IFNγ-dependent tumor-cell MHC-II is restrained by MAPK signaling and associated with immunotherapy responsiveness in pancreatic ductal adenocarcinoma

Abstract

Pancreatic ductal adenocarcinoma (PDAC) is characterized by an immunosuppressive tumor microenvironment (TME) and poor prognosis. Although major histocompatibility complex class II (MHC-II) is typically expressed by professional antigen-presenting cells, its regulation and significance in malignant PDAC cells remain unclear. Here, we integrated single-cell RNA sequencing, spatial transcriptomics, bulk transcriptomics, multiplex immunohistochemistry, and functional studies in murine PDAC models to investigate tumor-cell MHC-II. Across human PDAC datasets, elevated tumor-cell MHC-II expression was associated with increased CD4⁺ and CD8⁺ T-cell infiltration and localization within immune-rich tumor niches. In paired biopsies from patients treated with cobimetinib, malignant cells showed increased expression of MHC-II pathway genes, suggesting therapy-induced activation of this program. In preclinical KPC models, MHC-II^high tumors exhibited greater T-cell infiltration, whereas MHC-II^low tumors derived the greatest benefit from cobimetinib combined with immune checkpoint blockade. Mechanistically, direct tumor stimulation via in vitro administration of cobimetinib and recombinant interferon (IFN)-γ resulted in MHC-II induction in vitro by both MHCII^high and MHC-II^low PDAC cell lines. In vivo, combined MEK inhibition and immune checkpoint blockade-controlled tumor growth, including in MHC-II^low tumors. Deletion of the IFN-γ receptor abolished MHC-II induction and therapeutic benefit, demonstrating that this response is IFN-γ dependent. MEK inhibition enhanced downstream IFN-γ signaling and the combination was required for MHC II upregulation in vitro, indicating that MAPK activity restrains an IFN-γ-responsive MHC-II program. Collectively, these findings identify tumor-cell MHC-II as an IFN-γ-dependent, MAPK-restrained program that denotes immunotherapy sensitivity in PDAC.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) has the poorest survival rates among cancers for all stages combined, marked by a median survival of six months and a five-year survival rate < 10%(1). This aggressive malignancy is characterized by extensive local invasion, distant metastases, and a highly heterogeneous tumor microenvironment (TME) associated with early metastases, rendering most cases ineligible for surgical resection and complicating treatment strategies(2,3). Moreover, the lack of effective treatment options leads to reliance on cytotoxic therapeutic regimens that fail to achieve improved treatment outcomes(4–6).

Traditional genomic and transcriptomic analyses have historically relied on bulk tumor data, often neglecting the crucial role of the TME in PDAC progression and immune evasion(7,8). Additionally, PDAC tumors exhibit marked heterogeneity, characterized by diverse cellular clones and distinct immune landscapes that critically influence tumor behavior and therapeutic resistance(9). However, recent advancements in single-cell RNA sequencing (scRNA-seq) have provided unprecedented insights into this heterogeneity, enabling a more detailed exploration of the interactions between malignant PDAC cells and the immune microenvironment thus providing a granular understanding essential for refining therapeutic strategies(10–13).

PDAC tumors are highly complex and include not only malignant cells, which evolve, giving rise to distinct clones with diverse characteristics, but also a wide variety of stromal and immune cell populations(14). On one hand, there are cell types with established roles in promoting tumor cell survival and evasion from antitumor immune cells, such as cancer-associated fibroblasts (CAFs), tumor-associated macrophages (TAMs), and regulatory T cells (Tregs)(15). On the other hand, high densities of immune cells with cytotoxic potential, such as CD8⁺ cytotoxic T cells (albeit relatively uncommon in PDAC(16–18)) and CD4⁺ T cells, are associated with improved survival, reflecting a more robust antitumor immune response, which has emerged as a critical prognostic factor in PDAC(19–21).

Despite these insights, the influence of tumor cell-intrinsic factors, particularly major histocompatibility complex class II (MHC-II) expression, on modulating immune responses has been insufficiently explored in PDAC. MHC-II molecules, traditionally expressed on professional antigen-presenting cells (APCs), such as dendritic cells, B cells, and macrophages, present processed exogenous antigenic peptides to CD4⁺ T cells, thereby promoting activation and differentiation of several leukocyte subsets, some of which (such as T cells) can orchestrate adaptive immune responses(22). Notably, emerging evidence indicates that some malignant cells can express MHC-II, potentially enhancing their immunogenicity and altering the interactions of malignant cells with the immune TME(23,24).

To characterize malignant-cell expression of MHC-II, we employed a comprehensive multi-omics approach, integrating single-cell and spatial transcriptomic data from both human and mouse models of PDAC, to investigate the role of MHC-II expression. Specifically, we examined whether tumor-cell MHC-II in PDAC marks an IFN-γ-responsive immune-active state, how this program is regulated by MAPK signaling, and whether it is associated with therapeutic responsiveness. Our findings demonstrated that elevated MHC-II expression in malignant PDAC cells correlates with increased infiltration of CD4⁺ and CD8⁺ T cells in the TME, reflecting a more cytotoxic permissive immune microenvironment, which we validated by leveraging a high-resolution single-cell atlas comprising over a million cells across different stages of PDAC progression. Additionally, mechanistic experiments employing the PDAC (KPC) mouse models revealed that a shift from checkpoint inhibitor resistance to therapeutic responsiveness was associated with induction of MHC-II expression in malignant PDAC cells. Notably, we observed that the same KPC models, including those which previously showed resistance to combination chemo- and immunotherapy (gemcitabine, nab-paclitaxel, anti-CD40 agonist, anti-CTLA-4, and anti-PD-1) as reported in Li et al.(25), exhibited a marked therapeutic response when treated with cobimetinib plus dual PD-1 and CTLA-4 blockade, prompting us to investigate whether cobimetinib facilitates an IFN-γ-responsive tumor-cell program indicated by tumor cell expression of MHC II. Collectively, our findings identify tumor cell expression of MHC-II as a potential prognostic biomarker of immunotherapeutic response and altered immune landscape in the pancreatic cancer TME. These observations provide rationale for further investigation of the IFN-γ/MAPK/MHC-II axis as a biomarker of immunotherapy response in PDAC.

Materials and Methods

Human tissue samples

This study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki. Human tissue samples were obtained from a patient with pancreatic ductal adenocarcinoma (PDAC) enrolled in the window-of-opportunity SMMART clinical trial of cobimetinib (NCT04005690). For the present single-cell analysis, the patient was selected because serum CA19-9 levels decreased following cobimetinib treatment and matched pre-treatment and on-treatment tumor samples were available for analysis. No additional study-specific exclusion criteria were applied. The dataset comprised two matched metastatic PDAC tissue samples from patient ST00022941. The pre-treatment sample was collected before initiation of cobimetinib therapy, and the matched on-treatment sample was collected after 10 days of treatment. Samples were subsequently processed for single-cell RNA sequencing as described below. Collection and handling of human patient-derived materials were performed under protocols approved by the Oregon Health & Science University (OHSU) Institutional Review Board (IRB #00003609). Tissue samples were obtained through the Oregon Pancreas Tissue Registry program.

Tissue dissociation

Two matched biological specimens were analyzed, consisting of one pre-treatment and one on-treatment sample from the same patient; no technical replicates were generated. Patient tumor tissues were dissociated using gentleMACS Dissociator kits (Miltenyi Biotec, #130-095-929) according to the manufacturer’s instructions. The resulting cell suspensions were resuspended and filtered through a 70-μm cell strainer (Miltenyi Biotec, #130-098-462). Cells were collected by centrifugation at 300 × g for 7 minutes at 4°C and subsequently resuspended at a concentration of 700–1,200 cells/μL. Live cells were enriched using the EasySep^™ Dead Cell Removal (Annexin V) Kit (STEMCELL Technologies, #17899).

Single-Cell RNA Sequencing Library Construction

Single-cell suspensions were processed according to the 10x Genomics single-cell RNA-seq protocol using the Chromium Next GEM Single Cell 3’ Kit v3.1. No prior sorting or enrichment was performed, allowing analysis of the entire mixed cell population. Cell suspensions were loaded onto the Chromium Controller to generate Gel Beads-in-Emulsion (GEMs). Libraries were prepared from amplified cDNA using 14 PCR cycles. Quality control of intermediate products and final libraries was performed using a 1% E-Gel system (Thermo Fisher Scientific), and library concentrations were quantified using a Qubit Fluorometer (Thermo Fisher Scientific), following the manufacturer’s protocols. All libraries from a given patient were prepared within the same experimental batch to minimize batch effects. Sequencing was carried out on the Illumina NovaSeq 6000 platform at the OHSU Massively Parallel Sequencing Shared Resource, with a targeted sequencing depth of 20,000 read pairs per cell.

scRNA-seq data processing and quality control

Raw sequencing data were processed using Cell Ranger (10x Genomics, v7.1.0, RRID: SCR_017344). The mkfastq and count commands were executed with default parameters, aligning reads to the GRCh38 human reference genome to generate a matrix of unique molecular identifier (UMI) counts per gene and corresponding cell barcodes. Seurat objects were constructed from each dataset using the Seurat pipeline (version 5.0.3, RRID: SCR_016341).

Quality control filters were applied to exclude low-quality cells and potential doublets. Specifically, cells were removed if they met any of the following criteria: fewer than 300 detected genes or fewer than 1,200 UMIs; greater than 20% mitochondrial gene expression, indicating compromised cell integrity; greater than 7,500 UMI counts and more than 60,000 transcripts, suggestive of doublets. The filtered gene expression matrix was normalized using the NormalizeData function with default parameters. The merged and normalized matrix was used for downstream analysis. Doublets were computationally identified and removed using DoubletFinder (version 2.0.4, RRID: SCR_018771) with the following parameters: pN (proportion of artificial doublets): 0.25 (default value); nExP (expected doublet rate): 0.075 (based on the DoubletFinder pipeline). These settings were provided as inputs to the doubletFinder function, with the number of principal components (PCs) set to 10 for doublet identification. Cells were clustered using the Louvain algorithm and the FindNeighbors and FindClusters functions, with the top 10 PCA dimensions and a resolution parameter of 0.5. The Uniform Manifold Approximation and Projection (UMAP) for visualizing clusters was generated using the RunUMAP function with the same 10 principal components used for clustering.

scRNA-seq cell-type annotation

Cell-type annotation was performed at the cluster level following the clustering procedure described above. Clusters were assigned to major cell lineages based on the expression of established canonical marker genes. CD4⁺ T cells were identified by expression of CD3D, CD3E, and IL7R; regulatory T cells by FOXP3, IL2RA, and CTLA4; CD8⁺ T cells by CD8A and CD8B; proliferating T cells by MKI67, TOP2A, CDK1, PCNA, and BIRC5; NK cells by NCAM1, KLRD1, GNLY, and PRF1; B cells by CD19, CD79A, and MS4A1; myelomonocytic cells by CD14, LYZ, FCGR3A, CD68, CD163, and CD86; neutrophils by S100A8, S100A9, and CSF3R; mast cells by TPSAB1, CPA3, and KIT; endothelial cells by PECAM1, VWF, and CDH5; fibroblasts by FAP, ACTA2, and PDPN; malignant epithelial cells by EPCAM, MUC1, KRT19, KRT7, KRT18, KRT8, SPP1, CFTR, and CLU; and hepatocytes by ALB, TF, and APOA1. Marker-expression patterns for each cluster were manually reviewed to confirm consistency with the assigned cell lineage. Malignant epithelial-cell annotations were further supported by inferred copy-number variation profiles, as described below. All cluster annotations were performed by a single annotator to maintain consistency across the dataset.

Mice

C57BL/6 (Cat# 00664; RRID: IMSR_JAX:000664) mice were purchased from Jackson Laboratories (Bar Harbor, ME) and were bred and maintained under specific pathogen-free conditions at an American Association for the Accreditation of Laboratory Animal Care (AAALAC)-accredited animal facilities within Oregon Health and Science University. Mice were housed in compliance with the procedures that were reviewed and approved by the Institutional Animal Care and Use Committee of Oregon Health and Science University (Protocol #4153). Unless otherwise stated, sex and age-matched littermates (6-12 weeks of age, both sexes) were used for individual experiments.

In vivo reagents and treatments

Mice were treated with cobimetinib (Selleck Chemicals; Cat# S8041) by oral gavage at a dose of 5.0 mg/kg, in 10% PEG300/PBS. Treatment commenced after detectable tumors; between 7-14 days for orthotopically implanted tumors or 10-12 days for subcutaneously implanted tumors. Immune checkpoint blockade (‘ICB’) was treated by dual treatment with αPD-1/αCTLA-4 (‘PC’) as previously described(26). Briefly, αPD-1 (RMP1-14; BioXcell, Cat # BE0146; RRID: AB_10949053; 200 mg/dose) and αCTLA-4 (9H10; BioXcell, Cat # BE0131; RRID: AB_10950184; 200 mg/dose) were injected intraperitoneally (i.p.) on days 0, 3, and 6(26). For T cell depletion experiments, mice received treatments starting two days before cobimetinib/ICB was initiated using 200μg/dose of αCD8a (clone: 2.43; Bio X Cell, Cat# BE0061, RRID: AB_1125541) and 200μg/dose of αCD4 (clone GK1.5; Bio X Cell, Cat# BE0003-1, RRID: AB_1107636) every 4 or 3 days, respectively, for the duration of the experiment. For isotype controls, rat IgG2a (2A3; BioXcell, Cat # BE0089; RRID: AB_1107769; 200 mg/dose) was used. All antibodies were endotoxin free.

Implantation of tumor cell clones

Kras^LSL-G12D/+; Trp53^LSL-R172H/+; Pdx1-Cre; Rosa26^YFP/YFP tumor clones were generated as previously described in 2018 (25) and are also commercially available (Kerafast). Cell lines were examined by the Infectious Microbe PCR Amplification Test (IMPACT) and authenticated to be free of contamination by the Research Animal Diagnostic Laboratory (RADIL) at the University of Missouri.

Cells (passages 14-17 from the original derivation) were cultured in DMEM (high glucose without sodium pyruvate) with 7.5-10% FBS (Gibco) and glutamine (2.0 mM) and harvested after four days in culture when 70-80% confluent. Tumor cells had viability >90% for each experiment, and dissociated into single cells with 0.25% trypsin (Gibco), washed with serum-free Dulbecco’s Modified Eagle’s medium (DMEM) twice, and counted in preparation for subcutaneous implantation at a dose of 2.5×10⁵ PDAC tumor cells. For orthotopic tumors, 0.5-5×10⁵ PDAC cells were implanted into the pancreas via ultrasound guided injection (27) and monitored by ultrasound every 5-7 days. To generate Ifngr deficient lines, we utilized lentiviral transduction of with CRISPR/cas9 and sgRNA. Briefly, lentiCRISPR v2 (a gift from Feng Zhang; Addgene plasmid #52961 (28); RRID: Addgene_52961) was transfected into 293T cells using PEI reagent and packaged into lentivirus for transduction. Lentivirus was collected 48 and 72 hours after transfection and filtered for usage. Tumor cells were transduced with packaged virus with polybrene for 48 hours and selected with puromycin before expansion as we previously reported (29). Cells were then stimulated with 200 units of IFN-α or IFN-γ for 24 hours prior to flow sorting to confirm lack of IFN-γR expression. Ifngr guides: Ifngr1-sg1: CGACCGTATGTTTCGTATGT; Ifngr1-sg2: ACTTGAACCCTGTCGTATGC, Ifngr1-sg3: AGTATAGTCGAATGTGTAAC.

Tumor growth, regression, and animal survival assessment

Subcutaneous tumor sizes were measured every 2-3 days for tumor growth assessment experiments. Tumor length and width were measured with calipers and tumor volumes were then calculated as length*width²/2. Orthotopic tumors were detected and monitored longitudinally by ultrasound every 5-7 days using a Fujifilm VevoSonic 2100 ultrasound with 3D motor. Tumor regressions and waterfall plots were calculated using the initial tumor size at the start of treatment to tumor size at indicated later time point.

Flow cytometry of murine tumor samples

At the indicated time points, mice were sacrificed and tumors prepared for single cell suspension as previously described(26). Briefly, tumors were washed with PBS, then minced and incubated for 45 minutes in 1mg/mL collagenase XI with protease inhibitor (Sigma-Aldrich, Cat # C6079) at 37°C and filtered through a 70 μM cell strainer with cold PBS supplemented with 0.5% BSA and 2.0 mM ethylenediaminetetraacetic acid (EDTA). Cells were counted with Beckman Coulter Counter Z2 and stained with Ghost Dye UV450 (Cytek Biosciences Cat # 13-0868-T100). Cell surface molecules were assessed by incubating single cell suspensions from tissues with primary fluorophore-conjugated antibodies on ice for 45 minutes in PBS with 0.5% bovine serum albumin and 2.0 mM EDTA. Combinations of the following antibodies were used for flow cytometric analysis: I-A/I-E (MHC-II) (clone M5/114.15.2, Brilliant Violet 711, BD Horizons, Cat # 563414, RRID: AB_2738191), CD3e (clone 145-2C11, PE-Cy5, Biolegend, Cat # 100310, RRID: AB_312675), CD45 (clone 30-F11, UV486, BD Optibuild, Cat # 749889, RRID: AB_2874129). Flow cytometric analysis was performed on Fortessa flow cytometer (BD Biosciences, RRID: SCR_018655) using BD FACSDiva software (RRID: SCR_001456) and analyzed using FlowJo software version 10.10.1 (Treestar) (RRID: SCR_008520).

For in vitro tumor cell stimulations, the IC50 for each cell line after in vitro treatment with cobimetinib was determined by MTT Assay. Tumor cell lines were then cultured +/− cobimetinib at IC50 (0.0004-0.5 mg/mL) or vehicle control, +/− 10 ng/mL recombinant mouse IFN-γ, for 48 hours. Tumor cell expression of I-A (MHC-II) was then determined by flow cytometry as above.

Multiplexed Immunohistochemistry

Tumor specimens from 11 patients with PDAC were analyzed by multiplex immunohistochemistry, comprising 195 regions of interest (ROIs) and a total analyzed tissue area of 568 mm². Tumor specimen slides were processed and stained with a panel of 22 antibodies following previously described protocols(30,31). Briefly, mIHC staining was performed on 5.0 μm sections and imaged at 20x magnification on an Aperio At2 scanner (Leica Biosystems) (RRID: SCR_020993) prior to image processing. Hematoxylin and eosin staining of the same slides allowed for pathologist annotations that were used to select multiple 1.0 mm² regions of interest (ROIs) for sampling across each primary tumor resection. Human tonsil and spleen were also included in all rounds of mIHC as staining controls. Image registration was performed using the SURF algorithm in the Computer Vision Toolbox from MATLAB (The MathWorks) (RRID: SCR_017581), and the AEC chromogen signal was extracted using Fiji(32) (RRID: SCR_002285). Single-cell segmentation and labeling were carried out with StarDist 2D (RRID: SCR_027715), and marker mean signal intensities were quantified at the single-cell level using CellProfiler(33) (RRID: SCR_007358). Gating thresholds for marker positivity were established using FCS Express Image Cytometry (De Novo Software) (RRID: SCR_016431).

Single-cell data collection and processing

Public scRNA-seq data were sourced from the Deeply Integrated Single-Cell Omics (DISCO) database (https://www.immunesinglecell.org/). To ensure a reproducible and standardized dataset selection strategy, data retrieval was performed using the Python discotoolkit package with predefined filtering criteria: tissue = “pancreas”, platform = [“10×3’“, “10×5’“], cell_type_confidence = “high”, include_cell_type_children = True, and min_cell_per_sample = 300. Using this query framework, we collected 22 datasets comprising 264 pancreatic samples and 1,042,407 cells to construct the integrated atlas.

For tumor-focused analyses, we further restricted the dataset to 115 PDAC samples encompassing 557,304 cells. These samples were used to investigate the relationship between malignant-cell MHC-II pathway activity, or the proportion of predicted responsive malignant cells, and the infiltration of other cell types. This strategy was designed to generate a large, consistently processed reference cohort centered on malignant cells.

To address potential batch effects arising from differences in platforms and experimental protocols, data integration was performed in Scanpy (RRID: SCR_018139) using the BBKNN(34) (RRID: SCR_022807) algorithm with default parameters, specifying each study as the batch_key. Clustering and single-cell distributions were visualized using UMAP in combination with Leiden clustering. This approach effectively harmonized the datasets while preserving biological variability. A unified view of cell populations across PDAC stages was generated using sc.tl.umap with default parameters.

Malignant cells were identified using the R package scATOMIC (v2.0.2)(35), with known_cancer_type = “PAAD cell” and default settings for all other parameters. This rigorous annotation strategy enabled accurate classification of malignant populations and provided a robust foundation for downstream analyses of PDAC progression and heterogeneity.

Spearman correlation analysis was then performed to evaluate associations between malignant-cell MHC-II pathway activity, or the proportion of responsive malignant cells, and cell-type infiltration across samples.

Spatial transcriptomic data collection and processing

Public spatial transcriptomic RNA sequencing datasets were obtained from four sources. First, 14 Visium samples and their matched snRNA-seq data from the WUSTL Atlas were downloaded from the Human Tumor Atlas Network (HTAN) (RRID: SCR_023364) Data Coordinating Center Data Portal at the National Cancer Institute (https://data.humantumoratlas.org/). Second, 55 Visium samples, encompassing both primary and metastatic lesions, were retrieved from the Gene Expression Omnibus (GEO) (RRID: SCR_005012) under accession number GSE277783. As this second set lacked matched single-cell transcriptomic data, we utilized the same publicly available scRNA-seq dataset(36) employed in Pei’s study(37), derived from 23 patients with PDAC. This dataset contained 23,042 cells and captures all major cell populations, including T cells, B cells, plasma cells, myeloid cells, fibroblasts, endothelial cells, and malignant tumor cells. Additionally, four well-annotated Visium samples and nine ST sample were obtained from the SORC database (http://bio-bigdata.hrbmu.edu.cn/SORC/Browse.jsp#). Lastly, two Xenium slices from human pancreatic cancer tissue—generated using the pre-designed Xenium Human Multi-Tissue and Cancer Panel—were downloaded from the 10x Genomics database (RRID: SCR_023672). For downstream analysis, cells from each Xenium dataset were filtered to retain those with nFeature_Xenium > 10 and nCount_Xenium > 20.

The RCTD method was used to deconvolute cell-type composition within each spatial spot, utilizing scRNA-seq data as the reference for the 10x Visium samples from the HTAN WUSTL cohort. RCTD objects were generated from processed RDS files with the parameters max_cores = 20 and CELL_MIN_INSTANCE = 10. The RCTD pipeline was executed on these objects with doublet_mode = “full” for 10x Visium data. The deconvolution results were stored as matrices of cell-type weights for each spot. The cell-type weights for each spot were normalized to ensure that the sum of weights equaled 1. Proportion matrices were created, with rows representing spatial spots and columns representing cell types and were stored for further analysis. For rough annotation, each spatial spot was assigned a primary cell type based on the highest probabilistic proportion.

Bulk dataset collection and preprocessing

In this study, RNA-seq fastq data for 39 sorted mouse tumor cell samples (PRJNA432498) were downloaded from the GEO database (http://www.ncbi.nlm.nih.gov/geo/). The quality of the raw fastq files was assessed using FastQC (v0.11.8, RRID: SCR_014583) (http://www.bioinformatics.bbsrc.ac.uk/projects/fastqc/). Sequencing reads were aligned to the GRCm38 mouse reference genome using STAR (v2.6.0a, RRID: SCR_004463), and per-gene counts and transcripts per kilobase million (TPM) values were quantified using RSEM (v1.3.1, RRID: SCR_013027) based on the gencode.vM25.annotation.gtf gene annotation file (GENCODE, RRID: SCR_014966). For normalization, gene lengths were determined by the number of bases covered at least once across all exons for each gene. TPM values were calculated as follows: 1. Reads per kilobase (RPK) were computed for each gene; 2. The scaling factor was calculated as the sum of all RPK values $\text{Scaling}\text{factor}={\sum }_j{\text{RPK}}_j$; 3. TPM values were derived by normalizing each gene’s RPK to the scaling factor and multiplying by 10⁶: ${\text{TPM}}_i=\left(\frac{{\text{RPK}}_i}{\text{Scaling}\text{factor}}\right)\times {10}^6$. For cases with duplicate sequencing runs for the same patient, TPM counts were averaged. Gene expression values were log-transformed using the formula ${\log }_2\left(\text{TPM}+1\right)$ for each sample. To standardize expression data, z-scores were calculated for each gene across all samples using the formula: $z=\frac{x-\mathrm{\mu }}{\mathrm{\sigma }}$ where $x$ is the log-transformed TPM value, $\mathrm{\mu }$ is the average ${\log }_2\left(\text{TPM}+1\right)$ across all samples for a given gene, and $\mathrm{\sigma }$ is the standard deviation of ${\log }_2\left(\text{TPM}+1\right)$ for that gene.

Copy-number analysis

Copy-number variations in malignant epithelial cells were inferred using InferCNV (version 1.18.1, RRID: SCR_021140). Lymphoid immune cells, including CD8⁺ T cells, CD4⁺ T cells, NK cells, and B cells, were used as reference controls. The input data consisted of gene expression count matrices. Filtering, normalization, and centering based on normal gene expression were performed using the default parameters, followed by data scaling. A minimum average read count threshold of 0.1 per gene among reference cells was applied to ensure robust CNV detection.

Differential scRNA gene expression and pathway enrichment analysis

Differential gene expression analysis was performed using two-tailed Wilcoxon rank-sum tests implemented in the Seurat package via the FindMarkers function with default parameters. Results were visualized using dot plots to depict the relative expression levels of differentially expressed genes (DEGs) across experimental conditions. Gene Ontology (GO) analysis was performed using the msigdbr R package (v7.5.1, RRID: SCR_022870). Rank-based gene set enrichment analysis (GSEA) was conducted using the clusterProfiler R package (v4.10.0, RRID: SCR_016884). For bulk transcriptomic data and pseudo-bulk data (generated by aggregating gene expression counts from all malignant PDAC cells using the PseudobulkExpression module in the Seurat package), single-sample GSEA (ssGSEA) was performed with the GSVA R package (v1.50.0, RRID: SCR_021058). For scRNA-seq and stRNA-seq data, the AddModuleScore function in Seurat was utilized to account for technical variability and mitigate the impact of dropouts. Genes of interest were binned into 24 expression-level bins based on their average expression. Control genes were randomly selected from each bin using default parameters in the AddModuleScore function. The final module score was calculated as the difference between the average expression of the gene set of interest and the average expression of the control genes.

Determining cancer-cell lineage states

To characterize the heterogeneity of cancer cell lineage states, we assessed MHC-II expression across multiple platforms. For Visium data, the AddModuleScore function was used to calculate MHC-II pathway activity scores. Because AddModuleScore represents the average expression of the MHC-II gene set relative to matched control genes with similar overall expression levels, a score greater than 0 indicates relative enrichment of the MHC-II transcriptional program above the matched background, whereas a score of 0 or below indicates little or no relative enrichment of the program. Accordingly, spots with a score greater than 0 were classified as “MHC-II^high,” while those with a score of 0 or below were designated “MHC-II^low.” This threshold was used as an operational classification criterion rather than as an absolute biological cutoff. For Xenium data, due to the sparsity of gene expression signals, MHC-II status in cancer cells was defined based on the detectable expression of canonical MHC-II markers. For mIHC data, MHC-II protein expression status in malignant cells were determined using sample-specific thresholds adjusted for background noise, ensuring that marker positivity accurately reflected true expression.

Cellular neighborhood analyses

To characterize the spatial relationships between immune, stromal, and tumor cells, cellular neighborhoods were defined by placing a circular window with a fixed radius of 50 μm around each seed cell of a specified phenotype across Xenium and mIHC ROIs. All cells whose centers fell within this radius were considered neighbors and were included in the corresponding neighborhood of the seed cell. Neighborhoods across all malignant cell regions were then analyzed by clustering their normalized cellular compositions. This was performed using the dbscan R package in combination with the k-nearest neighbors (kNN) algorithm, followed by K-means clustering to identify distinct spatial neighborhood types. To assess compositional differences among the neighborhoods of malignant cells, we applied the Wilcoxon rank-sum test. Both absolute cell counts and relative proportions of neighboring cell types were evaluated to determine statistically significant variations. To ensure robust and reliable comparisons in the downstream cellular neighborhood analyses, each Visium sample was required to meet the following criteria: a minimum of 10 spots classified as “MHC-II^high” and 10 spots as “MHC-II^low”, as well as either at least 5 immune-enriched spots for immune neighborhood analyses or a minimum of 50 CAF spots for stromal neighborhood analyses, respectively. For mIHC colocalization analyses, to ensure sufficient and balanced representation of both MHC-II states, only ROIs containing more than 100 MHC-II^high malignant cells and more than 100 MHC-II^low malignant cells were included in downstream analyses.

Heterotypic interaction between spatially adjacent cell types

To quantify spatial interactions between distinct cell types, we computed a heterotypic interaction score based on spatial proximity within the Visium platform. First, spatial coordinates of each spot were scaled and mapped to a unified spatial object. For each sample, the six nearest spatial neighbors of each spot were identified using the kNN function like before, with a maximum distance threshold of 200 μm and a minimum of four ranked neighbors retained to ensure local connectivity. The resulting neighbor network was converted into an undirected graph containing weighted edges inversely proportional to spatial distances between spots. To examine heterotypic interactions, we utilized cell type deconvolution matrices obtained from previous analyses, focusing on spot pairs enriched for different cell types. A spot was considered positive for a given cell type if its deconvolution score exceeded 0.2. For each pair of distinct cell types, we identified all edges in the spatial network that connected spots enriched for one cell type to those enriched for another. These edges were used to define the heterotypic interaction network between the two cell types.

Statistical Analysis

All statistical analyses were conducted using R (v4.3.1, RRID: SCR_001905) and GraphPad Prism (versions 7) (RRID: SCR_002798). For comparisons between two independent groups, either a two-tailed unpaired Student’s t-test (for normally distributed data) or Wilcoxon rank-sum test (for non-normally distributed data) was applied. Kaplan–Meier survival analyses were assessed using the log-rank test. Tumor growth kinetics were analyzed using two-way ANOVA with mixed effects modeling. For analyses involving multiple comparisons, the Benjamini–Hochberg procedure was used to control the false discovery rate. Error bars in all figures represent either the standard deviation (SD) or the standard error of the mean (SEM), as indicated in the figure legends. Statistical significance was defined as p < 0.05. Significance levels are denoted as follows: p < 0.05 (*), p < 0.01 (**), and p < 0.001 (***). “ns” indicates not significant.

Results

Study design and overview of the study cohort.

Given the hypothesis that tumor-cell MHC-II expression is associated with immune contexture within the TME, we aimed to define its biological and clinical significance using a multi-omics framework. To begin, we analyzed paired pre- and post-treatment PDAC samples from a patient enrolled in a window-of-opportunity clinical trial of the MEK inhibitor cobimetinib (NCT04005690), where treatment led to reduced malignant cell proliferation (decreased Ki-67) and modulation of MAPK signaling, correlating with a therapeutic response (Fig. 1A). To molecularly confirm our observation from this patient, we integrated a range of datasets encompassing 264 pancreatic samples across human PDAC-relevant cohorts, including 195 1.0 mm² tumor regions evaluated by multiplex immunohistochemistry (mIHC) reflecting PDAC samples obtained from 11 patients, 39 murine PDAC RNA-seq samples, and multiple spatial transcriptomics datasets including two Xenium samples, nine Slide-seq samples (ST), and 73 Visium slides. For additional insights into PDAC biology and immune interactions, we included 22 publicly available PDAC scRNA-seq datasets, comprising 1,042,407 cells. Moreover, to validate the findings, data from the Cancer Cell Line Encyclopedia(38) (CCLE) (RRID: SCR_013836) database with 1,165 cancer cell lines (including 49 PDAC cell lines) were incorporated into the analysis. These comprehensive datasets enabled the integration of clinical, experimental, and computational analyses to decipher how neoplastic cell expression of MHC-II expression correlates with TME and leukocyte features in PDAC (Fig. 1A).

Figure 1. Properties of the cellular TME of pancreatic cancer metastases to the liver.

A. Study design and datasets overview. B. UMAP plot of pre- and on-treatment cell populations with distinct cell types annotated: cells were classified into epithelial (malignant) cells and hepatocytes, stromal cells (fibroblasts, endothelial cells), and immune cell types (CD4⁺ and CD8⁺ T cells, Tregs, natural killer (NK) cells, monocytes/macrophages (Mono/Mph), B cells, mast cells and neutrophils). Each cluster is annotated based on specific marker genes, with the number of cells per cell type indicated, n=1 patient samples at each time point. C. Dot plot showing average expression levels and percent expression of key marker genes across different cell populations: the dot size represents the percentage of cells expressing a specific gene within a given cell type, while color intensity indicates the average expression level. D. Fraction of malignant cells expressing CIITA, canonical MHC-II pathway genes, and IFN-γ-related lysosomal processing genes in matched pre- and on-treatment biopsies. Dot size indicates percent expressed. For each gene, blue denotes the condition with the higher fraction of expressing malignant cells, whereas gray denotes the comparison condition. Panel A was partially created using BioRender elements. Created in BioRender. Chen, C. (2026) https://BioRender.com/h7vcod9.

Single-cell RNA analysis of pancreatic ductal adenocarcinoma liver metastases

First, to evaluate transcriptomic reprogramming following MEK1/2 inhibition with cobimetinib, we conducted scRNA-seq on paired pre-treatment and on-treatment PDAC metastatic liver specimens biopsied from a patient (NCT04005690) who received 10 days of treatment with cobimetinib. Using the 10x Genomics platform, we generated high-quality single-cell transcriptomic profiles from a total of 6,187 cells, including 1,734 cells from the pre-treatment sample and 4,453 cells from the on-treatment biopsy sample. Notably, this patient, whose tumor molecularly responded to treatment as evidenced by a significant decrease of malignant cell Ki67 expression and increased infiltration of diverse leukocytes following cobimetinib therapy, provided an opportunity to investigate the underlying mechanisms of a tumor’s response to therapy.

To classify cellular heterogeneity in the metastatic PDAC liver, we combined pre- and on-treatment scRNA-seq data (Materials and Methods), normalized to count for sequencing depth, and to mitigate technical variations, and then applied principal component analysis (PCA). Potential batch effects were detected by using Uniform Manifold Approximation and Projection (UMAP), which demonstrated minimal batch effects, as most clusters contained cells from both pre- and on-treatment conditions indicating the tumor of origin was the dominant determinant in the UMAP (Supplementary Fig. 1A).

To investigate specific cell subsets, cell clusters were annotated based on expression of well-established cell-type lineage biomarkers(39,40). Primary cell types identified in the PDAC TME included epithelial cells, stromal cells, and immune cells. Secondary clustering further refined these classifications, revealing key subpopulations including CD4⁺ T cells (CD3D, CD3E, IL7R), regulatory T cells (Tregs) (FOXP3, IL2RA, CTLA4), CD8⁺ T cells (CD8A, CD8B), proliferative T cells (MKI67, TOP2A, CDK1, PCNA, BIRC5), NK cells (NCAM1, KLRD1, GNLY, PRF1), B cells (CD19, CD79A, MS4A1), myelomonocytic cells, e.g., monocytes or macrophage (CD14, LYZ, FCGR3A, CD68, CD163, CD86), neutrophils (S100A8, S100A9, CSF3R), mast cells (TPSAB1, CPA3, KIT), endothelial cells (PECAM1, VWF, CDH5), fibroblasts (FAP, ACTA2, PDPN), malignant tumor cells (EPCAM, MUC1, KRT19, KRT7, KRT18, KRT8, SPP1, CFTR, CLU) and hepatocytes (ALB, TF, APOA1). (Fig. 1B, 1C).

Given our interest in antigen presentation by tumor cells, we wanted to ensure high-confidence identification of malignant PDAC cells in subsequent treatment response analyses. To this end, we assessed chromosomal-scale copy number variations (CNVs) in malignant epithelial cells using the InferCNV program(41) and fastCNV algorithm(42). By comparing each gene expression of each cell across chromosomes to reference diploid cells, we identified somatic CNV changes. Malignant epithelial cells exhibited significant CNVs spanning entire p or q arms (Supplementary Fig. 1B), indicating a high degree of aneuploidy, a hallmark of PDAC(43,44). To visualize per-cell CNV burden across the dataset, we then projected all cells onto a UMAP and colored each cell by its inferred CNV fraction; high-burden cells localized to the malignant compartment, whereas immune and stromal cells showed near-zero CNV (Supplementary Fig. 1C). Notably, these chromosomal alterations were not observed in other cell types, validating the epithelial cells as malignant and underscoring the chromosomal instability characteristic of this metastatic PDAC.

Significant alteration in MHC-II pathway activity within malignant cell populations on-therapy

Using the clinical biopsy sample, we next investigated the expression profile change in the PDAC TME. After isolating the malignant cell population, we performed differential gene expression (DEG) analysis to assess treatment-induced functional changes. Pathway enrichment using the enricher function in the R package clusterProfiler(45) (Materials and Methods) revealed marked upregulation of immune-related pathways, indicating a robust immune activation signature within the tumor microenvironment(46). Notably, three MHC-II-related GO terms consistently ranked among the top enriched pathways—MHC class II protein complex binding, MHC protein complex binding, and MHC class II receptor activity (Supplementary Fig. 1D). Although CIITA itself was only sparsely detected, malignant cells from the on-treatment biopsy showed a broader representation of cells expressing multiple canonical MHC-II antigen-processing/presentation genes and selected IFN-γ-related lysosomal processing genes relative to the matched pre-treatment biopsy (Fig. 1D). Among these, “MHC class II protein complex binding,” the top-ranked MHC-II-related enriched pathway, was selected as the representative MHC-II program for subsequent analyses. These findings highlight the functional plasticity of malignant cells in response to therapy(47) —a critical factor contributing to intercellular heterogeneity in therapeutic response and immunogenicity, both of which can profoundly influence treatment efficacy and patient prognosis(48,49).

Upregulation of the MHC-II antigen presentation pathway activity in malignant cells during therapy, accompanied by shifts in immune and malignant cell proportions (Supplementary Fig. 1E) and a notable decline in the biomarker CA19-9 in this donor, indicated that PDAC cells may be responsive to cobimetinib treatment, potentially correlating with increased expression of antigen presentation machinery on the tumor cell surface.

Validation of MHC-II pathway gene expression in malignant tumor cells via multi-omics approaches

Our scRNA-seq analysis identified that MHC-II and correlated genes were expressed in PDAC cells using cell-type markers (Fig. 1C) and copy number variations (Supplementary Fig. 1B), but the possibility of signal contamination from non-malignant cells remained a concern. Therefore, we next assessed malignant cell-intrinsic MHC-II expression using transcriptomic profiles from the CCLE database, which exclusively comprises tumor cell lines and therefore minimizes confounding by immune cell-derived MHC-II transcripts. As shown in Fig. 2A, pancreatic cancer cell lines exhibited marked heterogeneity in the expression of canonical MHC-II antigen-processing and presentation genes, in contrast to lymphoma cell lines included as MHC-II-positive controls, which displayed consistently high expression of the same program. Overall, the lower and more heterogeneous MHC-II program expression observed across PDAC cell lines is consistent with the relatively immune-poor, immunosuppressive phenotype that characterizes PDAC in vivo(50), although these cell line data do not themselves capture the full tumor microenvironment. Comparable variability was also observed across the broader set of tumor cell lines (Supplementary Table 1), supporting the notion that tumor cell-intrinsic MHC-II programs vary substantially across cancer types and models.

Figure 2. Validation of malignant cell-intrinsic MHC-II expression across multi-omics platforms.

A. Heatmap showing z-scored expression of canonical MHC-II antigen-processing/presentation genes across CCLE cell lines (n=69 lymphoma cell lines, n=49 pancreatic cancer cell lines). Known MHC-II-positive lymphoma cell lines were included as positive controls, alongside pancreatic cancer cell lines, to illustrate the relative range and heterogeneity of tumor-cell MHC-II program expression. B. Visualization of cell type annotations mapped to Visium spots, inferred by the RCTD deconvolution algorithm. Shown is one representative image of the n=14 Visium samples from the HTAN WUSTL cohort analyzed in this study (specimen HT224P1, as labeled on the panel). C. Spatial distribution of key MHC-II-related genes (CD74, HLA-DRA, HLA-DPB1, HLA-DPA1, HLA-DRB1 and HLA-DMB) expression within the malignant cell spotsof the same representative section shown in (B) (specimen HT224P1), demonstrating variation in MHC-II pathway activation across the tumor microenvironment. D. Cell-type resolved spatial maps of n=2 tumor samples profiled by Xenium in situ platform: a Stage IIB, Grade 3 tumor (left) and a Grade I–II tumor with ~50% malignant content (right). E. Spatial gene expression heatmaps of CD74 (left) and HLA-DQB2 (right) in malignant regions from the corresponding tumor sections in (D). F. mIHC analysis of HLA-II expression in PDAC tissues (n = 11 patients; 195 ROIs analyzed): Representative images showing DAPI (nuclear stain, blue), cytokeratin (CK; epithelial marker, magenta), and HLA-DR (MHC-II marker, yellow) expression. Malignant cells co-staining for HLA-DR and CK indicate MHC-II-positive malignant cells. G. Boxplot illustrating comparison of MHC-II pathway activity between T-cell^high and T-cell^low regions. Each dot represents a single sample of ex vivo sorted tumor cells used for bulk RNASeq, with n=1-3 replicates across n=17 tumor cell clones, data analyzed from Li et al(25). Data analyzed by Wilcoxon rank sum test. H. Heatmap showing MHC-II-related gene expression across transcriptomic profiles of sorted malignant tumor cell clones derived from KPC tumors (n = 17 clones), using data from Li et al(25). Clone labels in red or blue color indicate MHC II^high or MHC II^low clones used in subsequent studies.

Further validation was performed using 14 Visium spatial transcriptomics samples and matched single-nucleus RNA sequencing (snRNA-seq) data from the HTAN WUSTL atlas(40) and 55 Visium spatial transcriptomics samples from GSE277783(37). Using the Robust Cell Type Decomposition (RCTD)(51) algorithm, we transferred cell-type labels from snRNA-seq or scRNA-seq to spatial transcriptomic spots, deconvolving the cell-type composition at each spot. A representative image (specimen HT224P1) is shown in Fig. 2B, with additional examples in Supplementary Fig. 2. Gene analysis revealed heterogeneous MHC-II expression in malignant cells, as illustrated by representative images in Fig. 2C. Pathway enrichment analysis further corroborated these findings, as shown in Supplementary Fig. 3. To strengthen these observations, we performed additional validation using three independent spatial transcriptomics PDAC datasets obtained from SORC database(52), accompanied by corresponding histologic images, which consistently confirmed MHC-II gene expression within malignant cell spots. (Supplementary Fig. 4).

To validate the above transcriptomic data, we next utilized two in situ(53) Xenium gene expression datasets (10x Genomics) derived from human PDAC tissue sections of differing grades, using a pre-designed multi-tissue and cancer panel. This allowed for additional validation of MHC-II on tumor cells at the single-cell level. The two datasets encompassed 184,273 cells with 380 genes and 190,962 cells with 474 genes, respectively, captured the expected pancreatic cell types (Fig. 2D). By focusing specifically on malignant tumor cell populations, we observed substantial heterogeneity in MHC-II gene expression among cells (Fig. 2E), thereby highlighting the variability of MHC-II expression across malignant cell regions and aligning with our scRNA-seq findings.

To further validate MHC-II protein expression on PDAC cells, we performed multiplex immunohistochemistry(30,54) —a technique that enables simultaneous detection of multiple protein markers within a single tissue section—on resection samples from 11 PDAC patients who had received standard-of-care neoadjuvant chemotherapy, analyzing 195 regions spanning diverse tissue areas and collectively covering 568 mm². Representative images showed DAPI (blue), CK (magenta), and HLA-DR (yellow) staining (Fig. 2F). Quantitative analysis of HLA-II expression confirmed the presence of MHC-II protein in malignant PDAC cells, as illustrated by a representative image in Supplementary Fig. 5. Together, these human data provide compelling transcriptomic and proteomic evidence that MHC-II is variably expressed on malignant PDAC cells, supporting the rationale to investigate the tumor-intrinsic regulation of MHC-II in controlled preclinical models.

Given that patient data showed that malignant PDAC cells expressed MHC-II in therapy-naïve tumors, with further enhancement during cobimetinib treatment (Fig. 1D), we next sought to examine tumor-intrinsic and microenvironmental contexts associated with this program in preclinical PDAC models. To this end, we analyzed previously published transcriptomic data from sorted malignant clonal tumor cells derived from KPC mice, as reported by Li et al(25). In that study, tumor clones were grouped according to the immune contexture of the TMEs they generated. Unsupervised analysis revealed two distinct patterns of MHC-II-related gene expression across malignant tumor clones, with elevated MHC-II expression associated with clones that gave rise to T cell-enriched TMEs (Fig. 2G). Consistent with this, the baseline expression of the MHC-II pathway activity was increased in malignant tumor clones associated with highly T-cell-infiltrated TMEs compared with those associated with T cell-low TMEs via transcriptomic analysis (Fig. 2H).

These results collectively underscore the heterogeneity of MHC-II expression across PDAC cells in different regions of the TME and the potential connection of tumor cell-expressed MHC-II with immune infiltration to the tumor site.

Integrated human PDAC atlas enables sample-level analysis of tumor-cell MHC-II-associated immune contexture

To determine whether the relationship between tumor-cell MHC-II expression and immune infiltration extended beyond the paired clinical biopsy, we assembled an integrated pancreas single-cell atlas from public scRNA-seq datasets retrieved through the Deeply Integrated Single-Cell Omics (DISCO) database(55) using predefined filtering criteria. This framework yielded 22 datasets comprising 264 pancreatic samples and 1,042,407 cells (Supplementary Table 2). The resulting atlas captures a broad spectrum of pancreatic disease states, including normal pancreas, adjacent normal tissue, primary PDAC, and metastatic PDAC (Fig. 3A). For the analyses presented here, we focused on 115 PDAC samples, encompassing both primary and metastatic tumors and totaling 557,304 cells. These samples were used for sample-level analysis of malignant-cell MHC-II pathway activity in relation to immune-cell abundance across a large and heterogeneous cohort. Details of batch correction and malignant-cell identification are provided in the Methods.

Figure 3. Spatial analysis of MHC-II positive malignant cells and their immune contexture in PDAC.

A. UMAP of over 1 million single cells from the PDAC Atlas, categorized into epithelial, stromal, and immune compartments. B. Correlation between MHC-II pathway activity in tumor cells (557,304 cells) and immune cell combinations within the tumor site: scatter plots showing positive correlations between MHC-II pathway activity in malignant cells and the proportions of CD8+ T cells and CD4+ T cells (left), CD8+ T cells and NK cells (middle), and combined CD8+ T cells, CD4+ T cells, and NK cells (right). Spearman correlation analyses were performed with indicated ρ and p values. C. Illustration of spatial neighborhoods around malignant cells: defining a specified 150 μm radius around each seed cell (orange) to assess the surrounding cellular composition. D-E. Quantification of immune cells within malignant cell neighborhoods with differential MHC-II expression (Xenium, n=2 samples). Bar plots comparing the average number of T cells (D) and B cells (E) per core in MHC-II^high versus MHC-II^low malignant regions across two tumor samples: Stage IIB, Grade 3 (left) and Grade I–II with ~50% tumor content (right). Indicated malignant cell counts shown below each bar. Data analyzed by two-sided Wilcoxon rank sum test. Bars represent the mean, and error bars indicate the standard error of the mean (SEM). F. Quantification of CD4⁺ T helper cells and CD8⁺ T cells within malignant-cell neighborhoods across mIHC ROIs (n = 4 patients; n = 7 ROIs), comparing MHC-II^high and MHC-II^low malignant-cell neighborhoods. Indicated malignant-cell counts are shown below the bars. Statistical comparisons were performed using two-sided Wilcoxon rank-sum tests. Bars represent the mean, and error bars indicate the standard error of the mean (SEM). G. Representative example of Visium spatial transcriptomics data from the same tissue section shown in Figure 2B, demonstrating that malignant cell spots colocalized with T cells are enriched for MHC-II pathway activity. Circled regions indicate representative areas of malignant–T cell colocalization. H. Feature-plot of MHC-II pathway activation scores across the same section (HT224P1), further highlights spatial heterogeneity and preferential immune proximity (within 150 μm) to MHC-II^high malignant cell regions. Significance for statistical analyses indicated by *: p < 0.05, **: p < 0.01 and ***: p < 0.001. The same significance thresholds apply to all subsequent figures unless otherwise specified. Panel C was partially created using BioRender elements (Created in BioRender. Chen, C. (2026) https://BioRender.com/2on5daf).

In parallel, analysis of sorted tumor cells from PDAC-bearing mice showed that MHC-II pathway activity was significantly higher in malignant cells derived from TMEs with high T-cell content than in those from TMEs with low T-cell content (Fig. 2G).

To determine whether this relationship also held in human PDAC, we next performed a Spearman correlation analysis to assess the association between malignant-cell MHC-II pathway activity and the abundance of T-cell and natural killer (NK) cell lineages across patient samples. For this analysis, malignant cells from each sample were aggregated to generate a pseudo-bulk profile, which was then used to quantify sample-level MHC-II pathway activity and relate it to immune-cell infiltration (Materials and Methods). The analysis revealed a strong positive correlation between MHC-II activity in malignant cells and the abundance of T cells in the corresponding patient TME (Fig. 3B). Notably, combinations of immune cells, such as CD8⁺ and CD4⁺ T cells, and NK cells, and combinations involving CD8⁺ T cells, CD4⁺ T cells, and NK cells, showed particularly high correlations (Fig. 3B). Additionally, MHC-II pathway activity in malignant cells was strongly correlated with the proportions of individual lymphocytes and combinations thereof, while showing a negative correlation with endothelial cells and fibroblasts (Supplementary Fig. 6A).

Despite the overall trend, we identified several samples with high MHC-II pathway activity in the absence of detectable lymphocyte infiltration. Notably, 54 out of 57 such cases originated from the GSE202051 dataset(56), which was generated using snRNA-seq of frozen tumor samples. Because snRNA-seq is particularly effective at profiling malignant and stromal cells but exhibits reduced sensitivity for capturing lymphocytes(57,58), this likely accounts for the observed discrepancy. Nonetheless, the strong positive association between MHC-II pathway activity and immune infiltration across multiple independent datasets underscores the robustness of MHC-II pathway activity as a surrogate indicator of immune presence within the tumor microenvironment.

Lymphocytic spatial neighborhoods reveal distinct spatial landscapes associated with malignant cells exhibiting different MHC-II pathway activity

Tumor cell expression of MHC-II pathway activity has not yet been systematically investigated for its potential role in shaping intratumoral spatial heterogeneity of T cell infiltration. Patients with distinct immune and stromal infiltration patterns often experience markedly different clinical outcomes, particularly in their response to immunotherapy(59,60). Despite this, current treatment strategies largely overlook the unique composition and spatial organization of tumor-infiltrating immune cells, even though immune contexture, including CD4⁺ and CD8⁺ T cell infiltration, has demonstrated stronger prognostic value than traditional TNM staging(61).

Given the observed correlation between MHC-II pathway activity in malignant PDAC cells and immune infiltration, we applied the K-nearest neighbor (KNN) algorithm to explore whether malignant cells with high MHC-II level were spatially closer to specific lymphocytic lineages. To do so, we defined “neighborhoods” as physical groupings of cells within a set distance (50 mm) from each malignant cell, referred to as the “seed cell” (Fig. 3C) (Materials and Methods).

Within the Xenium spatial transcriptomic datasets, which demonstrated robust T cell infiltration at single-cell resolution, patient malignant tumor cells were stratified into two groups based on MHC-II pathway activity: MHC-II^high and MHC-II^low. Neighborhood composition analysis revealed that T cell abundance was significantly higher in the microenvironments surrounding MHC-II^high malignant cells compared to those adjacent to MHC-II^low malignant cells (Fig. 3D). This trend remained consistent after correcting for total cell number (Supplementary Fig. 6B). A similar enrichment pattern was observed for B cells, with increased representation around MHC-II^high malignant cells (Fig. 3E and Supplementary Fig. 6C).

The robustness of spatial patterns observed in the Xenium data was next validated through neighborhood analyses at the proteomics levels using mIHC analyses (Supplementary Fig. 5 and Fig. 3F) and in Visium data sets (Fig. 3G), yielding consistent results across multiple omics platforms and different data sets and patient cohorts. Notably, even in tumor slides with only a limited number of T cells detected via Visium spatial transcriptomics, the malignant cells that co-localized with T cells were predominantly those exhibiting the highest levels of MHC-II gene expression within the tumor section (Fig. 3G – H).

Given the established role of plasma cells in modulating immune responses and their potential contribution to antitumor immunity in PDAC(37,62), we next sought to explore whether spatial proximity to malignant cells with high MHC-II activity might extend beyond T cells to include B cells and specifically, plasma cells. Notably, although the overall proportion of plasma cells did not exhibit a strong correlation with MHC-II pathway activity in malignant cells (Supplementary Fig. 6A), CellTrek analysis(63) further revealed that plasma cell–enriched regions were in significantly closer proximity to MHC-II^high malignant cell spots compared to MHC-II^low malignant cell spots across all Visium samples that had matched snRNA-seq data (Supplementary Fig. 7). This spatial association indicates a preferential interaction between plasma cells and MHC-II^high malignant cell regions, emphasizing the spatial heterogeneity of the TME and underscoring a consistent association between tumor-cell MHC-II activity and spatially organized immune features in PDAC.

CAF support for hypoxic environments in MHC-II^low malignant cell niches

Cancer-associated fibroblasts (CAFs) are one of the most abundant cell populations in the TME and have garnered significant attention for their multifaceted roles in tumor progression(64,65). Through their interactions with stromal components and immune cells, CAFs drive key processes such as angiogenesis, extracellular matrix (ECM) remodeling, and immune evasion(66).

Fibroblasts proportion was found to exhibit a negative correlation with MHC-II pathway activity in malignant cells (Supplementary Fig. 6A). To further dissect the complex spatial architecture of the TME, we applied nine distinct CAF subtype gene signatures defined in Cords’ research(64), classifying CAFs based on functional and phenotypic diversity, including vascular CAFs (vCAFs), matrix CAFs (mCAFs), interferon-responsive CAFs (ifnCAFs), tumor-like CAFs (tCAFs), inflammatory CAFs (iCAFs), dividing CAFs (dCAFs), reticular-like CAFs (rCAFs), and antigen-presenting CAFs (apCAFs).

Spatial analysis of these CAF subtypes across primary tumors and metastatic lesions (liver, lung, and peritoneum) revealed distinct patterns of co-localization with malignant cells exhibiting varying MHC-II pathway activity (Supplementary Fig. 8). Notably, apCAFs and iCAFs preferentially co-localized with MHC-II^high malignant cell spots in the PDAC TME. In particular, apCAF signature scores were consistently elevated in CAF spots adjacent to MHC-II^high malignant cell regions across nearly all samples. Given their capacity for antigen presentation, apCAFs may contribute to lymphocyte modulation in the PDAC TME(67). iCAFs, known for secreting pro-inflammatory cytokines, may similarly support immune modulation in the TME(64). Conversely, mCAFs and tCAFs were enriched around MHC-II^low malignant cell regions. mCAFs specialize in extracellular matrix production and remodeling—processes integral to structural support and tumor progression—while tCAFs were associated with tumor-like phenotypes, metabolic adaptation, and survival signaling.

Importantly, the spatial associations between CAF subtypes and malignant cell MHC-II status were consistent across metastatic sites, sexes, and distinct PDAC tumor subtypes in our analyses (including classical, intermediate, and basal) (Supplementary Fig. 8), indicating that these interactions are governed by a complex interplay between intrinsic tumor programs and microenvironmental cues.

While previous studies have reported that malignant cells associated with myCAFs contribute to the exclusion of plasma cells from the TME—a phenomenon linked to poor outcomes and resistance to chemotherapy—these findings have largely been limited to basal-like PDAC tumors(37,68). In contrast, our results reveal that MHC-II^high tumor cells reside within a more immunologically active niche, characterized by increased spatial proximity to various immune cell types, including T cells, B cells and plasma cells, as well as pro-inflammatory and immune-supportive CAF subtypes such as antigen-presenting CAFs and inflammatory CAFs, while exhibiting spatial exclusion from tumor-promoting CAF subsets (myCAFs and tCAFs). These findings indicate that tumor-cell MHC-II pathway activity is associated with distinct spatial immune–stromal interaction patterns in the PDAC TME: MHCII^high tumor cells correlate with a more permissive immune microenvironment at baseline, independent of classical, intermediate or basal-like PDAC subtype classifications.

Tumor-cell MHC-II identifies baseline immune infiltration and is inducible under IFN-γ-permissive conditions

Given the observed association between tumor cell MHC-II activity and lymphocyte infiltration, we next asked whether MHC-II expression could serve as a predictive biomarker for therapeutic responsiveness, particularly in the context of immunotherapy. Thus, we developed a model for predicting MHC-II activity to assess malignant cell responsiveness to drug treatment using the MHC-II pathway signature gene list. The model successfully distinguished sample heterogeneity (Fig. 4A) and demonstrated robust predictive performance, as evidenced by its strong correlation with lymphocyte infiltration. Specifically, the percentage of predicted responsive malignant cells showed highly positive associations with immune cells capable of producing the antitumor cytokine interferon (IFN)-γ. This included combinations of CD8⁺ and CD4⁺ T cells, as well as broader interactions involving CD8⁺ T cells, CD4⁺ T cells, and NK cells (Fig. 4B). Detailed correlations for individual cell types and other combinations are provided in Supplementary Fig. 9.

Figure 4. Tumor-cell MHC-II identifies baseline immune infiltration and is inducible under IFN-γ-permissive conditions.

A. Single-cell analysis of MHC-II pathway activity in malignant cells from the patient PDAC atlas (557, 304 cells). Bar plot depicting predicted response status based on MHC-II expression levels across different PDAC tumor samples. B. Correlation of MHC-II^high malignant cells with immune cell infiltration: scatter plots illustrating positive correlations between the proportion of MHC-II^high malignant cells and immune cell combinations, including left: CD8⁺ and CD4⁺ T cells and right: CD8⁺ T cells, CD4⁺ T cells, and NK cells combined. Data analyzed by Spearman correlations, ρ and p values shown. C. Experimental design for in vivo flow cytometric profiling of the tumor microenvironment using syngeneic MHC-II^low (6419c5) and MHC-II^high (7160c2) PDAC clones. D.Percentage of MHC-II⁺ malignant cells (left) and mean fluorescence intensity (MFI) of MHC-II on malignant cells (right) in MHC-II^low (6419c5) versus MHC-II^high (7160c2) tumor clones. E. Frequency of CD3⁺ T cells among CD45⁺ immune cells in 6419c5 and 7160c2 tumor clones. For D-E, Cohorts of n=4-6 mice were used. Data shown representative of three independent experiments. F. Schematic of in vitro stimulation with IFN-γ and cobimetinib to induce MHC-II expression, followed by flow cytometric analysis. G. MHC-II expression on tumor cells with varying MHC-II pathway levels following treatment with IFN-γ and cobimetinib in vitro. Histogram overlays showing MHC-II expression in tumor cells categorized as MHCII^low (6419c5, 6694c2, 2699c4) and MHCII^high (7160c2, 6555c3, 7160c5), also shown above and in right hand corner. Tumor cells were treated with IFN-γ (blue), cobimetinib (Cobi, red), or a combination of both (Cobi/IFN-γ, purple), with untreated control cells shown in gray. Data are normalized to mode to illustrate the shift in MHC-II expression across conditions. H. Experimental design for evaluating the impact of cobimetinib on MHC-II expression in orthotopic PDAC tumors. Mice were orthotopically implanted with either MHC-II^low or MHC-II^high PDAC clones and allowed to establish tumors for 7–14 days. Mice were then treated with either vehicle or cobimetinib for 7 days. I-K. Flow cytometric analysis of orthotopically injected tumor cells among live Cd45⁻YFP⁺ cells (I-J) or TME-T cells gated as CD3⁺CD45⁺live cells (K) +/− cobimetinib treatment from indicated MHC-II^high tumors (7160c2, red) or MHC-II^low tumors (6419c5, blue). L. Waterfall plots showing the percent change in tumor volume after 14 days of treatment as compared to baseline in individual mice treated with either control or cobimetinib in MHC-II^low (6419c5, blue) or MHC-II^high (7160c2, red) tumors, as shown. For I-L, n=5-6 mice per cohort were used, data representative of n=2 independent experiments. Each dot (D, I-K) or bar (L) represents an individual mouse, error bars indicate S.D., for (I-K), horizontal bars indicate the mean (D, I-K), and whiskers indicate the upper and lower quartiles (D). Analysis by unpaired t test (D), where ** indicates p<0.01. Samples in (G) analyzed in triplicate, representative of n=2-6 independent experiments.

To examine the relationship between tumor cell-intrinsic MHC-II expression, IFN-γ signaling, and treatment response, we next turned to preclinical mouse models of pancreatic cancer comparing clonal malignant cells with differing baseline MHC-II expression. In vitro and in vivo experiments were conducted using pancreatic orthotopic models with these clonal tumor cells(25). Tumor cells expressed the YFP lineage tag (KPCY), enabling tracking of tumor behavior and treatment responses. This model harbors the heterozygous Trp53 mutation (Trp53^R172H) and recapitulates key histopathological and molecular characteristics of human pancreatic ductal adenocarcinoma(25).

Flow cytometric analysis of PDAC TMEs (Fig. 4C–D) revealed that tumor cells with high MHC-II activity score from Fig. 2H (7160c2) exhibited a significantly higher proportion of MHC-II-positive malignant cells compared to tumor cells with a low MHC-II activity score (6419c5) (Fig. 4D). Correspondingly, tumor cells with elevated MHC-II activity in vivo (Fig. 4D) demonstrated significantly increased CD3⁺ T cell infiltration at baseline (Fig. 4E), in concordance with our earlier findings on ‘T cell high’ TMEs and supporting the role of MHC-II in fostering an immunologically active tumor site.

Given this positive correlation between MHC-II expression and lymphocyte infiltration, we next explored whether cobimetinib could induce MHC-II expression in MHC-II^low tumor cells as a means to potentially augment lymphocyte infiltration and thereby immunotherapeutic response. In vitro, flow cytometry showed that MHC-II expression was most strongly increased following cobimetinib – but only when combined with recombinant IFN-γ, particularly in MHC-II^low tumor cell lines (Fig. 4F–G). These data suggested MEK inhibition may enhance MHC-II upregulation under permissive cytokine conditions rather than demonstrating a uniform direct effect of cobimetinib alone.

To evaluate this relationship in vivo, mice bearing MHC-II^low or MHC-II^high PDAC tumors were treated with vehicle or cobimetinib, followed by flow cytometric analysis of tumor-cell MHC-II expression and T-cell infiltration (Fig. 4H). In MHC-II^low, but not MHC-II^high tumors, cobimetinib treatment induced a trend toward increased MHC-II expression on malignant PDAC cells (Fig. 4I–J). As expected from the baseline immune state, MHC-II^high tumors contained more T cells than MHC-II^low tumors (Fig. 4K). Cobimetinib treatment also reduced tumor volume in mice bearing MHC-II^low tumors, consistent with the idea that MHC-II^low tumors retain the capacity for therapeutic reprogramming when MAPK signaling is inhibited (Fig. 4L). Together, these findings indicate that tumor-cell MHC-II expression at baseline correlated with immune infiltration, but tumor cell expression of MHC II could potentially be modified in certain therapeutic contexts.

IFN-γ is required for cobimetinib-induced expression of MHC-II by tumor cells in vitro and therapeutic efficacy in vivo

Given that the baseline expression of MHC-II was higher on tumor cells from TMEs enriched for T cells (Fig. 4K), that we have shown that T cells are a source of IFN-γ in the PDAC TME(26), and that MHC-II^low tumor cell lines exhibited a significant upregulation of MHC-II in vitro after treatment with both cobimetinib and IFN-γ (Fig. 4G), we hypothesized that therapeutically promoting effector T cell function such as IFN-γ production, would potentiate the therapeutic benefits of cobimetinib treatment. Indeed, we find that in vitro, the expression of the IFN-γ receptor (IFN-γR, CD119) is downregulated on MHC-II^high or MHC-II^low tumor cells upon addition of rIFN-γ alone or with cobimetinib (Fig. 5A), indicating tumor cells are responding to rIFN-γ stimulation. In the absence of IFN-γR expression, we find that MHC-II^low tumor cells do not upregulate MHC-II, even when stimulated with cobimetinib and rIFN-γ in vitro (Fig. 5B), despite significant upregulation in parental (IFN-γR-sufficient) tumor cells (Fig. 2H). This suggested to us that tumor cell-intrinsic IFN-γ signaling in vivo would impact the efficacy of combined cobimetinib therapy.

Figure 5.

IFN-$\mathit{\gamma }$ signaling is required for malignant cell upregulation of MHC-II, and therapeutic efficacy of cobimetinib + dual ICB. A-B, Effect of cobimetinib +/− rIFN-γ treatment on IFN-γ (CD119) expression in vitro on indicated MHCII^low or MHC-II^hi malignant clonal cells that are IFN-γR sufficient (A) or deficient, with corresponding MHC-II expression (B). (C) experimental schema for in vivo treatment using cobimetinib + dual ICB. Experimental design: Mice were subcutaneously implanted with MHC-II^low or MHC-II^high PDAC clones (6419c5 or 7160c2 clones, respectively) and allowed to establish tumors (~50 mm²) over 7–14 days. Mice were then treated with cobimetinib in combination with either dual isotype controls or dual immune checkpoint blockade (ICB; anti-PD-1 and anti-CTLA-4) and monitored for tumor growth over time. Mice bearing the indicated PDAC clones shown in (D), or with mice bearing 6419c5 clonal tumor cells that lack IFN-γR expression (E), were treated as shown in (C). Tumor growth kinetics in (E) were run simultaneously but separated for ease of visualization. Cohorts of n=4-5 mice per group were used in D, data from one independent experiment, while cohorts of n=5-8 mice per group were used in E, data representative of one independent experiment. Schematic for testing individual therapeutic arms shown in (F) for indicated tumors injected subcutaneously (G) or orthotopically (H). Cohorts of n=5 mice per group were used in G, with data representative of three independent experiments, while cohorts of n=3-5 mice per group were used in (H), data representative of one independent experiment. Experimental design: Mice were implanted with MHC-II^low PDAC clones (6419c5 or 2669c4) and allowed to establish tumors. Mice were then treated with cobimetinib in combination with either dual isotype controls or dual immune checkpoint blockade (ICB; anti-PD-1 and anti-CTLA-4) and monitored for tumor growth over time (G) or survival (H). Each dot represents the mean of the cohort (D, E, G) or an individual mouse (H), error bars indicate S.E.M. Analysis by two-way ANOVA with mixed linear regression with Tukey’s post-test for indicated groups (D, E, G) or Log rank test (H). ** indicates p<0.01, *** indicates p<0.0001, ns indicates not significant. Data in A-B performed in triplicate and representative of n=1-3 independent experiments.

Knowing that dual immune checkpoint blockade (ICB: both anti-PD-1 and anti-CTLA-4) increases IFN-γ by T cells within the PDAC TME(26), we tested the therapeutic combination of cobimetinib with dual ICB (Fig. 5C). We observed that mice treated with cobimetinib and ICB exhibited substantially reduced tumor growth as compared to compared to vehicle + isotype control treated mice across both MHC-II^high and MHC-II^low tumor bearing mice (Fig. 5D). Although MHC-II^high tumor clones showed partial MHC-II upregulation after cobimetinib stimulation in vitro (Fig. 4G) and did not exhibit further upregulation of MHC-II gene expression following cobimetinib treatment in vivo (Fig. 4I), mice bearing tumors from MHC-II^high PDAC clones still responded to cobimetinib + ICB therapy (Fig. 5D). However, the most pronounced therapeutic benefit of combination cobimetinib + ICB was observed in mice bearing MHC-II^low tumors, where mice had significantly reduced tumor sizes as compared to mice treated with vehicle controls (Fig. 5D). As most PDAC TMEs are treatment-resistant and generally lack T cells (25), this finding indicates that baseline MHC-II activity and sensitivity to IFN-γ in malignant cells may serve as a predictive and prognostic biomarker for patient responses to immunotherapy(69).

Additionally, these data align with our observation that in the correct contexts, MHC-II^low tumor cells strongly upregulated MHC-II expression. Here, we observed MHC II upregulation after MEK inhibition and IFN-γ stimulation (Fig. 4G, 4I) and are associated with therapeutic sensitivity when ICB was in vivo used in lieu of IFN-γ (Fig. 5D). In recent reports, tumor cell-intrinsic IFN-γ signaling is not required for treatment efficacy when using ICB in combination therapies for PDAC (29,70). Here, we observed that the therapeutic efficacy of combined cobimetinib + ICB is completely lost in mice bearing IFN-γR deficient, MHC-II^low tumor cells (Fig. 5E, Supplementary Fig. 10A). Additionally, mice that lack T cells, especially MHC-II-reactive CD4 T cells, do not show slowed tumor growth after treatment with cobi + ICB (Supplementary Fig. 10B). These data support the notion that the IFN-γ pathway – which a cytostatic but not cytotoxic – is required for effective tumor control after MEK inhibition, and improved therapeutic response could be detected via increased expression of MHC II by the tumor cells.

The MHC-II^low tumor clone 6419c5, along with many of the other MHC-II^low clones from Fig. 2H, are known to be highly therapeutically resistant(25). Given that 6149c5 was sensitive to the combination of cobimetinib + ICB, we hypothesized that sensitivity was uniquely due to the combination treatment and not due to individual ‘single arms of the therapy (outlined in Fig. 5F). We found that indeed, the single therapies were ineffective while cobimetinib + ICB significantly reduced tumor volume and improved survival relative in both subcutaneous and orthotopic models of MHC-II^low tumors (Fig. 5G–H, Supplementary Fig. 10C–E). In contrast to the significantly improved outcomes after treatment in mice bearing MHC-II^low tumors, combination cobimetinib + ICB therapy yielded minimal benefit over ICB alone in MHC-II^high tumors (Supplementary Fig. 11A–B), consistent with the lack of MHC-II expression change in these tumor cells observed following cobimetinib treatment (Fig. 4I–J). These data indicate a marked therapeutic response accompanied by induction of MHC-II expression in previously therapy-resistant, MHC-II^low tumor clones where MHC-II levels are permissive to treatment modulate. Tumor-cell IFN-γR expression was required for MHC II modulation and therapeutic benefit, and furthermore, MHC-II expression correlated with therapeutic outcomes. These data suggest MHC-II expression may be a key predictor of immunotherapy responsiveness in PDAC: TMEs with baseline MHC-II^high tumor cells mounted more effective immune-mediated tumor control, but TMEs with MHC-II^low tumor cells overcame resistance to immune checkpoint therapy when MEK inhibition is enforced.

Discussion

Pancreatic ductal adenocarcinoma remains one of the deadliest malignancies, characterized by its dismal prognosis and highly T cell-suppressive tumor microenvironment. The complexities associated with PDAC, such as its extensive stromal component and heterogeneous immune landscape(71), necessitate a deeper understanding of how immune dynamics influence tumor progression and therapeutic response. By integrating single-cell RNA sequencing, spatial transcriptomics, bulk transcriptomics, multiplex immunohistochemistry, and mechanistic studies in mouse PDAC models, our work positions malignant-cell MHC-II as both a readout of tumor immune state and an indicator of treatment sensitivity.

A central finding of this study is that elevated malignant-cell MHC-II expression is consistently associated with a more immune-active TME in human PDAC. Across datasets and platforms, tumors enriched for MHC-II-expressing malignant cells showed greater CD4⁺ and CD8⁺ T-cell infiltration, supporting the idea that tumor-cell MHC-II marks a state that is permissive to adaptive immune engagement. This extends observations from melanoma(23), non-small cell lung cancer (NSCLC)(72), breast cancer(73) and colorectal cancer(74) to PDAC, a disease classically considered poorly immunogenic and highly resistant to T cell-mediated immunotherapy. In this context, tumor-cell MHC-II may serve as an indicator of whether malignant cells reside within an immune-inflamed niche.

At the same time, our analyses highlight that this biology cannot be fully captured by dissociative single-cell profiling alone. We observed substantial intratumoral heterogeneity, with MHC-II^high and MHC-II^low malignant niches coexisting within the same tumor in both human and mouse samples. Although conventional scRNA-seq analyses showed limited correlation between plasma-cell abundance and MHC-II pathway activity, spatial transcriptomic profiling revealed pronounced spatial proximity between plasma cells and MHC-II^high malignant regions. These findings underscore the importance of tissue context and suggest that localized immune–malignant cell interactions may be obscured when tissue architecture is disrupted(75). More broadly, they reinforce the value of spatially resolved approaches for studying PDAC, where ecological organization within the TME is likely to be as important as cell-state composition itself.

Our data also refine this ecological view by linking malignant-cell MHC-II states with distinct fibroblast-rich microenvironments. Pro-inflammatory CAF subtypes, particularly apCAFs and iCAFs, were preferentially associated with MHC-II^high malignant niches, whereas mCAFs and tCAFs, which are more commonly linked to tumor progression and stromal restraint of immunity, were enriched around MHC-II^low malignant regions. These patterns support a model in which malignant-cell MHC-II is not an isolated tumor-cell feature, but rather part of a broader spatially organized immune-stromal state. In this sense, tumor-cell MHC-II may serve as an integrative marker of niche identity, reflecting how malignant cells, fibroblasts, and immune populations collectively shape local tumor behavior(76).

Importantly, our study suggests that tumor-cell MHC-II is not only a baseline marker of immune contexture, but also a dynamically regulated program during therapy. In matched pre- and on-treatment clinical biopsies from a patient treated with cobimetinib, malignant cells displayed increased expression of MHC-II pathway genes during therapy, indicating that this program can be induced in the clinical setting. This observation was mirrored in preclinical KPC models, where MHC-II^high tumors exhibited greater T-cell infiltration at baseline, whereas selected therapy-resistant MHC-II^low tumors showed therapeutic benefit after cobimetinib plus immune checkpoint blockade. Together, these findings support a model in which MHC-II^low tumor responses under the tested combination treatment are accompanied by induction of the MHC-II program. This distinction is important because baseline MHC-II status may capture the current immune state of the tumor, whereas inducibility of the MHC-II program may reflect therapeutic opportunity.

The mechanistic experiments further strengthen this framework. First, genetic deletion of IFN-γR in mouse PDAC cell lines abolished tumor cell MHC-II induction in vitro, demonstrating that the MEK inhibition/MHC-II program is dependent on intact IFN-γ signaling. Second, MEK inhibition caused a slight reduction of IFN-γR expression in vitro, supporting the idea that MAPK pathway activity restrains tumor-cell responsiveness to IFN-γ. In vivo, IFN-γR expression was required on tumor cells for the therapeutic efficacy of Cobi + ICB in MHC-II^low TMEs – as were T cells, a putative source of IFN-γ after ICB(26). These findings provide direct experimental support for a model in which oncogenic MAPK signaling suppresses an IFN-γ-responsive antigen-presentation program, whereas MEK inhibition relieves this restraint and sensitizes tumor cells to inflammatory cues in the TME. Rather than acting as a uniformly sufficient inducer on its own, MEK inhibition appears to promote MHC-II activation in part by enhancing the capacity of tumor cells to respond to microenvironmental IFN-γ.

Despite these advances, several limitations remain. First, MHC-II expression was heterogeneous across datasets and platforms, and technical differences, particularly in single-nucleus datasets, likely influenced recovery of immune populations and quantification of tumor-cell states(67,68). Standardized approaches for measuring malignant-cell MHC-II in clinical samples will therefore be important for future translational application. Second, although our data support an IFN-γ-dependent and MAPK-restrained model, additional regulatory layers almost certainly contribute to variability in MHC-II expression, including epigenetic state, tumor evolutionary history, and differences in cytokine exposure across tumor niches. Third, because we did not directly perturb CIITA or MHC-II in the tumor cells, our data do not establish a causal role for tumor-cell MHC-II in therapeutic response. MHC-II presents peptide antigens to CD4+ T cells and does not itself transmit a tumor-cell signaling response; therefore, any functional contribution of tumor-cell MHC-II would be expected to involve CD4+ T-cell recognition(22,23). In the present study, combined CD4/CD8 depletion prevented tumor control, but this design does not isolate the contribution of CD4+ T cells or tumor-cell antigen presentation. We and others have previously shown that CD4+ T cells are required for immunotherapeutic response to pancreatic tumors in multiple different therapeutic combinations(29,70). However, this combination of MEK inhibition and ICB is the first to show that IFN-γR signaling in tumor cells is required for treatment efficacy, and that therapeutic response correlates with tumor cell expression of MHC II. Future experiments using CIITA-deficient tumors will reveal if tumor-cell MHC-II-mediated antigen presentation contributes to specific therapeutic contexts.

Our study also has potential clinical implications. Incorporating malignant-cell MHC-II expression, or the inducibility of the MHC-II program, into translational workflows may improve patient stratification for rational immunotherapy combinations. The present findings support evaluating tumor-cell MHC-II as a readout of local IFN-γ-responsive immune activity rather than as a therapeutic target. Future clinical studies will be needed to determine whether tumor-cell MHC-II can serve as a predictive biomarker of response to MEK inhibitor-based immunotherapy combinations in patients with PDAC.

In summary, our study identifies tumor-cell MHC-II as an immune-associated and therapeutically inducible program in PDAC. Elevated malignant-cell MHC-II marks immune-inflamed tumor niches, while MHC-II^low tumors can be reprogrammed under conditions of MEK inhibition and IFN-γ signaling. Tumor-cell IFN-γR is required for MHC-II induction and for therapeutic benefit in MHC-II^low models treated with cobimetinib plus immune checkpoint blockade. Together, these findings support a model in which tumor-cell MHC-II reflects both baseline immune contexture and inducible therapeutic vulnerability, positioning this program as a promising candidate biomarker for rational combination immunotherapy in PDAC.

Synopsis:

The authors identify tumor-cell MHC-II as an IFN-γ-dependent, MAPK-restrained program that marks immune-active tumor states and immunotherapy responsiveness in pancreatic cancer, establishing tumor-cell MHC-II as a candidate biomarker for therapeutic sensitivity and combination immunotherapy in pancreatic cancer.

Data availability.

Single-cell RNA sequencing data generated in this study from paired pre- and on-treatment tumor biopsies (patient enrolled in NCT04005690) have been deposited in dbGaP under accession phs003597. Publicly available datasets used in this study include the Xenium spatial transcriptomics datasets from 10x Genomics, and Visium samples from the HTAN WUSTL atlas, which are accessible via the Human Tumor Atlas Network (HTAN) dbGaP Study Accession: phs002371.v1.p1 (https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs002371.v1.p1), as well as through the Gene Expression Omnibus (GEO) under accession number GSE277783. Additional spatial data were obtained from the SORC(52) database. Cancer cell line expression data were retrieved from the Cancer Cell Line Encyclopedia (https://sites.broadinstitute.org/ccle/datasets). The human genome reference (GRCh38, version 4.0.0) used for single-cell analyses was obtained from public sources, with genome builds and annotation files provided by 10x Genomics (https://support.10xgenomics.com/single-cell-gene-expression/software/release-notes/build). For mouse transcriptomic analyses, the GRCm38 reference genome and GENCODE vM25 annotation were downloaded from the GENCODE portal (https://www.gencodegenes.org/mouse/). The mIHC-derived single-cell phenotyping and spatial localization data have been deposited in Synapse and are accessible at https://www.synapse.org/#!Synapse:syn51078766.

Code availability.

All packages used in this study are publicly available. The data, scripts, and code used for data preprocessing, statistical analyses, and figure generation are also available through a Code Ocean Compute Capsule (https://codeocean.com/capsule/5765946/tree/v1; provisional DOI: 10.24433/CO.8181629.v1).