them with bulk RNA‐seq data. Then we log‐normalized the data before running the prcomp function in R (RRID:SCR_001905) to run principal component analysis (PCA). We then created plots to compare our bulk RNA‐seq results with scRNA‐seq clusters. For that, bulk RNA‐seq from primary fibroblasts or T cells cultured alone or co‐cultured with cancer PDOs were combined with scRNA‐seq cluster centroids (T cells, epithelial, CAF, endothelial). Gene IDs were mapped to ENSEMBL. Spearman correlations were computed between primary fibroblasts and the CAF centroid. The significant gene differences in the CAF cluster were defined as a PDAC CAF reference from the ROC markers with AUC >0.70. To calculate myCAF program score, the canonical myCAF set (ACTA2, TAGLN, MYL9, COL1A1, COL1A2, ITGA11, THBS2, MMP2, PDGFRB, FAP, COL11A1) was scored on bulk RNA only as the mean column‐wise z‐score of genes present after log₂ normalization.

5.20. Study Approval

The Institutional Review Board (IRB) at Rush University Medical Center authorized the collection of pancreatic cancer biopsies and blood from patients undergoing diagnostic endoscopic ultrasound as per research protocol (IRB Number: 22013102). All individuals who participated provided an informed consent form (ICF). The IRB allowed the acquisition of post‐mortem cadaveric pancreas, categorizing the organ as non‐human tissue (118) from the Gift of Hope (GOH) organization that manages organ donation in Illinois, US. The GOH secured approval to utilize the donor pancreas for research from the donors/families, as part of the general consent acquired for transplantation and research purposes. The pancreas was available for research only when the organ was considered “not suitable” for transplantation or when the transplant list was exhausted.

5.21. Cell Trafficking Dye

T cells were stained with red trafficking dye (Abcam, Catalogue #AB269446) for 30 min and washed three times before adding to the upper chamber (vessel) of the Chips.

Funding

NIH/NCI grants (CA277110, CA279487, and OD039982) to F.B.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File: advs73536‐sup‐0001‐SuppMat.pdf.

Data Availability Statement

All sequencing reads generated for this study have been deposited in the National Center for Biotechnology Information (NCBI) BioProject database under accession number PRJNA1415166. Other datasets are available to editors, reviewers, and readers from the corresponding author upon request.