Quiescent tumor cells shape the immunosuppressive microenvironment in pancreatic cancer

Abstract

Immunotherapy, including chimeric antigen receptor (CAR) T-cell therapy, has limited activity in pancreatic ductal adenocarcinoma (PDAC). Using orthotopic PDAC mouse models, we identified a rare population of quiescent PDAC cells that increases after CAR-T cell therapy and exhibits relatively higher clonogenic growth and self-renewal potential than bulk tumor cells. These quiescent cells express high levels of Epiregulin (EREG), a secreted ligand for EGFR and ErbB4, and induce an immunosuppressive tumor microenvironment by increasing the frequency of ErbB4-expressing tumor-associated macrophages. Using complementary genetic and pharmacologic approaches, we demonstrate that targeting EREG enhances the sensitivity of quiescent tumor cells and PDAC tumors to CAR T-cell therapy, resulting in reduced relapse and improved overall survival. These findings support a model in which rare quiescent tumor cells contribute to remodeling of the PDAC tumor microenvironment through EREG-associated signaling and suggest that EREG inhibition may enhance the efficacy of adoptive cellular immunotherapy in this disease.

CAR T cell therapy remains largely ineffective in solid tumors, where quiescent cancer cells contribute to treatment resistance. Here, the authors identify a population of quiescent pancreatic cancer cells that upregulate epiregulin (EREG) and mediate immunosuppression upon CAR T treatment.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is projected to be the second leading cause of cancer-related deaths in the United States by 2030¹. A major factor in these poor clinical outcomes is resistance to both cytotoxic chemotherapy and targeted agents. Similarly, immune checkpoint inhibitors have limited clinical activity, except in rare cases of PDAC with microsatellite instability (MSI)^2,3. Adoptive immunotherapy with CAR-T cells has demonstrated clinical efficacy in PDAC, but initial tumor responses are not sustained^4,5. Therefore, further insights into the mechanisms of resistance may lead to improved immune-based approaches for this disease.

Multiple mechanisms contribute to immunotherapy resistance in PDAC, including low neoantigen burden, the dense desmoplastic stroma, and the accumulation of immunosuppressive cell populations within the tumor microenvironment (TME)^6,7. Myeloid-derived cells, particularly macrophages, are central to establishing and maintaining the immunosuppressive TME, thereby limiting the activity of both endogenous and adoptively transferred cytotoxic T cells^6–13. While the immunosuppressive nature of the PDAC TME is well recognized, its adaptation to active anti-tumor immunity is not well understood. However, further insights into these changes may identify specific mechanisms and therapeutic targets to improve the durability of immune-based therapies.

Functionally distinct cell populations have been identified in PDAC with relatively enhanced clonogenic growth, tumor initiation, and metastatic potential compared to bulk tumor cells^14–17. These properties suggest that cancer stem cells (CSCs) play a central role in tumor regrowth and progression¹⁸. Both normal adult stem cells and CSCs often reside in a quiescent state^19–24, which confers protection from immune surveillance by down-regulating MHC class I and NK cell-activating ligand expression^25–28. Furthermore, quiescent cancer cells can actively promote local immune suppression within the TME by inducing hypoxia and recruiting immunosuppressive cells through secreted factors^{25–27,29,30}.

Given the intrinsic resistance of quiescent cancer cells, we developed a PDAC model to examine the impact of CAR T-cells targeting tumor cells in an MHC-independent manner^28,31–33. Using a live-cell quiescence reporter, we found that quiescent cancer cells with CSC properties are significantly enriched after CAR T-cell therapy. Quiescent cancer cells also expressed high levels of Epiregulin (EREG), an EGF-family ligand for EGFR/ErbB4 with established roles in normal tissue repair, regeneration, and inflammation^31,34–38. Further EREG expression is associated with low CD8⁺ T-cell and high M2 macrophage infiltration in human PDAC³⁹. Targeting EREG in tumors enhanced CAR T-cell activity, including improved killing of quiescent tumor cells, reduced tumor regrowth, and extended survival. These EREG-mediated effects were dependent on ErbB4⁺ macrophages, indicating that quiescent tumor cells orchestrate an immunosuppressive TME through macrophages and that targeting EREG may improve adoptive cell therapy in PDAC.

Method

Ethics statement

All experiments were conducted in accordance with relevant ethical regulations. Animal studies were approved by the Institutional Animal Care and Use Committee at the University of Texas at Austin (Protocol No. IPROTO202500000255). Human PDAC specimens were collected with written informed consent under protocols approved by the University of Texas at Austin Institutional Review Board (Protocol No. 2017060074). No participants received compensation.

The maximum permitted tumor burden under the approved Institutional Animal Care and Use protocol was 1.5 cm in diameter. Tumor growth was monitored throughout the study, and no animal exceeded this maximum permitted tumor size. Both male and female mice were used throughout the study. Sex was considered during study design; however, experiments were not powered to detect sex-specific effects and data from both sexes were pooled for analysis.

Cell culture

mT4-2D murine PDAC cell lines were provided by David Tuveson at Cold Spring Harbor Laboratory, while KPC-/- and KxPxC GFP-Luc were provided by Cassien Yee at The University of Texas MD Anderson Cancer Center^40,41. All three cell lines were maintained in DMEM with 10% FBS and 100 μg ml⁻¹ Primocin (Invivogen, antpm05). HEK 293FT cells were purchased from ATCC and cultured in DMEM with 10% FBS and 100 μg ml⁻¹ Primocin. Cells were tested for mycoplasma by PCR (Southern Biotech, 1310001) every 2 months and not passaged more than 15 times.

Patient-derived PDAC xenografts

PDAC surgical specimens were collected with written consent under the University of Texas at Austin IRB-approved-protocol (Protocol No. 2017060074). Patient-derived xenografts (PDX) were established and passaged in Foxn1^Nu/J (Nude; strain no. 000819, Jackson Laboratory) mice, cryopreserved between passages F3-F6, and stored in liquid nitrogen.

Animals

All experiments were conducted within AAALAC guidelines and approved by the University of Texas Institutional Animal Care and Use Committee (Protocol No. IPROTO202500000255). C57/BL6J (strain no. 000664) and NSG (strain no. 005557) mice were either purchased from the Jackson Laboratory or bred in-house. Mice were housed in a specific pathogen-free facility maintained at 20–26 °C with 30–70% relative humidity under a 12 h light/12 h dark cycle with ad libitum access to food and water.

Molecular cloning and generation of DNA constructs

All cloning was performed using gBlocks from Integrated DNA Technologies and NEB HIFI assembly master mix (New England Biolabs). All restriction enzymes used were from New England Biolabs. All plasmids were confirmed by NGS whole plasmid sequencing through plasmidsaurus (https://plasmidsaurus.com). To generate the human CD19 (hCD19) tdTomato-luc2 construct, the pCDH-EF1-luc2-P2A-tdTomato plasmid (Addgene, 72486) was digested with SalI-EcoRI and subjected to HIFI assembly with an hCD19 gBlock containing a 40 bp homology to the digested vector. The hCD19 sequence was obtained from MSCV-hCD19 (Addgene, 127889). To generate the p27k-mVenus-BlastR and p27k-miRFP670 constructs, part of the 2 A sequence and puro resistance gene was removed from the p27k-mVenus plasmid (Addgene, 176651) with BamHI-SalI and replaced in frame with a gBlock containing a 40 bp homology with the T2A-BlastR cassette. The p27k-miRFP670 vector was constructed by digesting the p27k-mVenus vector with XbaI-EcoRV and assembling with a miRFP670 gBlock containing a 40 bp homology to the digested vector. Both sequences (BlastR and miRFP670) were obtained using Snapgene. All cloning schemes were first tested using the Snapgene NEB HIFI assembly tool.

Constitutive gene knockdown was carried out using shRNAs (Millipore Sigma): Non-target shControl: cat #: SHC016, GCGCGATAGCGCTAATAATTT; shEREG: cat #: TRCN0000250419, TGCGATGTGAGCACTTCTTTC; shNGF: cat #: TRCN0000351084, CCTGAAGCCCACTGGACTAAA; shSPRR1a: cat #: TRCN0000098406, CAACTGTCACTCCATCACCAT, shERBB4-1: cat# TRCN0000360621, GAGAATAGTGATCCGAGATAA; shERBB4-2: cat #: TRCN0000023405, CCACATAACTTCGTGGTAGAT.

Inducible EREG knockdown was carried out by cloning an EREG shRNA, TRCN00002419 (TGCGATGTGAGCACTTCTTTC), into the TetON-pLKO-puro vector (Addgene, 21915) using the manufacturer’s protocol, except transforming plasmids into NEB stable competent cells instead of Stbl3. The scramble control shRNA (shScramble) was obtained from TetON-pLKO-puro-Scramble (Addgene, 47541).

The anti-hCD19 CAR vector was constructed using the CD19 scFV sequence (Addgene, 194457) derived from the anti-CD19 clone FMC63, and the anti-murine mesothelin scFV sequence was obtained from SS scFV⁴². To generate anti-mesothelin CAR T-cells expressing mTagBFP2, the eGFP gene was replaced with mTagBFP2 using a gBlock with sequence from Snapgene. Both scFV sequences for the murine CAR T-cells contain the same leader sequence and linkers followed by murine CD28 and CD3z endodomains.

The anti-human mesothelin CAR was assembled using the lentiviral expression vector pRRL.SIN.EF1A.CD19-FMC63.218.CAR-2G (Addgene, 194457), which was digested with EcoRI and NotI, and the anti-mesothelin scFV from the Amatuximab sequence (PDB: 8CXC_A). The vector contains a GMCSF signal peptide, CD8A hinge transmembrane domain, and 4-1BB, and CD3 signaling domains as well as bicistronic expression of a dual GFP-Bleomycin selection cassette. The mesothelin scFV was cloned downstream of the GMSF signal peptide and in frame with the CD8A hinge to generate anti-mesothelin human CAR T-cells.

Subcutaneous tumors

NSG or C57/BL6J mice were shaved and ear tagged under isoflurane anesthesia. Cell suspensions (150,000 for NSG, 250,000 in C57BL6J) in a volume of 100 μl growth factor-reduced Matrigel (Corning, CLS356230) diluted 1:1 with PBS were implanted under the skin of left and/or right flanks using a 28-gauge syringe. The syringe was left under the skin for 10 s to allow the Matrigel to solidify.

Orthotopic pancreas tumors

Mice were prepped for surgery the day before by shaving and ear tagging. The day of surgery, mice received pre-operative analgesia (Ethiqa XR extended-release buprenorphine), administered by the University of Texas veterinary staff. Carprofen (MedVet International, 503045269) was diluted to a final concentration of 2 mg ml⁻¹ in sterile saline and administered pre-operatively (10 mg kg⁻¹). Mice were anesthetized with 3% isoflurane and prepped for surgery by cleaning the abdomen with betadine solution and isopropanol three times each. Sterile eye lubricant (MedVet International, 730979) was applied before surgery. An initial 2 cm abdominal incision was made to expose the peritoneum, followed by a second 2 cm incision lateral to the spleen. The spleen and pancreas were extracted and placed on a sterile pre-wet cotton swab, and a 1:1 mixture of Matrigel:PBS (50 μl) containing 50,000 mT2-2D, 200,000 KPC-/-, or 500,000 KxPxC GFP-luc cells was injected into the pancreas. The needle was held for 20 s to allow the Matrigel to solidify, and successful injections were verified by a small transparent bubble in the tail of the pancreas. The spleen and pancreas were carefully replaced, and the peritoneum was closed using 5–0 vicryl sutures (MedVet International, VCP593G) and outer skin with 9 mm wound clips (AutoClips, Braintree Scientific). Mice received post-operative analgesia using carprofen (10 mg kg⁻¹) every 24 h for up to 72 h post-surgery and were monitored two times per day for the initial 48-h recovery period. Mice with extensive tumors outside of the pancreas but lacking pancreatic tumors (wound tumors from cell suspension spillover) were removed from the study.

For adoptive transfer of CAR T-cells, mice were placed on a doxycycline chow diet (Bio-Serv, 14-727-450) for 3 days prior to and 11 days after administration of CAR T-cells (For TetON-shRNA tumors). Mice received a single dose of cyclophosphamide I.P. (180 mg kg⁻¹), followed 24 h later by an I.P. injection of 0.25–5 million CAR T-cells. Tumors were collected 3 days post-CAR T-cell administration for spatial transcriptomics to limit the extent of cell death complicating the analysis. For other studies, tumors were harvested 1 week after CAR T-cell treatment; this time point was earlier than S.C. tumor studies to overcome the low and inadequate recovery of tumor cells for downstream analyses.

To deplete macrophages, the same adoptive transfer/doxycycline protocol was followed, except that clodronate liposomes (APExBio, K2721) or PBS-loaded vehicle control liposomes (APExBio, K2721) were additionally injected I.P. on the day of and 72 h after CAR T-cell administration. Each treatment consisted of 200 μl of vehicle or clodronate-loaded liposomes (5 mg ml⁻¹).

Animal imaging

For bioluminescence (BLI) imaging, mice were injected intraperitoneally with 100 μl of D-luciferin (Syd Labs, MB000102-R70170) dissolved in PBS (30 mg ml⁻¹). Animals were anesthetized with isoflurane for imaging using the IVIS instrument (Revvity) in the Biomedical Imaging Center (RRID:SCR_021898) within the Center for Biomedical Research Support at the University of Texas at Austin. Images were analyzed and quantified using the ROI function of the BLI signal in Living Image software v4.8.2 (IVIS imaging systems).

EREG neutralizing antibody production, purification

Neutralizing antibodies were produced by the Advanced Protein Therapeutics Core Facility (RRID # SCR_023740) at the University of Texas at Austin using clone H321 as the template⁴³. VH and VL sequences were codon-optimized for CHO expression using the Twist Bioscience platform and synthesized with flanking regions for Gibson assembly into AbVec human IgG1 and human Igκ expression vectors. Constructs were assembled by Gibson cloning, transformed into DH5α cells, and selected on 2 × YT agar containing 100 µg/mL carbenicillin.

For expression, anti-EREG VH (huIgG1) and VL (huIgK) plasmids were co-transfected into ExpiCHO cells (Thermo Fisher Scientific) following the manufacturer’s High Titer protocol (7 days at 32 °C, 5% CO₂). Clarified supernatants were diluted in Protein A binding buffer (25 mM Tris-HCl, 25 mM NaCl, pH 7.4) and purified using a HiTrap Protein A HP column (Cytiva) on an ÄKTA Pure FPLC system. Antibody was eluted with 100 mM sodium citrate, 100 mM NaCl (pH 3.0) and immediately neutralized with 1 M Tris (pH 8.0). Eluted fractions were pooled, concentrated, and buffer-exchanged into PBS using a 30 kDa MWCO centrifugal filter (Amicon Ultra-4).

To confirm EREG antibody binding, high-binding plates were coated overnight at 4 °C with 2 µg/mL EREG-Fc (Sino Biological) in PBS. Wells were washed with PBS/0.05% Tween-20 (PBS-T) and blocked with PBS-T containing 5% non-fat milk for 1 h at room temperature. Antibodies were added in duplicate starting at 5 µg/mL with serial 10-fold dilutions and incubated for 1 h. After washing, goat anti-human κ-HRP (1:2000; Southern Biotech) was added for 1 h. Plates were developed with TMB substrate, stopped with 1 M HCl, and read at 450 nm. Binding curves were fit using four-parameter logistic regression.

In vivo EREG neutralizing antibody treatment

For neutralizing antibody treatment, mice were treated with either control mouse IgG2a (InvivoGen, mabg2a-ctrlm) or anti-EREG neutralizing antibody (10 mg kg⁻¹) diluted in PBS, which was administered intraperitoneally twice weekly for 3 weeks for a total of six treatments. To initiate studies, mice bearing orthotopic tumors were randomized based on tumor BLI, treated with cyclophosphamide (180 mg kg⁻¹), and adoptively transferred with 5e6 anti-mesothelin CAR T-cells the following day. Control IgG or EREG neutralizing antibody (EREG NAb) treatment was initiated during the CAR T-cell transfer and continued for 3 weeks.

The primary study endpoints were therapeutic response and relapse, while secondary endpoints included confirmation of on-target EREG neutralization activity. Response rates and Relapse-free survival (RFS) were assessed in the KxPxC GFP-Luc model for up to 50 days following adoptive transfer, corresponding to the time point at which 100% of IgG-treated control mice had relapsed. In the mT4-2D model, tumors were harvested following a complete 3-week treatment course to evaluate therapeutic response and confirm in vivo EREG neutralization activity. On-target inhibition was assessed by immunostaining tumor tissues collected at the end of treatment using a phospho-ErbB4/EGFR antibody (p-ErbB4/EGFR), followed by quantitative image analysis with HALO software to determine the total number of p-ErbB4/EGFR-positive cells per tumor. Tissue processing, staining, and image analysis were performed by iHisto (https://www.ihisto.io/).

Responses to CAR T-cell therapy were assessed by comparing pre-treatment and post-treatment tumor BLI measurements. Percent change in tumor BLI was calculated as:

$$\%\mathrm{\Delta }\mathrm{BLI}=\frac{\mathrm{Post}-\mathrm{ACT}-\mathrm{Pre}-\mathrm{ACT}}{\mathrm{Pre}-\mathrm{ACT}}\times 100$$

RECIST-like response criteria were defined as follows: complete response (CR), no detectable tumor or measurable BLI signal; partial response (PR), $\le -30\%$ change from baseline BLI; stable disease (SD), between $-30\%$ and$+20\%$ change from baseline; and progressive disease (PD), $\ge +20\%$ increase from baseline BLI.

RFS in KxPxC GFP-Luc mice was monitored through the 50-day endpoint. Relapse was defined as any increase in tumor BLI greater than $+20\%$ relative to pre-ACT baseline measurements. Tumor BLI measurements were acquired every 10 days following adoptive transfer.

Murine T-cell culture and transduction

The day before spleen harvesting, 24-well dishes were incubated overnight at 4 °C with 1 ml of PBS containing 5 μg ml⁻¹ of anti-CD3 antibody (Biolegend, 100239). The following day, spleens were harvested from 6 to 8 week old C57/BL6 mice, minced with a razor blade, and forced through a 70 μM cell strainer. Single-cell splenocyte suspensions were treated with red blood lysis solution (eBioscience), neutralized with PBS, and re-suspended in 1 ml of mojo sort buffer (Biolegend) at a concentration of 1 × 10⁸ ml⁻¹. To generate sufficient numbers of CAR T-cells, spleens from two or three mice were pooled together. CD3+ T-cells were isolated using mojo sort CD3 mouse T-cell isolation kit (Biolegend) according to the manufacturer’s instructions. T-cells (1 × 10⁶ cells ml⁻¹) were active by culturing overnight in T-cell media (RPMI-1640, 20% FBS, 50 μM B-ME, 1 mM sodium pyruvate (Corning), 1x non-essential amino acids (Cytivia), and 100 μg ml⁻¹ Primocin) with 100 IU ml⁻¹ recombinant murine IL-2 (R&D systems, 402 ML) and 2 μg ml⁻¹ anti-CD28 antibody (Biolegend, 102115) on the CD3 antibody-coated plates. The following day, 100× concentrated retrovirus was added to each well, and plates were covered with parafilm and spun for 90 min at 500 × g and 32 °C. Cells were then incubated for two additional hours at 37 °C and briefly spun down at 500 × g for 3 min at room temperature to remove virus. Fresh T-cell media containing 100 IU ml⁻¹ IL2 was added, and cells were cultured for an additional 48 h followed by the addition of 10 ng ml⁻¹ recombinant human IL-7 (Cell signaling technologies, 50425) and IL-15 (Cell signaling technologies, 73786). Cells were cultured for up to 10 days, then directly used or cryopreserved. Transduction efficiency was assessed by GFP, mTagBFP2, or staining with an anti-G4S linker antibody, and flow cytometry. T-cells were resuspended in PBS and co-cultured with tumor cells or injected into mice I.P. at a dose of 1–5 × 10⁶ cells as described in the Orthotopic pancreas tumors section.

Generation of CAR retrovirus

All plasmids used to generate retrovirus or lentivirus were cloned and expanded in NEB Stable competent E. coli cells and isolated with a ZymoPURE II Plasmid midiprep kit (Zymo Research). HEK 293FT cells were seeded at 4 × 10⁶ml⁻¹ in 150 mm dishes coated with collagen I (Gibco) to prevent detachment. The following day, cells were transfected with 4 μg of pCl-Eco (Addgene, 12371) and 4 μg of anti-CD19 or anti-mesothelin CAR MSCV using Fugene HD (Promega). Specifically, plasmids were mixed with 400 μl Optimem (Gibco) and 24 μl Fugene and incubated for 20 min. Cells were transfected, and the media was replaced 24 h later with 30 ml of DMEM containing 1% FBS and 25 mM HEPES (Corning). After 72 h, media was collected, filtered through 0.45 μM syringe filters, and concentrated using a 4× precipitation solution containing of 40% PEG-8000 (Thermo Fisher) and 1.2 M NaCl (Thermo Fisher) in PBS. Precipitated virus was re-suspended in PBS at a 100× concentration, and media from three 150 mm dishes was pooled and concentrated to transduce 10⁷ T-cells.

Stable cell line generation

HEK293FT cells were seeded on collagen I-coated 100 mm dishes at 4 × 10⁶ cells per dish for 24 h, then transfected with 2 μg of transfer plasmid along with the second-generation packaging plasmids: 1.5 μg psPax2 (Addgene, 12260) and 0.5 μg pMD2.G (Addgene, 12259) diluted in Optimem to a final volume of 40 μl. The Fugene HD reagent (12 μl) was mixed with 145 μl of Optimem for 5 min before adding plasmids and incubating for 15 min at room temperature. The DNA-Fugene mix was added to HEK293FT cells overnight, and the media replaced with DMEM containing 1% FBS, 25 mM HEPES, and 1% penicillin/streptomycin. Media was collected 48 and 72 h post-transfection, pooled, passed through a 0.45 μM filter, and concentrated using 100 K centrifugal units (Millipore Sigma, UFC9100) and centrifuging for 40 min at 1500 × g or a 4x PEG-8000-based concentration solution. For the PEG-8000 solution, 4-fold viral media was incubated 12–24 h and centrifuged at 1500 × g for 30 min. Pelleted virus was re-suspended 100X in PBS for 10 min at room temperature, followed by vigorous pipetting and centrifugation at maximum speed for 1 min to pellet carryover debris. Viral supernatant was used immediately or stored at −80 °C.

Cells were transduced for 24 h and selected by FACS or antibiotics 48–72 h post-transduction. Cells with multiple constructs were serially infected and selected by antibiotics or FACS with single constructs. For antibiotic selection, 5 μg ml⁻¹ of puromycin (Invivogen) and blasticidin (Invivogen) was used.

Tumor digestion for FACS and flow cytometry

Subcutaneous and orthotopic tumors were harvested and placed in a culture dish containing DMEM containing 10% FBS and 1% penicillin/streptomycin on ice and minced with a razor blade. In some cases, a portion of the tumor was placed in 4% paraformaldehyde (PFA) prior to mincing. Minced tumors were collected in 5 ml of digestion media (DMEM containing 1% Penicillin/streptomycin, 0.125 mg ml⁻¹ of collegnase IV (Stem Cell Technologies, 07909), 0.125 mg ml⁻¹ Dispase (Stem Cell Technologies, 07913), and 10 μg ml⁻¹ of DNAse I (Stem Cell Technologies, 07900)) and incubated for 1.5 h at 37 °C. Digested tumors were collected by centrifugation at 300 × g for 5 min and then re-suspended in 5 ml of TryplE (Gibco) and 10 μg ml⁻¹ DNase I for an additional 10 min at 37 °C. Undigested tumor fragments were allowed to settle for 1 min, and the supernatant was transferred to a new tube with excess PBS to neutralize TrypLE. Cells were filtered through 70 and 40 μM cell strainers to generate single-cell suspensions. Prior to FACS or flow cytometry cells were incubated with viability dyes diluted in PBS (zombie aqua (1:1000) (Biolegend, NC0476349), Ghost Red 710 (1:1000) (Cell Signaling Technology, 90232), or (1 μg ml⁻¹ of 4’, 6-diamidino-2-phenylindole, dihydrochloride (DAPI) (Cell signaling Technology, 4083)), washed, and then re-suspended in PBS with 1% FBS prior to FACS. For identification of mVenus^High quiescent tumor cells, gating thresholds were established based on serum-starved positive control populations, which reproducibly enriched for p27-mVenus reporter activity. mVenus^High cells were operationally defined as the top 20% of mVenus fluorescence in serum-starved conditions. p27-mVenus-negative cells were additionally included to establish background fluorescence and validate reporter specificity. FACS using BD Aria Fusion or SONY SH800 and flow cytometry using Cytek Aurora was performed at the Center for Biomedical Research Support Microscopy and Flow Cytometry Facility at UT Austin (RRID:SCR_021756)

Organoid formation

Organoids were generated by isolating and processing tumors as described in the Tumor digestion for FACS and flow cytometry section. Cells were re-suspended in PBS with 1% FBS and sorted based on negative or high mVenus expression in tdTomato+ cells. Cells were plated in 20 μl Matrigel domes at 1000 cells per dome in pre-heated 24-well dishes. For orthotopic tumors, 100 cells were plated per well. Epcam ⁺ tumor cells were isolated using the mojosort mouse Epcam isolation kit (Biolegned), incubated with CD45-APC and DAPI for 20 min on ice, washed twice, and sorted in single cell mode (Sony MA900 sorter) directly into chilled 96-well dishes with non-solidified Matrigel (50 μl per well-kept cold). Isolated tdTomato+ tumor cells were incubated at 37 °C for 10 min before adding media. Organoids containing ≥50 cells were counted 10 days after plating.

Limiting dilution analysis for tumor-initiating cell frequency

Tumors were isolated, digested, and subjected to FACS as described in the Tumor digestion for FACS and flow cytometry section. mVenus^negative and mVenus^high cells were mixed with PBS:Matrigel (1:1) and injected at 2–4 sites containing 1000, 500, or 100 cells in 100 μl. Tumor formation was assessed by palpation, and tumors were weighed 2 months after injection. For serial tumor formation, primary tumors were generated with 100 mVenus^High cells and harvested 2 months later. Tumor cells were FACS sorted into mVenus^Negative and mVenus^High populations, and 500, 100, or 10 cells were injected into secondary recipients. Two months later, tumors were quantified and imaged. Tumor-initiating cell frequencies were calculated using the Extreme limiting dilution software⁴⁴.

In vitro 3D spheroid co-culture

p27k-mVenus and tdTomato-luc2-hCD19 KPC cells were mixed with anti-hCD19 CAR T cells at effector-to-target (E:T) ratios of 0:1 (5000 KPC cells only) or 5:1 (5000 KPC cells:25,000 CAR T cells). Cell mixtures were pelleted at 300 g for 5 min and resuspended in 100% growth factor–reduced Matrigel. Suspensions were plated as 10 µl domes in pre-warmed 24-well plates and allowed to solidify for 10 min at 37 °C. Complete growth medium (DMEM with 10% FBS and 1× Primocin) containing 100 IU ml⁻¹ IL-2 was added to each well. Cells were incubated for 48 h before analysis. For spheroid flow cytometry analysis, cells were recovered using TrypLE, stained with Zombie Aqua viability dye for 15 min on ice, and analyzed immediately.

Ex Vivo KPC: CAR T-cell co-culture

p27k-mVenus and tdTomato-luc2-hCD19 cells (250,000) were implanted into NSG or C57/BL6 mice. After 10 days C57/BL6 mice received 5 million anti-hCD19 CAR T-cells, whereas NSG mice received no treatment. Two-weeks later, tumors from both NSG and C57/BL6 mice were isolated and digested into single-cell suspensions as described in the Tumor digestion for FACS and flow cytometry section. Viable, tdTomato+ cells were sorted from both NSG and C57/BL6 mice into 15 ml tubes containing DMEM with 10% FBS. mVenus^Negative cells were isolated from NSG mice, and mVenus^High cells from C57/BL6 mice. Cells were plated (500 cells/well) in 96-well plates. Three hours later, 5000 CD3/28 activated CAR T-cells were added in media supplemented with 100 IU ml⁻¹ of recombinant IL-2 and incubated for 48 h. Cell viability was then quantified by luciferase activity by adding 100 μl of 50 μg ml⁻¹ D-luciferin, incubating for 5 min, then recording luminescent signal (BioTek Synergy H1). All values were relative to tumor cells without CAR T-cells. Luc2 null cells and no-luciferin media conditions were used for background correction.

Ex Vivo Macrophage: CAR T cell conditioning

TAMs were isolated from three independent orthotopic tumors by mincing and digesting tissue for 45 min in 0.125 mg ml⁻¹ collagenase, 0.125 mg ml⁻¹ dispase, 10 μg ml ⁻¹ DNAse, and 5% FBS. Tissues were passed through a 40 μM cell strainer and incubated with anti-F4/80-biotin antibody, followed by streptavidin enrichment using the mojosort mouse F4/80 selection kit according to manufacturer’s instructions (Biolegend, 480170). F4/80⁺ cells (25,000) were cultured on fibronectin (Corning, 354008) coated (10 μg ml⁻¹) coverslip-bottomed 24-well dishes (CellVis, P24-1.5H-N) for 2 h to allow macrophage adherence before non-adherent cells were washed away with three PBS washes. For ErbB4 silencing in TAMs, cells were transduced with high-titer lentivirus (>100 MOI) packaged with either shControl or shERBB4 RNAi (see “DNA Constructs” section) for 24 h. To maximize knockdown efficiency, two independent shRNAs targeting ERBB4 were pooled during transduction. Seventy-two hours following transduction, TAMs were subjected to the EREG conditioning and co-culture protocol described below. Anti-mesothelin CAR T-cells were isolated by FACS 5 days after transduction using an Alexa-fluor 647 conjugated G4S linker specific antibody (Cell Signaling Technologies, 69782) and cultured with or without macrophages at a 1:3 CAR T:TAM ratio (8,333 CAR T-cells:25,000 TAMs per well) with or without 100 ng ml⁻¹ recombinant EREG (rEREG) for 72 h. All conditions were cultured in macrophage media (RPMI 1640 with 20% FBS, 2 mM L-glutamine, 2 mM sodium pyruvate, 55 μM 2-mercapthoethanol, 10 ng ml⁻¹ M-CSF, and Primocin (1X)). Media was supplemented with 10 ng ml⁻¹ recombinant human IL-7 (Cell Signaling Technology, 73786) and IL-15 (Cell Signaling Technology, 50425) prior to the addition of CAR T-cells. After 72 h, CAR T-cells were harvested and co-cultured with tdTomato-luciferease (Luc2) KPC target cells in a 96-well plate at a 1:1 (E:T) ratio for 48 h. Luciferase activity was quantified as described in the Ex Vivo KPC:CAR T-cell co-culture section before the addition of CAR T-cells. After 24 h, the CellEvent caspase-3/7 green detection reagent (Invitrogen, R37111) was added to each well according to the manufacturer’s instructions and quantified using a plate reader (BioTek Synergy H1). IFNγ was quantified 48 h after the addition of CAR T-cells by collecting 100 μl of media and assaying with an IFNγ Enzyme Linked Immuno-sorbent Assay (ELISA) (Biolegend, 430807).

Cells were then washed three times with PBS, and the remaining viable KPC cells were quantified based on luciferase signal. Readings (relative luciferase units, RLU) were normalized to their respective pre-CAR T-cell values. Relative percent CAR T-cell killing was calculated using the following formula:

$$\%\mathrm{Killing}=\left(1-\frac{\mathrm{Normalized}\mathrm{residual}\mathrm{tumor}{\mathrm{cell}}_{\mathrm{streatment}}}{\mathrm{Normalized}\mathrm{residual}\mathrm{tumor}{\mathrm{cells}}_{\mathrm{control}}}\right)\times 100$$

where the PBS condition was used as the control condition.

Flow cytometry

Tumors were digested as described in Tumor digestion for FACS and flow cytometry and cells were re-suspended in staining buffer containing 1% BSA. Cells were stained with antibodies for 15 min on ice, washed twice with PBS, and re-suspended in 400 μl of PBS for flow sorting. Single stain and unstained controls were used. For GFP⁺ T-cells, fluorescent minus one controls were used. All flow cytometry was performed on the Cytek Aurora and analyzed using FlowJo.

For intracellular antigens, cells were fixed in 4% PFA for 15 min, washed in PBS, and incubated on ice in 90% methanol for 10 min. Cells were then washed with PBS and incubated with antibodies for 1 h, followed by two more washes. For EREG staining, cells were also stained with anti-mouse Alexa Fluor 647 secondary antibody for an additional 30 min. For CAR T-cell staining, whole blood was incubated with 1xRBC lysis buffer (eBioscience), washed twice, then stained and analyzed as above.

Western blotting

Cells were lysed with RIPA buffer (Thermo Scientific) for 5 min and scraped into 1.5 mL collection tubes and incubated on ice for 30 min. Lysates were centrifuged for 15 min at 16,000 × g at 4 °C, and protein concentration was measured using the Pierce BCA assay (Thermo Scientific). Lysates were diluted to 30 μg, resolubilized in 5x LDS loading buffer (Thermo Scientific), and heated to 95 °C for 7 min. Samples were resolved on a 10% SDS-PAGE gel (Bio-Rad) and transferred to a nitrocellulose membrane using semi-wet transfer (Bio-Rad). Transferred membranes were blocked for 1 h in 5% non-fat dairy milk. Blocked membranes were washed with 1X Tris-buffered saline (TBS) + 0.1% tween-20 (TBST), then incubated with the diluted primary antibody overnight at 4 °C. The following day, membranes were washed three times with TBST followed by the addition of rabbit or mouse IgG secondary horse radish peroxidase (HRP) conjugated antibody diluted in 5% non-fat milk for 1 h. Membranes were washed three times with TBST and developed using the West Super Femto substrate (Thermo Scientific).

RT-qPCR

Viable tdTomato⁺ mVenus^Negative and mVenus^High cells were sorted directly into Trizol (Invitrogen), and RNA isolated using the Zymo Direct-zol RNA micro or mini kit (Zymo Research). For TAM shRNA knockdown confirmation, 150 μl of Trizol was added directly to the wells of a 24 well dish followed by RNA isolation using the Direct-Zol RNA mini kit. cDNA synthesis was carried out (100 ng of RNA) using the iscript cDNA synthesis kit (Bio-Rad). Undiluted cDNA (1 μl) was used for each PCR reaction using Power up SYBR green master mix (Thermo Scientific, A25741) according to the manufacturer’s instructions. Gene expression was normalized to GAPDH and analyzed relative to mVenus^Negative cells. Relative expression was calculated using the 2^−ΔΔCT method, where ΔΔCT = ΔCT_sample − mean ΔCT_mVenus^Neg.

Primers (5 ′−3′):

EREG:

Forward: CAGGCAGTTATCAGCACAACCG

Reverse: CATGCAAGCAGTAGCCGTCCAT

GAPDH:

Forward: CATCACTGCCACCCAGAAGACTG

Revers: ATGCCAGTGAGCTTCCCGTTCAG

NGF:

Forward: AGTTTTGGCCTGTGGTCGT

Reverse: GGACATTACGCTATGCACCTC

SPRR1A:

Forward: CAAGGCACCTGAGCCCTGCAA

Reverse: AGGCTCTGGTGCCTTAGGTTGG

ERBB4:

Forward: GGCAATATCTACATCACTG

Reverse: CCAACAACCATCATTTGAA

Ultra-low input RNA-sequencing

Tumors from NSG mice with or without hCD19 expression were harvested 2 weeks after CAR T-cell therapy. Tumors were digested as described in Tumor digestion for FACS and flow cytometry but using RNAase-free DNase (Thermo Scientific). Samples were incubated with zombie aqua viability stain for 20 min, washed, and re-suspended in PBS with 1% FBS before FACS isolating mVenus^Negative zombie aqua negative, tdTomato+ cells from hCD19^– tumors or viable mVenus^High tumor cells from hCD19⁺ tumors. Cells were directly sorted into 300 μl of Trizol and frozen at −80 °C until ultra-low input RNA-seq was carried out by GeneWiz. Count matrices were normalized using Deseq2 and Log2Fold changes were filtered to absolute ±5 and adjusted p-values < 0.01 between mVenus^Negative and mVenus^High cells to identify significantly altered genes. These thresholds were selected to prioritize robust and reproducible transcriptional changes while minimizing false positives inherent to high-dimensional transcriptomic data. Importantly, these criteria were applied uniformly across all comparisons. Gene names were copied from this data set and pasted into EnrichR gene ontology for pathway analysis⁴⁵.

Human single-cell data analysis

Human patient data from obtained from gene expression omnibus GSE205013 was downloaded as a count matrix and processed using the Seurat package. Samples P04, P05, P07, P15, P19, and P23 were selected as they represented treatment-naive primary PDAC. Count matrices were uploaded to R studio and used to create a Seurat object. Cells expressing greater than 10% mitochondrial genes were filtered out. Data normalization and scaling, variable feature selection, and cell clustering were performed using the Seurat package in R. The data were normalized to total expression per cell and scaled by a factor of 10,000. The default Seurat “vst” method was used to select the top 2000 variable genes. These were subsequently subjected to principal component analysis (PCA) for dimensionality reduction. The first 30 principal components were selected using the elbow method. 16 clusters were detected using the Louvain algorithm, which is the default Seurat method with a resolution set to 0.1. The resulting data was visualized using the Uniform Manifold Approximation and Projection (UMAP) algorithm. This data was then integrated, and a subset was generated based on Keratin 19 (KRT19) expression. This subset was then subjected to cell cycle analysis using the Seurat package, cell cycle score. The results were visualized using a dot plot to show the EREG, MKI67, and TOP2A features within the different cell cycle-based clusters.

Single-cell RNA sequencing (scRNA-seq) data from human PDAC tumors were obtained from publicly available datasets deposited in the Gene Expression Omnibus (GSE205013). Raw count matrices in Matrix Market format were imported into R (v4.x) using the Matrix package. Gene identifiers were converted to uppercase gene symbols, and cells were retained based on the availability of corresponding barcode annotations. Downstream analyses were performed on log-transformed expression values using a natural log transformation (log1p) applied directly to raw counts.

Epithelial tumor cells were identified based on expression of canonical epithelial markers, including EPCAM and KRT19. Cells expressing one or more of these markers (non-zero counts) were classified as tumor cells. Samples containing fewer than 50 tumor cells were excluded to ensure robust estimation of tumor cell–specific features. Within the tumor cell compartment, expression of epiregulin (EREG) was quantified, and a proliferation score was computed for each cell as the average log-transformed expression of established cell cycle genes (CDK1 and TOP2A). Tumor cells were then stratified within each sample to identify a quiescent EREG-enriched population, defined as cells within the upper quantile (top 10%) of EREG expression and the lower quantile of proliferation score (bottom 20%). The fraction of these quiescent EREG⁺ tumor cells was calculated per sample relative to the total tumor cell population. Cytotoxic T cells were identified across all cells using expression of canonical cytotoxic markers (CD8A, CD8B, GZMB). The fraction of cytotoxic T cells was computed for each sample relative to the total number of cells profiled.

For each tumor sample, summary metrics including the fraction of quiescent EREG⁺ tumor cells, cytotoxic T cells were computed. Associations between tumor cell state and immune composition were evaluated at the sample level. Group-based comparisons were performed using non-parametric Wilcoxon rank-sum tests following stratification of samples based on the tumor cell metric. All statistical analyses and visualizations were performed in R using standard packages including ggplot2.

Immunofluorescence of OCT-embedded and FFPE tissues

Tumors were harvested and placed in 4% Paraformaldehyde (PFA) for 30 min at room temperature. After 30 min, tissues were washed in PBS and a 30% sucrose solution overnight at 4 °C. Tissues were embedded in OCT (Tissue-Tek) and kept at −80 °C until sectioning. For sectioning, 10 μM thick sections (Leica Cryostat) were cut onto slides pre-treated with Poly-L-lysine (Advanced Biomatrix, 5048) according to the manufacturer’s instructions. Sections were dried for 1 h, then kept at −20 °C. Sections were brought to room temperature for 2 h, then washed three times with PBS to remove OCT. Slides were blocked for 1 h in blocking buffer (PBS with 5% FBS) and permeabilized for 20 min with 0.3% Triton-x-100 in PBS. For nuclear staining (Ki67 and mVenus reporter), slides were permeabilized with 1% Trixon-x-100 in PBS for 1 h. Blocked slides were incubated with primary antibodies for 15 h at 4 °C in a modified humidity chamber, washed three times with PBS, and incubated with secondary antibodies for 2 h at room temperature protected from light. After a second washing step, slides were mounted with DAPI-containing mounting media (Prolong Gold Antifade, Invitrogen). Slides were sealed with nail polish and allowed to dry overnight prior to imaging or subsequently stored at 4 °C. Imaging was obtained within 1 week of slide prep. A Nikon A1R confocal microscope was used to acquire images, and ImageJ was used for quantification. To quantify the intensity of different regions of interest, single channels were used to first select cancer cells (RFP signal) using the ROI manager tool. This selection was overlayed onto GFP channel images (mVenus reporter) to identify reporter high and low regions using a threshold function. Fluorescent intensity of the EREG channel (Cy5) was measured within the regions of interest to obtain a % intensity/ROI. For whole-pancreas tumor images, slides were imaged with the upright Nikon Ni-E upright light microscope for large-image stitching. For whole pancreas image quantification, a threshold was set for each channel, and cells meeting the threshold were counted using Nikon Elements software. Only mVenus^High cells (GFP/AF488) that surpassed the pan-cytokeratin (pan-cytokeratin/AF647) threshold were counted to prevent the inclusion of GFP + CAR T-cells. Imaging was performed at the Center for Biomedical Research Support Microscopy and Flow Cytometry Facility at UT Austin (RRID:SCR_021756).

Processing, staining, imaging, and analysis of FFPE whole-slide tumors was carried out by iHisto (https://www.ihisto.io/). Sections (4 μM FFPE) from three independent tumors were mounted on a single slide and subjected to LiecaBond automated staining using tyramide signal amplification (TSA). Staining included: anti-ErbB4 (1:100) + rabbit-HRP + Cy5 TSA (1:100); anti-F4/80 (1:500) (provided by iHisto, cat #70076, clone: D2S9R) + rabbit-HRP + FITC TSA, and anti-EREG (1:25) + horse anti-mouse HRP + Texas Red TSA. Whole-slide scanning was performed using a 3DHISTECH Pannoramic MIDI II and quantified with HALO analysis software. Total ErbB4⁺ F4/80⁺ cells were quantified from the entire slide (n = 3 tumors) based on EREG high and low regions of interest. iHisto also performed H&E staining and analysis by a board-certified DVM pathologist on three shEREG tumors isolated at day 150.

Immunofluorescence images of pancreatic cancer microarrays

(BioMax, cat#PA484a) were scanned as a whole slide and exported as single-channel TIFF images corresponding to pan-cytokeratin (FITC), EREG (AF555), and p27 (AF647). Channels were merged into a multi-channel OME-TIFF using Fiji (ImageJ) and imported into QuPath for analysis. In QuPath, tissue microarray cores were identified using the TMA dearrayer tool, and cell detection was performed using fluorescence-based segmentation based on DAPI signal. Cell boundaries were expanded to approximate whole-cell regions for cytoplasmic marker quantification.

Per-cell fluorescence intensities were extracted for each channel. Tumor cells were defined based on high pan-cytokeratin (Pan-CK) expression, using an upper quantile threshold (top 20%) of Pan-CK intensity to restrict analysis to epithelial tumor populations. Within pan-CK⁺ tumor cells, p27 expression was stratified into high and low groups using normalized z-scores whereby >3 was considered p27^High and <1 p27 ^Low. EREG expression was quantified at the single-cell level using mean cellular AF555 intensity in pan-CK ⁺, p27^High and ^Low cells.

To account for inter-core variability inherent to tissue microarrays, EREG intensity was normalized within each core by dividing by the mean EREG signal per core. Statistical comparisons between p27 ^High and ^Low tumor cells were performed using a two-sided Wilcoxon rank-sum test. Data visualization was performed using box plots to represent the distribution of EREG expression across groups.

For spatial association between ErbB4 (HER4)-positive macrophages and CD8 abundance using multiplex immunofluorescence imaging, TMA slides (BioMax, cat #PA082b) were sent off for staining and acquired as individual fluorescence channels corresponding to HER4 (Texas Red), CD68 (Cy7), CD8 (Cy5), and DAPI by iHisto (https://www.ihisto.io/). Square annotations corresponding to individual TMA cores were manually defined and exported as coordinate files in QuPath. Using Bio-Formats (bfconvert), each channel was cropped into individual core images based on identical x/y coordinates.

Image analysis was performed in ImageJ. For HER4/CD68 overlap analysis, individual HER4 and CD68 core images were imported using Bio-Formats and converted into maximum-intensity projections. Per-core contrast normalization was performed using the Fiji “Enhance Contrast” function with normalization enabled. Thresholds were empirically optimized across representative cores and applied uniformly to all samples (CD68 minimum 10 and maximum 46), CD8 (minimum 20 and maximum 60), and HER4 (minimum 4 and maximum 10). Binary masks were generated independently for HER4 and CD68 channels using fixed intensity thresholds. The percentage overlap area between HER4 and CD68 signal was quantified as the fraction of overlapping positive pixels relative to total core area.

CD8 abundance was quantified separately using particle-based segmentation. Following maximum-intensity projection and per-core contrast normalization, thresholded binary masks were generated and analyzed using the Fiji “Analyze Particles” function. CD8 abundance was quantified as percent area fraction of positive staining within each core. Cores with extreme CD8 outlier values attributable to segmentation artifacts were excluded prior to downstream analysis.

A total of 43 patient cores were analyzed. Statistical analyses were performed in R (v4.5.1). For group-wise comparisons, samples were stratified into HER4/CD68 ^Low and HER4/CD68^High groups based on median overlap values, and CD8 abundance was compared using two-sided Wilcoxon rank-sum testing. Graphs were generated in base R.

Human CAR T-cell culture and transduction

PBMCs were isolated from whole blood using a Ficoll-paque density gradient according to manufacturer’s instructions (Millipore Sigma, GE17-1440-02). CD3 cells were then isolated from PBMCs using CD3 microbeads (Miltenyi, 130097043) and resuspended at 1 × 10⁶ per ml in T-cell media (Immunocult, Stem Cell Technologies) containing 100 IU ml⁻¹ human recombinant IL-2 (Corning) and 2 μg ml⁻¹ Ultra-LEAF purified anti-CD28 (Biolegend, 350902). Cells were plated on pre-coated plates (5 μg ml⁻¹ anti-CD3 Ultra-LEAF purified anti-CD3, Biolegend, 300332) for 24 h, then transduced overnight with anti-mesothelin CAR T-cell lentivirus. Media was replaced after 24 h, and 2 days later, cells were isolated by FACS based on GFP expression.

Human organoid co-culture

Low-passage PDX tumors (F3-F6) were minced and digested in 10 ml of DMEM containing 2% FBS, 1% penicillin/streptomycin, 0.125 mg ml⁻¹ collagenase IV, 0.125 mg ml⁻¹ Dispase II, and 10 μg ml⁻¹ DNAse I and incubated for 45 min at 37 °C. Digested tumors were further treated with TrypLE supplemented with 10 μg ml⁻¹ of DNAse I for 10 min at 37 °C. Digested tumors were neutralized with DMEM and passed through 100 μM filters. To ensure that residual mouse stromal and immune cells were depleted, PDX-based organoids were passaged three times before experimentation. To obtain CD14⁺ monocytes, PBMCs (obtained through We are Blood) were incubated with anti-CD14-FITC antibody (Miltenyi, 130113153) followed by anti-FITC beads (Miltenyi, 130097050) and magnetic selection using LS columns (Miltenyi) according to the manufacturer’s instructions. CD14⁺ cells and PDX tumor cells were mixed at 3:1 CD14: PDX (60 K:20 K) cell ratio and resuspended in a 1:1 mixture of Matrigel (4 mg ml⁻¹) and collagen I (3 mg ml⁻¹) pre-neutralized with sodium bicarbonate (7.5%) until the phenol red indicator in the Matrigel turned salmon pink. Domes (10 μl) were pipetted into the center of pre-warmed 24-well plates and allowed to solidify for 10 min at 37 °C, followed by the addition of 500 μl of DMEM containing 15% FBS, L-glutamine, 0.01 mg ml⁻¹insulin, 0.01 mg ml⁻¹ hydrocortisone, and Primocin. For the first 2 days of culture, 10 μM of a ROCK inhibitor (Y-27632, Millipore Sigma) was added to the media. After 72 h of culture, 100,000 GFP⁺ anti-mesothelin CAR T-cells were placed on top of the domes with PBS or 2.5 μg ml⁻¹ of EREG-neutralizing antibody. The media was supplemented with IL-7 and IL-15 (10 ng/ml) during this time. Matrigel domes were digested 48 h later with TrypLE for 20 min. Cells were harvested, fixed, permeabilized, and stained for 1 h using antibodies as described in the Flow Cytometry section.

Spatial transcriptomics

Spatial transcriptomics was performed by the Genomic Sequencing and Analysis Facility at UT Austin, Center for Biomedical Research Support (RRID#: SCR_021713) using the Visium HD 3’ Spatial Transcriptomics platform (10x Genomics) and the manufacturer’s protocols (CG000803: Planner, CG000804: Tissue Prep (Rev A), CG000805: User guide (Rev A)). Tumors were collected and immediately placed in OCT molds on top of a dry ice/100% ethanol bath, then kept at −80 °C. Tumors were sectioned at −20 °C (thickness: 10 µm) and mounted on super frost gold slides (Fisherbrand). All surfaces were pretreated with RNAseZap (Thermo Fisher) prior to sectioning. Serial sections were also collected in 1.5 ml tubes to confirm RNA integrity through Qubit analysis (Thermo Fisher) prior to processing. Tissue sections were processed using the CytAssist Instrument (10x Genomics) to spatially barcode mRNA transcripts. Tissue sections were first fixed in chilled methanol and stained with Hematoxylin and Eosin (H&E) for histological visualization. Brightfield imaging was performed prior to permeabilization using a Nikon Ti2 microscope at 20× magnification using a plan apochromat λ 20x; NA − 0.75; WD − 1.0 mm objective. Images were later used for tissue alignment and annotation during downstream analysis. Following imaging, slides were loaded into the CytAssist instrument, which transferred spatial barcodes from the Visium HD slide to the tissue section, enabling capture of spatially resolved mRNA. After permeabilization, reverse transcription was carried out to synthesize cDNA with spatial barcodes. The resulting cDNA was amplified, and sequencing libraries were prepared according to the Visium HD 3’ Gene Expression User Guide. Sequencing was performed on an Illumina NextSeq 1000 platform using paired-end reads at 240 M reads. Raw data were processed using Space Ranger (V4.0) and aligned to the mouse reference genome (mm10).

Spatial transcriptomic analysis

Visium 3’ HD data were processed using Scanpy v1.10⁴⁶, Squidpy v1.3⁴⁷, and Python 3.10. Nuclear segmentation was conducted using Space Ranger v4.0.1 and the representative images were generated in Lupe Browser using the space ranger Lupe browser output file. T cells were identified by “CD8a”and “CD247”, EREG⁺ tumor cells by “EREG”,“MSLN”, and “KRT19”, and macrophages by “ITGAM” and “ADGRE1” in the 10x Genomics Lupe browser application (v9.0). To quantify spatial interactions, macrophage–tumor cell proximity was defined using a distance-based approach. Macrophages were classified as “proximal” if their segmented nuclei were located within ≤20 μm of an EREG⁺ tumor cell and were compared to the remaining tumor associated macrophages. Differential expression analysis was performed comparing proximal macrophages to total tumor macrophages within the same tissue section to control for local microenvironmental variation. Although these approaches enabled consistent identification of localized cellular interactions using nuclear-segmented samples, we acknowledge that manual region selection and fixed-distance thresholding may introduce bias and may not fully capture the complexity of spatial cellular organization within the TME. Importantly, spatial differential expression findings were further supported by independent bulk RNA-seq analyses of EREG-conditioned macrophages

For further analysis, 8 μM binned Space Ranger (v4.0.1) output files were used. Outputs were imported into AnnData⁴⁸, filtered to remove low-quality spots (<200 genes or <500 UMIs), and log-normalized to 10,000 transcripts per spot. Highly variable genes were selected using the Seurat-like method (3000 highly variable genes (HVGs)), followed by scaling and principal component analysis (PCA) (50 components). Neighborhood graphs were constructed using 15 nearest neighbors, and embeddings were generated with uniform manifold approximation and projection (UMAP). Clustering was performed using the Leiden algorithm (resolution 0.8)⁴⁹. Cell states were annotated using canonical marker genes from pancreas single-cell atlases⁵⁰ and by scoring lineage-specific signatures (epithelial, immune, endothelial, stromal) using Scanpy. Scores for macrophage, monocyte, T, B, NK, fibroblast, endothelial, acinar, ductal, and endocrine programs were computed using curated gene lists as follows: Macrophage: [“CD68”, “CD163”, “MRC1”, “MARCO”, “CSF1R”, “ITGAM”, “LYZ”], Monocyte: [“LY6C2”, “LYZ”, “S100A8”, “S100A9”, “VCAN”, “IL1B”], Fibroblast: [“COL1A1”, “COL1A2”, “COL3A1”, “FAP”, “PDGFRA”, “PDGFRB”, “ACTA2”, “TAGLN”, “DCN”, “LUM”, “MMP2”, “MMP11”], Endothelial: [“PECAM1”, “VWF”, “ESAM”, “KDR”, “ENG”], Pericytes: [“PDGFRB”, “RGS5”, “MCAM”, “CSPG4”], Ductal: [“KRT19”, “EPCAM”, “MUC1”, “KRT7”, “SOX9”], Tumor: [“KRT19”, “EPCAM”, “MUC1”, “KRT8”, “KRT18”, “TACSTD2”, “SOX9”], Alpha:[“GCG”, “TTR”, “IRX2”], Beta: [“INS”, “IAPP”, “PDX1”], T Cells: [“CD3D”, “CD3E”, “CD3G”, “CD2”, “TRAC”, “IL7R”, “CCR7”], NK Cells: [“NKG7”, “GNLY”, “GZMB”, “PRF1”, “KLRD1”], Dendritic Cells: [“ITGAX”, “HLA-DPA1”, “HLA-DPB1”, “HLA-DRA”, “HLA-DRB1”, “CLEC9A”, “IRF8”], Hypoxia Score: [“HIF1A”, “CA9”, “ENO1”, “ALDOA”, “LDHA”]. Spatial neighborhood and proximity analyses were performed on normalized log1p-transformed expression data using pixel coordinates. Tumor anchors were identified by epithelial markers (EPCAM, KRT19, MSLN) and EREG-high tumor anchors were defined as tumor/epithelial spots in the top 5% (95th percentile) of normalized EREG expression. Microenvironment cells were annotated by signature scoring (macrophage, fibroblast, lymphoid) and were considered localized to an EREG anchor when they lay within 50 pixels of any EREG+ anchor. For macrophage- and fibroblast-centric analyses, localized cells (within 50 pixels of EREG+ anchor) were subset and clustered using 2000 HVGs, PCA (30 PCs), neighbors (n = 10) and Leiden clustering (resolution = 0.8) to delineate macrophage-like and fibroblast-like subclusters. Neighborhood composition around EREG-high and matched EREG-low anchors was computed by KD-Tree query⁵¹ (radius = 50 pixels) and expressed as mean fractional composition per anchor; enrichment was calculated as the difference in mean fractions (High–Low) and standardized by z-score. Distance-to-anchor distributions were compared with Mann–Whitney U tests and proportions of localized cells with Fisher’s exact test; p-values were adjusted using Benjamini–Hochberg where applicable. Tumor cell EREG expression was quantified and Z-normalized. Spots were classified as EREG-high when above 1 standard deviation from the mean. Spatial neighbors were computed using Squidpy⁴⁷. Heatmaps, UMAPs, and spatial feature plots were generated using Matplotlib⁵², Scanpy, and Squidpy. Core packages included Scanpy 1.10, Squidpy 1.3, Anndata 0.10, NumPy 1.26, Pandas 2.1, and Matplotlib 3.8.

Generation of macrophage EREG-ErbB4 transcriptional signature

Orthotopic tumors expressing TetOn-shScramble or shEREG were harvested 1 week following CAR T-cell therapy, dissociated into single-cell suspensions, and subjected to FACS as described above to isolate CD11b⁺ F4/80⁺ tumor-associated macrophages (TAMs). Cells were directly sorted into 300 µL of TRIzol, lysed for 10 min at room temperature, and stored at −80 °C prior to ultra-low input RNA sequencing (GENEWIZ). For in vitro studies, TAMs were isolated and cultured as described in Ex Vivo Macrophage: CAR T cell conditioning, then treated with recombinant EREG (100 ng mL⁻¹) or PBS for 24 h prior to sorting (F4/80⁺) into TRIzol for RNA sequencing. To assess ErbB4-dependent transcriptional programs, bone marrow-derived macrophages (BMDMs) were generated by culturing bone marrow cells in 50 ng mL⁻¹ M-CSF for 7 days⁵³. BMDMs were plated on fibronectin-coated 24-well plates, pre-treated with DMSO or Neratinib (50 nM) for 6 h, and subsequently stimulated with recombinant EREG (100 ng mL⁻¹) for 24 h. Cells were then collected in Zymo DNA/RNA Shield (Zymo, R1100-50) and processed for RNA sequencing (Plasmidsaurus). Gene expression count matrices were analyzed using DESeq2, and transcripts exhibiting an absolute log2 fold-change ≥1.2 in EREG knockdown, EREG-treated, or tumor-derived macrophages were defined as EREG-modulated. Genes consistently downregulated upon Neratinib treatment vs vehicle in BMDMs were further selected to define the ErbB4-dependent macrophage signature used for downstream analyses, including survival studies.

TCGA survival analysis with EREG-ErbB4 macrophage gene signature

Bulk RNA-seq and clinical data for TCGA-PAAD were obtained using the TCGAbiolinks R package. Survival analyses were performed in R using the survival R package and survminer R package. Gene expression quantification files generated using the STAR-counts workflow were downloaded from the Genomic Data Commons and processed with GDCprepare. Transcript abundance values were extracted from the tpm_unstrand assay and log2 transformed [log2(TPM + 1)] prior to downstream analyses.

Genes included in the transcriptional signature were CD209, BCL3, IL31RA, SPIB, and BMP8A. Gene identifiers were converted to uppercase symbols, duplicate entries were removed, and only genes present in the expression matrix were retained for analysis. A composite expression score was calculated by averaging the z-score normalized expression values of the five genes across tumors. Specifically, expression values for IL31RA, CD209, BCL3, SPIB, and BMP8A were individually standardized using the scale() function in R and combined into a single mean signature score for each patient.

Clinical annotations were extracted from TCGA metadata and matched to expression profiles using patient barcodes truncated to the first 12 characters. Overall survival time was defined as days to death or, for censored patients, days to last follow-up. Vital status was binarized as deceased versus alive.

For Kaplan–Meier survival analyses, patients were stratified into “high” and “low” signature groups using predefined score thresholds (high ≥ 0.9; low ≤ −0.9). Patients with intermediate scores were excluded from dichotomized survival analyses. Survival curves were generated using the survival and survminer packages in R. Statistical significance between groups was assessed using the log-rank test. Hazard ratios and continuous score associations with survival were evaluated using univariate Cox proportional hazards regression models implemented with the coxph function. Survival plots were visualized with ggsurvplot using 95% confidence intervals and risk tables.

Detailed information regarding reagents and antibodies can be found in Supplementary Dataset 1.

Results

Quiescent PDAC cells are enriched after CAR T-cell therapy

In order to identify and study quiescent PDAC cells, we engineered mT4-2D cells derived from a tumor formed in the KPC (Pdx1-Cre; KRAS^G12D/+, p53^R172H/+) mouse model⁴¹ to express the live-cell quiescence reporter, p27k–mVenus. This reporter expresses the stabilized fusion fluorescent protein p27k-mVenus, which accumulates in quiescent cells and is degraded on cell-cycle entry³² (Fig. 1a), and mVenus^High cells were quiescent since they lacked the proliferation antigen Ki67 (Fig. 1b) and down-regulated cell-cycle transcriptional signatures (e.g., “G2–M checkpoint”, “E2F targets”) (Fig. 1c). Given that quiescent cancer cells can evade T cell recognition by down-regulating MHC class I^25–27, we also expressed human CD19 (hCD19) to serve as a defined target for CAR T-cells, as well as tdTomato and luciferase to track cells both ex vivo and in vivo (Supplementary Fig. 1a). We initially studied subcutaneous tumors in NSG mice, which lack functional B, T, and NK cells, to limit anti-tumor immunity to CAR T-cells. Two weeks after treatment, anti-hCD19 CAR T-cells (5e⁶ cells) significantly decreased hCD19⁺ tumors, whereas the growth of hCD19⁻ and PBS-treated hCD19⁺ tumors was not impacted (Supplementary Fig. 1b, c).

Fig. 1. Quiescent PDAC cells are enriched after CAR T-cell therapy.

a Experimental design schematic illustrating tumor implantation, treatment, and downstream analyses. b Immunofluorescence of p27k-mVenus (anti-GFP, Alexa Fluor 488, green) and Ki67 (Alexa Fluor 647, red) in tumor sections. Co-localization of mVenus^High and Ki67 was quantified using the Pearson correlation coefficient. c Principal component analysis (PCA) of RNA-seq samples from mVenus^Negative and mVenus^High tumor cells (n = 3 per group) with gene ontology analysis of transcripts up- or down-regulated in mVenus^High cells. d Percentage of mVenus^High cells within viable (Zombie Aqua⁻), tdTomato⁺ tumor cells (PBS, n = 3 biologically independent tumors; CAR T-cell treated, n = 4 biologically independent tumors). Data are presented as mean ± s.d. Statistical significance was assessed using two-sided two-way ANOVA with Tukey’s multiple-comparison test. e Quantification of tdTomato+ mVenus High cells in NSG tumors (n = 3 biologically independent tumors), BL6 1× tumors (n = 4 biologically independent tumors), and BL6 5× tumors (n = 3 biologically independent tumors). Data are presented as mean ± s.d. Statistical significance was determined using a two-sided Kruskal–Wallis test with Dunn’s multiple-comparison correction. f Immunofluorescence of orthotopic pancreatic tumors 1 week after PBS or CAR T-cell treatment. Dot plots show mVenus^High (GFP/Alexa Fluor 488/green) cells per 1000 pan-cytokeratin⁺ cells (Alexa Fluor 647/red) (mean ± s.d., n = 3 tumors per group; unpaired two-tailed t-test). RNA-seq data from (c) were also used in Fig. 5a and Extended Data Fig. 4a. Statistical significance was assessed as indicated for each panel. Data are presented as mean ± s.d. unless otherwise indicated. All figures and illustrations were created in https://BioRender.com. Source data is included for all figures.

We quantified quiescent cancer cells in residual tumors after CAR T-cell treatment, and the frequency of mVenus^High cells was significantly higher in hCD19⁺ tumors compared to tumors generated from hCD19⁻ cells or PBS-treated animals (42% ± 5.5% vs ~6.4% in hCD19^- with CAR T-cells, ~4% in hCD19^- with PBS, and ~3% hCD19⁺ with PBS; p = 0.009; Fig. 1d). We similarly found that mVenus^High cells were enriched in subcutaneous tumors formed in immunocompetent C57BL/6 (BL6) mice by 2.4- and 5.2-fold after treatment with 1 × 10⁶ or 5 × 10⁶ CAR T-cells, respectively (Fig. 1e). Orthotopic tumors better recapitulate the PDAC TME and are more resistant to CAR-T cells than subcutaneous tumors⁵⁴, and CAR-T cell treatment increased mVenus^High cells by 3.5-fold compared to PBS in this model (Fig. 1f). mVenus^High cells also emerged from mVenus^Negative cells in vitro and in NSG tumors; however, frequencies were significantly higher in tumors generated in immunocompetent mice treated with CAR T cells (~2.9-fold vs in vitro, p = 0.001; ~1.5-fold vs NSG, p = 0.025) (Supplementary Fig. 1d). Collectively, these findings indicate that the quiescent state is highly plastic, and the frequency of p27-expressing cells is increased following adoptive immunotherapy.

Quiescent cancer cells are not intrinsically resistant to CAR-T cells in vitro

To determine whether the quiescent phenotype is associated with resistance to adoptive immunotherapy, we compared in vitro CAR T-cell cytotoxicity of mVenus^High cells isolated from CAR T-cell treated BL6 mice and mVenus^Negative cells from untreated NSG mice. Tumor cells from NSG mice without CAR T-cell treatment were used as a comparator as they represent a relatively immune-sensitive cell population. Isolated cells were treated in vitro with CAR T-cells, and despite surviving therapy in vivo, mVenus^High cells displayed no differences in sensitivity to CAR T-cell mediated killing compared to mVenus^Negative cells from untreated tumors (Supplementary Fig. 2a, b). We also treated PDAC cells grown as spheroids in 3D cultures with or without CAR T-cells and found that treatment did not impact the frequency of mVenus^High cells (Supplementary Fig. 2c). Therefore, quiescent tumor cells are not intrinsically resistant to immune-mediated killing in vitro, suggesting that this cellular phenotype may instead confer a selective advantage in vivo through extrinsic mechanisms, such as modulation of the TME.

EREG is upregulated in quiescent cancer cells

The relative lack of intrinsic resistance to CAR-T cells in vitro suggests the enrichment of mVenus^High cells in vivo may be mediated by secreted factors that alter the surrounding TME. RNA-seq identified several secreted factors highly expressed by mVenus^High cells, which are involved in inflammation or immunity, including epiregulin (EREG), nerve growth factor (NGF), and small proline-rich protein 1a (SPRR1a) (Supplementary Fig. 3a). We knocked down each of these factors in PDAC cells, formed tumors in BL6 mice, and treated with CAR T-cells (Supplementary Fig. 3b, c). The loss of EREG, but not NGF or SPRR1a, drastically reduced the burden of quiescent cancer cells as quantified by p27k-miRFP670 expression compared to shRNA controls (Supplementary Fig. 3d–f). Notably, EREG knockdown did not impact quiescent cancer cells in the absence of CAR-T cell therapy (Supplementary Fig. 3e, f). We also confirmed that EREG expression was significantly higher in mVenus^High cells than mVenus ^Negative cells isolated from tumors post-CAR-T cell therapy at the mRNA and protein levels (Fig. 2a, b). To assess the clinical relevance of these findings, we examined a human PDAC single-cell RNA-seq dataset (GSE205013, n = 6 treatment naïve patients) and found that EREG expression was increased in epithelial cells (Keratin 19, log2FC > 2.5) classified as G0/G1 (quiescent) compared to cycling cells expressing MKI67 (Ki67) and TOP2A (Fig. 2c). EREG expression was validated at the protein level in human PDAC tumors (n = 23) using a tissue microarray where p27^High tumor cells (Pan-Cytokeratin⁺) exhibited significantly (p < 2.2 × 10⁻¹⁵) higher EREG protein expression compared to p27 ^Low cells, indicating that EREG is selectively upregulated in quiescent cancer cells in vivo in both treatment-naïve and CAR T-cell–treated settings (Fig. 2d. Supplementary Dataset 2).

Fig. 2. EREG is upregulated in quiescent cancer cells.

a RT–qPCR analysis of EREG mRNA in mVenus^Negative and mVenus^High tdTomato+ tumor cells isolated by FACS from CAR T-cell-treated tumors (n = 3 biologically independent tumors per condition, two-tailed unpaired t-test). Transcript levels were normalized to GAPDH and expressed relative to the mean of mVenus^Negative cells. Data are presented as mean ± s.d. Statistical significance was determined using a two-sided unpaired t-test. b EREG intensity (red) measured in tdTomato⁺ (white) mVenus^Negative and mVenus^High regions (green) from three biologically independent tumors (n = 3). Three to four regions were analyzed per tumor and averaged to generate a single biological replicate. Data are presented as mean ± s.d of each biological replicate. Statistical significance was determined using a two-sided unpaired t-test. c Analysis of EREG expression across distinct cycling states in epithelial (KRT19⁺, log₂FC > 2.5) cells from primary PDAC patient samples (n = 6). Cell-cycle stratification was validated using MKI67 and TOP2A transcripts. d Immunofluorescence staining of human PDAC tissue microarray (TMA; n = 23 individual patients) for pan-cytokeratin (FITC, green), EREG (AF555, red), and p27 (AF647, white). Box plot shows EREG expression in pan-cytokeratin⁺ tumor cells stratified by p27 status using normalized z-scores of AF 647 cellular fluorescence (High > 3 p27 z-score and Low < 1 p27 z-score) with values normalized per tumor core. Box plots show center line, median; box limits, 25th and 75th percentiles; whiskers, minimum and maximum values. Statistical significance was determined using a Wilcoxon rank-sum test (p = 2.2e⁻¹⁶). Data are presented as mean ± s.d. unless otherwise indicated. All figures except for c were created in https://BioRender.com. Source data is included for all figures.

Tumor-derived EREG impacts CAR-T cell activity in vivo

EREG has been identified as part of a five-gene prognostic signature (MET, ERAP2, IL20RB, EREG, and SHC2) in human PDAC, associated with reduced CD8⁺ T-cell infiltration and increased T-cell dysfunction³⁹. Furthermore, we found that relatively high EREG expression is associated with significantly worse overall survival in human PDAC, with a 5-year survival rate of 0% in EREG-high patients compared with 33% in EREG-low patients (p < 0.0001; Fig. 3a), implicating EREG as a potential regulator of anti-tumor immunity and a determinant of clinical outcome in PDAC.

Fig. 3. Tumor-derived EREG impacts CAR-T cell activity in vivo.

a, Kaplan–Meier survival of PDAC patients stratified by EREG mRNA using a pTPM cutoff of 4.43. Patients were grouped into high (EREG > 4.43, n = 40) and low (EREG < 4.43, n = 136) cohorts. Significance was determined by two-sided log-rank test. b, Kaplan–Meier curves of mice bearing orthotopic TetOn-shScramble tumors treated with PBS (n = 3), TetOn-shScramble tumors with CAR T-cells (n = 6), or TetOn-shEREG tumors with CAR T-cells (n = 6). Significance was assessed using two-sided log-rank test. c, Representative H&E staining of TetOn-shEREG tumors at day 150 post-implantation (n = 3). Histopathology to confirm absence of disease was performed by a board-certified DVM in anatomic pathology. d, Quantification of CD3⁺ T cells in tumors 1 week post adoptive transfer. CD3 (Alexa Fluor 700, purple) was measured using FIJI 3D Cell Counter. Data represent mean per tumor (n = 3; unpaired two-tailed t-test). e, Tumor-infiltrating CAR T cells (CD8⁺ GFP⁺) analyzed for IFNγ and TNFα 1 week after transfer. Data are mean ± s.d. (n = 3 tumors, unpaired two-tailed t-test). f Immunofluorescence of tumors 1 week post CAR T therapy stained for CD3 (red), granzyme B (GZMB, green), and DAPI (blue). Three tumors per condition were analyzed with ≥3 regions per tumor. Dot plots show mean CD3⁺GZMB⁺ cells per tumor (n = 3; unpaired two-tailed t-test). Data are presented as mean ± s.d. unless otherwise indicated. The same stained tissues from (d) were used for analysis in Fig. 4d by including mVenus (GFP) stained cells. c–f were created in https://BioRender.com. Source data is included for all figures.

To determine whether EREG modulates CAR T-cell activity in vivo, we knocked down its expression in cancer cells using doxycycline-inducible EREG shRNA. While our initial studies used hCD19 as a model target antigen, we studied anti-mesothelin CAR T-cells, which are currently under clinical testing for advanced PDAC⁴. Notably, EREG knockdown did not impact orthotopic tumor growth in NSG mice or sensitivity to anti–mouse mesothelin CAR-T cells in vitro (Supplementary Fig. 4a, c, e, f), indicating that EREG does not directly regulate tumor cell growth or intrinsic CAR T-cell resistance.

Mice with inducible EREG (shEREG) or a scramble control (shScramble) shRNA were administered doxycycline starting 3 days prior to anti-mesothelin CAR T cells (5e6) and continued for a total of 14 days. Compared to both untreated controls (shScramble + PBS) and CAR T-cell–treated controls (shScramble + CAR T), EREG knockdown in cancer cells significantly increased median survival from 36–38 days to 136 days (p < 0.0009; Fig. 3b). Strikingly, EREG knockdown resulted in complete tumor eradication in 50% of treated mice (3/6), with histological analysis confirming no residual microscopic disease in pancreatic tissues at day 150 (Fig. 3c). Prolongation of overall survival was also observed in an independent KPC-derived orthotopic model (KPC-/-) (Supplementary Fig. 5a). To determine whether these effects extended beyond genetic silencing, mice were treated with an EREG-neutralizing antibody (EREG NAb). Compared to IgG control, EREG NAb treatment markedly suppressed p-EREG/ErbB4 signaling in vivo, as confirmed by tumor immunostaining, demonstrating on-target activity (Supplementary Fig. 5b). EREG NAb treatment also enhanced CAR T-cell efficacy against the mT4-2D and a third KPC-derived orthotopic model (KxPxC GFP-luc), with increased objective response rates (ORR) and the prevention of disease relapse (Supplementary Fig. 5c–e).

Consistent with its role as a predictor of T-cell dysfunction and infiltration in human PDAC, EREG knockdown significantly increased overall CD3⁺ T-cell infiltration, including the frequency of IFNγ⁺ TNFα⁺ CAR T-cells (GFP⁺ CD8⁺) and Granzyme B (GZMB)⁺ CD3⁺ T-cells compared with control tumors (Fig. 3d–f). Together, these results demonstrate that EREG expression in PDAC cells shapes both the distribution and functional activity of T-cells.

EREG loss enhances quiescent cell targeting by CAR T-cells

Given the elevation of EREG expression in quiescent cancer cells following CAR T-cell therapy, we next tested whether EREG knockdown enhances the sensitivity of these otherwise (in vivo) immune-resistant cells. Compared with control (shControl) tumors, EREG knockdown (shEREG) tumors exhibited significantly reduced frequencies of mVenus^High cells after CAR T-cell treatment, an effect not observed with PBS treatment (Fig. 4a, b). Furthermore, a higher proportion of mVenus^High cells from shEREG tumors were apoptotic and expressed cleaved caspase-3 (CC3⁺) (Fig. 4c) post-therapy, indicating that their reduction is due to increased killing rather than impaired entry into quiescence. Consistent with enhanced CAR T-cell mediated killing, shEREG tumors contained higher numbers of CD3⁺ T-cells in direct contact with mVenus^High cells compared to shControl tumors (Fig. 4d). EREG knockdown similarly reduced the frequency and increased killing of mVenus^High cells in a second independent KPC derived model (KPC-/-) (Supplementary Fig. 6a–c).

Fig. 4. EREG loss enhances quiescent cell targeting by CAR T-cells.

a Western blot showing EREG knockdown (shEREG) versus control (shCon) tumor cells repeated three times with similar results. Uncropped blots are in the Supplementary Information. b Flow cytometry of orthotopic pancreatic tumors 1 week after PBS or 5 × 10⁶ anti-mesothelin CAR T-cells. Contour plots show mVenus^High cells within viable (DAPI⁻), lineage-negative, tdTomato⁺ cells. Graphs quantify mVenus^High cells per 10⁶ viable cells (^# denotes cell number vs percentage), comparing shControl and shEREG tumors (n = 3 per group, two-way ANOVA with Bonferroni’s multiple comparisons test). c Immunofluorescence of inducible shScramble and shEREG tumors 1 week after CAR T-cell therapy using mVenus (GFP/Alexa Fluor 488, green) and cleaved caspase-3 (CC3; Alexa Fluor 647, red), with DAPI (blue). Graph shows the mean mVenus⁺ and CC3⁺ cells from 2–3 fields per section across three tumors (n = 3; unpaired two-tailed t-test). d Immunofluorescence of tumors stained for CD3 (Alexa Fluor 700, purple) and mVenus (GFP/Alexa Fluor 488, green). Dot plots show mean CD3–mVenus⁺ interactions per field (3 fields per tumor; n = 3; unpaired two-tailed t-test). e Schematic of PDX-derived organoid co-culture. CD14⁺ monocytes were co-cultured with PDX-derived tumor cells and CAR T cells ± EREG neutralizing antibody (EREG NAb). f Flow cytometry of quiescent cancer cells 48 h post CAR T-cell addition ± EREG NAb. Contour plots show apoptotic (CC3⁺), p27^High cells. Bar graph shows % CC3⁺ p27^High cells from four independent organoid models (n = 4; mean ± s.d.; two-tailed Mann–Whitney U test). Data are presented as mean ± s.d. unless otherwise indicated. All figures, except contour plots, were created in https://BioRender.com. Source data is included for all figures.

To validate these findings in human PDAC, we established a 3D co-culture model using mesothelin-expressing PDX-derived tumor cells (Supplementary Dataset 2). The PDAC TME is characterized by dense fibrosis and extensive immune infiltration, with TAMs representing the dominant population⁷. As such, we co-cultured PDAC tumor cells with CD14⁺-derived macrophages and embedded them in collagen I to recapitulate the native TME (Fig. 4e, Fig. 7a). Treatment with anti-mesothelin CAR T cells and an EREG-neutralizing antibody (EREG NAb; Supplementary Fig. 7b) significantly increased quiescent cancer cell apoptosis (EpCAM⁺ CC3⁺ p27^High) compared to CAR T-cells alone (Fig. 4f). Consistent with our EREG knockdown murine models, EREG neutralization had no effect when tumor cells were cultured directly with CAR T-cells, indicating that EREG does not mediate intrinsic resistance to CAR T-cell cytotoxicity (Supplementary Fig. 7c). Collectively, these data demonstrate that EREG indirectly protects quiescent cancer cells from CAR T-cell–mediated killing.

Fig. 7. EREG–ErbB4 tumor-associated macrophages correlate with T-cell exclusion and patient survival in human PDAC.

a Human PDAC scRNA-seq datasets were analyzed by stratifying patients based on abundance of slow-cycling EREG-high tumor cells (CDK1^low, TOP2A^low, EPCAM⁺, Keratin 19⁺, EREG^High) followed by quantification of cytotoxic GZMB⁺ CD8 T-cell abundance. Each dot represents an individual patient. Statistical significance was determined using an unpaired two-tailed t-test. b Human PDAC tumor microarrays (TMAs) were stained for CD8 (white), ErbB4 (HER4) (red), and CD68 (green) followed by multiplex immunofluorescence analysis. TMA cores were stratified based on ErbB4/CD68 overlap staining intensity and total CD8 abundance was quantified per core. Box plots represent CD8⁺ area fraction in HER4/CD68 ^Low (n = 22 patient cores) versus HER4/CD68^High (n = 21 patient cores). Statistical significance was determined using a two-sided Wilcoxon rank-sum test. c Three independent approaches were used to identify EREG-regulated macrophage transcriptional programs: (1) ex vivo rEREG stimulation, (2) macrophages isolated from EREG-knockdown tumors (shEREG), and (3) macrophages localized adjacent to EREG-expressing tumor cells in vivo (localized TAMs). Shared transcripts across all conditions were used to generate an EREG-modulated macrophage signature. d Bone marrow-derived macrophages (BMDM)s were treated with recombinant EREG in the presence or absence of pan-ErbB inhibition using neratinib. Heatmap depicts EREG-induced macrophage genes suppressed following ErbB4 inhibition. e Shared EREG- and HER4-regulated macrophage genes were converted to human orthologs and applied to TCGA PDAC datasets for Kaplan–Meier survival analysis using high- and low-signature groups with log-rank testing. Data are presented as mean ± s.d. unless otherwise indicated. BioRender (https://BioRender.com) was used to generate schematics in (a, c, d). Venn diagrams were generated using Venny 2.1. Source data is included for all figures.

Quiescent cancer cells exhibit stem-like, tumor-propagating properties

Quiescent cancer cells have been implicated in tumor regrowth and disease relapse following treatment^18,55, and we found that mVenus^High cells expressed higher levels of transcripts associated with CSCs compared to mVenus^Negative cells, including PROM1 (CD133), ALDH1A3, F3, and AQP5 (Fig. 5a)^18,56. Functionally, mVenus^High cells exhibited increased organoid formation in vitro and tumor-initiating cell frequency in vivo, produced larger tumors, and enhanced self-renewal during serial transplantation (Fig. 5b–e). We also found that cells isolated from shEREG tumors treated with CAR T-cells formed significantly fewer organoids compared to shControl tumors (Supplementary Fig. 8a, b). In contrast, EREG knockdown in vitro did not affect organoid-forming capacity, even with continuous doxycycline exposure for 10 days prior to analysis (Supplementary Fig. 8c). Together, these results suggest that quiescent tumor cells possess functional stem-like properties and EREG does not directly affect these properties but rather promotes their resistance to CAR-T cells in vivo.

Fig. 5. Quiescent cancer cells exhibit stem-like, tumor-propagating properties.

a Heatmap of differentially expressed genes related to cell cycle regulation and cancer stem cell (CSC) signatures in p27k-mVenus^Negative (Neg) and ^High (High) cells from three independent experiments (n = 3). b Organoid formation efficiency shown as number of organoids per 1000 cells plated (n = 4 independent experiments). c Tumor-initiating capacity presented as the number of tumors formed per total injections following limiting dilution. CSC frequency was calculated using a chi-square goodness-of-fit test. d Tumor weights from mice injected with p27k-mVenus Negative or High cells at 100 (n = 7 Negative, n = 8 High, tumors), 500 (n = 5 tumors per condition), or 1000 (n = 6 tumors per condition) cells per injection. Data are presented as mean tumor weight ± SEM. e Summary of tumors formed per injection in secondary limiting dilution assays. All statistical analyses, except (c), were performed using unpaired two-tailed t-tests. Data are presented as mean ± s.d. unless otherwise indicated. All figures and illustrations were created in https://BioRender.com. Source data is included for all figures.

We also studied the role of EREG in tumor recurrence following treatment by treating mice with 2.5e⁶ CAR T-cells. We utilized this lower CAR T-cell dose since we found that 5 × 10⁶ cells completely eradicated tumors in 50% of mice (Fig. 3b, c). While 67% of the shControl tumor-bearing mice relapsed by 40 days after treatment, 100% of the mice with shEREG tumors remained relapse-free at day 60 (Supplementary Fig. 8d, e). Together, these findings indicate that EREG knockdown delays tumor recurrence, potentially through CAR T-cell targeting of the CSC pool.

ErbB4 expressing tumor associated macrophages are spatially localized to EREG

To better understand the mechanism by which EREG impacts the TME, we performed spatial transcriptomic profiling of orthotopic tumors 3 days after CAR-T cell therapy (Fig. 6a). Tumor regions with high EREG expression (EREG^High), defined as the top 5% of normalized EREG expression in tumor cells, were enriched for macrophages, monocytes, and fibroblasts (Fig. 6a, b). Re-clustering of cells within these EREG^High regions revealed preferential enrichment of macrophage gene signatures relative to fibroblasts (Supplementary Fig. 9a). We also found that macrophages proximal to EREG⁺ tumor cells (localized tumor-associated macrophages, TAMs) expressed the EREG receptor ErbB4 at 4.5-fold higher levels than bulk macrophages, whereas the expression of EGFR and other ErbB family members was not significantly altered (Fig. 6c). We confirmed these finding by immunostaining and found increased ErbB4⁺ F4/80⁺ macrophages within EREG^High tumor regions, in contrast to EREG ^Low regions, which consisted of F4/80⁺ macrophages lacking ErbB4 expression (Fig. 6d). Localized induction of ErbB4 by macrophages can be induced by EGF-family ligands, such as Neuregulin-4⁵⁷, suggesting that EREG itself may drive ErbB4. We quantified ErbB4⁺ macrophages (F4/80) by flow cytometry and found that tumors with EREG knockdown had a significantly lower frequency of ErbB4⁺ F4/80⁺ macrophages than control tumors in both the mT4-2D and KPC-/- models (Fig. 6e and Supplementary Fig. 9b). Taken together, these results indicate that ErbB4⁺ macrophages are selectively enriched within EREG-rich tumor regions.

Fig. 6. ErbB4 expressing tumor associated macrophages are spatially localized to EREG.

a Schematic of tumor microenvironment evaluation using spatial transcriptomics. Tumors were collected 3 days post CAR T-cell therapy and analyzed with Visium HD 3’ spatial transcriptomics. UMAP plot shows clusters annotated by predicted cell type. b Comparison of predicted cell type enrichment in EREG-high versus EREG-low regions within a murine pancreatic tumor. c Segmented tumor map showing localization of EREG⁺ cancer cells (black), tumor-associated macrophages (TAMs; red), and CD8⁺ T cells (blue). Differential expression of potential EREG receptors (EGFR, ERBB2–4) was analyzed in macrophages proximal to EREG-expressing cancer cells (localized TAMs) versus bulk TAMs. d Whole-slide immunostaining of orthotopic tumors (n = 3, independent tumors). EREG regions were classified as high or low (purple pseudocolor). Double-stained ErbB4⁺ (red) F4/80⁺ (green) cells were quantified in each region and normalized to % of total cells using HALO analysis software (iHisto.io). e Flow cytometric quantification of ErbB4⁺ macrophages (F4/80⁺) in orthotopic tumors with or without EREG knockdown 7 days post CAR T-cell transfer (n = 3, independent tumors; mean ± SD; two-tailed unpaired t-test). f Flow cytometry quantification of macrophage depletion post clodronate treatment, expressed as CD11b⁺F4/80⁺ cells per million sorted cells (n = 2 for vehicle, n = 3 clodronate, independent tumors; mean ± SD). g Flow cytometry of mVenus^High cancer cells from vehicle or clodronate-treated tumors 1 week post CAR T-cell therapy (n = 6 vehicle and n = 7 clodronate independent tumors). h Schematic of the experimental design. i Validation of ErbB4 silencing in cultured TAMs by RT–qPCR (n = 2, biological replicates, p = 1.0e⁻⁵)). j Quantification of target cell killing 72 h following TAM pre-conditioning with or without rEREG. Viable tumor cells were measured by luciferase activity 48 h after co-culture with conditioned CAR T-cells, normalized to baseline seeding, and converted to relative % killing values using the shControl TAM + PBS condition as reference. Statistical analyses were performed using two-way ANOVA with Tukey’s multiple comparisons test, except for (i), which used an two-tailed unpaired t-test and g which used Mann-Whitney U test. Data are presented as mean ± s.d. unless otherwise indicated. All figures and illustrations, with exception to contour plots and spatial transcriptomic map, were created in https://BioRender.com. Source data is included for all figures.

EREG requires ErbB4 tumor-associated macrophages to elicit immune suppression

To determine whether macrophages are required for EREG-mediated immune suppression, we depleted macrophages in vivo using clodronate liposomes (Fig. 6f). Consistent with our previous observations, shEREG tumors exhibited significant reductions in p27k-mVenus^High cells in control mice treated with PBS-loaded liposomes (Fig. 6g). In contrast, macrophage depletion with clodronate liposomes abolished the difference in mVenus^High cell frequency between shControl and shEREG tumors, indicating that macrophages are required for EREG-mediated immune protection of quiescent tumor cells (Fig. 6g).

To further define this mechanism, anti-mesothelin CAR T cells were cultured in vitro with or without TAMs and/or recombinant EREG (rEREG) (Supplementary Fig. 9c). Exposure of CAR T-cells to rEREG did not alter intrinsic cytotoxic function or phenotype, as assessed by target cell killing, caspase-3/7 activation, and IFNγ secretion (Supplementary Fig. 9d–f). By contrast, rEREG significantly potentiated TAM-mediated suppression of CAR T-cell cytotoxicity and IFNγ production (Supplementary Fig. 9d–f). To determine whether EREG-induced immune suppression requires intact ErbB4 in macrophages, ErbB4 was silenced in TAMs using RNAi and 72 h later the rEREG-TAM conditioning assay was repeated (Fig. 6h–i). This analysis revealed ErbB4 knockdown to abrogate EREG induced TAM inhibition of CAR T-cell functional killing, demonstrating that EREG-mediated immune suppression through TAMs is an ErbB4-dependent process (Fig. 6j). Collectively, these findings indicate that EREG impairs CAR T-cell function indirectly through ErbB4 in TAMs rather than through direct effects on CAR T cells.

EREG–ErbB4 tumor-associated macrophages correlate with T-cell exclusion and patient survival in human pancreatic tumors

We next sought to determine whether an immunosuppressive EREG–ErbB4 TAM axis exists in human PDAC. To this end, we queried a human PDAC scRNA-seq dataset (GSE205013) and stratified patients based on the relative abundance of slow-cycling EREG ⁺ tumor cells. Patients with high EREG tumor abundance exhibited reduced cytotoxic T-cell infiltration, as defined by frequency of CD8⁺ GZMB ⁺ cells (Fig. 7a). Given that ErbB4-expressing macrophages mediate EREG-dependent immune suppression in murine models, we next evaluated whether ErbB4 (HER4) TAM abundance similarly correlated with anti-tumor immunity in human PDAC. Immunostaining of 43 human PDAC TMA cores revealed significantly reduced CD8 T-cell infiltration in tumors enriched for HER4⁺CD68⁺ macrophage regions identified by overlapping staining intensity of both HER4 and CD68 (Fig. 7b, Supplementary Dataset 2).

To determine whether this macrophage state was associated with clinical outcomes, we generated an EREG–ErbB4 macrophage transcriptional signature by integrating RNA-sequencing datasets from rEREG-stimulated macrophages, macrophages isolated from EREG-knockdown tumors, and macrophages spatially localized adjacent to EREG⁺ tumor regions. This analysis identified a conserved set of EREG-regulated macrophage transcripts that were induced by EREG exposure in vitro and enriched in localized macrophages in vivo, but diminished in EREG-depleted tumors (Fig. 7c). Unexpectedly, pathway analysis revealed strong enrichment of neuronal-like transcriptional programs alongside select immunomodulatory transcripts (Supplementary Fig. 10a). To evaluate ErbB4 dependence, bone marrow-derived macrophages were stimulated with rEREG in the presence or absence of the pan-ErbB inhibitor neratinib, which potently inhibits ErbB4 at nanomolar concentrations⁵⁸. Compared with vehicle-treated macrophages, neratinib attenuated expression of multiple EREG-induced immunomodulatory transcripts. Notably, no consistent changes were observed in canonical M1 or M2 polarization markers, suggesting that EREG induces a macrophage transcriptional program distinct from traditional polarization states (Fig. 7d and Supplementary Fig. 10b).

Application of the conserved EREG–ErbB4 (HER4) macrophage signature to TCGA PDAC datasets demonstrated that low signature expression was associated with significantly improved patient survival (Fig. 7e). Collectively, these findings support the existence of an EREG–ErbB4 (HER4) macrophage axis linked to immune exclusion and poor clinical outcome in human PDAC.

Discussion

The PDAC TME remains a major barrier to effective immunotherapy. Here, we identify quiescent tumor cells as a dynamic tumor cell state that is enriched following CAR T-cell therapy and associated with macrophage-dependent immune suppression. Consistent with prior studies demonstrating that quiescent, stem-like tumor cells can shape the TME toward immune suppression^33,59,60, our findings support a model in which therapy-associated tumor cell plasticity contributes to impaired anti-tumor immunity in the context of CAR T-cell therapy. Mechanistically, we identify epiregulin (EREG) as a mediator linking quiescent tumor cells to suppression of CAR T-cell activity through ErbB4-expressing TAMs. Importantly, targeting EREG enhanced CAR T-cell activity, reduced persistence of quiescent tumor cells, and improved therapeutic responses across multiple orthotopic PDAC models.

We further show that quiescent tumor cells are associated with the production of epiregulin (EREG), which in turn correlates with the presence and activity of ErbB4⁺ TAMs. Under physiological conditions, EREG is low or absent in most adult tissues but can be induced during inflammation and tissue regeneration^{34–36,61,62} via NF-κB and IL-1β signaling^61,63. In this context, the inflammatory microenvironment generated during active anti-tumor immunity may provide conditions permissive for EREG induction and enrichment of quiescent tumor cell states. Consistent with this possibility, we observed conversion of non-quiescent tumor cells into quiescent cells across multiple conditions, although this transition was greatly increased following CAR T-cell treatment in vivo. Notably, EREG expression was enriched in quiescent tumor cells at both the mRNA and protein levels even in treatment-naïve settings, suggesting that inflammatory signaling associated with CAR T-cell therapy is not strictly required for EREG expression or quiescence. Together, these findings suggest that both the baseline PDAC TME and therapy-associated inflammatory signals may contribute to the generation of an EREG-associated quiescent tumor cell state. Defining the upstream regulators governing EREG expression and tumor cell quiescence, such as inflammatory cytokines, will be an important area for future investigation.

An important consideration is the relationship between quiescence and stem-like tumor states. Consistent with prior studies, quiescent tumor cells displayed enhanced clonogenic growth and tumor-initiating capacity relative to bulk tumor cells. These observations suggest that quiescence and stemness may overlap within PDAC. However, whether quiescence directly confers stem-like properties or instead enriches for a broader stress-adapted tumor cell population remains unclear. We therefore view quiescence as a dynamic cellular state that may coexist with stem-like phenotypes rather than a completely distinct biological entity. Future studies will be required to further resolve the relationship between these overlapping cellular programs.

Our study also identifies a role for the EREG–ErbB4 (HER4) macrophage axis in human PDAC. EREG expression was enriched in quiescent human PDAC tumor cells and associated with poor overall survival in TCGA datasets. Furthermore, tumors enriched for EREG⁺ slow-cycling cells exhibited reduced cytotoxic T-cell infiltration, while HER4⁺ CD68⁺ macrophage-rich regions correlated with T-cell exclusion in human PDAC tissue microarrays. Importantly, the conserved EREG–HER4 macrophage transcriptional program was associated with poor patient survival, supporting the clinical relevance of this signaling axis beyond murine models. Given the limited efficacy of current immunotherapies in PDAC, these findings suggest that EREG or ErbB4 may serve as biomarkers of immune-excluded disease and therapeutic targets to enhance responsiveness to adoptive cell therapies. The observation that EREG neutralization phenocopied genetic knockdown further supports the feasibility of pharmacologic targeting of this pathway in preclinical models. Additional studies in clinically relevant therapeutic settings will be required to determine the translational potential of targeting the EREG–ErbB4 axis, including whether this pathway may also contribute to resistance to immunotherapy, such as immune checkpoint blockade.

We also identified an EREG-associated macrophage program distinct from canonical M1/M2 polarization paradigms. EREG-conditioned macrophages displayed a transcriptional profile enriched for neuronal-associated and immunomodulatory programs, suggesting that EREG promotes a macrophage state not adequately captured by conventional polarization frameworks. Spatial analyses further demonstrated that ErbB4-expressing macrophages were selectively enriched adjacent to EREG⁺ tumor cells, consistent with highly localized paracrine signaling within discrete tumor regions. Although these findings provide insights into EREG–ErbB4 signaling, additional studies utilizing higher-resolution spatial approaches will be required to further define the spatial relationship between EREG-rich tumor niches and ErbB4 signaling in macrophages.

EREG enhanced macrophage-mediated suppression of CAR T-cell activity in an ErbB4-dependent manner without directly impairing CAR T-cell cytotoxicity, supporting macrophages as the primary mediator of EREG-dependent immune suppression. Defining the downstream consequences of ErbB4 signaling in macrophages, including effects on cytokine production, antigen presentation, metabolism, and epigenetic regulation, will be important areas for future investigation.

Collectively, our findings identify an EREG–ErbB4 signaling axis linking therapy-associated tumor cell plasticity to macrophage-mediated immune suppression in PDAC. These data support a model in which quiescent, EREG-expressing tumor cells contribute to remodeling of the local immune microenvironment to impair CAR T-cell activity, highlighting this pathway as a candidate therapeutic vulnerability in immune-resistant PDAC.

Reporting summary

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

Author contributions

B.M. designed and performed all experiments and wrote the manuscript. Q.W. assisted with cell culture and in vitro experiments. K.A. provided clinical expertise, experimental guidance, and manuscript editing. W.M. provided experimental guidance, secured funding, and edited the manuscript. All authors reviewed and edited the manuscript.

Peer review

Peer review information

Nature Communications thanks anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by the Robert E. Askew, Sr., M.D. Chair in Oncology Endowment at the University of Texas at Austin Dell Medical School.

Data availability

Bulk RNA-seq datasets generated in this study have been deposited in the Gene Expression Omnibus (GEO) under accession codes GSE311363, and GSE328717. Spatial transcriptomic datasets generated in this study have been deposited in GEO under accession code GSE31136. These datasets are publicly available and can be accessed without restriction. Previously published single-cell RNA-seq data (GSE205013) (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE205013) used in this study were obtained from the Gene Expression Omnibus. Bulk RNA-seq data were analyzed using DESeq2 in RStudio. Single-cell RNA-seq data were processed using the Cell Ranger (v10.0) pipeline. Source data underlying all figures and Supplementary Figures. are provided with this paper, including uncropped and unprocessed immunoblot scans with molecular weight markers. No restrictions apply to access to the data associated with this study. Source data are provided with this paper.

Code availability

All programming was carried out using the specified packages. No code was produced.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-75883-z.