Asynchronous Evolution of Epithelium and Stroma Differentiates Precursor Lesions from Pancreatic Cancer

Spatial analysis of human donor pancreas identifies asynchronous changes in the epithelium and microenvironment of pancreatic precancerous lesions.

Abstract

Pancreatic intraepithelial neoplasia (PanIN) precedes pancreatic cancer, a deadly disease characterized by an extensive tumor microenvironment. How the microenvironment evolves during cancer progression is largely unknown, as PanINs are microscopic and nondiseased pancreas samples are exceedingly rare, while adjacent normal samples are disrupted by the presence of malignancy. Leveraging donor organs and spatial technologies, we mapped the evolution of PanIN to cancer. The PanIN epithelial component falls on a continuum with cancer, whereas the PanIN microenvironment is drastically distinct. Progression to cancer is accompanied by profound geographic reorganization of myeloid cells and lymphocytes and the formation of a cancer-specific fibroblast population characterized by high levels of smooth muscle actin, LRRC15, and the WNT signaling component LEF1. Together, our data show asynchronous evolution of epithelial and stromal components during pancreatic carcinogenesis. Lack of stromal reprogramming might explain why most PanINs do not progress to cancer. Compiled data available at https://www.pancreas-precancer-atlas.com/.

Significance:

Pancreatic cancer precursor lesions are widely prevalent in the population, but progression to invasive cancer is uncommon. Identifying barriers to cancer progression is critical for cancer interception. Our findings describe unique features of the precancer microenvironment that are likely responsible for keeping precursor lesions from progressing to malignancy.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is one of the most lethal cancers, often diagnosed at late stages because of the asymptomatic nature of early disease and the absence of routine screening and early detection tools [for review see (1)]. Two main types of precursor lesions give rise to pancreatic cancer in humans. Cystic lesions, most often classified as intrapapillary mucinous neoplasms (IPMN), can be visualized with abdominal imaging techniques; pancreatic intraepithelial neoplasia (PanIN), the most common precursor to pancreatic cancer [for review see (2)], is microscopic and cannot be detected noninvasively. At the genetic level, more than 90% of human pancreatic cancer samples are associated with oncogenic mutations in the KRAS gene; mutant KRAS is also a feature of nearly all human PanIN lesions (3). Most of our understanding of human PanIN lesions derives from the study of tumor-adjacent samples, which might harbor field defects caused by proximity to cancer and typically are collected from older individuals (4). Through our collaboration with Gift of Life Michigan, we procured healthy donor pancreata from more than 150 individuals across age ranges (20–70 years). In initial analysis of 30 pancreata, we discovered that more than 60% of the healthy population harbors PanIN lesions regardless of age (5), contrary to the long-held belief that PanINs are rare and accumulate with age (6). The observation that PanIN are extremely common implies that progression to malignancy is rare.

Pancreatic cancer is associated with an extensive tumor microenvironment (TME) that constitutes up to 90% of the tumor mass [for review see (2)]. Fibroblasts and immune cells are key components of the TME and play complex roles in the biology of pancreatic cancer. Although a large body of literature has dissected the nature and functional status of TME components in pancreatic cancer, our understanding of the early stages of carcinogenesis and TME formation relies on genetically engineered mice, due to the paucity of available human samples. Although mouse models are an essential tool for basic discovery and preclinical work, mouse PanINs differ in fundamental ways from their human counterpart. Mice very rarely develop spontaneous PanIN, even in old age; furthermore, in most available models, oncogenic KRAS—a hallmark mutation in pancreatic cancer—is activated during embryonic development of the pancreas (7, 8). Even in models that allow adult activation of oncogenic KRAS (9, 10), expression is driven simultaneously in every epithelial cell of the pancreas, which in turn might cause field effects that are not observed in human spontaneous PanIN. Finally, mice are genetically homogeneous and are maintained in a constant and highly controlled environment throughout their lifespan, in sharp contrast with the constant change that characterizes human life. To understand the nature of human PanIN and map the evolution of the microenvironment surrounding PanINs, it is essential to leverage human samples.

Human PanIN lesions, like their mouse counterpart, are surrounded by an accumulation of fibroblasts, immune cells, and extracellular matrix (ECM). The nature of the cells composing the PanIN microenvironment and their specific characteristics remain unknown. A key challenge to studying PanIN lesions and their microenvironment is the microscopic nature of lesions, which cannot be dissected and studied in isolation from surrounding healthy tissue. Previous studies of human PanIN largely relied on “adjacent normal” (4, 11–15). This poses a challenge as adjacent normal tissue is not, in fact, normal but is disrupted by cancer-driven field effects such as extensive inflammation (5, 16). To obtain sporadic PanIN in otherwise healthy pancreata, we leveraged our unique collaboration with Gift of Life Michigan, which allowed us to accrue more than 150 donor pancreata; we also leveraged the pancreatic cancer biobank at the University of Michigan. Together, we accrued a comprehensive cohort of samples that enabled direct mapping of the epithelial and stromal components during the progression from indolent PanIN to cancer.

Given that PanIN form a microscopic portion of otherwise healthy organs, single-cell RNA sequencing (scRNA-seq) is not suitable for analyzing their evolution. To specifically analyze the transcriptional status of PanINs and their microenvironment, we employed spatial transcriptomics, together with histopathologic annotation of epithelial structures, followed by protein validation. Our samples encompassed nonneoplastic epithelial cell types, such as acinar and ductal cells, and partially dedifferentiated acinar clusters co-expressing acinar and ductal markers, which we refer to as acinar–ductal metaplasia (ADM), a response to tissue damage that is usually transient. Donor samples also contained sporadic PanIN, and tumor samples contained PanIN, areas of well-differentiated or glandular tumor (GLTumor), and areas of poorly differentiated tumor (PDTumor). In addition, though not the focus of the current study, our sections included endocrine islets. We discovered that PanIN epithelial domains are distinct from normal pancreas domains (both acini and ducts) and that they fall on a continuum with neoplastic samples and are closest to GLTumor, although clear differences exist. Key characteristics of malignancy, such as upregulation of signaling in RAS/MAPK, JAK/STAT3, and other inflammatory pathways were slightly upregulated in PanIN compared with normal pancreas structures, higher in GLTumor, and drastically higher in PDTumor. In contrast, the microenvironments of PanIN and invasive cancer differed markedly. The PanIN microenvironment was abundantly infiltrated by plasma cells, whereas macrophages tended to be located away from epithelial cells; this spatial organization was reversed in tumor samples, in which plasma cells were scattered and macrophages were near tumor epithelium. Furthermore, tumor samples were associated with the de novo appearance of a fibroblast population characterized by high expression of smooth muscle actin (SMA), LRRC15, and LEF1, a WNT signaling component and target gene. We utilized 3D cocultures and observed that cancer organoids drive LEF1 expression in cancer-associated fibroblasts (CAF), but not in fibroblasts from the normal pancreas, functionally validating our key findings. Together, our data show that the PanIN microenvironment lacks key characteristics of the TME, a finding that might in part explain the lack of progression of the vast majority of PanIN to cancer.

Results

Integration of Single-Cell RNA-seq and Spatial Transcriptomics Defines Tumor-Specific Domains

To overcome the limitation posed by the scarcity of PanIN within a pancreas, we developed a spatial domain identification workflow that integrates scRNA-seq and spatial transcriptomics using spatially informed clustering and spot deconvolution along with histologic annotation (Fig. 1A). Utilizing this workflow, we compared the transcriptional status of epithelial and stromal compartments of PanIN and pancreatic cancer (Fig. 1B). To validate our findings, we utilized independent Visium datasets, publicly available Xenium datasets, The Cancer Genome Atlas (TCGA) pancreatic adenocarcinoma (PAAD) dataset, donor and PDAC tissue immunostaining, and tumor organoid/fibroblast cocultures (Fig. 1C).

Figure 1.

Overview of study design and sample acquisition. A, Schematic representation of the analysis workflow. Healthy donor pancreas and treatment-naïve PDAC samples were profiled using 10x Genomics scRNA-seq and Visium CytAssist. An scRNA-seq atlas including donor pancreas, cancer, and adjacent normal (AdjNormal) samples was generated following data processing, integration, and CNV inference. A combination of cell-type deconvolution, joint spatially informed clustering, and pathologist annotation was used to help define different spatial domains. B, We traced the neoplastic epithelial compartment evolution from healthy donor–associated PanINs to cancer. We next investigated the surrounding stroma co-evolution, including immune cells and fibroblasts. C, We validated our findings utilizing publicly available Xenium and CosMx datasets, TCGA-PAAD, and an independent set of 6 Visium samples. We then validated our findings at the protein level using IF staining of donor PanIN and cancer samples. Finally, we conducted functional validation using tumor organoid–fibroblast cocultures. D, Hematoxylin and eosin (H&E) staining showing the capture areas for the healthy donor samples profiled using 10x Visium. SP, Spatial. E, H&E staining showing the capture areas for the PDAC samples profiled using 10x Visium.

To compare transcriptional features of sporadic PanIN and pancreatic cancer, we performed spatial transcriptomics using the 10x Genomics Visium platform on 11 PanIN-containing donor samples, 7 pancreatic adenocarcinoma samples, and 2 adjacent normal pancreas samples from patients with cancer (Fig. 1D and E; Supplementary Fig. S1A; Supplementary Table S1). All samples were annotated by board-certified gastrointestinal pathologists (J. Shi, J.M. Szczepanski, and A. Maitra). Prior to analyzing spatial transcriptomics, we generated a reference scRNA-seq atlas of both previously published (12, 17) and new samples of healthy donor pancreatic tissue (n = 24), adjacent normal to cancer (n = 3), and pancreatic adenocarcinoma (n = 18; Supplementary Table S1). Together, the atlas contained more than 200k cells, representing a variety of epithelial and stromal populations (Fig. 2A). Different cell clusters were annotated according to published known lineage markers (Supplementary Fig. S2A; Supplementary Table S2) and were present at expected proportions (Supplementary Fig. S2B and S2C; ref. 5). As expected, we identified several epithelial cell populations of the exocrine pancreas, which we classified as acinar, ADM, acinar/ADM, and ductal. To identify cancer cells, we used the Numbat algorithm that allows prediction of copy-number variation (CNV; as a proxy of genomic aberrations) from scRNA-seq data (18). A cluster expressing markers including KRT17, KRT19, CLDN18, and TFF1 exhibited the highest CNV probability, allowing us to identify it as tumor epithelial cells (Fig. 2B; Supplementary Fig. S3A). Interestingly, the CNV profile derived from transcriptomic data was consistent with that in the TCGA-PAAD dataset, including recurrent deletions in chromosomes 6, 17, and 18 (Supplementary Fig. S3B; ref. 19). Among components of the microenvironment, we identified immune cell types (macrophages, granulocytes, T cells, NK cells, B cells, plasma cells, and mast cells) and nonimmune stromal cell types (endothelial, lymphatic endothelial, endocrine, fibroblasts, pericytes, and neurons). The single-cell atlas included both normal pancreas and cancer samples; however, cells belonging to PanIN lesions cannot be readily identified in dissociated tissue because of the loss of histology, relative transcriptional similarity to neoplastic cells in the tumor, and their relative rarity.

Figure 2.

Identification of spatial domains and cell-type composition in donor and cancer samples. A, UMAP representation of integrated scRNA-seq data from healthy, adjacent normal (AdjNormal), and PDAC samples. Color represents annotated cell type. B, UMAP representation of scRNA-seq data of PDAC samples colored by expected CNV in each cell to identify cancer cells. C, Deconvolution of each spot to identify the fraction of different cell types (acinar, tumor epithelial, ducts, and fibroblasts) for each spot. The same H&E images correspond to a panel each in Fig. 1D and E. D, Spatially informed clustering results for all samples (healthy, adjacent normal, and PDAC), showing 17 different spatial domains based on gene expression from all the cells in the spot. E, Violin plot showing the distribution of transcripts based on deconvolution results to attribute them to specific cell types. F, High-magnification hematoxylin and eosin image with the Visium spots overlaid, colored by the annotated spatial domain. G, Top, UMAP representation of all spots colored by annotated spatial domain; bottom, histogram showing the relative proportion of different spatial domains in each tissue type.

To study PanIN lesions in the healthy pancreas and compare their characteristics with neoplasms in cancer samples, we utilized spatial transcriptomics with the 10x Genomics CytAssist Visium platform. We included the following samples: healthy donor (n = 5), treatment-naïve PDAC (n = 7), and adjacent normal to cancer (n = 2). For healthy donor pancreas samples, we chose sections that included PanIN lesions based on histologic evaluation. We then generated an additional independent validation dataset including healthy donor pancreas (n = 6 samples from 5 donor pancreata, head and tail samples in one case), choosing samples with no/low PanIN burden. The Visium platform sequences “spots” that encompass up to 10 cells each. A key challenge is to attribute transcripts to specific cell types within each area (spot) sampled. To estimate the specific cell–type composition of each Visium spot (Supplementary Fig. S4A), we performed spot deconvolution using the robust cell-type deconvolution (RCTD) algorithm (20). We observed a strong correlation between the estimated cell-type enrichment map of each cell type within each slide with pathologist annotations of histologic phenotypes (e.g., acinar, ductal, and tumor) and with stromal components (e.g., spots with a high fraction of acinar cells fell within regions of histologically defined acinar tissue). This supported the use of deconvolution to accurately delineate the composition of stromal components (e.g., fibroblasts and macrophages) that co-localize in a domain (Fig. 2C; Supplementary Fig. S4B).

Although deconvolution estimates the cell composition of each spot, we used BayesSpace (21, 22) to define different spatial domains, each composed of 1 or more cell type(s). We integrated spots from all samples and defined 17 distinct spatial domains based on the transcriptome of each spot (Fig. 2D; Supplementary Fig. S4C and S4D). To annotate different domains, we leveraged a combination of estimated cell-type enrichment (Fig. 2E), marker expression (Supplementary Fig. S5A; Supplementary Table S3A and S3B), and histology (Fig. 2F).

Three domains (D1, red; D2, blue; D3, green) were predominantly enriched in acinar and acinar/ADM transcript fractions with varying abundance of other cell-type transcripts (Fig. 2E). Domain 4 (D4, purple) was enriched in ductal transcripts and acinar transcripts, reflecting tissue histology (Fig. 2F).

Independent histologic analysis of tissue revealed distinct neoplastic structures: PanINs (present across all tissue types: healthy, adjacent normal, and PDAC) and both GLTumor (well differentiated) and PDTumor structures within PDAC. Two domains (D5, orange; D6, pink) were enriched in the tumor epithelial cell fraction. Upon investigation of the corresponding histologic phenotype, we noted that the (D5) domain was found in adjacent normal, healthy, and cancer samples and overlapped with histologically defined PanINs and GLTumor. When we analyzed transcriptomic domains, tumor-adjacent PanIN was indistinguishable from GLTumor; of note, cancerization of the ducts can result in PanIN-like areas that are in fact cancerization of the ducts (23). For this reason, we elected to combine tumor-associated PanIN and GLTumor in a single category for further analysis, referred to as GLTumor. Donor PanINs are referred to simply as PanIN. Lastly, the (D6) domain overlapped with regions of PDTumor with extensive fibrosis.

Next, we identified fibroblast-enriched domains (D7–10), hereby (Fibro1–4); a pericyte-enriched domain (D11) that overlapped stromal regions around vessels; and a spatial domain that overlapped adipose tissue (D12). We detected a domain overlapping islets, enriched for endocrine transcripts (D13), and a domain overlapping neurons (D14). We also detected 3 immune cell–enriched domains: a T cell/B cell–enriched domain (D15), which histologically corresponded to tertiary lymphoid structures (TLS); a plasma cell–enriched domain (D16); and a macrophage-enriched domain (D17).

The relative changes in spatial domain proportion between different tissue types concorded with the changes observed in our scRNA-seq data, supporting our data analysis approach (Fig. 2G; Supplementary Fig. S5B). When we compared healthy and PDAC samples, we observed a prevalence of acinar-rich domains in healthy and adjacent normal tissues, and, conversely, an enrichment in stromal domains in tumor samples (Fig. 2G; Supplementary Fig. S5C). Other domains, such as PDTumor, were only present in cancer samples, as expected.

Overall, the integration of scRNA-seq and spatial transcriptomics using spot deconvolution and spatially informed clustering provided complementary views and identified spatial domains conserved in healthy donor and cancer tissues, as well as cancer-specific ones. With this approach, we will be able to compare the co-evolution of the transcriptional status of neoplastic lesions and their surrounding neighborhood in healthy donor pancreata and cancer.

Neoplastic Domains Are Distinct from Benign Acinar and Ductal Domains and Display an Evolutionary Trajectory

Pancreatic cancer arises from the exocrine pancreas; studies in mice have identified both ductal and acinar cells as a possible origin of malignancy (24). Spatial transcriptomics allowed us to compare domains across benign normal epithelium (acinar and ductal domains) and neoplastic epithelium (PanIN, GLTumor, and PDTumor). We hypothesized that a continuum exists from acinar/ductal cells to PanIN and cancer with gradual transcriptomic changes. In contrast, transcriptomic data from spatial domains distinguished 2 separate groups, indicating that neoplastic spatial domains (PanINs, GLTumor, and PDTumor) have drastically distinct transcriptional profiles compared with benign epithelial domains (acinar and ducts; Fig. 3A). Pairwise gene expression correlation supported the notion that acinar and ductal domains cluster together, whereas donor PanIN, tumor-associated PanIN, GLTumor, and PDTumor occupy distinct domains that fall on a continuum (Supplementary Fig. S6A). Mapping known basal and classical gene signatures (25) showed high expression of classical markers in the PanIN/GLTumor domain and conversely high expression of basal markers in the PDTumor domain; thus, transcriptomic data matched expected histologic characterization (Supplementary Fig. S6B).

Figure 3.

Neoplastic spatial domains are transcriptionally distinct from benign domains, whereas a continuous evolutionary pattern connects PanIN to GLTumor and PDTumor. A, UMAP representation of spots from neoplastic and benign domains, colored by annotated spatial domain. B and C, Schematic representation of Visium spot deconvolution. AdjNormal, adjacent normal. D, PCA of epithelium-specific gene expression showing a trajectory from donor PanIN to GLTumor and PDTumor. PC, principal component. E, Differential gene expression analysis comparing donor PanIN and GLTumor. Dot size represents square root [Sqrt (baseMean)] expression. F, Differential gene expression analysis comparing GLTumor and PDTumor showing upregulation of basal genes (KRT17). Dot size represents Sqrt (baseMean) expression. G, Ridge plot showing GSEA comparing donor PanINs and GLTumor (PDAC) using the hallmark MSigDB pathways. H, Ridge plot showing GSEA comparing GLTumor and PDTumor using the hallmark MSigDB pathways. I, Representative images of IF staining for pERK (red) together with lineage markers pan-cytokeratin (PanCK; white) for epithelial cells (n > 10/sample type). Nuclei are stained with 4′,6-diamidino-2-phenylindole (DAPI; blue) in donor PanIN and cancer samples. J, Representative images of IF staining for pSTAT1 (green) together with lineage markers PanCK (white) and E-cadherin (ECAD; red) for epithelial cells (n = 5/sample type). Nuclei are stained with DAPI (blue) in donor PanIN and cancer samples. K, Representative images of IF staining for pSTAT3 (red) together with lineage markers ECAD (white) for epithelial cells and PDGFRa (green) for fibroblasts. Nuclei are stained with DAPI (blue) in donor PanIN and cancer samples, showing both GLTumor and PDTumor (n > 10/sample type). The IF image in the last panel is also shown as the representative image in the workflow graph in Fig. 1C.

To gain a more granular understanding of neoplastic spatial domains, we isolated them and generated a new Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP) representation (Supplementary Fig. S6C); we were then able to apply pseudotime trajectory analysis to the data using the Palantir algorithm (26), with PanINs as the starting root point. We observed a clear evolutionary trajectory from PanINs in organ donors to GLTumor and ending with PDTumor (Supplementary Fig. S6D). Further analysis using the whole-spot readout confounded efforts to trace epithelium-specific changes, as their composition is driven by heterogeneous cell types within each domain. For instance, healthy donor PanIN domains also contain acinar transcripts, due to acinar proximity, whereas GLTumor and PDTumor domains contain fibroblast transcripts, consistent with the accumulation of a desmoplastic microenvironment (Supplementary Fig. S6E). Therefore, to specifically interrogate changes in epithelial cells within each domain and determine whether epithelial transcripts drive domain heterogeneity and evolution, we applied a deconvolution framework (Fig. 3B and C; ref. 27). We thus obtained the tumor epithelium–specific expression for each domain and each sample. By principal component analysis (PCA) of tumor epithelium–specific expression data, we identified distinct clusters for PanIN, GLTumor, and PDTumor, indicating that differences between domains are, at least in part, driven by changes in epithelial cells (Fig. 3D).

To identify specific gene expression changes over the course of neoplastic evolution, we conducted pairwise comparisons utilizing the extracted tumor epithelium–specific expression. Among the upregulated genes in GLTumor compared with PanIN, we observed genes in the KLK family (e.g., KLK6, KLK7, and KLK8), which encode proteolytic kallikrein-related peptidases involved in remodeling the TME (28, 29). We also detected upregulation of CD55, known as decay-accelerating factor, which is involved in inhibiting complement-mediated attacks by accelerating the decay of C3 and C5 proteins of the innate immunity complement system and plays noncomplement roles in several downstream signaling pathways such as JAK/STAT (30). Finally, expression of GLUT1 (SLC2A1), a key component of glucose metabolism, was upregulated in GLTumor compared with PanIN; GLUT1 is a glucose transporter that has been associated with higher grades of dysplasia (Fig. 3E; Supplementary Table S4; ref. 31).

We also detected several genes downregulated in the GLTumor spatial domain compared with PanIN, including TFF1, MUC5AC, NKX6-2, GKN1, and GKN2. Of these, TFF1, a gastric tumor suppressor, and MUC5AC are known PanIN markers that are downregulated in tumors (5, 32). NKX6-2 is a key determinant of gastric cell identity in low-grade IPMN (33). GKN1 and GKN2, secretory proteins associated with gastric mucosa that maintain homeostasis, are associated with early precursor lesions in human and mouse tissue and impede tumor progression (34, 35). Less expectedly, the cytokine IL33 was downregulated in GLTumor compared with PanIN; IL33 is expressed in both pancreatic cancer cells, with stage-dependent tumor-restraining roles (36, 37), and fibroblasts, in which it promotes cancer growth (38).

We then compared GLTumor and PDTumor domains to focus on genes and transcriptional programs associated with the higher grade of dysplasia. In PDTumor domains, we observed upregulation of genes associated with a basal PDAC phenotype (e.g., KRT6A, KRT17, and LAMC2). We also observed upregulation of TGM2, which has been associated with poor survival due to immunosuppressive activity through T-cell suppression (39). WNT7A, a WNT signaling ligand (40), was also among the top upregulated genes (Fig. 3F; Supplementary Table S5). WNT7A and basal markers, including KRT17, also emerged as upregulated genes when comparing PDTumor with donor PanIN (Supplementary Fig. S6F; Supplementary Table S6).

Gene set enrichment analysis (GSEA) using hallmark gene sets (MSigDB; ref. 41) showed that multiple inflammation-related pathways such as those related to interferon (IFN) α and IFNγ responses were upregulated in GLTumor domains compared with PanIN domains and further upregulated in PDTumor domains; other pathways such as IL6/JAK/STAT3 signaling were upregulated exclusively in PDTumor domains (Fig. 3G and H; Supplementary Fig. S6G). This observation is concordant with upregulation of inflammatory pathways in tumor samples compared with PanIN in proteomic data (16). In addition, pathways previously linked to PDAC, such as KRAS signaling, hypoxia, and NF-κB, were upregulated. Conversely, we observed downregulation of metabolic pathways, such as oxidative phosphorylation and fatty acid metabolism (Fig. 3G and H; Supplementary Figs. S7A, S7B, and S8A; Supplementary Tables S7–S9).

To validate these findings at the protein level, we stained donor PanIN and cancer samples for phosphorylated ERK (pERK), a readout of MAPK signaling, and phosphorylated STAT1 (pSTAT1), a readout of inflammatory pathway activation including IFN signaling. Although PanINs presented with low-level pERK staining, expression was higher in tumor samples, both in terms of intensity and relative number of positive cells (Fig. 3I). STAT1 phosphorylation was present in occasional cells in PanIN but more prevalent in tumor samples (Fig. 3J).

Epithelial IL6/STAT3 signaling is required for progression to malignant disease in mouse models (42–44). To validate activation of JAK/STAT3 signaling, we stained organ donor PanIN and tumor tissue for pSTAT3, a readout of pathway activity. In PanIN, we observed no epithelial pSTAT3, although some expression was present in the surrounding microenvironment, where, based on mouse model studies, it also supports carcinogenesis (45–47). Conversely, in tumor samples, we observed robust epithelial expression of pSTAT3 in PDTumor regions, as well as in the surrounding microenvironment (Fig. 3K).

Overall, analysis of epithelial domains comparing neoplasm in donor (PanIN) and tumor (GLTumor and PDTumor) samples showed profound transcriptional reprogramming relative to normal epithelial cells in the pancreas (both acinar cells and ducts), as well as a clearly distinct transcriptional status compared with tumor samples. We observed continuous evolution from donor PanIN to GLTumor to PDTumor, including a progressive increase in inflammatory pathways, epithelial-to-mesenchymal transition, STAT1 and STAT3 signaling, and dysregulation of oxidative phosphorylation and fatty acid metabolism.

Asynchronous Evolution of Epithelial and Stromal Compartments during the Progression of Pancreatic Cancer

A key advantage of spatial transcriptomics over traditional scRNA-seq approaches is the ability to analyze components of the microenvironment based on proximity to specific epithelial structures. We thus analyzed benign and neoplastic epithelial neighborhoods, defined as spots within 150 μm of histologically and transcriptionally defined normal ducts (present in both donor and cancer samples), PanINs and GLTumor, and PDTumor. Neighborhood analysis showed similar composition of domains around ducts in donor and tumor samples, with a slight loss of acinar domains and a slight increase in other cell types, consistent with tissue disruptions in areas adjacent to tumors (Fig. 4A). Interestingly, neighborhood composition around PanINs was similar to ducts, which contrasted sharply with epithelial domain transcriptomic analysis. Although the epithelial component of PanINs clustered with tumors, and away from healthy ducts (Fig. 2A), the makeup of the PanIN microenvironment in otherwise healthy samples was similar to that of ducts and distinct from the TME. On the other hand, the neighborhood composition of malignant domains (GLTumor and PDTumor) was strikingly different from normal ducts and PanIN and characterized by immune cell spatial redistribution as well as the appearance of a distinct fibroblast domain (Fibro2) and TLSs (Fig. 4A).

Figure 4.

Distinct spatial domains characterize the microenvironment of PanIN and pancreatic cancer. A, Distribution of different spatial domains within 150 μm of benign (ducts) and neoplastic (PanIN, GLTumor, and PDTumor) epithelial domains. B, Spatial map showing the distribution of epithelial domains (ducts, PanIN, GLTumor, and PDTumor) and domains rich in plasma cells and macrophages. Note that plasma cells are in proximity of PanIN but not GLTumor and PDTumor and, conversely, macrophages are enriched near tumor epithelia. Fraction of plasma cell–enriched domains (C) and macrophage-enriched domains (D) in the neighborhood of benign and neoplastic epithelial domains. Numbers above the bars indicate P values from pairwise Wilcoxon rank-sum tests. E, IF staining for the epithelial marker E-cadherin (ECAD; white), the plasma cell marker CD27 (red), and the antigen presentation marker CD74 (green) together with nuclear staining [4′,6-diamidino-2-phenylindole (DAPI), blue] in donor PanIN and PDAC samples. F, Spatial map showing the distribution of tertiary lymphoid structures in healthy and PDAC tissue. G, Fraction of different spatial domains within 150 μm of TLSs. H, Expression of CCL19, CCL5, and CCL21, cytokines associated with TLSs, across spatial domains. Note the enrichment in TLSs. I, Normalized expression product of different chemokines with their respective receptors (CCL19–CCR7, CCL5–CCR4, and CCL21–CCR7), indicating potential spatial interaction in each domain and showing an enrichment in TLSs. J, Spatial map showing CCL19–CCR7 normalized potential interaction in PDAC. K, Expression of TLS-associated cytokines and their respective receptors on a per-cell basis using scRNA-seq.

With respect to the immune compartment, we observed an increase in macrophage-enriched spatial domains around GLTumor and PDTumor compared with both PanINs and ducts. Conversely, we observed a decrease in plasma cell–enriched domains in PDTumor and GLTumor compared with PanINs and ducts (Fig. 4B–D; Supplementary Fig. S9A). For validation, we performed immunostaining for the plasma cell marker (CD27), comparing PanINs and tumor samples. Consistent with the spatial sequencing data, we observed abundant CD27⁺ cells in close proximity to PanIN but not near tumors (Fig. 4E).

Another distinction between donor and tumor samples was the nearly exclusive presence of TLS domains, defined histologically and transcriptionally in the latter. TLSs have been previously described in pancreatic cancer (37, 48). Here, we show that although present, they are geographically separated from tumor domains (Fig. 4F; Supplementary Fig. S10A), with few lymphocytes proximal to tumor cells, which might undermine their ability to support immune responses. Using neighborhood analysis, we observed that TLSs are surrounded by fibroblast- and macrophage-enriched domains (Fig. 4G). TLS domains had high expression of CCL19, CCL21, and CCL5 (Fig. 4H), cytokines previously associated with TLS structures (48–50). We applied LIANA+ ligand–receptor (LR) analysis (51) and observed that TLS domains uniquely show high interaction of CCL19, CCL5, and CCL19 with their cognate receptors CCR7 and CCR4 (Fig. 4I and J; Supplementary Fig. S10B). To define potential sources of these chemokines, we utilized our scRNA-seq data and determined that CCL19 was highly expressed in a subset of fibroblasts, whereas CCL21 was expressed by lymphatic endothelial cells and CCL5 was expressed by T and NK cells (Fig. 4K), potentially all involved in TLS formation.

Overall, our data show that the microenvironment of PanINs is comparable with that of normal ducts and has not acquired features characteristic of the microenvironment of neoplastic domains in malignant tissue. This contrasts sharply with our observation that the transcriptome of the epithelial domains of PanIN closely resembles those of neoplastic ones (especially GLTumor) and suggests asynchronous evolution of epithelial and stromal compartments.

Tumor-Specific Fibroblasts Have a Distinct Transcriptional Signature and Are Associated with Unique Macrophage Populations

Fibroblasts are an abundant cellular component of the microenvironments surrounding PanIN and pancreatic cancer. Our initial analysis identified 4 distinct fibroblast domains (Fig. 2G). Fibro1, 3, and 4 were common to all tissue types (i.e., healthy donor, adjacent normal, and PDAC), and Fibro2 was enriched in tumor (Fig. 5A). Fibro1 was enriched in stromal areas surrounding adipose tissue, and Fibro4 was enriched in spots with high mitochondrial genes, indicating potential low-quality spots (Supplementary Fig. S11A and S11B). Based on neighborhood analysis (Fig. 3A), we observed that Fibro2 and Fibro3 were the main fibroblast-enriched domains around benign and neoplastic epithelial domains. Fibro3 was enriched near epithelial structures (ducts and PanIN) but became less abundant near GLTumor domains and was further depleted near PDTumor domains. Conversely, Fibro2 was absent near ducts and PanIN and highly enriched near tumor domains, with the most around the PDTumor domain (Fig. 5B and C). We thus focused on Fibro2 and Fibro3 for further analysis.

Figure 5.

Identification of a tumor-specific fibroblast-enriched domain (Fibro2) absent in PanIN. A, Left and middle, spatial map showing the distribution of different fibroblast-enriched domains; right, histogram showing the relative proportion of different domains to the total fibroblast-enriched domains in each tissue type. Note Fibro2 (brown) is specifically enriched in cancer samples, whereas Fibro1, Fibro3, and Fibro4 are also detected in PanIN-bearing samples. Fibro1 and Fibro4 are not in proximity to epithelial structures. The same slide is shown in Figs. 2D, 4B, and 5A, highlighting, respectively, all transcriptional domains (Fig. 2D), immune cell domains (Fig. 4B), and fibroblast domains (Fig. 5A). AdjNormal, adjacent normal. B, Bar plot showing the fraction of Fibro3 and Fibro2 domains in the neighborhood of epithelial domains. Note the relatively homogeneous distribution of Fibro3, whereas Fibro2 is cancer specific. Numbers above the bars indicate P values from pairwise Wilcoxon rank-sum tests. C, High-magnification hematoxylin and eosin image with Visium spots overlaid, showing proximity of Fibro2 and Fibro3 domains to duct-, PanIN-, and tumor-containing domains. The same slide is shown in Fig. 2F (all domains), Fig. 4A (neighborhood analysis), and Fig. 5C (fibroblast domains). D, UMAP representation of Visium spots from Fibro2 and Fibro3 domains and schematic representation of Visium spots overlaying multiple cells. E, PCA of extracted fibroblast-specific expression following deconvolution of Fibro2 and Fibro3 spatial domains. PC, principal component. F, Differential gene expression analysis comparing fibroblast-specific expression of Fibro2 and Fibro3 spatial domains. Dot size represents the square root [Sqrt (baseMean)] expression. G, UMAP visualization of scRNA-seq fibroblast subtypes colored by annotated subtype. H, Relative abundance of each fibroblast population by tissue type. I, Markers defining each fibroblast subtype. J, Fraction of each scRNA-seq–defined fibroblast subtype in the extracted fibroblast-specific expression of Fibro2 and Fibro3 spatial domains. K, UMAP representation of macrophage subtypes identified in scRNA-seq data. L, Markers defining each macrophage subtype. M, Fraction of each scRNA-seq–defined macrophage subtype in the extracted macrophage-specific expression of macrophage-enriched spatial domains in PDAC tissue. N, Scatter plot showing positive correlation between single-sample GSEA scores for alternatively activated (AltAct) macrophages and LRRC15^High in TCGA-PAAD dataset.

UMAP representation showed Fibro2 and Fibro3 are clearly distinct and independent of the sample from which they are derived (Fig. 5D). To determine whether transcriptomic differences were driven specifically by fibroblasts in each domain (rather than by a composition of different cell types), we applied a deconvolution algorithm similar to the one used for analyzing the tumor epithelial fraction of neoplastic domains. PCA plots of fibroblast-specific expression from the pseudobulk expression profile showed separate clustering of Fibro2 and Fibro3 domains (Fig. 5E). To understand which unique features characterize Fibro2, we compared it with the Fibro3 population common to normal and tumor samples by running differential expression analysis of the fibroblast-specific expression data. Fibro2 had lower expression of GPX3, C7, ADH1B, and EDNRB but higher expression of COL11A1, MMP11, CTHRC1, and POSTN, genes associated with ECM remodeling (52–59), a key feature of pancreatic cancer, and IGFL2, a gene associated with worse prognosis across cancer types (Fig. 5F; ref. 60). Analysis of TCGA database showed that genes highly expressed in Fibro2 were associated with tumor progression and metastasis (60–71), whereas those lower in Fibro2, such as GPX3, EDNRB, C7, and ADH1B, were associated with better prognosis (Fig. 5F; Supplementary Table S10; refs. 72–78).

Spatial transcriptomic profiling with 10x Genomics Visium is constrained by its relatively low resolution, which hampers the accurate delineation of specific cell types within a given domain due to mixed transcriptional signals. To gain further insights into the nature of Fibro2, we thus leveraged our scRNA-seq atlas. Utilizing a hierarchical deconvolution workflow, we further deconvoluted the extracted fibroblast-specific expression in Fibro2 and Fibro3 domains using scRNA-seq-defined fibroblast signatures. This methodology allowed us to define fibroblast subtypes enriched in our spatial domains. We isolated fibroblasts and identified 7 different populations with distinct gene signatures (Fig. 5G–I; Supplementary Table S11). In healthy and adjacent normal tissues, 3 populations, namely, IGF1^High (expressing high levels of PDGFRA and C7), HAS1^High [also expressing IL6 and the universal fibroblast marker PI16 (79)], and RORA^High, were prevalent. In PDAC samples, fibroblast populations shifted drastically, with LRRC15^High and SFRP4^High fibroblasts, both COL1A1^High, as well as a distinct LRRC15^LowSFRP4^LowCOL1A1^High population becoming prevalent. LRRC15^High fibroblasts expressed high levels of ACTA2, encoding for SMA, and the WNT transcription factor LEF1. Lastly, we defined a CD74^High population similar to previously described antigen-presenting CAFs (80).

We then deconvoluted Fibro2 and Fibro3 spatial domains and found that Fibro3 was enriched for IGF1^High and HAS1^High fibroblasts, whereas LRRC15^High and SFRP4^High fibroblasts were enriched in Fibro2 (Fig. 5J; Supplementary Fig. S11C).

We next investigated macrophage populations across samples. Macrophages were specifically abundant near tumor areas and even higher near PDTumor domains, compared with normal ducts and PanIN (Fig. 4A). To determine whether populations were transcriptionally different across samples, we again utilized a hierarchical deconvolution workflow to define macrophage subtypes in the macrophage-enriched domains proximal to the neoplastic domain in PDAC samples (GLTumor and PDTumor). Subclustering of macrophages from scRNA-seq data defined 6 different subtypes (classical, nonclassical, resident, alternatively activated, THBS1^High, and cycling) and a dendritic cell cluster (Fig. 5K and L; Supplementary Table S12). When we examined the prevalence of each cluster based on tissue type, we observed prevalent resident, alternatively activated, and THBS1^High macrophages in donor pancreas samples (Supplementary Fig. S11D). Adjacent normal and tumor samples had near-identical macrophage distribution, distinct from normal pancreas, likely because of tumor-induced inflammatory field effects. Classical, nonclassical, and alternatively activated macrophages were prevalent in both adjacent normal and tumor samples (Supplementary Fig. S11D). We found that alternatively activated macrophages had the highest enrichment in macrophage-enriched domains proximal to tumor domains, followed by resident and THBS1^High macrophages (Fig. 5M). Lastly, to evaluate the potential prognostic impact of these fibroblast signatures, we applied the top 20 marker genes that define each signature to TCGA-PAAD dataset (19). Although each signature could be detected in the database, the LRRC15^High fibroblast signature was the only one associated with poor survival, whereas the others did not show significant correlation with prognosis (Supplementary Fig. S11E). We also leveraged TCGA-PAAD data and observed that the LRRC15^High fibroblast signature had the highest correlation with alternatively activated macrophage signatures (Fig. 5N; Supplementary Fig. S11F). To exclude artifacts in the spatial sequencing run, we independently ran 6 new donor sample slides (from 5 donors; for 1 donor, we included separate slides from the head and tail of the pancreas), using the same 10x Genomics Visium platform as a validation cohort. We limited the resolution of clustering to main cell types (acinar, ducts, PanINs, endocrine, fibroblasts, pericytes, macrophages and adipose; Supplementary Fig. S12A–S12C; Supplementary Table S3B). We mapped the previously defined Fibro2 and Fibro3 signatures on fibroblast-enriched domains. Across the samples, we observed enrichment for the Fibro3 signature, but not Fibro2, consistent with lack of tumor in these samples (Supplementary Fig. S13A and S13B). Even though we did not define a distinct plasma cell–enriched domain in the validation dataset, we observed that the plasma cell marker (JCHAIN) was spatially correlated with PanIN (Supplementary Fig. S13C), consistent with our previous observations linking PanIN and plasma cell accumulation.

Taken together, our data show that a distinct fibroblast population, here named Fibro2 and characterized by high expression of LRRC15, ACTA2, and LEF1, is present in tumors but not around PanINs, supporting our initial observation that the PanIN precancer microenvironment is distinct from that of malignant areas.

Single-Cell Resolution Spatial Transcriptomics and Immunofluorescence Validate Fibroblast Reprogramming in Tumor Samples

We sought to validate the signatures defined from our Visium data analysis, utilizing 2 Xenium datasets published by Bell and colleagues (13) and Fiorini and colleagues (81) and 1 CosMx dataset published by Shiau and colleagues (82). Both Visium and Xenium pose limitations; Visium is not a single-cell platform, and although Xenium provides subcellular transcription localization, defining cell boundaries is challenging. We thus developed an analytic workflow to investigate our Visium-defined signatures in Xenium data (Fig. 6A). In brief, we used the Proseg unsupervised segmentation algorithm (83) to re-segment cells in the datasets; we then processed and integrated samples from each set independently because of variation in gene panels used in the 2 studies. In each dataset, we detected tumor/PanIN cells, fibroblasts, macrophages, plasma cells, and T cells; some cell populations, such as B cells, were only detected in one dataset, as the other did not contain population-specific markers (Supplementary Fig. S14A–S14D). Each dataset provided enough fibroblast markers to score the Fibro2 and Fibro3 signatures, with the caveat that we use different components of each fibroblast signature in the various datasets, due to the limited size of the gene panels (see “Methods”). Fibro2 and Fibro3 populations were mapped on fibroblasts in the tissue, and the radial distance from epithelial cells was calculated. Cells expressing the Fibro2 score were enriched close to cancer cells in the Fiorini and colleagues (81) dataset and declined as a function of distance from tumor cells. Conversely, Fibro3 was lower near tumor cells and increased with distance (Fig. 6B and C). In the Bell and colleagues (13) dataset, we were able to use KRT17 (a marker of poorly differentiated pancreatic cancer) and MUC5AC, a marker of PanIN, to identify epithelial structures and measure Fibro2 and Fibro3 signatures in proximity. The Fibro2 score was specifically enriched in proximity to KRT17 (Fig. 6D and E). Lastly, we analyzed a published CosMx dataset (82) that included 164 tumor-adjacent fields of view captured from 9 treatment-naïve PDAC samples. We scored Fibro 2 and Fibro 3 signatures on tumor-adjacent fibroblasts (Fig. 6F; Supplementary Fig. S14E) and observed an enrichment for the Fibro2 signature compared with the Fibro3 signature (Fig. 6G). We also observed high expression of alternatively activated macrophage markers (SPP1 and MARCO) in tumor-adjacent macrophages (Fig. 6H). Together, these data are consistent with the association between Fibro2 cells and cancer cells in Visium data. To validate the association of Fibro2 and PDAC at the protein level, we performed co-immunofluorescent staining for PDGFR (expressed by most fibroblasts) and αSMA/LRRC15 (enriched in Fibro2) on PanIN and tumor sections. Although abundant PDGFR⁺ fibroblasts surrounded PanIN, they were mostly αSMA negative and did not express LRRC15. In contrast, in tumor samples we observed abundant co-expression of all 3 markers, supporting transcriptional data (Fig. 6I; Supplementary Fig. S14F and S14G). Furthermore, immunostaining showed proximity of SPP1+ macrophages and LRRC15⁺ fibroblasts (Fig. 6J).

Figure 6.

Validation of fibroblast evolution during PanIN–cancer progression using Xenium spatial transcriptomic data. A, Workflow for Xenium data analysis. B, Fibro2 and Fibro3 scores in fibroblasts as a function of square root of radial distance from tumor cells in the Fiorini and colleagues Xenium dataset (81). Y-axis represents scaled (0–1) Fibro2 (brown) and Fibro3 (blue) scores, and x-axis represents square root of radial distance from tumor epithelium. C, Low- and high-magnification spatial maps showing the Fibro2 signature score (green to red) in fibroblasts surrounding tumor cells (purple). D, Fibro2 and Fibro3 scores in fibroblasts as a function of square root of radial distance from MUC5AC+ or KRT17+ epithelial cells in the Bell and colleagues Xenium dataset (13). Y-axis represents scaled (0–1) Fibro2 (brown) and Fibro3 (blue) scores, and x-axis represents square root of radial distance from MUC5AC+ or KRT17+ epithelial cells. E, Low- and high-magnification spatial maps showing the Fibro2 signature score (green to red) in fibroblasts surrounding MUC5AC+ (black) or KRT17+ (purple) epithelial cells. F, Workflow for the analysis of the CosMx dataset from Shiau and colleagues (82). G, Violin plot showing Fibro2 and Fibro3 signature scores in fibroblasts of the CosMx dataset. H, Dot plot showing expression of alternatively activated macrophage markers in different cell types in the CosMx dataset. I, Immunostaining for the fibroblast marker PDGFR (white) and Fibro2 markers SMA (red) and LRRC15 (green) in donor PanIN and cancer. Note that epithelial cells are not counterstained, but they are identified based on histology. J, Immunostaining for the fibroblast marker LRRC15 (green), macrophage lineage marker CD68 (pink), and alternatively activated macrophage marker SPP1 (yellow) in cancer.

Cancer-Restricted Fibroblasts Are Associated with Elevated WNT Signaling

We next explored potential signaling pathways driving cross-talk between tumor cells and fibroblasts. Differential expression analysis of fibroblast-specific expression in Fibro2 and Fibro3 spatial domains showed that LEF1 was significantly upregulated in Fibro2 (Fig. 7A). LEF1 encodes for a key transcriptional mediator and target of canonical WNT signaling [for review, see (84, 85)]. We then investigated expression of different WNT ligands in our scRNA-seq and spatial transcriptomic datasets (Supplementary Fig. S15A and S15B). We observed that WNT2 (a canonical WNT ligand) and WNT5A (a noncanonical WNT ligand) were higher in Fibro2 domains compared with Fibro3 (Fig. 7A). We observed that WNT7A and WNT7B [which can act as canonical or noncanonical ligands (40)] and WNT10A (a canonical WNT ligand) were highest in PDTumor domains compared with other neoplastic domains (Fig. 7B). scRNA-seq data confirmed that LEF1, WNT2, and WNT5A were expressed by fibroblasts, whereas WNT7A, WNT7B, and WNT10A were expressed by tumor epithelial cells (Fig. 7C). Together, these data support a model whereby fibroblasts both produce WNT ligands and respond to them (through autocrine or paracrine mechanisms) to activate downstream signaling. WNT signaling is associated with pancreatic cancer (86), and functional studies support a requirement for WNT activation during tumor formation and progression (87–89). Our data show that WNT signaling uniquely distinguishes CAFs from the fibroblasts surrounding sporadic PanIN.

Figure 7.

Pancreatic cancer–specific Fibro2 fibroblasts are characterized by elevated expression of WNT components and targets. A, Expression of WNT signaling components LEF1, WNT2, and WNT5A in estimated fibroblast-specific expression of Fibro2 and Fibro3 spatial domains. B, Estimated epithelial expression of WNT ligands WNT7A, WNT7B, and WNT10A. AdjNormal, adjacent normal. C, Dot plot showing scaled expression of LEF1, WNT2, WNT5A, WNT7A, WNT7B, and WNT10A in specific cell types in the scRNA-seq atlas. D, Experimental design for tumor organoid, normal fibroblasts, and CAF mono- and cocultures. E, Hematoxylin and eosin staining of tumor organoid, normal fibroblasts, and CAF mono- and cocultures. F, Dot plot showing scaled expression of WNT target/component LEF1 and all detected WNT ligands and classical and basal PDAC markers by cell type and culture condition. Note an increase in LEF1 expression in fibroblasts in the coculture and increased expression of known basal PDAC markers in epithelial cells in the coculture of tumor organoids with CAFs. G, IF for LEF1 (red), SMA (green), and nuclear staining [4′,6-diamidino-2-phenylindole (DAPI), blue] in donor PanIN and PDAC samples showing higher expression of LEF1 in tumor-adjacent fibroblasts. H, Bar plot showing quantification of LEF1 expression in αSMA+ fibroblasts in PanIN and cancer. I, Bar plot showing the expression of SPP1 by qPCR in macrophages with and without tumor organoid CM. Org. CM, organoid CM. J, Schematic representation of the asynchronous evolution of epithelium and stroma. The PanIN epithelium is distinct from healthy ducts and acini and acquires cancer-like characteristics such as elevation of pSTAT3, pSTAT1, and pERK signaling; these characteristics become more pronounced over progression to GLTumor and PDTumor, symbolized by deepening magenta color. In contrast, the microenvironment of PanIN is largely similar to that of healthy ducts, with the main distinction being accumulation of plasma cells. In cancer, the immune cell geography is rearranged, with a prevalence of macrophages in proximity to epithelial structures. Furthermore, a fibroblast population high in SMA, LRRC15, and LEF1 only arises in cancer samples.

To functionally investigate how epithelial cell–fibroblast cross-talk regulates WNT signaling, we utilized patient-derived organoids from 3 different PDAC samples and matched CAFs or 3 different pancreatic primary fibroblast lines from donor pancreata. We cultured organoids, CAFs, or normal fibroblasts separately or together (Fig. 7D) in 3D. Although organoids, CAFs, and fibroblasts each grow in Cultrex, cocultures acquire more complex tissue organization and abundant ECM (Fig. 7E). We performed scRNA-seq on dissociated cultures and identified separate tumor cell and fibroblast clusters (Supplementary Fig. S16A and S16B). We observed that LEF1 was not expressed at baseline in normal pancreas fibroblasts and was only slightly upregulated when they were cocultured with tumor organoids. Conversely, CAFs had higher basal expression of LEF1 and more prominent upregulation when cocultured with tumor organoids, consistent with cancer cells promoting WNT signaling in CAFs. Compared with monoculture, some WNT ligands decreased, such as WNT2 and WNT5A in CAFs and WNT10A in organoids, but organoid expression of WNT7A and WNT7B was higher under coculture conditions (Fig. 7F); together with the increase in LEF1 expression, these data support paracrine activation of WNT signaling. Furthermore, we observed that tumor organoids cocultured with CAFs, but not with normal pancreas fibroblasts, showed higher expression of several basal PDAC markers (KRT17, S100A2, and KRT6A) and lower expression of classical markers (MUC5AC, TFF1, and GATA6), suggesting that coculture with CAFs drives tumor organoids toward a basal fate (Fig. 7F). Together, our data show that intrinsic differences exist between CAFs and pancreatic fibroblasts, some of which are retained in vitro.

We then performed IF staining for LEF1, a WNT target gene and transcriptional mediator, together with αSMA on donor PanIN and tumor tissue samples. In PanIN samples, we observed rare LEF1 expression. In contrast, fibroblasts adjacent to tumor cells expressed LEF1 at high frequency (Fig. 7G and H), matching the spatial transcriptomic data. Given that macrophages surrounding tumor epithelium have specifically high expression of markers of alternative activation, we sought to determine whether tumor epithelium can directly affect macrophage polarization. For this purpose, we derived monocytes from healthy human peripheral blood mononuclear cells (PBMC), cultured them with macrophage colony-stimulating factor to differentiate M0 macrophages, and then cultured them with control medium or medium from tumor organoids (Supplementary Fig. S16C). We observed upregulation of SPP1 (Fig. 7I) and trending upregulation of MARCO (Supplementary Fig. S16D) with conditioned medium (CM) from distinct organoid lines.

To evaluate interactions between cell types in PanIN and tumor samples, we also ran spatially aware cell–cell communication analysis using CellChat (90). We observed that PDAC tissue had a higher number and strength of inferred interactions between spatial domains compared with donor tissue (Supplementary Fig. S16E). GLTumor, PDTumor, macrophages, and Fibro2 domains had the highest incoming and outgoing interaction strength (Supplementary Fig. S16F), with the caveat that spatial data are not ideal to identify signal source. We then plotted specific pathways of interest. Laminin, collagen, and fibronectin signaling, all mediating interaction with the extracellular domain (Supplementary Fig. S16G–S16I), were all higher in tumor than PanIN samples, as was CCL signaling (Supplementary Fig. S16J), a readout of inflammation. Interestingly, collagen and fibronectin 1 were in the Fibro2 domain.

Together, our data show that although the epithelial components of PanINs in donor pancreata transcriptionally resemble tumor samples, their microenvironment lacks key features of the pancreatic cancer microenvironment. Although human sample studies are by nature correlative, these data might explain why the vast majority of donor-associated PanINs do not progress to cancer; furthermore, they may set the stage for interception strategies aimed at preventing the formation of a tumor-permissive environment even in the presence of precursor lesions.

Discussion

Pancreatic cancer is a deadly disease, in part due to late diagnosis. As such, there is a strong interest in early detection. Different precursor lesions have been described; most pancreatic cancer derives from PanIN, a microscopic neoplasia that cannot be detected noninvasively (2). As PanINs are microscopic and not accessible in individuals who do not have pancreatic disease, their characteristics have remained poorly understood. Through a unique collaboration with the Gift of Life Michigan organ donation program, we obtained donor pancreata not needed or suitable for transplantation. As most individuals, even at a young age, bear spontaneous PanIN (5), we set out to understand how spontaneous PanIN compares with pancreatic malignancy.

We leveraged spatial and single-cell technologies to study the epithelial and stromal components of human sporadic PanIN and compare them with pancreatic cancer, applying both domain-based analysis and deconvolution strategies. We show that benign epithelium (acinar and ducts) is transcriptionally distinct from neoplastic epithelium (PanIN, GLTumor, and PDTumor). Conversely, donor PanIN, GLTumor, and PDTumor fall on a continuum, in which PanINs are closest to GLTumor. Key characteristics of malignancy, such as high KRAS signaling and elevated inflammatory pathways, show an upward trajectory from PanIN to GLTumor and PDTumor. We then compared transcriptional domains associated with the microenvironment of normal ducts, PanIN, and tumor. The PanIN microenvironment resembles that of normal ducts and is drastically distinct from the microenvironment of both GLTumor and PDTumor. Specifically, the geographic organization of immune cells differs sharply, whereby plasma cells are proximal to the PanIN epithelium but scattered through the tissue in tumor samples; in contrast, macrophages are concentrated in proximity to the tumor epithelium but do not accumulate near PanIN. Due to their relatively small size and low transcript content, we were not able to assess immature myeloid cells/granulocytes (and myeloid-derived suppressor cell populations) or regulatory T cells, which are important in the TME [for review see (2)], and we plan to investigate these in future studies. Furthermore, malignant tissue is characterized by the emergence of a distinct fibroblast population rich in SMA, LRRC15, and LEF1 (Fig. 7J). These findings show that transcriptional reprogramming in the epithelium and the stromal compartments occur in an asynchronous manner and might partially explain the indolent nature of most PanIN.

Our understanding of the TME has rapidly progressed due to the advent of single-cell technologies, which allow us to map its composition at the single-cell level [for review see (91)]. Key characteristics of the TME are the accumulation of fibroblasts and the infiltration of immune cells, with a prevalence of suppressive myeloid cells (2). A downside of single-cell technologies is the loss of spatial information, which makes it impossible to characterize heterogeneity within a tissue. More recently, spatial analysis has improved our understanding of how cells are organized within a tumor sample (11, 13, 14). Although spatial transcriptomics preserves the architecture of the tumor of origin, it necessitates a tradeoff between the number of genes analyzed and the ability to generate single-cell data. Here, we elected to use the Visium platform which, although not single cell, allowed us to explore tissue neighborhoods and relate them to the histology of each sample. We also used single-cell data from matched or equivalent samples to optimize deconvolution and obtain a higher number of gene reads. The picture that emerged was one of stepwise transformation, in which PanIN epithelial cells have acquired cancer-like characteristics, whereas their microenvironment has not. In tumor samples, even within the same section, we observed a correlation between tumor histology (which we divided into GLTumor and PDTumor) and matched microenvironment. In a different study, heterogeneity between different areas of the same tumor was observed, with some areas being rich in cellular components and others, labeled as “deserted,” being poor in immune infiltration and fibroblast density (92).

Among the epithelial changes that progress in intensity between PanIN and PDTumor, we noted multiple inflammatory pathways; of note, inflammatory pathways are already upregulated when comparing incidental, organ donor PanIN to tumor-adjacent PanIN, possibly due to a field effect of the cancer itself. In addition, we observed metabolic reprogramming and an increase in hypoxia. These pathways resemble the differences observed at the protein level between incidental PanIN and cancer-associated PanIN, indicating that the latter might be influenced by the cancer-induced field effect (16). Analysis of the fibroblast component revealed a specific population of CAFs expressing SMA, LRRC15, and LEF1, likely similar to LRRC15+ fibroblasts identified in mouse models (79). Interestingly, in a previous spatial study comparing tumor-adjacent PanIN with cancer, similar fibroblast populations were identified in PanIN and cancer, likely due to the use of tumor-adjacent samples (13). Through the use of pancreas samples from healthy individuals, we discovered that cancer-like fibroblast populations do not arise in precursor lesions, although they accompany malignant progression. What drives the transition from normal fibroblasts to CAFs and whether this transition is tumor promoting remains an area of future investigation. Mouse model studies suggest that fibroblasts restrain tumorigenesis until they acquire tumor-promoting characteristics (93).

A key open question is what signals drive the reprogramming of the microenvironment in the transition from PanIN to malignancy. Based on the previous work (3–5), PanINs have KRAS mutations but lack mutations in TP53 and other tumor suppressors. It is possible that cells that acquire additional mutations activate a new complement of microenvironment-reprogramming factors or that changes in the microenvironment facilitate epithelial proliferation, in turn causing the accumulation of additional lesions. One of the key factors associated with the cancer-specific fibroblast population was LEF1, which is both a mediator and target of WNT signaling. At the same time, WNT ligands are higher in GLTumor compared with PanIN and higher still in PDTumor, opening the possibility that WNT signaling, previously associated with cancer cells themselves (89), might be a key regulator of the TME.

In future studies, we plan to implement single-cell spatial transcriptomic techniques such as Xenium (and, in the future, genome-wide approaches as they become available) to further characterize the nature of human PanIN. A key outstanding question is how heterogeneous these lesions are; addressing this question will require analyzing lesions across a large number of donors; although currently difficult due to limitations in both funding and computational power, large-scale PanIN studies will no doubt become possible in the near future.

Ultimately, our study sets the stage for future functional work addressing the causes of microenvironment reprogramming and, at the same time, determining whether the microenvironment might serve to restrain progression of most lesions. Furthermore, the factors that might break the microenvironmental barrier and allow malignant progression might constitute a target for future interception approaches.

Methods

Tissue Procurement

Donor pancreata not eligible for transplant or for whom there were no eligible recipients were collected at the Gift of Life Michigan Donor Care Center. While maintaining blood flow, the superior mesenteric arteries and celiac arteries were cross-clamped, and the pancreas was recovered and immediately placed in a physiologic organ preservation solution on ice and transported to the University of Michigan, approximately 6 miles away. Acquisition of donor pancreata for research purposes was approved by the Gift of Life Research Review Group. Acquisition of patient tumor samples through formal written informed consent obtained following in person or phone conversation with patients or donor families was described previously (5) and approved by the University of Michigan Institutional Review Board (HUM00025339, HUM00098128, and HUM00041280). For cancer samples, patients with a diagnosis of PDAC were consented prior to surgical resection or diagnostic biopsy. In addition, deidentified samples from the University of Michigan Pathology Biobank were acquired for additional immunostaining validation. Adult patients of any age were included as pancreatic cancer is not a pediatric disease. Both male and female patient samples were included. Donor pancreata were recovered from individuals on life support following certification of brain death and confirmed donor status. Donor age was 20 to 70 years; donors were about two thirds male and one third female. About two thirds of the donors identified as Caucasian/white, whereas one third identified as African American or Black, with the remaining unknown or mixed race. For the specific samples used for single-cell or spatial sequencing, demographics data are included in Supplementary Table S1.

Tissue Processing

Upon arrival at the laboratory, donor pancreata were dissected by a clinically trained pancreatobiliary surgeon. Organs were dissected into the head, body, and tail, with portions of each placed into DMEM with 1% BSA/10 μmol/L Y27632 or 10% formalin for single-cell sequencing or paraffin embedding, respectively. For single-cell processing, tissue was minced into 1-mm³ pieces and then digested with 1 mg/mL collagenase P for 20 to 30 minutes at 37°C with gentle agitation. Digested tissue was rinsed 3 times with DMEM/1% BSA/10 μmol/L Y27632 and then filtered through a 40-μm mesh. Resulting cells were submitted to the University of Michigan Advanced Genomics Core (RRID: SCR_025788) for single-cell sequencing using the 10x Genomics Platform.

scRNA-seq Processing

Newly acquired samples (n = 14) together with previously published samples from (5, 12, 17) were processed using Cell Ranger (v7.0.1; RRID: SCR_017344), and then ambient RNA contamination was corrected using CellBender (v0.3.0; RRID: SCR_025990; ref. 94) and Python (v3.7.12; RRID: SCR_008394). We used R (v4.4.0; RRID: SCR_001905). Samples were processed using Seurat (v5.2.1; RRID: SCR_016341; ref. 95); each sample gene expression matrix was log-normalized to library size and scaled, and the top 2,000 variable genes were identified for downstream analysis. We used Scanpy (v1.10.4; RRID: SCR_018139; ref. 96) and Python (v3.12.8), utilizing the scanpy.pp.scrublet() (RRID: SCR_018098; ref. 97) function to estimate a doublet score for each cell in each sample. To account for batch effects, we used scVI (v1.3.3; RRID: SCR_026673; ref. 98) and Python (v3.12.8) to integrate all samples. Louvain graph-based clustering was done on the shared latent scVI embedding at a resolution of 0.2, which showed stable clustering results based on Clustree analysis (v0.5.1; RRID: SCR_016293; ref. 99). Marker genes for each cluster were defined using the Wilcoxon rank-sum test and annotated based on known lineage markers for each cell type. CNV inference was done on the PDAC samples using Numbat (v1.4.2; RRID: SCR_019207; ref. 18), which leverages both gene expression and germline heterozygous SNPs to generate CNV probability for each cell. Healthy normal pancreas samples were used as a reference. We also used the SeuratExtend (v1.2.0; RRID: SCR_026143) R package for visualization (100).

Spatial Transcriptomic Processing

Donor (n = 5), adjacent normal to the tumor (n = 2), and PDAC (n = 7) samples were profiled using the 10x Genomics Visium CytAssist platform using the 11 mm × 11 mm slide (RRID: SCR_024570). Sample processing, library preparation, and sequencing were performed at the University of Michigan Advanced Genomics Core (RRID: SCR_025788). Visium data were analyzed using Seurat (v5.2.1; ref. 95). Each spot was log-normalized and scaled. To account for batch effects, we integrated the spots of all samples from different tissues (healthy, adjacent normal, and PDAC) using reciprocal PCA (rPCA), which is known to work well with heterogeneous datasets with non-overlapping cell types (e.g., tumor and normal tissue samples) without overcorrection. We used the RCTD module (20) from the spaceXr package (v2.1.7) to perform spotwise cell-type deconvolution using our established scRNA-seq reference. We used the “multi” deconvolution mode to account for having more than 2 cell types in each spot.

We used BayesSpace (v1.16.0; RRID: SCR_026309; ref. 21) to perform spatially informed clustering. We applied BayesSpace clustering utilizing joint rPCA embedding, defining distinct spatial domains by the transcriptome of all cells in their spots. BayesSpace provides a qTune() function that runs BayesSpace at multiple values of q (desired number of clusters) and computes the average pseudo log-likelihood. Our dataset plateaued at q = 23; therefore, we generated 23 different clusters and then collapsed them to 17 clusters by merging highly correlated domains. Different clusters (spatial domains) were annotated utilizing a combination of estimated cell-type fraction, markers, and a histology-guided approach. We used miloR (v2.2.0; RRID: SCR_025630; ref. 101) to perform differential abundance analysis between donor and tumor tissue. Seurat (v5.2.1; ref. 95) and SPATA2 (v3.1.4; RRID: SCR_026167; ref. 102) were used for visualization.

A similar workflow was used for the analysis of the donor sample validation dataset, in which we defined 9 spatial domains only. We projected the signatures defined from the integrative analysis of donor and PDAC samples (Fibro2 and Fibro3) onto the validation dataset using UCell scoring.

Neighborhood Analysis

To assess the neighborhood composition surrounding specific epithelial structures, we utilized semla (v1.3.1; ref. 103) to determine the radial distance from spots belonging to designated domains (ducts, PanINs, GLTumor, and PDTumor). Within each sample, we identified spots located within a 150-μm radius from spots of the epithelial structure of interest. We then calculated the proportion of various spatial domains surrounding these structures in each sample, as well as the average proportion across all samples. The nonparametric Wilcoxon rank-sum test was used to test differential enrichment of certain spatial domains around different epithelial structures.

Pseudotime Analysis

We used Palantir (v1.3.6; RRID: SCR_027194; ref. 26), Scanpy (v1.10.4; ref. 96), and Python (v3.9.21) to run pseudotime analysis on the subsetted neoplastic domains (PanINs, GLTumor, and PDTumor). Diffusion maps were constructed based on integrated rPCA spot embedding, generating 5 diffusion components that were used to estimate pseudotime. A cell from a healthy donor PanIN cluster was chosen as a root node.

Pseudobulk Deconvolution

We used BayesPrism (v2.0; RRID: SCR_027499; ref. 27) to extract epithelium-specific or fibroblast-specific expression from different spatial domains utilizing our established scRNA-seq atlas. We generated a pseudobulk matrix for each spatial domain across samples, generating a (genes × sample spatial domain) matrix, and then extracted epithelium- or fibroblast-specific expression using BayesPrism deconvolution. Variance-stabilizing transformation from DESeq2 (v1.46.0; RRID: SCR_015687; ref. 104) was used to normalize the estimated cell type–specific expression. We ran PCA on the normalized expression matrix using the top 2,000 highly variable genes. We ran differential gene expression analysis following the DESeq2 standard workflow. Although deconvolution substantially improved the cell-type specificity of our differential gene expression analyses, we observed that some differentially expressed (DE) genes detected by DESeq were not exclusive to the cell type of interest. Notably, these nonspecific genes typically exhibited low average expression (baseMean) values. To address this and enhance interpretability, we adjusted the sizes of dots in our volcano plots to reflect the baseMean value, thereby visually emphasizing genes with higher and more cell type–specific expression. This approach allowed us to prioritize biologically meaningful, cell type–specific differential expression results in our analysis. GSEA (105) of MSigDB hallmark pathways (RRID: SCR_016863; ref. 41) was subsequently run using log₂ fold change (FC)–based ranks using clusterProfiler (v4.14.4; RRID: SCR_016884; ref. 106) and fgsea (v1.32.0; RRID: SCR_020938), and results were visualized as ridge plots using enrichplot (v1.26.6; RRID: SCR_026996). Gene sets were downloaded using msigdbr (v7.5.1; RRID: SCR_022870).

Fibroblast Hierarchical Deconvolution

We implemented a hierarchical deconvolution workflow to further resolve cell type–specific expression profiles to scRNA-seq–defined subtypes. Using scRNA-seq data, we isolated fibroblast and macrophage populations and then performed reclustering to identify distinct subtypes within each subpopulation. We used fibroblast-specific expression profiles from the Fibro2 and Fibro3 domains (initially estimated using BayesPrism), along with a fibroblast scRNA-seq reference, as inputs for BayesPrism to estimate the relative proportions of different scRNA-seq–defined fibroblast subpopulations within the fibroblast fraction of these spatial domains. We followed the same approach to estimate relative proportions of scRNA-seq–defined macrophage subpopulations in the macrophage-enriched domains in PDAC samples.

TCGA-PAAD Survival Analysis

We evaluated the signatures of various scRNA-seq–defined fibroblast and macrophage subsets within the TCGA-PAAD dataset. Data were obtained via TCGAbiolinks (v2.34.1; RRID: SCR_017683; ref. 107). For subclusters within each subset (fibroblasts and macrophages), a marker score was calculated by multiplying the log₂ FC by the percentage of cells expressing each gene within the cluster (pct.1). Subcluster signatures were then defined by selecting the top 30 genes with the highest scores. Enrichment scores for different signatures were computed using single-sample GSEA implemented in the GSVA package (v2.0.7; RRID: SCR_021058; ref. 108). Finally, pairwise Pearson correlations between enrichment scores were calculated. Overall survival of TCGA-PAAD (19) patients stratified by high versus low scores of scRNA-seq–defined fibroblast subtype signatures. Plots were generated using SurvivalGenie (109).

LR Interaction Analysis

LR interaction analysis was done using LIANA+ (v1.5.0; ref. 51) and Python (v3.8.20). Global Moran I within each sample and local normalized expression product within each spot were calculated for each LR pair. CellChat (RRID: SCR_021946; ref. 90) analysis was done using the gene expression and spatial information to define spatially proximal cell–cell communication using the default parameters. The analysis was done for the donor sample dataset and the PDAC sample dataset separately.

Isolation and Culturing of Organoids and Fibroblasts from PDAC Tumor

Matching organoids and CAFs were derived from surgically resected patient tumors as previously described (17, 38). For organoid culture, cells were embedded in Matrigel and maintained in a complete human organoid medium consisting of 50% L-WRN CM and 50% DMEM/F12, supplemented with B27 (1× final), nicotinamide (10 mmol/L), N-acetylcysteine (1.25 mmol/L), murine EGF (50 ng/mL), human FGF (100 ng/mL), human Gastrin 1 (10 nmol/L), Y-27532 (10.5 mmol/L), and penicillin/streptomycin (2×; ref. 17). CAFs were cultured in DMEM/F12 containing 20% fetal bovine serum (FBS) and 1× penicillin/streptomycin. All primary cell cultures were routinely tested for Mycoplasma contamination, and tumor organoids underwent characterization by whole-exome sequencing as previously described (17).

Organoid–Fibroblast Coculture

To establish cocultures, tumor organoids and matched CAFs or normal fibroblasts were combined at a 1:2 ratio (organoid:CAF or organoid:fibroblast) in Type II Cultrex (#3533-010-02, R&D Systems) in a 12-well plate. Separately, organoids and CAFs or normal fibroblasts were also plated as monocultures in Type II Cultrex. Both cocultures and monocultures were maintained in DMEM/F12 containing 5% FBS and 1× penicillin/streptomycin. After 3 days, cultures were dissociated into single cells using AccuMax (Invitrogen) according to the manufacturer’s instructions. Cells were then washed twice with DMEM/F12 with 0.05% BSA, filtered through a 40-μm mesh cell strainer, and counted. Sequencing and single-cell transcriptomic cDNA library preparation were completed at the University of Michigan Advanced Genomics Core (RRID: SCR_025788). Samples were run using 50-cycle paired-end reads on the HiSeq 4000 (Illumina) or NovaSeq 6000 (RRID: SCR_016387) to a depth of 100,000 reads per sample. Data processing and alignment to the GRCh38 genome were performed by the University of Michigan Advanced Genomics Core using Cell Ranger (version 7.1.0) with an expected cell count of 10,000. Samples were processed following the same workflow used above for human scRNA-seq data. For the specific samples used to generate organoids, CAFs, or normal fibroblasts, demographic data are included in Supplementary Table S1.

Xenium and CosMx Data Analyses

We obtained Xenium data from 2 previously published studies (13, 81) and CosMx data (82). For Xenium data, we used Proseg (83) for unsupervised segmentation of transcripts. We analyzed each dataset independently because of mismatch in the respective gene panels. Samples were processed using Seurat (v5.2.1; ref. 95); gene expression was log-normalized, and samples were integrated using Harmony (v1.2.3; RRID: SCR_022206; ref. 110) and clustered using graph-based Louvain clustering. Separately, we ran RCTD (20) to estimate the type of each segmented cell. Cell-type annotation was done based on combined clustering information and the RCTD-estimated cell type. For the CosMx data, we utilized cell segmentation as described in the original study. Although the CosMx data included both treatment-naïve and treated samples, we only included the former in our analysis. Fibro2 and Fibro3 signature scores were calculated in fibroblasts in each dataset. Specifically, using the extracted fibroblast-specific expression from Fibro2 and Fibro3, we generated Fibro2 and Fibro3 signatures based on DE genes between both groups with an adjusted P value < 0.05 and log₂ FC >0.5. Then, we intersected these gene sets with the gene panel in each dataset. Although BayesPrism deconvolution enriches fibroblast-specific genes, certain DE genes were not fibroblast specific; therefore, we curated this gene set to include fibroblast-specific genes (Supplementary Table S13). The signature scores in each dataset were calculated using UCell (v2.10.1; RRID: SCR_027109; ref. 111) and then scaled to fit between (0–1). Radial distance from tumor cells was calculated using semla (v1.3.1), and then a spline was fitted to regress Fibro2 and Fibro3 scores in fibroblasts as a function of distance from tumor cells using a generalized linear model.

Immunofluorescence Staining

Donor tissue or tumor tissue fixed in 10% neutral-buffered formalin were serially sectioned 5 μm thick using the microtome system. Slides were baked for 60 minutes, deparaffinized, and rehydrated by serial incubation in xylene (10 minutes, 2 times), 100% ethanol (5 minutes, 2 times), 95% ethanol (2 minutes, 2 times), 70% ethanol (2 minutes, 2 times), and dionized water (5 minutes). All antigens were retrieved at 96°C with 10 mmol/L sodium citrate (pH 6) for an additional 8 minutes, except for pSTAT3 antigen which was retrieved with ethylenediaminetetraacetic acid solution (pH 8; Cell Signaling Technology). Once sections were cooled to room temperature (RT), they were washed 3 times with PBS (2 minutes per wash). Slides were then incubated in 0.3% H₂O₂ in methanol for 15 minutes before blocking in 10% goat serum (Invitrogen) for 30 minutes at room temperature. After blocking, slides were incubated in a primary antibody diluted in PBS in a humidified chamber overnight at 4°C. Antibody information and specific dilution factors are listed in Supplementary Table S14. The following day, slides were incubated with an anti-mouse or anti-rabbit secondary antibody diluted in PBS for 1 hour at room temperature. Concentrations are listed in Supplementary Table S14. Alexa Fluor 488 Tyramide SuperBoost Kit (Invitrogen) was utilized for primary antibodies raised in the same species. Lastly, tissue sections were stained with 4′,6-diamidino-2-phenylindole for 5 minutes and washed 3 times with PBS before being mounted with ProLong Gold Antifade. High-magnification images at 63× were obtained using a Leica STELLARIS 8 microscope (RRID: SCR_024664) using Leica Application Suite X (RRID: SCR_013673). Area of the immunofluorescence (IF) staining was quantified under the Otsu threshold using ImageJ (RRID: SCR_003070) analytic software. For LRRC15 quantification, the LRRC15 median fluorescence intensity (MFI) was normalized to the PDGFR MFI per region of interest.

Harvesting CM from Organoid Lines

To generate tumor organoid CM, dissociated organoids were seeded at 50% confluence and cultured in 2 mL of RPMI with 5% FBS media for 72 hours. CM was harvested and spun down at 400 relative centrifugal force and filtered through a 40-μm filter to remove cell debris and stored at −80°C in 1-mL aliquots. For the specific samples used to harvest CM, demographic data are included in Supplementary Table S1.

Isolation of Monocytes from Blood and Generation of M0 Macrophages

Human CD14⁺ monocyte cells were obtained from PBMCs of healthy donors using positive magnetic selection with a CD14⁺ monocyte cell isolation kit (Miltenyi Biotec) as per the manufacturer’s instructions. Cells were seeded at a density of 0.5 × 10⁶ cells/mL per well in a non–tissue culture–treated 6-well plate in 2 mL of RPMI complete media for 5 days. The RPMI complete media was supplemented with 20% FBS, 100 U/mL penicillin/streptomycin, 1 mmol/L sodium pyruvate, and 1× minimum essential medium non-essential amino acids.

Treating M0 Macrophages with Tumor Organoid CM, RNA Extraction, and Quantitative Real-time PCR

Five-day-old M0 macrophages were lifted using Accumax and seeded at 1 × 10⁶ cells per well in a non–tissue culture–treated 6-well plate. Cells were cultured with (i) 50% organoid CM and 50% RPMI complete media or (ii) RPMI complete media for 48 hours. Total RNA was extracted using the RNeasy Plus Mini kit and cDNAΔ generated using reverse transcriptase kit (Invitrogen). Gene expression changes in macrophages were assessed via quantitative real-time PCR (QuantStudio 6 Flex, Life Technologies; RRID: SCR_020239) using TaqMan Fast Advanced Master mix (Applied Biosystems). Relative gene expression was calculated using the 2^−ΔΔCt approach and normalized against GAPDH. TaqMan primers for SPP1 (Hs00959010_m1), MARCO (Hs00198937), and GAPDH (Hs02786624_g1) were acquired from Thermo Fisher Scientific.

Supplementary Material

Supp Table 1

Clinical Data

Supp Table 2

scRef Clusters Markers

Supp Table 3A

Spatial Domain Markers

Supp Table 3B

Spatial Domains Markers (Validation Dataset)

Supp Table 4

GLTumor vs Donor PanINs DEGs

Supp Table 5

PDTumor vs GLTumor DEGs

Supp Table 6

PDTumor vs Donor PanINs DEGs

Supp Table 7

GLTumor vs Donor PanINs GSEA

Supp Table 8

PDTumor vs GLTumor GSEA

Supp Table 9

PDTumor vs Donor PanINS GSEA

Supp Table 10

Fibro2 vs Fibro3 DEGs

Supp Table 11

scRNAseq Fibroblasts Subset Markers

Supp Table 12

scRNAseq Macrophages Subset Markers

Supp Table 13

Fibro2 and Fibro3 Genesets for Xenium Projection

Supp Table 14

Antibodies Dilutions

Supp Figs S1-S10

Supplementary Figure S1. Healthy Donor Pancreas (Validation Dataset). Supplementary Figure S2. A. Dot plot showing marker genes for each cell type in scRNAseq data. Supplementary Figure S3. A. UMAP representation of scRNAseq data of PDAC individual samples colored by Numbat joint CNV probability based on gene expression and heterozygous SNPs beta-allele frequency (BAF) ratio. Supplementary Figure S4. A. Hematoxylin and Eosin (H&E) of healthy and PDAC samples profiled with 10x Genomics CytAssist Visium platform, with overlaid Visium spots. Supplementary Figure S5. A. Dot plot showing marker genes for each spatial domain in spatial transcriptomics data. Supplementary Figure S6. A. Heatmap showing pairwise Pearson correlation between different spatial domains in different tissue types annotated by spatial domain information and tissue type. Supplementary Figure S7. A. Heatmap showing scaled expression of leading edge genes from GSEA of OXIDATIVE_PHOSPHORYLATION, FATTY_ACID_METABOLISM and BILE_ACID_METABOLISM. Supplementary Figure S8. A. Heatmap showing scaled expression of leading edge genes from GSEA of TNFA_SIGNALING_VIA_NFKB, EPITHELIAL_MESENCHYMAL_TRANSITION, INTERFERON_ ALPHA_RESPONSE, HYPOXIA, KRAS_SIGANLING_UP and IL6_JAK_STAT3_SIGNALING. Supplementary Figure S9. A. Spatial map showing the distribution of plasma cells and macrophages around neoplastic domains (PanIN, GLTumor and PDTumor) in healthy, adjacent normal and PDAC samples. Supplementary Figure S10. A. Spatial map showing the abundance and distribution of tertiary lymphoid structures around neoplastic domains in healthy, adjacent normal and PDAC tissue samples. Colors represent annotated spatial domains.

Supp Figs S11-S16

Supplementary Figure S11. A. High magnification hematoxylin and eosin (H&E) image with the Visium spots overlaid, showing adjacency of Fibro1 and Adipose domains, colored by the annotated spatial domain. Supplementary Figure S12. A. Spatially-informed clustering results from donor samples validation dataset, showing 9 different spatial domains based on gene expression from all cells in the spot. Supplementary Figure S13. A. Violin plot showing the Fibro2 and Fibro3 signature score in the fibroblast spatial domain of the donor samples validation dataset. Supplementary Figure S14. A. UMAP representation of integrated Xenium samples from Fiorini et al. (81) dataset, colored by annotated cell type. Supplementary Figure S15. A. Dot plot showing the expression of WNT ligands, receptors and downstream targets in different spatial domains from spatial transcriptomics data. Supplementary Figure S16. A. UMAP representation of scRNAseq data from the mono- and co-culture experiment.

Data Availability

Data from this analysis can be interactively explored at: https://www.pancreas-precancer-atlas.com/. 10x Visium spatial transcriptomic raw data can be found at: https://doi.org/10.5281/zenodo.19304282. New donor scRNA-seq samples are available at: https://doi.org/10.5281/zenodo.17538974. scRNA-seq data for organoid monoculture and coculture are available at: https://doi.org/10.5281/zenodo.19305784. Processed data from human pancreas scRNA-seq, spatial transcriptomics, organoids, sRNA-seq, and reprocessed Xenium data (13, 81) can be downloaded at: https://doi.org/10.5281/zenodo.21867195.

Code Availability

Code used for the data analysis can be found at: https://pascadimagliano-lab.github.io/PancAtlas_Methods/.

Authors’ Disclosures

A.W. Bray reports grants from NCI Oncology Research Training Grant at the University of Michigan during the conduct of the study. A. Rao reports grants from NIH, National Science Foundation, and Department of Defense during the conduct of the study, as well as other support from Voxel Analytics LLC, Telperian, TCS Ltd., and American Heart Association and nonfinancial support and other support from Indian Institute of Science outside the submitted work. A. Maitra reports other support from Exact Sciences (Abbott Labs) during the conduct of the study. E.J. Fertig reports grants from NIH/NCI and Lustgarten Foundation during the conduct of the study, as well as grants from Break Through Cancer and National Foundation for Cancer Research and personal fees from Piper Sandler, Mestag Therapeutics, and ResistanceBio/Viosera Therapeutics outside the submitted work. M. Pasca di Magliano reports grants from NIH/NCI, NCI/National Institute of Diabetes and Digestive and Kidney Diseases, and grants from Department of Veterans Affairs during the conduct of the study. No disclosures were reported by the other authors.

Authors’ Contributions

A.M. Elhossiny: Conceptualization, data curation, software, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. P. Kadiyala: Conceptualization, data curation, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. J.O. Okoye: Data curation, formal analysis, validation, investigation, methodology. H.L. Hiraki: Validation. M.C. Procario: Validation, investigation. T. Giridharan: Validation, investigation. H.R. Watkoske: Validation, investigation. M. Tannus Ruckert: Validation, investigation. J. Wang: Investigation. B.D. Griffith: Data curation, investigation. A.W. Bray: Data curation, validation, investigation. J.N. Mills: Resources, data curation. C.E. Espinoza: Investigation. J. Zeller: Data curation, validation, investigation. N. Peterson: Data curation, investigation. F. Bednar: Resources, data curation, formal analysis, writing–review and editing. Y. Zhang: Data curation. A. Rao: Data curation, software, methodology. C.A. Lyssiotis: Resources, funding acquisition, investigation, methodology. J.M. Szczepanski: Data curation, writing–review and editing. J. Shi: Data curation, investigation. A. Deshpande: Resources, methodology. A. Maitra: Validation, writing–review and editing. E.J. Fertig: Conceptualization, supervision, investigation, methodology, project administration, writing–review and editing. E.S. Carpenter: Conceptualization, data curation, supervision, investigation, methodology, writing–original draft, writing–review and editing. T.L. Frankel: Conceptualization, resources, supervision, funding acquisition, investigation, methodology, writing–original draft, project administration, writing–review and editing. M. Pasca di Magliano: Conceptualization, resources, data curation, supervision, funding acquisition, investigation, methodology, writing–original draft, project administration, writing–review and editing.