Novel Acinar Metaplastic States Uncovered in Exocrine Pancreas Disease

Abstract

Background & Aims

In response to injury, pancreatic acinar cells undergo acinar-to-ductal metaplasia (ADM), marked by loss of acinar identity and acquisition of ductal features. Although ADM can resolve to support tissue repair, it may also persist and serve as a precursor to pancreatic cancer. Whether diverse pancreatic stressors drive a shared or context-specific ADM program remains unclear. We sought to comprehensively define metaplastic responses to clinically relevant exocrine pancreas diseases known to increase cancer risk.

Methods

We profiled ADM and the surrounding microenvironment across mouse models of exocrine disease—including acute, recurrent, and chronic pancreatitis, as well as in the setting of oncogenic Kras—capturing over 300,000 single cells. To enable high-quality transcriptomic profiling in enzyme-rich tissue, we leveraged FixNCut, a method that preserves RNA integrity in the exocrine pancreas. Findings were validated in human pancreas tissue using CosMx spatial transcriptomics.

Results

We identify a conserved acinar response across disease contexts that gives rise to previously unrecognized distinct metaplastic states, including a “gateway” ADM population that precedes more advanced metaplastic states marked by complete loss of acinar identity. In pancreatic intraepithelial neoplasia (PanIN) precancerous lesions, we detect classical-like and basal-like states, suggesting that pancreatic cancer subtypes are specified much earlier than previously appreciated. In Kras-mutant tissue, we identify a second wave of inflammation and the emergence of an immunosuppressive niche, coinciding with PanIN formation.

Conclusions

Our findings define a conserved program of acinar plasticity across exocrine pancreas diseases. We further link unresolved ADM to immune remodeling during precursor lesion formation and observe the emergence of pancreatic cancer subtypes in early PanIN lesions.

Graphical abstract

Summary.

This work maps acinar metaplastic cell states across diverse pancreatic disease models and reveals a conserved plasticity program. We uncover previously unrecognized metaplastic states, including a transitional “gateway” state that precedes advanced metaplasia. We also find that classical- and basal-like pancreatic cancer programs and coordinated immune remodeling emerge early in precursor lesions, well before invasive cancer develops.

What You Need to Know.

Background

Acinar-to-ductal metaplasia is a cellular response to pancreatic injury that can support regeneration or lead to cancer. The underlying responses and cell states across exocrine disease contexts remain poorly understood.

Impact

We uncover novel metaplastic states and a plasticity program conserved across pancreatic disease. We reveal cancer subtype specification early in pancreatic intraepithelial neoplasia precursors and coordinated transcriptional shifts in epithelial and immune compartments.

Future Directions

These findings establish a framework to study mechanisms of acinar plasticity and progression to cancer, advancing efforts to develop strategies for early detection, risk stratification, and interception of pancreatic cancer.

Metaplasia is a stress-induced change in cell identity. Across organs including the stomach, lung, and pancreas, metaplasia allows epithelial cells to acquire distinct differentiated features as an adaptive response to inflammation or stress.¹ In the pancreas, acinar cells exposed to stressors such as oncogenic mutations or inflammation acquire duct-like characteristics in a process classically termed acinar-to-ductal metaplasia (ADM). Mechanistically, ADM can be driven by inflammatory cytokines or stress that activate nuclear factor-κB (NF-κB), Janus kinase (JAK)-signal transducer and activator of transcription (STAT), NOTCH, and epidermal growth factor receptor (EGFR)-mitogen-activated protein kinase (MAPK) signaling pathways,2, 3, 4 which suppress acinar programs and transcription factors including pancreas transcription factor 1a (PTF1A), MIST1, and nuclear receptor subfamily 5 group A member 2 (NR5A2) and induce ductal regulators like SRY-box transcription factor 9 (SOX9) and HNF1 homeobox B (HNF1B).5, 6, 7, 8, 9 This transition can be reinforced by metabolic and epigenetic reprogramming that stabilizes the ADM state.10, 11, 12

This plasticity can serve as a protective mechanism, allowing acinar cells to transiently adopt more resilient metaplastic states that mitigate damage and promote regeneration. However, persistent metaplasia, as seen in the context of oncogenic KRAS mutations, can give rise to precancerous pancreatic intraepithelial neoplasia (PanIN) lesions that may progress to aggressive, often fatal pancreatic ductal adenocarcinoma (PDAC). ADM can be observed in diverse exocrine disease contexts including acute pancreatitis (AP), chronic pancreatitis (CP), and oncogenic KRAS-driven cellular stress—all of which are associated with elevated risk of PDAC13, 14, 15—reinforcing the potential of ADM to serve as a precursor to neoplasia.

Recent single-cell and lineage tracing studies have revealed that ADM encompasses a spectrum of plasticity states.16, 17, 18, 19, 20, 21 These include gastric-like (pit and neck), chief-like, enteroendocrine (EEC), Tuft-like, proliferative, and senescent populations, each defined by distinct transcriptional programs.16, 17, 18, 19, 20, 21, 22, 23 Among these, EEC-like cells are thought to secrete factors that promote PanIN growth,²⁴ whereas Tuft cells play a more complex and context-specific role, having protective effects in some settings,^16,19 yet capable of malignant transformation in others.^17,20 Gastric pit-like cells have been detected in precancerous and cancerous lesions.¹⁸ The remaining ADM populations have been reported in various models of chronic pancreatitis and oncogenic KRAS-driven metaplasia,^17,18 yet comparative analysis of these cell states across disease contexts has not been performed.

The states and fates of ADM populations are shaped by the environmental context. Transient ADM seen during acute injury typically resolves as inflammation subsides and acinar identity is restored, whereas persistent inflammatory signaling or oncogenic KRAS activation promotes prolonged ADM that can progress to PanIN.²⁵ This raises the critical question of whether transient ADM that supports regeneration and persistent ADM that precedes neoplastic transformation represent distinct, context-specific programs of metaplasia.

Understanding acinar metaplasia across disease contexts has been challenged by the high levels of digestive enzymes physiologically produced by acinar cells, which interfere with transcriptome quality. This technical barrier extends to ADM populations, which retain high enzyme expression. Although prior studies have characterized the metaplastic response in individual disease models,^17,18,26 it remains difficult to determine whether differences reflect biology or technical variability in cell capture. To overcome these challenges, we leverage the recently pioneered single-cell approach, “FixNCut,” that achieves unprecedented transcriptome quality from exocrine pancreas, including RNase-rich acinar cells and immune cells.²⁷ We investigate acinar plasticity across exocrine pancreas diseases that increase risk of PDAC, enabling direct comparison of plasticity states and deeper insight into the biology of metaplasia in disease and precancerous contexts. We identify a common acinar plasticity response across diverse exocrine pancreas diseases. We also find that features of advanced pancreatic cancer—including molecular subtypes and immune evasion—are observed during the formation of precursor lesions. By mapping these early events, we advance understanding of how epithelial metaplasia and the microenvironment converge to drive pancreatic disease and tumor initiation.

Results

FixNCut Enables Comprehensive Characterization of Acinar Metaplasia Across Diverse Pancreatic Injury Conditions

To investigate acinar metaplasia, we leveraged FixNCut to transcriptionally profile 4 exocrine disease contexts—AP, recurrent acute pancreatitis (RAP), CP, and acute pancreatitis in the setting of an oncogenic Kras mutation (K-AP), where ADM is known to progress to PanINs. Using established cerulein-injection protocols, we profiled the pancreas over time across these conditions (Figures 1 and 2). For consistency, RAP (2-, 4-, and 6-week) and CP (4-week) samples were collected 3 days after the final cerulein injection. AP samples were collected until 14 days post-injury, and K-AP samples were collected out to 12 weeks to capture the full trajectory of ADM to PanIN formation.

Figure 1.

Experimental models and cellular composition of exocrine pancreas disease. (A) Schematic illustrating acinar cell metaplasia, which occurs during cellular stress such as pancreatitis or an oncogenic Kras^G12D mutation. Under chronic stress, ADM can progress to PanIN formation. (B) Summary of disease models and associated epithelial states (acinar, ADM, PanIN), with arrows indicating the direction of acinar cell fate changes. Experimental design in mouse models showing timepoints for sample collection and cerulein injections (red arrows); asterisks indicate timepoints analyzed 3 days after the final injection. (C) UMAP projections of all cells from each condition (AP, RAP, CP, and K-AP), colored by cell type annotation, with the number of cells indicated for each condition.

Figure 2.

Histological progression of exocrine pancreatic injury and metaplasia across disease models in mice. (A–D) Representative H&E-stained pancreatic tissue sections from each timepoint of (A) AP, (B) Kras^G12D + AP, (C) RAP, and (D) CP mouse models. Timepoints reflect days or hours post-injury. In RAP and CP, pancreata were harvested 3 days after a final cerulein dose.

This experimental setup generated a comprehensive single-cell dataset of over 300,000 total cells across the 4 disease models. We integrated epithelial cells from all conditions, initially focusing on acinar, ductal, ADM, and PanIN populations (Figure 3A and B). This allowed direct comparison of heterogeneous ADM states across multiple exocrine pancreas disease contexts and their evolution over time, revealing shared and divergent metaplastic pathways.

Figure 3.

Single-cell analysis reveals common ADM programs across exocrine pancreatic diseases. (A and B) UMAP of epithelial cells from all conditions, color-coded by cell type (A) or disease (B). (C) Heatmap showing scaled expression of acinar, ductal, and ADM-associated genes across epithelial states. (D) Schematic illustrating the early ADM response to injury. (E) Z-scores of the ADM Index (ADMI) markers in non-ductal epithelial populations (acinar, ADM, and PanIN cells) across AP and RAP conditions. (F) Dot plot showing the expression of individual ADMI markers, including Reg and ductal genes as well as non-canonical trypsinogens, across non-ductal epithelial cells from AP and RAP samples.

Diverse Biological Stressors Deploy Common Acinar Metaplasia Responses

Immediately after injury, acinar cells transition into Acute-ADM at day 1 followed by Resolving-ADM at day 3. These closely related states exhibit reduced expression of acinar genes and induction of ductal genes, as well as upregulation of noncanonical trypsinogens (Prss1, Prss3) and Reg genes (Reg2 and Reg3 family) (Figure 3C and D). By day 3, we also observe a distinct Acinar-Reg population, marked by high Reg genes and notably the absence of ductal and noncanonical trypsinogen expression, distinguishing it from both Acute-ADM and Resolving-ADM. The Acinar-Reg state has been previously reported in mouse and human in both healthy and PDAC contexts, and it is the only acinar subtype observed in human CP.^28,29

Following a single episode of AP, the pancreas demonstrates remarkable regenerative capacity. By day 14, transcriptional and histologic profiles return to baseline, consistent with full recovery and restoration of acinar identity (Figure 2, Figure 3D, and 4A).²⁷ Notably, relative to a single episode, RAP elicits a progressive increase of ADM markers²⁷ with each additional episode, suggesting escalating impact of injury (Figure 3E and F).

Figure 4.

Advanced-ADM states observed across pancreas diseases. (A) Bar plots (left) show Advanced-ADM states by relative proportion of epithelial cell type by sample. Bubble plot (right) shows relative proportion for Advanced-ADM populations only, with size of dot proportional to population percentage. ∗ indicates states observed in subset of samples: EEC after RAP 6× (6 weeks), Tuft cells after K-AP 14 days. ∗∗ denotes a transcriptional PanIN state in the absence of histologic lesions. (B) Hierarchical clustering and transcriptional correlation of epithelial states with yellow box denoting Advanced-ADM populations. (C) Dot plot showing the expression of marker genes across epithelial subtypes.

Whereas Acute-ADM and Resolving-ADM represent an adaptive response to AP (Figure 3D), recurrent or persistent injury—as in CP, RAP, and K-AP—drives a transition into Advanced-ADM states, characterized by near-complete loss of acinar identity and increased transcriptional similarity to ductal and PanIN states (Figure 4A and B; Supplementary Tables 1 and 2). In addition to previously described metaplastic states including EEC-like and Tuft-like cells,16, 17, 18, 19, 20, 21 our study identifies additional Advanced-ADM subtypes: polo-like kinase 2 (Plk2)⁺ ADM, syntabulin (Sybu)⁺ ADM, and gateway ADM (gADM). These Advanced-ADM subtypes are shared across conditions, although their prevalence varies: CP and K-AP exhibit all advanced ADM subtypes, whereas RAP lacks Tuft cells (Figure 4A). Consistent with prior reports, we identified Tuft and EEC cells in CP and K-AP, and notably extend these observations to recurrent AP, where EEC cells were detected by single-cell RNA sequencing (scRNA-seq) and confirmed at the protein level by immunohistochemistry (IHC) (Figure 5A), suggesting EEC differentiation occurs broadly in sustained inflammatory conditions. In K-AP, Tuft and EEC cells were predominantly localized within PanIN lesions (Figure 5B). Interestingly, in CP, PanIN-like transcriptional states were observed even without histologic PanINs, indicating that sustained inflammation alone can drive precursor-like reprogramming (Figure 5C). Together, these findings suggest that sustained inflammatory stress—independent of oncogenic KRAS status—is sufficient to elicit common metaplastic trajectories.

Figure 5.

Validation of metaplastic cell types and PanIN-like transcriptional programs. (A) Representative IF of RAP 6-week pancreas shows presence of EEC cells marked by CHGA (yellow) within REG3B+ (ADM, red). (B) Representative IF of K-AP pancreas tissue at 6 weeks, stained for REG3B (ADM, red), DCLK1 (Tuft cells, green), and CHGA (yellow). (C) PanIN-like cells marked by SERPINB5 (green) are detected in CP and early K-AP (72-hour) samples. These SERPINB5+ cells co-localize with REG3B (yellow) and are highlighted by arrowheads. All sections are counterstained with DAPI (blue).

Following the immediate response to injury, cells in Acute-ADM diverge along 2 broad programs. One is a recovery-associated program, characterized by elevated expression of noncanonical trypsinogens and Reg family genes (Figure 6), which remain elevated in Resolving-ADM populations. The other is a reprogramming-associated program, marked by activation of immune, metabolic, and cytoskeletal remodeling genes, which are rapidly downregulated in Resolving-ADM, but persists in Advanced-ADM and PanIN states. These findings suggest that sustained early reprogramming responses to injury may contribute to neoplastic progression.

Figure 6.

Injury-induced gene program and a transitional gADM state. (A) Line plot with scaled expression of gene programs upregulated in Acute-ADM, indicating recovery-associated (blue and purple) and reprogramming-associated (orange, red, and yellow) programs across pseudotime. Cell type color code shared with Figure 4A. (B) Dot plot with expression of injury-induced genes, grouped by functional category.

Oncogenic KRAS Drives Early Reprogramming and Defines Unique PanIN Biomarkers

Acinar metaplasia initially manifests as a response shared across pancreatic disease models, but in the presence of oncogenic KRAS it is rapidly redirected toward a precancerous PanIN fate. In the K-AP model, histologic PanINs begin to appear at day 7 and expand markedly by day 14 and beyond (Figure 2). Strikingly, however, a PanIN transcriptional signature is detectable as early as day 3 (Figure 5C), well before histologic lesions appear, indicating that oncogenic KRAS initiates preneoplastic reprogramming early in the metaplastic response.

Comparative analysis of ADM reveals that 14 hours following injury, ADM cells activate acute stress and injury pathways, including mechanistic target of rapamycin (mTOR), unfolded protein response, p53, and oxidative phosphorylation pathways (Supplementary Table 3). However, cells harboring oncogenic KRAS show greater activation of interleukin (IL)-2/STAT5, transforming growth factor beta (TGFβ), and tumor necrosis factor alpha (TNFα) via NF-κB, and Notch pathways, suggesting early reprogramming for sustained metaplastic progression. By 72 hours, the divergence between K-AP and AP becomes more pronounced, with AP ADM shifting to repair and cell cycle programs, whereas K-AP ADM maintains activation of p53, TGFβ, apoptosis, and TNFα via NF-κB, hypoxia responses, and IL-2/STAT5 (Supplementary Table 3). These sustained signatures reflect an emerging pre-neoplastic state, presaging the earliest histologic evidence of PanINs. In the absence of oncogenic KRAS, prolonged inflammation during CP induces ADM, and cells instead activate p53 and ‘KRAS down’ programs, suggesting engagement of tumor-suppressive barriers to PanIN formation (Supplementary Table 3).

We next investigated whether oncogenic KRAS-specific markers could be identified within ADM states. Strikingly, despite pathway-level distinctions, no K-AP specific markers emerged during ADM, underscoring a key biological insight: diverse stressors, including oncogenic KRAS, deploy a common metaplastic response. Only at the PanIN stage did we observe oncogenic KRAS-specific biomarkers, highlighting PanINs as a state where oncogenic programs diverge from injury associated ADM programs. Importantly, we identified Mcpt1, Mucl3, and Muc5ac to be uniquely expressed in PanINs and absent in non-oncogenic pancreas tissue (Figure 7A–C). We validated these findings across multiple independent single-cell datasets spanning diverse pancreas diseases models (Figure 7B).^17,18,30 These candidate biomarkers could aid in distinguishing oncogenic KRAS-driven precancerous lesions from other pancreatic injury contexts, with implications for early detection and clinical risk stratification.

Figure 7.

PanIN biomarkers across cell types and datasets. (A) Dot plot showing expression of oncogenic Kras biomarkers across cell types in K-AP. (B) Dot plot showing expression of same biomarkers across epithelial cells from multiple exocrine pancreas disease datasets, including the current study and published datasets. (C) UMAP plots showing biomarker expression in epithelial cells, highlighting their specificity in PanIN lesions. (D) Bar plot of acinar and acinar-derived metaplastic cells from K-AP ordered by pseudotime, reflecting transcriptional similarity. Cell type color code shared with Figure 4A. (E) Heatmap of TF activity (Z-score) across acinar, ADM, and PanIN populations. (F) mKate2 expression on UMAP. (G and H) PAGA pseudotime analysis (G) and RNA velocity analysis (H) identify gADM as a key node between acinar-like and Advanced-ADM states. (I) Dot plot of gADM marker genes across epithelial subsets.

Identification of Gateway ADM as a Transitional State in Pancreatic Metaplasia

We next focused on a previously unrecognized transitional state—gADM—which marks an inflection point along the ADM continuum. Pseudotime analysis revealed that gADM and upstream states (Acute- and Resolving-ADM, Acinar-REG) retain acinar identity, whereas downstream Advanced-ADM states lose acinar markers and progressively acquire PDAC-associated signatures (Figure 7D). Transcription factor (TF) activity corroborates this shift, with a decline in acinar identity TFs (PTF1A) and a rise in oncogenic reprogramming TFs beyond the gADM state (Figure 7E).

To elucidate relationships between metaplastic states, complementary RNA velocity and pseudotime connectivity analyses was performed, focusing on acinar-derived (mKate2⁺) cells (Figure 7F), which showed gADM bridging early, acinar-like ADM populations (Acute- and Resolving-ADM) and Advanced-ADM subtypes, further supporting that this state may serve as a pivotal gateway to further metaplastic progression (Figure 7G–I).

We validated the presence of gADM in both mouse and human pancreas tissue by demonstrating co-expression of markers associated with both acinar identity and Advanced-ADM states, consistent with its role as a transitional population (Figure 8A–C).

Figure 8.

Analysis of gADM. (A) CosMx spatial transcriptomics of human pancreas with metaplasia shows co-expression of acinar (pink) and ADM/PanIN (orange) markers. DAPI (nuclei, blue), PanCK (ductal/PanIN structures, green), and cell segmentation (cyan) determined by CosMx cell segmentation algorithm. Enlarged view (lower) highlights co-expressing cells (arrowheads). (B) IF of mouse tissue confirms the gADM population co-expresses acinar (PRSS, yellow) and advanced ADM (EGFR, green) markers. Samples from 72 hours post-injury in AP, RAP, CP, and K-AP shown. (C) Higher magnification of K-AP with arrowheads highlighting examples of co-expressing gADM cells. (D) Pearson correlation matrix comparing transcriptional similarity of gADM populations across timepoints in AP and K-AP. (E) PCA of gADM samples by individual mouse, colored by condition and timepoint. (F) Monocle UMAP projections of acinar-derived cells from early timepoints (14 hours and 72 hours) in AP (left) and K-AP (right).

Notably, the fate of cells reaching the gADM state depends strongly on biological context, with resolution in AP and progression in K-AP. Surprisingly, despite these divergent outcomes, static transcriptomic profiles of gADM cells in AP and K-AP are remarkably similar (Figure 8D and E). This raises the possibility that transcriptomic momentum,³¹ rather than gene expression alone, may distinguish recovery from transformation.

To investigate the dynamics of gADM, we employed Monocle, which infers cell-state trajectories by ordering cells based on transcriptional similarity. To capture condition-specific dynamics without interference from later-stage metaplasia, we performed analyses separately for AP and K-AP, using only early timepoints. In AP, gADM lies between Acute- and Resolving-ADM (Figure 8F). In contrast, K-AP gADM occupies an intermediate position between Acute- and Advanced-ADM states, supporting its proposed role as a ‘gateway’ to further metaplasia and PanIN formation.

RNA velocity analysis, which models cell state dynamics based on unspliced vs spliced transcripts, further supports this model: gADM cells in K-AP exhibit strong velocity vectors towards advanced ADM, suggesting more commitment toward pathologic progression (Figure 9), whereas in AP, velocity is minimal, consistent with resolution (Supplementary Table 4).

Figure 9.

Divergent transcriptional dynamics of gADM in pancreas diseases. (A) RNA velocity analysis comparing all conditions. (B and C) Select K-AP velocity genes grouped by TFs, tumor suppressors, and chromatin regulators, showing UMAP expression patterns (top) and spliced-to-unspliced RNA velocity plots for K-AP (middle) and AP (bottom), with gADM cells highlighted in red.

To gain insight into candidate genes potentially driving these divergent gADM trajectories, we identified the top ranked genes contributing to K-AP RNA velocity signals. Genes with high levels of unspliced to spliced transcript ratios typically reflect recent transcriptional activation, and although these genes cannot conclusively be designated as drivers, they represent candidate regulators influencing gADM fate decisions. In K-AP gADM, top velocity genes included several TFs and coactivators involved in lineage and identity regulation, as well as epigenetic modulators involved in chromatin remodeling (Figure 9). Interestingly, several tumor suppressors also emerged as top-ranked velocity-associated genes, potentially representing a protective feedback mechanism activated in early metaplastic contexts. Notably, many of the K-AP gADM velocity-associated genes are preferentially expressed in Advanced-ADM, reinforcing their potential roles in advanced metaplastic fates. In contrast, AP gADM cells exhibit fewer high-velocity genes, most of which were broadly expressed in acinar cells, consistent with a trajectory of recovery rather than transformation.

Together, these findings establish gADM as a pivotal transitional state in the metaplastic continuum. Our data nominate specific TFs and coactivators as well as chromatin remodelers as regulators of progression to advanced metaplastic states in the context of oncogenic KRAS.

Advanced-ADM Subtypes Exhibit Distinct Molecular Profiles and Trajectories in Pancreatic Disease

Having established Advanced-ADM as a collection of metaplastic states downstream of gADM, we next sought to understand how these populations contribute to PanIN development. Using Monocle-based pseudotime analysis, we re-clustered cells within the Advanced-ADM and PanIN compartment to resolve metaplastic subtypes and their transcriptional trajectories (Figure 10A and B).

Figure 10.

Transcriptionally distinct Advanced-ADM subtypes reveal novel reprogramming states. (A and B) UMAP of Acute-ADM and Advanced-ADM populations, color-coded by cell-type (A) or disease (B). (C) IF staining validates Plk2⁺ ADM, marked by ARC (green) in RAP, CP, and K-AP pancreas. REG3B (yellow) marks ADM, and co-localization is indicated by arrowheads. (D) IF validates a distinct Sybu⁺ ADM population, marked by STNB1 (green), which is similarly observed across RAP, CP, and K-AP tissue. Arrowheads highlight STNB1+ REG3B+ cells. All samples are counterstained with DAPI (blue).

This analysis revealed multiple transcriptionally distinct branches of Advanced-ADM, including known EEC and Tuft cell populations, as well as 2 previously unreported subtypes: Plk2⁺ ADM and Sybu⁺ ADM, which we identified using single-cell transcriptomics and confirmed via IHC in mouse pancreas tissue (Figure 10C and D).

The ‘Plk2⁺ ADM’ population is marked by high expression Plk2, a known p53 target (Figure 11A). In addition to high p53 activity (Figure 11B), these cells show strong signatures for stress-associated programs, including DNA damage response, metabolic rewiring, and epigenetic remodeling, which was also reflected in pathway activation (Figure 11C–F). Furthermore, Plk2⁺ ADM exhibited elevated TF activity linked to unfolded protein response and cell stress (Xbp1, Atf4, Smad2, Klf6) (Figure 7E). These stress-associated Plk2⁺ ADM cells do not appear to directly progress to PanIN lesions, instead positioned as a distinct branch point within the ADM continuum consistently observed across RAP, CP, and K-AP conditions. These findings suggest that Plk2⁺ ADM represents a conserved, high stress-associated metaplastic state.

Figure 11.

Plk2⁺ and Sybu⁺ ADM populations. (A) Plk2 expression on UMAP; arrowhead indicates Plk2⁺ ADM population. (B) Violin plot showing p53 activity scores across ADM subtypes. (C) Dot plot of representative marker genes enriched in Plk2⁺ and Sybu⁺ ADM populations, grouped by functional category. (D) Heatmap of GO biological process (BP) enrichment scores across epithelial subtypes. (E) UMAPs showing expression of select neural-associated genes across epithelial subtypes. (F) UMAP projections showing PROGENy pathway enrichment scores for select pathways. (G) Sybu expression on UMAP; arrowhead indicates Sybu⁺ ADM population. (H) Violin plot showing NRP scores across ADM and PanIN populations. (I) UMAP of epithelial cells colored by NRP score.

In contrast, the ‘Sybu⁺ ADM’ population exhibits a distinct transcriptional signature suggestive of reprogramming towards a neural-like phenotype (Figure 11G). These cells uniquely upregulate Sybu and other neural-associated genes and pathways (Figure 11C–F). To investigate the neural aspects of this population further, we scored cells against the recently characterized ‘neural-like progenitor’ (NRP) signature identified in PDAC.²⁹ Although Sybu⁺ ADM harbor an ADM identity and thus do not fully recapitulate all NRP markers, they consistently exhibit elevated NRP scores (Figure 11H and I). This suggests that Sybu⁺ ADM cells are partially reprogrammed towards a neural-like identity but are not fully differentiated NRPs. Thus, Sybu⁺ ADM constitutes another distinct trajectory within Advanced-ADM, observed across various contexts of prolonged pancreas injury (RAP, CP, and K-AP).

Monocle trajectory analysis suggests that neither Plk2⁺ nor Sybu⁺ ADM populations directly transition into PanIN. Instead, PanINs arise from a more truncal state of gADM that precedes fate commitment, suggesting that only less differentiated ADM cells retain neoplastic potential. Although not reported previously, Plk2⁺ and Sybu⁺ ADM populations are detectable in independently published CP and K-AP single-cell datasets (Figure 12), supporting their reproducibility across studies.

Figure 12.

Cross-dataset validation of Plk2⁺ and Sybu⁺ populations in CP and Kras^G12D models. (A and B) UMAPs and representative marker gene expression from published single-cell datasets of (A) CP (GSE172380) and (B) Kras^G12D (GSE141017). (C) Expression of Plk2⁺ ADM (top) and Sybu⁺ ADM markers (bottom) across the current study and published datasets, showing conservation of these transcriptionally distinct ADM subtypes across inflammatory and oncogenic contexts, localized to a subset of cells (indicated by arrowheads).

Together, these findings reveal distinct and previously unrecognized metaplastic states embedded within the broader Advanced-ADM continuum.

Identification of Distinct PanIN Subtypes Underscores Precancerous Heterogeneity

PanIN lesions have been extensively described in human samples and mouse models, yet their cellular heterogeneity at the single-cell level remains poorly characterized. In human studies, limited by availability of biopsy or cadaveric specimens, spatial transcriptomics have determined that PanINs have signatures of the ‘Classical’ PDAC subtype.^32,33

Our single-cell analysis uncovers previously unrecognized transcriptional heterogeneity within PanIN lesions. Unexpectedly, we detect distinct ‘Classical-like’ and ‘Basal-like’ PanIN populations, by scoring individual cells against established PDAC classification signatures (Figure 13A and B). We validated using IHC in K-AP, showing spatially distinct expression of classical (TFF1) and basal (KRT17) markers, with individual PanINs often containing both subtypes side-by-side (Figure 13C). Using CosMx spatial transcriptomics, we confirm that this PanIN heterogeneity is conserved in human tissue based on subtype-specific markers of classical and basal states (Figure 13D–F). ‘Classical-like’ PanINs are enriched for known classical-associated TF Gata6, whereas ‘Basal-like’ PanINs are enriched for EMT-related pathways and TF Snai1, known features of the basal, mesenchymal-like state (Figure 14A). Although field effects cannot be excluded in non-cancerous PDAC-adjacent samples, CosMx nonetheless validates the spatial heterogeneity of PanIN subtypes in the human pancreas.

Figure 13.

PanIN lesions exhibit transcriptional heterogeneity with classical- and basal-like subtypes. (A) UMAP highlighting Advanced-ADM and PanIN states. (B) Heatmap of PDAC subtype classification signature scores across epithelial cell types. (C) Detection of TFF1 and KRT17 markers in 6-week K-AP PanINs by IHC. (D) CosMx spatial transcriptomics of human PanIN showing gene signatures corresponding to classical and basal PDAC subtypes (left). Zoomed-in view of a PanIN lesion with CosMx cell segmentation overlay (right). (E) Expression of classical and basal markers from the same human PanIN region. Markers shown correspond to the combined signature scores in (D). For (D and E), cells with a green overlay represent those in the unsupervised InSituType cluster identified as ‘PanIN’ by transcriptomic marker genes and cell morphology. Gray cells belong to other phenotype clusters. Cyan outlines indicate cell segmentation boundaries determined by the CosMx cell segmentation algorithm. DAPI (blue) marks nuclei, and PanCK (green) labels ductal/PanIN structures. (F) Additional representative regions of human PanINs showing merged expression of basal and classical markers.

Figure 14.

PanIN subtype characterization. (A) Hallmark pathway enrichment (top) and TF activity (bottom) projected onto UMAP. (B) UMAPs showing normalized expression of metaplastic markers within PanINs. (C) Dot plot showing expression of marker genes across epithelial cell and PanIN subtypes.

Beyond the classical-basal dichotomy, we observe further heterogeneity consistent with previously described metaplasia programs, including spasmolytic polypeptide-expressing metaplasia (SPEM)/gastric pit-like, gastric neck/SPEM-like, and chief-like (Figure 14B).^18,22 We also identify a population of cycling PanIN cells (Figure 14C).

These data uncover previously unrecognized transcriptional heterogeneity in pancreatic neoplasia, with both classical- and basal-like programs already present at the PanIN stage. This early dichotomy suggests that the well-known PDAC subtypes may originate during precursor formation, well before initiation of invasive cancer, raising the possibility that precursor subtype may influence future clinical behavior. These findings may have implications for precursor monitoring as the basal PDAC subtype is typically more therapy resistant.³⁴ To our knowledge, this is the first report of a single-cell classical vs basal dichotomy in early PanINs.

Conserved Early Microenvironmental Responses to Pancreatic Injury

The transition from reversible metaplasia to precancerous lesions is not solely dictated by epithelial-intrinsic programs, but is critically shaped by cues from the surrounding microenvironment. To understand this process, we characterized immune and stromal populations across injury models that resolve (AP), recur (RAP), persist (CP), or progress to precancerous lesions (K-AP) (Figure 15).

Figure 15.

Immune and stromal populations characterized in pancreas injury. (A and B) UMAP of all immune cells across conditions, color-coded by cell-type (A) or disease (B). (C) Dot plot showing expression of marker genes used to define cell types. (D) UMAPs showing expression of representative marker genes for macrophage subsets.

Despite ultimately divergent outcomes, the initial response to pancreatic injury in AP and K-AP unfolds in a remarkably conserved manner, suggesting that a regenerative program dominates the initial tissue response. At 14 hours post-injury, both conditions exhibit robust immune infiltration and comparable fibroblast transcriptional profiles (Figure 16). The early immune response includes neutrophils, Arg1⁺ and cycling macrophages, dendritic cells, and T cells. These results indicate that the early pancreatic microenvironment response to injury is consistent with a shared regenerative program.

Figure 16.

Immune and fibroblast responses following injury in AP and K-AP. (A) UMAP projections of immune cells by timepoint and condition, with populations annotated as in Figure 15A. (B) UMAP projections of fibroblasts by timepoint and condition.

Oncogenic KRAS Hijacks Tissue Repair to Trigger a Second Wave of Inflammation

Following early inflammation, both AP and K-AP show signs of resolution: Arg1⁺ macrophages shift to a Trem2⁺ phenotype by 24 hours and neutrophils clear by 72 hours (Figure 17). In AP, this resolution continues over the next week (5 and 14 days), with sustained decline in inflammation and restoration of acinar cells. In contrast, metaplasia in K-AP persists, and by day 7, a ‘second wave’ of inflammation emerges, marked by renewed infiltration of neutrophils, Arg1⁺ macrophages, and cycling macrophages (Figure 18). The second inflammatory wave was validated by IHC, which showed reproducible neutrophil clearance by 72 hours, followed by a striking resurgence at day 7 (Figure 18C).

Figure 17.

Proportional shifts in immune cell types across pancreatitis models and timepoints. (A–D) Stacked bar plots showing the relative abundance of immune cell types across timepoints in (A) AP, (B) Kras^G12D + AP, (C) RAP, and (D) CP. UMAPs are shown alongside RAP and CP samples to illustrate immune population distributions on the UMAP from Figure 15A.

Figure 18.

Second wave of inflammatory mediators in K-AP. (A) Quantification of acute inflammatory populations over time in AP (top) and K-AP (bottom), highlighting the emergence of a “second wave” of neutrophil and macrophage infiltration in K-AP. (B) Dot plot showing expression of inflammatory mediator genes across macrophages, fibroblasts, and endothelial cells. (C) IHC staining of LY6G (neutrophils) by timepoint in K-AP tissue histologically validates the second wave.

Notably, day 7 also corresponds to the first histologic appearance of early PanIN lesions. During the second wave, macrophages, fibroblasts, and endothelial cells in K-AP re-express inflammatory mediators initially induced in early inflammation including Ccl2/7/8, Cxcl1/2/5, and Tgfb1 (Figure 18B). Thus, the ‘second wave’ reflects a partial reactivation of the original injury response. However, this reactivation occurs in an altered tissue environment, now dominated by oncogenic KRAS signaling and widespread metaplasia, favoring sustained inflammation and epithelial plasticity.

Following injury, pancreatic fibroblasts in both AP and K-AP activate inflammatory and myofibroblast programs, with markers that are shared with iCAF and myCAF (‘cancer-associated fibroblasts’) populations in PDAC, highlighting that these markers are not exclusive to the tumor microenvironment, but can also be broadly injury-inducible (Figures 16B and 18B). During the K-AP ‘second wave’ 7 days following injury, these fibroblast transcriptional states are re-engaged. Thus, persistent epithelial plasticity and oncogenic KRAS signaling may promote these fibroblast programs beyond the normal resolution phase, fostering a pro-fibrotic and pro-tumorigenic microenvironment.

In summary, a second wave of inflammation unique to the oncogenic KRAS context reflects a partial reactivation of the initial injury response. Now in the setting of widespread metaplasia and oncogenic KRAS, this reactivation may promote sustained plasticity, inflammation, and precursor formation.

The Emergence of an Immunosuppressive Precancerous Niche

A hallmark of PDAC and PanINs is an immunosuppressive microenvironment, characterized by T cell dysfunction, expansion of regulatory populations such as regulatory T (Tregs) and myeloid-derived suppressor cells (MDSCs), and immune checkpoint upregulation.³⁵ TGF-β1––a potent immunosuppressive cytokine secreted by both epithelial and immune cells––plays a central role in establishing this niche.³⁶ Indeed, reflecting these established mechanisms, we observe upregulation of Tgfb1 in PanINs and in microenvironment populations during the second wave in K-AP (Figure 11F).

We next assessed whether this immunosuppressive program is initiated at the earliest stages of pancreatic tumorigenesis. To examine the adaptive immune response, we profiled 13,028 lymphoid cells across pancreatic disease states using FixNCut (Figure 19). Most notably, Tregs significantly expanded in K-AP starting at day 7 (Figure 19E), coinciding with both the second wave of inflammation and the emergence of histologic PanINs. In cancers, Tregs expansion often reflects persistent antigenic stimulation.³⁷ In this context, their expansion likely indicates ongoing epithelial injury and/or the emergence of neoantigens from developing PanINs. Although Tregs are essential for restraining excessive inflammation, they can also shield pre-neoplastic lesions from immune clearance, supporting an immune-tolerant microenvironment that can ultimately enable PanIN progression.

Figure 19.

Single-cell profiling of lymphoid populations reveals diverse T, B, NK, and ILC subsets across pancreatic diseases. (A and B) UMAP of lymphoid cells across all conditions, color-coded by cell-type (A) or disease (B). (C) Dot plot showing representative marker gene expression across identified lymphoid subsets. (D) UMAPs showing spatial expression of representative lymphoid markers. (E) Quantification of lymphoid cell-types by condition and timepoint.

Importantly, Treg enrichment is absent in all other pancreatic injury states where PanINs are not observed. Chronic pancreatitis, for example, contains abundant T cells, but exhibits a more balanced Treg-to-CD8⁺ T cell ratio and an elevated population of IL-17-producing γδ T cells (Figure 19D and E), indicating persistent inflammation rather than immune suppression.

The uniquely immunosuppressive microenvironment of K-AP has additional distinctions beyond Treg enrichment that set it apart from CP and RAP. Although CP and RAP are abundant in plasma cells (Figure 19A and B), these cells are notably rare in K-AP, suggesting that the Treg-rich milieu limits plasma cell accumulation, further reinforcing a tolerogenic setting conducive to tumor initiation. Additionally, dendritic cells (DCs) adopt an immunosuppressive state around day 7 in K-AP (Figure 20), marked by gene expression changes linked to apoptotic cell clearance, immune tolerance, and survival, along with a sustained increase in a tolerogenic DC score (Figure 20B and C).³⁸

Figure 20.

Populations of dendritic cells across inflammatory and precancerous conditions. (A) UMAPs of DC subtypes, color-coded by cell type (top) and by disease condition (bottom). (B) Heatmap showing normalized DC expression of early inflammatory and 7-day K-AP activated genes; tolerogenic DC score for each condition shown on the right. (C) UMAP showing expression of genes from the tolerogenic DC signature. (D) Marker genes for DC populations.

In sum, oncogenic KRAS, epithelial metaplasia, and emerging PanINs are associated with a shift in the immune landscape toward tolerance, marked by the accumulation of Tregs and immunosuppressive DCs—features that may contribute to or reflect early immune evasion.

Neutrophils have historically been challenging to isolate for single-cell analysis due to enzyme-rich granules.³⁹ Here, we leveraged FixNCut to capture 6192 neutrophils across exocrine pancreas diseases, allowing us to identify multiple transcriptional states (Figure 21). Following injury, we observe mostly Lrg1⁺ neutrophils, previously reported in both homeostatic and tumor contexts.³⁹ A smaller subset of neutrophils express cathepsins, complement genes, and Apoe, features associated with immunosuppressive monocytic MDSCs that can suppress T cell responses and promote tumor progression. Although the early immune landscape is broadly conserved between AP and K-AP, we observe subtle differences in neutrophil states in K-AP, with modest upregulation of interferon-induced transmembrane (Ifitm) genes and select pro-inflammatory and pro-survival genes (S100 family and Mcl1) (Figure 21B). These nuanced differences may reflect early cues of oncogenic KRAS influence, emerging within an otherwise shared regenerative response.

Figure 21.

Neutrophil heterogeneity and activation dynamics across inflammatory conditions. (A) UMAP showing transcriptionally distinct neutrophil states. (B) Violin plots comparing gene expression between AP and K-AP at 14 hours. (C) UMAPs showing temporal dynamics of neutrophil subtypes across AP and K-AP conditions. (D) Dot plot showing expression of marker genes across neutrophil subtypes. (E) UMAP overlaid with NF-κB pathway enrichment scores.

Corroborating a second wave of inflammation uniquely occurring in K-AP, neutrophils infiltrate robustly in both AP and K-AP but are largely cleared by 72 hours, consistent with normal resolution of inflammation. However, in K-AP, neutrophils re-emerge by day 7 (Figures 18A, C, and 21C)—coinciding with onset of the second wave of inflammation and first appearance of PanIN lesions. Even more, these second-wave neutrophils in K-AP are transcriptionally distinct, marked by Cxcl2 expression and strong NF-κB activation (Figure 21C and E). These neutrophils resemble tumor-associated neutrophils,^39,40 and suggest early reprogramming toward immunosuppressive and tumor-supporting states.

Taken together, our findings demonstrate how oncogenic KRAS co-opts an otherwise conserved regenerative pancreatic injury program to drive a second wave of inflammation, leading to the accumulation of Tregs, tolerogenic dendritic cells, and non-resolving macrophages that persist in the microenvironment and establish a lasting immunosuppressive niche. This environment is further reinforced by late-stage transcriptional shifts in fibroblasts and endothelial cells consistent with sustained immune suppression and tissue remodeling (Figures 18B and 22). Importantly, these microenvironmental changes arise alongside epithelial plasticity, with the emergence of gADM in all disease contexts. Prolonged or recurrent metaplasia leads to progression into Advanced-ADM states, with early divergence into classical- and basal-like PanIN subtypes in K-AP. Together, these findings highlight a coordinated reprogramming of epithelium and microenvironment that lays the foundation for pancreatic tumor initiation, setting the stage for functional interrogation of regulatory mechanisms that separate tissue repair from tumor initiation.

Figure 22.

Endothelial cell gene expression reveals dynamic states during inflammation, recovery, and precancerous progression. (A) Dot plot showing expression of endothelial marker genes associated with acute inflammation and recovery across all conditions. (B) Dot plot showing expression of control- and precancerous-associated endothelial genes across the same conditions.

Discussion

Pancreatic acinar metaplasia is a universal injury-induced response that, when sustained, can lead to neoplasia. Our single-cell time-course data across models of injury- and oncogene-driven pancreatic metaplasia reveal that these diverse exocrine insults all trigger a conserved program of acinar plasticity.

A key subpopulation that emerges across all disease contexts is gateway ADM, or gADM. Although gADM cells from regenerating (AP) and progressing (K-AP) conditions are transcriptionally indistinguishable at the static level, RNA velocity analysis reveals divergence in their trajectories, nominating gADM as a critical gateway for fate decisions in acinar plasticity. gADM cells in disease contexts with Advanced-ADM exhibit a transcriptional burst of chromatin remodelers, transcription factors, and tumor suppressors—programs expressed in Advanced-ADM and PanIN states. These findings identify gADM as a transitional intermediate, where fate is defined not by static transcriptional identity, but by the activation kinetics of key regulatory programs. We anticipate that epigenetic processes in part govern these activation kinetics, and thus integrating transcriptomic with epigenomic profiling will be instrumental in defining gADM trajectories across disease contexts, as well as understanding the regulatory barriers that separate regeneration from neoplastic progression.

Downstream of gADM, we identify additional distinct ADM states. Plk2⁺ ADM activates p53 targets and DNA damage checkpoints. Given that p53 is known to restrain both ADM and PanIN formation in the context of oncogenic KRAS,^41,42 Plk2⁺ ADM may function as a protective failsafe that diverts metaplastic cells from a PanIN fate. We also observe Sybu⁺ ADM, which acquires neuronal and developmental markers. Neuronal reprogramming appears to be a recurring feature in acinar-to-PDAC progression, as it is observed both in Sybu⁺ ADM and in the recently characterized neural-like progenitor PDAC state.^20,29

Importantly, Plk2⁺ and Sybu⁺ ADM appear distinct from the PanIN lineage in trajectory analyses and are also observed in non-neoplastic RAP and CP models. Although these observations raise the possibility that such metaplastic states represent adaptations to chronic epithelial injury and may serve as cellular safeguards against tumorigenesis, it is also possible that they contribute to cancer initiation. For instance, metaplastic EEC-like cells have been shown to secrete factors that stimulate PanIN growth,²⁴ whereas Tuft cells exhibit a more complex role—providing protection against tumorigenesis in some contexts,^16,19 yet also harboring the potential to transform into PDAC.^17,20 Together, these examples highlight the need to further investigate whether distinct metaplastic programs ultimately constrain or promote neoplastic progression.

Compared with AP in wild-type tissue, oncogenic KRAS alters epithelial responses and gives rise to additional ADM subtypes—including Plk2⁺ ADM, Sybu⁺ ADM, and EECs—yet the initial immune and stromal responses remain largely similar. This suggests that tissue injury response programs initially dominate over oncogenic KRAS effects. At day 7 following injury, oncogenic KRAS reactivates this injury-induced program but with additional tolerogenic features, including Cxcl2⁺ neutrophils resembling tumor-associated populations,^39,40 Tregs, and tolerogenic dendritic cells. This ‘second wave’ coincides with the onset of PanIN formation and likely further promotes tumor initiation.

Classical and basal molecular subtypes are well characterized in advanced PDAC, with basal states linked to poorer prognosis. Here, we show that this dichotomy emerges in the earliest stages of neoplasia—with classical- and basal-like transcriptional states already present within PanINs. Interestingly, a recent study by Söderqvist et al⁴³ also identified classical- and basal-like cell states in the pancreas preceding PDAC, with classical markers enriched in injured lobules that contained abundant ADM, and basal markers more prominent in desmoplastic, stroma-rich regions, complementing findings of our study. Early emergence of a basal-like program is consistent with prior lineage tracing studies showing that metaplastic cells can acquire epithelial-to-mesenchymal transition (EMT) features and disseminate to distal organs before the development of invasive PDAC,⁴⁴ suggesting that invasive potential may be encoded early in precancerous lesions, rather than acquired in late-stage PDAC. Whether these early PanIN subtypes remain stable as classical and basal-like states that give rise to corresponding PDAC subtypes or retain plasticity during tumor evolution remains an open question. Nevertheless, the early coexistence of divergent transcriptional programs within PanINs suggests that phenotypic heterogeneity—and potentially metastatic potential—is established far earlier than previously recognized.

PanINs are increasingly recognized as a common finding in the pancreas of otherwise healthy adults,⁴⁵ yet the features that distinguish indolent lesions from those that progress to PDAC remain poorly understood. Our work elaborates the diverse transcriptional landscape within acinar metaplasia and PanIN lesions, with cell state heterogeneity that may be critical for understanding early precursor biology and informing future efforts for early detection and risk assessment. These findings provide a biological foundation for future risk stratification approaches—moving beyond histologic grading to incorporate the enrichment of high-risk transcriptional programs. Shifting focus from lesion presence to cellular identity offers a framework for improving early detection and illuminating the trajectories that drive both regeneration and tumor initiation across exocrine pancreas diseases.

Materials and Methods

Mouse Models of Pancreatitis

Pancreatitis was induced in male mice fed ad libitum at 8 to 9 weeks of age by administering cerulein via intraperitoneal injection. All animal procedures were performed according to Institutional Animal Care and Use Committee and relevant guidelines. For AP, male C57BL/6 mice from The Jackson Laboratory were injected with cerulein (0.1 μg/g diluted in saline; Millipore-Sigma) on 2 consecutive days once every hour for 8 hours, as previously described.³ For RAP mice, this 2-day injection protocol was repeated weekly for 6 weeks.

Mice for K-AP were generated by crossing Ptf1a-CreERT/+, CAGs-LSL-rtTA-IRES-mKate2/+, and LSL-Kras^G12D, referred to as Kras^G12D mice. One week before AP induction, Kras^G12D mice were administered tamoxifen (0.25 mg/g; Cayman Chemical) delivered in corn oil via oral gavage for 3 consecutive days. For CP, mice were crossed to generate Ptf1a-CreERT/+, CAGs-LSL-rtTA-IRES-mKate2/+ (called mKate2 mice), to enable acinar cell lineage tracing by mKate2⁺ fluorescence. CP was induced via a 4-week protocol of cerulein injections (0.25 μg/g) administered 2 times per day for 5 days per week as previously described.¹⁷

Timepoints for collection are outlined in Figure 1B. RAP (2, 4, and 6-week) and CP (4-week) samples were collected 3 days following the final cerulein injection. In K-AP and AP conditions, the final injection of cerulein was defined as hour 0. To increase mKate2⁺ cells for CP and K-AP (14-day, 6-week, and 12-week timepoints), mKate2⁺ cells were sorted using a BD FACSAria III Sorter using the standard cell preparation outlined below.

FixNCut scRNA-seq: DSP Fixation Buffer and Sample Preparation

FixNCut scRNA-seq was performed by preparing a fresh 1 mg/mL solution of dithiobis (succinimidyl propionate) (DSP, Thermo Scientific) as previously described.²⁷ For each sample, a ∼3 × 3 mm piece from the head of the pancreas was fixed in 500 μL of DSP at room temperature (RT) for 30 minutes, followed by quenching with 10 μL of 1 M Tris-HCl pH 7.5. Pancreas tissue was dissociated using 1 mL of 200 μg/mL Liberase TM (Millipore-Sigma) in phosphate-buffered saline (PBS) at 37°C with agitation at 800 rpm for 30 minutes, followed by filtering the cell suspension (pluriStrainer Mini 70 μm #43-10070-40). Cells were washed 3 times in 10 mL of cold PBS with 1% bovine serum albumin (BSA; Millipore-Sigma) and pelleted using centrifugation at 500 × g for 5 minutes at 4°C. Cells were resuspended in solution with 2 U/μL RNase Inhibitor (Millipore-Sigma) and filtered using 40-μm Flowmi Cell Strainer (BAH136800040-50EA, Millipore-Sigma). Cells were counted using a LUNA-FL (LogosBiosystem) and loaded onto the 10x Chromium controller (10x Genomics) to achieve a target capture of 10,000 cells. Libraries were prepared using Chromium Single-Cell 3ʹ Reagent Kit V3.1 following manufacturer’s instructions. Libraries were sequenced using the Illumina platforms (100 cycles, paired-end).

Data Pre-processing and Clustering

Cell Ranger Count (v7.0.0/v8.0.1) was used to demultiplex FASTQ reads, and align to the mouse mm10 2020-A genome to generate a feature-barcode matrix for each sample. For samples with mKate2 transgene, counts were aligned to a genome with the mKate2 gene added. Downstream analysis was performed in python (v3.10) using rpy2 (v3.4.5) to integrate R tools. CellBender (v0.2.0)⁴⁶ ‘remove-background’ was used to estimate and remove ambient RNA to improve cell-type markers with AP previously described²⁷ and K-AP, CP, and RAP settings optimized (expected-cells = 10,000, total-droplets-included = 40,000, model = full, fpr = 0.01, epochs = 150). Filtering was performed to remove droplets with fewer than 1000 unique molecular identifier (UMI) counts or 200 genes, or greater than 20% of reads from mitochondrial genes (genes starting with “mt-”). We also removed droplets with greater than 100,000 counts and droplets detected as doublets by scDblFinder (v1.13.7)⁴⁷ using default parameters. Genes were filtered out if they were detected in fewer than 20 cells. Downstream analysis was performed using the python package scanpy (v1.9.3).⁴⁸ Normalization was performed with proportional fitting (PF), followed by log1p transformation and an additional PF, shown to be a superior normalization method for variance stability, depth normalization, and monotonicity, as well as minimize false-positive differentially expressed genes.⁴⁹

Dimensionality was reduced using principal component analysis (PCA) with scanpy ‘pca’ tool using default settings with n_comps = 30 and use_highly_variable = False. Timepoints were integrated using Harmony⁵⁰ via python compatible tools scanpy.external (v1.9.3) and harmonypy (v0.0.9) with parameters max_iter_harmony = 50, epsilon_cluster = 1e-4, and epsilon_harmony = 1e-5. Initial neighbors were identified using the scanpy ‘neighbors’ tool, followed by an identification of highly variable genes (HVGs) using Triku (v2.1.6)⁵¹ with n_windows = 100. Neighbors were recalculated with use_rep = “X_triku”. When calculating neighbors, metric = ‘cosine’ was used and defined the size of the local neighborhood, or number of neighbors using the Triku recommended formula:

$n\_neighbors=\lfloor (\frac{0.5\ast \left(\sqrt{n_{cells}}\right)}{2}\rfloor$

A Uniform Manifold Approximation and Projection (UMAP) was used to visualize cells in 2-dimensional space. Clustering was performed using scanpy ‘leiden’ tool and scanpy ‘rank_genes_groups’ was performed using default settings (method = ’wilcoxon’ and corr_method = ‘bonferroni’) on the normalized layer to find differentially expressed genes and cluster specific markers. Cell-types were determined using top cluster markers. Scanpy ‘score_genes_cell_cycle’ was used to determine the cell cycle stage for each cell.

Advanced-ADM, PanIN, and Acute-ADM populations were analyzed using the same pipeline described above up to HVG selection. For inferring lineage relationships, cell clustering was performed with triku selected HVGs using monocle3 (v1.3.7)52, 53, 54, 55 using 25 principal components and UMAP analysis was performed (umap.min_dist = 0.2, umap.n_neighbors = 75, umap.metric = 'manhattan'). Louvain based cell clustering was used to refine cell type naming, and clustering information was merged back with the original scanpy object for downstream analysis.

To integrate epithelial cells from AP, K-AP, CP, and RAP, several changes were made to the pipeline. To ensure only high-quality cells, filtering thresholds were increased to cells with >1700 read counts. Variance was stabilized across cells with varying sequencing depths by scaling cell counts by the median total count across all cells and applying a log(1+x) transformation. To prevent losing informative features from individual datasets, 1500 HVGs were identified separately for each dataset, which were subsequently merged to comprehensive set of variable genes for downstream PCA. The Batch Balanced K-Nearest Neighbors (BBKNN) algorithm⁵⁶ (key = “donor” for individual mice, neighbors_within_batch = 3) was used for batch correction as we found it better preserves local neighborhood structure and was more robust to sparse data compared to Harmony. A UMAP was generated for visualization and downstream analysis. Of note, one RAP replicate was removed from this epithelial analysis as there was only one epithelial cell that passed quality filtering.

Comparison to Other RNA-seq Datasets

scRNA-seq datasets for mouse pancreas metaplasia were downloaded for comparison, using National Center for Biotechnology Information (NCBI) Sequence Read Archive (SRA) to download GSE172380¹⁷ and GSE141017¹⁸ (control, 17-day, 6-week, 3-month). Acinar and acinar-derived cells were filtered based on marker expression and were processed and analyzed as described in the pipeline above, with GSE172380 analyzed using Monocle given that this was performed in the original paper. For biomarker analysis, the above datasets, as well as GSE207943,³⁰ were used after processing using the previously described pipeline.

Gene Scoring, Pseudotime Calculations, and PCA Plots

All single-cell scoring was performed using scanpy ‘score_genes’ (v1.9.3), which computes the average expression of gene sets relative to a reference distribution. Several curated gene lists used for scoring and analysis were employed, which can be found below. ADM Index genes, previously defined in acute pancreatitis ADM.²⁷ For scoring individual cells in Figure 6A, reported PDAC genes were used from mouse¹¹ and human,⁵⁷ as well as known mouse acinar genes.²⁷ Genes for tolerogenic DCs³⁸ and the NRP signature²⁹ were previously reported. Full gene lists can be found below. Classical and basal gene signatures were obtained from prior studies,58, 59, 60, 61 and gene lists were filtered for those present in mouse and our datasets.

To identify genes upregulated during acute ADM used in Figure 6A, average gene expression was calculated, and a series of filters was applied to identify genes with the highest expression differences, statistical significance, and magnitude of change, with the goal of identifying genes most robust induction during acute ADM over all disease contexts. These genes were grouped into functional categories for visualization in Figure 6A, with individual genes shown in Figure 6B.

For Figures 6A and 7D, only epithelial cells from K-AP unsorted pancreas samples were used. Pseudotime ordering was performed using the scanpy with standard settings, with diffusion maps (diffmap) construction, using ‘Acinar’ as the root cell for trajectory, and computing diffusion pseudotime (dpt) to order cells along the inferred trajectory. Cells were ordered by their pseudotime values for plotting. To generate a smoothed quantification reflecting the expression dynamics for acute ADM gene programs, the pseudotime scale was divided into 20 bins, and average expression of gene categories was computed for each bin using scanpy score_genes tool. Binned scores were plotted against pseudotime to visualize transcriptional changes with cellular state transitions.

For comparing cell types and timepoints, scanpy was used to generate dendrograms and correlation matrices using Pearson correlation as the distance metric. A custom tool for python (plot_confidence_ellipse) was used to make PCA plots to represent the variability and distribution of cellular clusters in 2-dimensional projection spaces. The tool fits a confidence ellipse around each cluster based on the covariance structure of the data points. Proportion plots were generated from whole pancreas epithelial samples rather than fluorescence-activated cell (FAC)-sorted mKate2⁺ cells to avoid bias from cell sorting skewing proportionality.

Pathway, TF, and Gene Set Analysis

To infer functional activity programs across epithelial cell types, we applied the Python implementation of decoupleR (v1.9.2),⁶² a framework for estimating pathway and regulator activity from transcriptomic data. All analyses were performed on normalized expression values stored in the specified AnnData layer. Inferred pathway and TF activity scores were visualized.

Over-representation analysis (ORA) was performed using Hallmark and Gene Ontology (GO) Biological Process gene sets from MSigDB to generate per cell enrichment scores based on the representation of gene sets among expressed genes. For pathway activity inference, a multivariate linear model (MLM) was applied using the PROGENy⁶³ framework, which estimates activity based on expression of downstream target genes for the 14 canonical PROGENy signaling pathways—Androgen, EGFR, Estrogen, Hypoxia, JAK-STAT, MAPK, NFkB, PI3K, TGF-β, TNFα, TNF-related apoptosis-inducing ligand (TRAIL), vascular endothelial growth factor (VEGF), WNT, and p53. TF activities were estimated using a univariate linear model (ULM) and the CollecTRI network, returning activity scores and P values based on weighted expression of known TF targets. These scores were used for comparative analysis between cell populations and visualization.

To identify pathway changes in metaplastic cells across different samples and conditions, ORA was performed using upregulated genes in acinar and acinar derived cells from early timepoints (14 hours and 72 hours from AP, K-AP, and CP). Upregulated genes were calculated on normalized data using scanpy (rank_genes_groups using a t-test), and genes with P values < .05 were used for gseapy (v.0.10.8) enrichr with the MSigDB Hallmark 2020 gene set.⁶⁴

Bubble-plot Visualization

For the bubble plot shown in Figure 4A, the relative abundance of each cell state as a fraction of all epithelial cells (including acinar, ADM, PanIN, and ductal cells) was calculated for unsorted samples. For AP, RAP, and CP, values from the 72-hour timepoint were used, with RAP representing the average across 2-, 4-, and 6-week samples. For K-AP, values across samples from 72 hours to 12 weeks were averaged to capture the full spectrum of transformation. To visualize differences across conditions, a bubble plot was generated in which the area of each circle is proportional to the relative abundance of the corresponding cell state.

Trajectory Inference and RNA Velocity Analysis

To model lineage relationships and infer cellular trajectories across metaplastic acinar states, we employed a combination of PAGA graph abstraction and RNA velocity analysis. For Figure 7G and H, AP epithelial cells were excluded from analysis, as these populations did not progress toward Advanced-ADM states. First, scanpy PAGA was applied to the UMAP embedding, with connectivity visualized using the paga_compare plotting function to highlight transition probabilities between clusters.

To infer transcriptional dynamics, RNA velocity analysis was performed using Velocyto (v0.17.17)³¹ and scVelo (v0.2.4)⁶⁵ (Velocyto was used to generate loom formatted files (http://loompy.org; loompy v3.0.7) of unspliced and spliced reads from each sample’s filtered count matrix and aligned bam file. scVelo was used to merge loom files with the processed anndata objects, and moments were computed (using 30 principal components and 30 neighbors), followed by velocity estimation in deterministic mode. Velocity stream plots were generated on UMAP embeddings with the min_mass parameter set to 3.5. scVelo was also performed on each condition separately to estimate disease context-specific velocity using the same general framework. Stream plots were rendered on the UMAP using a custom visualization function (plot_velocity_on_full_umap) to overlay embedding across the full dataset to ensure consistent comparison, using min_mass parameter set to 4.0 and arrow sizes adjusted to improve visualization.

To identify key genes driving transcriptional dynamics in gADM cells, scVelo’s rank_velocity_genes was used to rank genes by correlation between RNA velocity and gene expression trends across the UMAP manifold. This analysis was performed separately for the K-AP and AP datasets, using a minimum correlation threshold of 0.3.

Plotted K-AP gADM genes were selected based on Spearman correlation and velocity scores. Normalized expression was visualized across the full epithelial UMAP, and spliced vs unspliced counts were plotted using scVelo’s scatter function, with gADM cells highlighted in red to localize their contribution to the trajectory. To compare AP and K-AP gADM velocity-contributing genes, scanpy score_genes was used to compute module scores for the 50 top-ranked AP and K-AP velocity genes. Scores were visualized on the UMAP embedding.

To resolve dynamics specific to early metaplasia in AP and K-AP conditions, condition-specific re-clustering and analysis on a subset of 14- and 72-hour timepoints were performed for each condition independently, using Monocle3 as described above. scVelo was run on each Monocle-derived UMAP embedding, to generate velocity stream plots as detailed above.

Biomarker Identification

To identify candidate K-AP biomarkers, we performed differential expression analysis across experimental conditions (K-AP vs CP, RAP, and AP) for each cell type using scanpy rank_genes_groups (t-test with Bonferroni correction). To ensure markers are specific, they were further filtered for genes with moderate to high expression (mean normalized expression in K-AP cell type >0.1), specificity to that cell type (compared with full dataset), and enrichment in K-AP over other datasets above a certain threshold (test statistic >5).

The only cell type that yielded K-AP specific biomarkers were PanINs. To ensure markers are specific to oncogenic Kras-driven PanIN—and not expressed in non-neoplastic or injured states—candidate genes were cross-validated against independent single-cell datasets from non-oncogenic Kras conditions, including normal pancreas and injury models. Genes with mean normalized expression >0.01 in any non-oncogenic Kras datasets were excluded. This ensured that the final marker set was specific to neoplastic transformation and not confounded by signatures of metaplasia from inflammation or injury alone.

Histology and IHC

Tissues were fixed in 10% neutral-buffered formalin (Fisher Scientific), embedded in paraffin, and sectioned at 5 μm. Sections were stained with hematoxylin and eosin (H&E) per the manufacturer’s instructions (Abcam). For IHC, slides were deparaffinized and rehydrated through xylene and graded alcohols, followed by heat-induced epitope retrieval in a citrate-based antigen unmasking solution (Vector Laboratories). Endogenous peroxidase was quenched in BLOXALLTM Endogenous Peroxidase (Vector Laboratories) at RT for 10 minutes.

Diaminobenzidine (DAB) IHC was performed using primary antibodies against LY6G (Cell Signaling), TFF1 (Proteintech), and KRT17 (Proteintech) and visualized with the VECTASTAIN Elite ABC HRP Kit (Vector Laboratories). Fluorescent immunohistochemistry (IF-IHC) was performed using primary antibodies against DCLK1 (Abcam), CHGA (Santa Cruz) PRSS1 (R&D systems), EGFR (Abcam), Serpinb5 (Abcam), Arc (Abcam), Stnb1 (Atlas Antibodies, Sigma), and Reg3B (R&D Systems) and then visualized with Alexa Fluor 488, 555, or 647 labeled secondary antibodies (Invitrogen).

NanoString CosMx Spatial Molecular Imager profiling: Sample Preparation

Human pancreas was profiled using CosMx Spatial Molecular Imager (SMI; Bruker Spatial Biology) with the Human 6k Discovery Panel.^66,67 Specimens were in compliance with institutional ethical guidelines. Specimens were selected from regions adjacent to resected PDAC tumors to capture areas with metaplasia and PanIN lesions, identified by viewing H&E-stained slides under microscope. Five-μm sections were cut from the selected formalin-fixed, paraffin-embedded (FFPE) tissue blocks and adhered to Leica BOND Plus slides, then baked overnight at 60°C. Slides underwent deparaffinization, target retrieval, and tissue permeabilization using the Bruker Semi-Automated Slide Preparation protocol. Briefly, slides were loaded directly into the Leica BOND RX Research Stainer with all distilled water reagents replaced with diethyl pyrocarbonate (DEPC)-treated water. According to protocol guidance, slides were baked at 60°C for 30 minutes and dewaxed using BOND Dewax Solution. Slides then underwent heat-induced epitope retrieval (HIER) at 100°C for 20 minutes using Epitope Retrieval (ER) 1 solution. Mild digestion was conducted with 3 μg/mL proteinase K (Bruker Spatial Biology) for 30 minutes. Slides were then washed for 5 minutes at ambient temperature using DEPC-treated BOND Wash Solution.

After rinsing with DEPC-treated water and application of an adhesive incubation frame, slides were incubated with 0.001% fiducial solution (Bruker Spatial Biology) in 2× saline sodium citrate with 10% Tween-20 buffer for 5 minutes at RT. After 1× PBS wash to remove excess fiducials, slides were fixed with 10% neutral buffered formalin for 1 minute at RT. After quenching fixation with 2 washes of Tris-glycine buffer (0.1 M glycine, 0.1 M Tris-base) and a 5 minute wash with 1× PBS, samples were blocked using 100 mM NHS-acetate in CosMx NHS-acetate buffer (Bruker Spatial Biology) for 15 minutes at RT. Samples were washed with 2 rounds of 2× saline sodium citrate (SSC) for 5 minutes each. In preparation for in situ hybridization (ISH), the CosMx Human 6k Discovery Panel probes were denatured at 95°C for 2 minutes, then snap-cooled on ice. The ISH probe working mix was prepared using 1 nM ISH probes in DEPC-treated water with CosMx Buffer R and CosMx RNase Inhibitor (both from Bruker Spatial Biology). Samples were incubated with the probe mix at 37°C overnight using coverslips and a HybEZ II hybridization oven (ACDBio).

On the next day, slides were washed 2 times with 50% formamide/50% 4× SSC at 37°C in a water bath (25 minutes each wash), and rinsed twice with 2× SSC for 2 minutes, at RT. Before antibody incubation, slides were incubated with 4′,6-diamidino-2-phenylindole (DAPI) nuclear stain diluted in blocking buffer (both from Bruker Spatial Biology) for 15 minutes followed by 1× PBS wash. For visualization of cell morphology, antibodies against CD298/B2M, PanCK, CD45, and CD68, which are conjugated to a readout domain that binds fluorophores on-instrument (CosMx Segmentation Markers Kit, Bruker Spatial Biology), were incubated on the slides at RT for 1 hour. After washing slides 3 times with 1× PBS, CosMx flow cell coverslips (Bruker Spatial Biology) were attached to the sample slide to form a flow cell.

NanoString CosMx SMI Profiling: Readout on CosMx SMI

RNA target readout on the CosMx SMI instrument was performed as described.^66,67 Reporter Wash, Imaging, and Strip Wash Buffers all supplied by Bruker Spatial Biology. Briefly, the assembled flow cell was loaded onto the instrument, and Reporter Wash Buffer was flowed to remove air bubbles. A preview scan of the entire flow cell was taken, and multiple fields of view (FOVs), each measuring 0.5 mm × 0.5 mm, were placed on the tissue to define the region(s) to be assayed. RNA readout began by flowing 100 μL of Reporter Pool 1 into the flow cell and incubating for 15 minutes. Reporter Wash Buffer (1 mL) was flowed to wash unbound reporter probes, and Imaging Buffer was added to the flow cell for imaging. Eight Z-stack images (0.8 μm step size) for each FOV were acquired, and photocleavable linkers on the fluorophores of the reporter probes were released by UV illumination and washed with Strip Wash buffer. The fluidic and imaging procedure was repeated for the 27 reporter pools, and the 27 rounds of reporter hybridization imaging were repeated multiple times to increase RNA detection sensitivity. Cell morphology was imaged on-instrument prior to RNA readout by adding fluorophore-bound reporters to the flow cell and capturing eight z-stack images in the channels 488 nm (CD298/B2M), 532 nm (PanCK), 594 nm (CD45), 647 nm (CD68), and 385 nm (DAPI).

NanoString CosM:x Image Analysis

Following standard data upload, processing, and cell segmentation on the AtoMx Spatial Informatics Platform, a Cell Typing pipeline was performed using InSituType⁶⁸ to generate 20 unsupervised clusters. Manual cluster annotation was performed by a trained pathologist to determine cluster identities based on histology, cell morphology, and top marker genes, on simultaneously viewing scanned tissue slides overlaid with the unsupervised InSituType cell clusters, using the AtoMx Image Viewer. Representative tissue regions then were selected for Figure generation. PanCK marker intensity was set to a minimum of 796 and maximum of 20,000, and DAPI marker intensity was set to a minimum of 370 and a maximum of 5,000. Individual classical and basal marker probes were imaged overlaid on the cell phenotype mask at 50% transparency to ensure probe expression was occurring in PanIN cells specifically (classical probe dots scaled to 300%, basal probe dots scaled to 400%). gADM images were captured with the cell segmentation map overlay to ensure accurate spatial localization, with a probe dot scaling of 200%.

Data Visualization and Data and Code Availability

Figures were created using BioRender.com, matplotlib.pyplot (v3.7.1), seaborn (v0.13.2), and scanpy (v1.9.3). Datasets are available on NCBI’s Gene Expression Omnibus and can be accessed through GSE235874 (AP) and GSE314765 (K-AP, RAP, CP). All analysis and figures generating code will be available on GitHub (https://github.com/nissimlab).

Gene lists

NRP score²⁹: Kcnj16, Zbtb16, Ctnnd2, Pde3a, Pdgfd, Cntn4, Cftr, Flrt2, Adcy5, C6, Crisp3, Ralyl, Nr1h4, Bcl2, Esrrg, Slc4a4, Csmd2, Rgs17, Crp, Slc17a4, Reln, Pah, Pkhd1, Linc01320, Acsm3, Dscaml1, Ar, Znf208, Ttll7, Sox6, Spp1, Asxl3, Pou6f2, Dzip1, Itih5, Prkg1, Il1r1, Sema3e, Gucy1a2, Ac092535.3, Rcan2, Slc3a1, Man1a1, Wdr72, Acss3, Onecut1, Nrp1, Akap7, Ldlrad4, Ac012593.1, Caln1, Ugt2b15, Agbl4, Glis3, Trpv6, Abcb1, Plxdc2, Nlgn4y, Nek10, Tacc1, Homer2, Snap25, Sctr, Slc2a2, Mir99ahg, Trabd2b, Nr5a2, Fam135b, Fgg, Dcdc2, Syne1, Crisp2, Sema5a, Bicc1, Tns1, Chst9, Ncam1, Rora, Adamts9-as2, Apcs, Mum1l1, Setbp1, Nrcam, Cfap221, Atp13a4, Lin7a, Ttc28, Stxbp6, Kctd16, Nr2f2-as1, Dock8, Lrrk2, Rerg, Ac018742.1, Pcdh9, Ptchd4, Atp10a, Wnk2, Dpyd, Ac019117.1, Sncaip, Limch1, Anxa4, Bmpr1b, Kcnma1, Ptp4a1, Dtna, Nbea, Ttn, Dlg2, Ptprm, Scn9a, Tmem132c, Hif1a, Khdrbs2, Slc16a7, Ajap1, Kif12, Sema6a, Grm8, Lrat, Trim5, Nfib, Pde7a, Onecut2, Prickle2, Cdh6, Ac124312.1, Mapk10, Rbpms, Apcdd1, Ces1, Linc01266, Serpina6, Adgrl2, Linc00671, Tenm3, Kcnj15, Meis2, Fign, Gabrb3, Sdk1, Kcnt2, Znf503-as1, Cfh, Arhgap44, Anks1b, Thsd4, Nr3c2, Adarb2, Nrxn3, Tox, Nfia, Anpep, Grb10, Tusc3, Habp2, Dach1, Mpp6, Camk1d, Slc1a1, Serping1, Mef2c, Muc5b, Fhit, Slc5a1, Pdgfc, Asrgl1, Sult1c4, Cacna1h, Nrep, Dsel, Slc4a7, Reg1a, Mllt3, Tdrp, Dlgap1, St8sia3, Fxyd2, Epb41l4a, Iqca1, Prkce, Mpp7, Nrg1, Itpr2, Znf667, Auts2, Lrp1b, Bco2, Pbx1, Rassf8, Hydin, Prkd1, Zscan18, Klkb1, Znf676, Pbx3, Cep112, Cys1, Galnt18

PDAC human genes⁵⁷: Arl4c, Anxa1, Mki67, Asph, Cdh3, Capg, Ceacam5, Ceacam6, Ctsl2, Cav2, Cdk1, Cdc2l1, Ccnb1, Ccnd1, Dbn1, Ereg, Ago2, Fer1l3, Fosl1, Gjb2, Hst3st1, Hoxb6, Igf2bp3, Imup, Inhba, Igfbp3, Itgb1, Cxcl1, Cxcl2, Cxcl5, Krt17, Krt19, Krt7, Kiaa1199, Kiaa1265, Kiaa1363, Lamc2, Lgals1, Lcn2, Mmp11, Mmp7, Msln, Met, Map4k4, Muc4, Muc5b, Muc5ac, Muc5c, Nnmt, Plau, Plaur, Phldb1, Plec1, Psca, Rap2b, Rai3, Runx1, Ruvbl1, S100a11, S100a4, S100p, Slpi, Serpinh1, Serpinh2, Sik1, Sfn, Slc16a3, Slc16a12, Tspan1, Thbs2, Top2a, Bmal2, Tmprss3, Tmem45a, Tff2, Talc2

PDAC mouse genes¹¹: Egfr, Krt19, Cd44, Tgfb1, Smad3, Myc, Pparg, Runx1, Hes1, Mmp14, Itgb1, Igf1r, Erbb2, Akt3, Tgfbr2, Notch2

Acinar genes²⁷: Cela1, Pnlip, 2210010C04Rik, Zg16, Ctrl, Cela3b, Pla2g1b, Rnase1, Sycn, Tff2, Cela2a, Prss2, Cpa1, Cpb1, Ctrc, Try4, Clps, Gp2, Hamp2

Tolerogenic DC genes³⁸: Dram1, Nrip1, Cebpb, Smpdl3a, Nod2, Cd14, Papss2, St3gal1, Sema6b, Cd300lf, Acsl1, Trem1, Ninj1, Ncf1c, Rgs18, Tspan14, Ms4a4a, Cd93, Ncoa4, Brd8, C1qa, Gk, C5ar1, Epb41l3