A classical epithelial state drives acute resistance to KRAS inhibition in pancreas cancer

Abstract

Intra-tumoral heterogeneity in pancreatic ductal adenocarcinoma (PDAC) is characterized by a balance between basal and classical epithelial cancer cell states, with basal dominance associating with chemoresistance and a dismal prognosis. Targeting oncogenic KRAS, the primary driver of pancreatic cancer, shows early promise in clinical trials but efficacy is limited by acquired resistance. Using genetically engineered mouse models and patient-derived xenografts, we find that basal PDAC cells are highly sensitive to KRAS inhibitors. Employing fluorescent and bioluminescent reporter systems, we longitudinally track cell-state dynamics in vivo and reveal a rapid, KRAS inhibitor-induced enrichment of the classical state. Lineage-tracing uncovers these enriched classical PDAC cells are a reservoir for disease relapse. Genetic or chemotherapy-mediated ablation of the classical cell-state is synergistic with KRAS inhibition, providing a pre-clinical proof-of-concept for this therapeutic strategy. Our findings motivate combining classical-state directed therapies with KRAS inhibitors to deepen responses and counteract resistance in pancreas cancer.

Introduction

Oncoprotein-targeted therapies have revolutionized the landscape of cancer treatment. After decades of arduous drug development, mutationally activated KRAS is now a viable drug target. KRAS is a particularly exciting target in pancreatic ductal adenocarcinoma (PDAC), a disease predominantly driven by oncogenic KRAS. PDAC has a dismal prognosis and limited treatment options, with most patients succumbing within 12 months of diagnosis despite intensive chemotherapy regimens. Early data demonstrates clinical benefit for adagrasib and sotorasib in KRAS-G12C-mutant PDAC (<1% of cases); similarly, pan-RAS inhibitors have recently shown an early signal of clinical activity (1–3). A plethora of innovative KRAS inhibitors, which target the common alleles found in PDAC, including MRTX1133—a KRAS-G12D-selective inhibitor—are currently advancing into clinical trials (4). Thus, therapies targeting KRAS are anticipated to enter the therapeutic paradigm for PDAC in the near future. Yet, initial clinical data suggest extent and rate of response are far from complete (1–3,5,6), warranting mechanistic studies to advance understanding of how KRAS inhibitors function in vivo.

While KRAS inhibitors can drive potent initial responses, acquired resistance limiting the durability of response remains a key clinical challenge. At the time of tumor relapse on KRAS-targeted therapies, clinical data from colon and lung cancers demonstrates a high incidence of secondary resistance mutations. Moreover, multiple types of mutations within signaling components upstream or downstream of KRAS, parallel to KRAS, in KRAS itself, or in other RAS genes have been observed, and individual tumors often harbor many co-incident alterations (7,8). In PDAC, emerging data from pre-clinical studies has shown that diverse genetic alterations drive acquired resistance to pharmacologic or genetic inactivation of KRAS, including amplification of MYC or alterations that activate YAP-TAZ signaling, among others (9–11). Thus, devising sequential or combination therapies that address the broad diversity of genetic alterations seen at the time of acquired KRAS inhibitor resistance is likely to be infeasible.

In this context, targeting cancer cell populations that are acutely resistant to KRAS inhibition—before secondary genetic mutations take place—represents an untapped therapeutic opportunity. Acutely resistant cancer cells have the potential to serve as reservoirs for resistance, allowing cells to evade therapeutic pressure until acquired mutations emerge. In PDAC, where repeated biopsies are infeasible, our understanding of the acute responses to therapies is particularly limited, and little is known about the biological properties of residual disease. In vivo models that faithfully approximate the pathophysiology of human PDAC and allow elucidation of the temporal dynamics of response to KRAS inhibition are needed to bridge this knowledge gap.

Despite the relatively uniform landscape of genetic alterations across tumors (KRAS, TP53, CDKN2A, SMAD4/TGFBR2), PDAC cells exhibit considerable intra-tumoral phenotypic diversity in vivo (12–14). This diversity is transcriptionally defined and can be broadly classified into two cell states that recur across patients: (i) a classical state, characterized by molecular and histopathological features of epithelial differentiation, and (ii) a basal/quasi-mesenchymal state associated with invasive morphology, chemoresistance, and poor prognosis (13–16). While the relative abundance of these cell states within tumors serves as a prognostic biomarker, the dynamic lineage relationships between them and their sensitivity to KRAS-targeted agents and other therapies remains a critical open question. Moreover, combining therapies in order to suppress distinct cell states has the potential to eliminate disease reservoirs and extend therapeutic benefit.

Although great strides are being made with in vitro culture methods (17), these models fail to fully recapitulate cancer cell states present in human PDAC tumors (12). In contrast, genetically engineered mouse models (GEMMs) provide a unique platform for functional interrogation of cancer cell states in situ in autochthonous tumors that, like human cancers, evolve in the context of the relevant tissue site (12). In commonly used PDAC GEMMs, Cre or Flp recombinase is directed to pancreatic epithelial cells, allowing somatic activation of oncogenic KrasG12D and inactivation of Trp53 (hereforth KPC or KPF models, respectively) (18–20). These models mimic KRAS-driven human PDAC at molecular and histopathological levels, including key hallmarks such as desmoplastic stroma, an immunosuppressive microenvironment, and presence of classical, basal, and mesenchymal cancer cell states (18,19,21,22). Importantly, our recent results indicate orthotopic transplantation of cell lines derived from the KPC and KPF models faithfully recapitulate tissue architecture and cell state heterogeneity of autochthonous mouse and human PDAC tumors (12). Here, we sought to identify cell states acutely resistant to KRAS inhibition by applying and further engineering these models to investigate the temporal dynamics and lineage relationships of PDAC cell states.

Results

KRAS inhibition enriches for a malignant classical epithelial state in PDAC

We previously characterized cell state composition in the KPC model via single-cell RNA sequencing (scRNA-seq), and found that the mouse model of PDAC harbors (i) a classical state, which adopts a luminal, columnar cellular morphology; (ii) a basal state, invasive in appearance and expressing laminins/keratins; and (iii) a mesenchymal state with fibroblast-like morphology that typically intermingles with the fibrous stroma (Fig. 1A) (12). Importantly, these cancer cell states express markers common to human disease. Galectin-4 is the most highly correlated transcriptional marker with classical identity in a recent human scRNA-seq dataset (13) and is a core marker of the classical subtype (15,16) (Fig. 1B). Galectin-4 strongly co-localizes with the commonly used classical state marker GATA6 in mouse and human PDAC (Supplementary Fig. 1A). Keratin 17 (KRT17) is a commonly used marker to identify the basal subtype in human disease (Fig. 1B) (13,15,16,23). Moreover, transcriptional signatures and markers of human PDAC classical vs. basal subtypes, as drawn from bulk-RNA sequencing datasets in foundational studies, show strong concordance with states observed within the GEMM (15,16) (Fig. 1A, Supplementary Fig. 1B, C).

Figure 1. KRAS inhibition enriches for a classical epithelial state in PDAC.

(A) (top) UMAP embedding of autochthonous PDAC cell transcriptomes, classified as classical (purple), basal (green), or mesenchymal (orange) cancer cell populations, from Pitter et al. (12) (middle) Expression of previously described human classical, basal, and mesenchymal expression signatures projected along the pseudotemporal axis (13,46) (bottom). Experimental design for investigating KRAS(G12D) inhibition in autochthonous KPCT PDAC tumors using the KRAS(G12D) inhibitor MRTX1133. (B) UMAP projection of Lgals4 (left) and Krt-17 (right) gene expression. (C) Quantification of classical tumor states [galectin-4⁺/pan-cytokeratin⁺] and basal tumor states [Krt-17⁺/pan-cytokeratin⁺] with 7 days on MRTX1133 or vehicle control (n = 5 in each group). (D) Representative images of immunofluorescent staining for galectin-4 (red), KRT17 (green), and pan-cytokeratin (orange) in an autochthonous KPT PDAC tumor treated with MRTX1133 vs vehicle. Scale bar: 100 μm. (E) Expression of Epcam and CD44 along pseudotemporal classical-basal-mesenchymal axis. Expression of markers segregates classical, basal, and mesenchymal PDAC cell states. (F) Representative flow cytometry plots of cancer cells (tdTomato⁺/CD45⁻/CD31⁻/CD11b⁻/F480⁻/TER119⁻/DAPI⁻) depicting expression of Epcam (x-axis) and CD44 (y-axis) in PDAC tumors from KPCT GEMMs after MRTX1133 (7 days) compared with vehicle control. (G) Quantification of the proportion of classical, basal, and mesenchymal states using flow cytometry after MRTX1133 compared with vehicle control, n = 5 mice/group. (H) Experimental design to investigate KRAS(G12D) inhibition in orthotopic KRAS(G12D);P;Rosa26-mTmG PDAC tumors. (I) Principal component analysis (PCA) of (tdTomato⁺/CD45⁻/CD31⁻/CD11b⁻/F480⁻/TER119⁻/DAPI⁻) single PDAC cell transcriptomes from MRTX1133-treated vs control tumors (n =3 tumor in each group; n = ~3000 cells per condition) (top). Unsupervised clustering of PDAC single cell transcriptomes, colored and annotated based on human gene signatures (13) (middle, bottom). Embedding density of MRTX1133 vs. vehicle treated tumors within PCA plot. (J) Pearson correlation of gene weights within each principal component (PC) with their presence in human classical/basal gene signatures (13). (K) Heatmap depicting the difference in gene signature scores between MRTX1133-treated orthotopic KP tumors, as compared with vehicle-treated tumors. Classical and basal signatures drawn from four independent human data sets (13–16). Difference calculated with the average signature score over all single-cell transcriptomes per condition; n = 3 tumors/condition; ~3000 cells/condition. Statistical significance is assessed over all single-cell transcriptomes between treatments by a two-sided Wilcoxon rank-sum test. (L) Representative image of immunofluorescent staining for phospho-ERK across galectin-4+ (classical) states and Keratin-17⁺ (basal) states in an untreated KPT GEMM PDAC tumor. Scale bar: 40 μm (M) Quantification of phospho-ERK intensity in classical (galectin-4⁺) and basal (KRT17⁺) cells (n = >2000 cells/group; n = 3 tumors) from unperturbed KPT GEMM PDAC tumors. Dark line represents mean value, and dashed lines the interquartile value. Statistical significance was assessed using an unpaired t-test. (N) Hallmark transcriptional gene sets significantly enriched or depleted in (left) MRTX1133-treated resistant classical cells vs. vehicle-treated classical cells, and in (right) MRTX1133-treated resistant basal cells vs. vehicle-treated basal cells. The normalized enrichment score from gene set enrichment analysis (GSEA) is shown on the x-axis. Unpaired t tests were used to test significance compared to vehicle in (C), (G). Error bars indicate standard deviation.

We investigated the response of autochthonous KPC PDAC tumors to KRAS inhibition with MRTX1133, a selective KRAS-G12D inhibitor (24,25). To mark all cancer cells in these mice, we had crossed in a Rosa26^{LSL-tdTomato/+} reporter allele (KPCT mice). KPCT mice were randomized to receive KRAS inhibitor (MRTX1133) or vehicle at approximately 8 weeks of age, when tumors became palpable, and were harvested after 1 week of therapy (Fig. 1A). Tumor weights measured after treatment revealed a robust response to KRAS-G12D inhibition (Supplementary Fig. 1D). Strikingly, KRAS inhibition led to a dramatic decrease in basal cells and an enrichment of the classical cells (Fig. 1C, D). Moreover, residual cancer cells harbored more epithelial and glandular features as compared with vehicle-treated tumors (Fig. 1D). To further validate our observations, we developed a rapid flow cytometry-based phenotyping method for PDAC cells along the classical-basal-mesenchymal axis. Leveraging scRNA-seq data, we noted that EpCAM, a marker of epithelial identity across tumor types, was highly expressed in classical cells but diminished in basal/mesenchymal states. Conversely, CD44 showed enrichment in the basal state, with lower expression at the phenotypic extremes of classical and mesenchymal identity (Fig. 1E). Employing these two markers, we classified cells into the three principal cell states (Fig. 1E) and confirmed the enrichment of state-specific marker genes in these populations by quantitative PCR (qPCR) for FACS-sorted cell states (Supplementary Fig. 1E). Using this phenotypic readout, we observed a strong shift towards classical identity under MRTX1133 therapy (Fig. 1F, G, Supplementary Fig. 1F), consistent with our observations using immunofluorescence.

Given that analyses based on a small number of markers may overlook additional phenotypic changes in the PDAC cells, we next examined shifts in cancer cell states in response to MRTX1133 using single-cell RNA sequencing (scRNA-seq). Given the dramatic response to MRTX1133 in the GEMM models, in line with prior reports (26), residual cancer cells were often very limited in abundance and difficult to profile. Therefore, we chose to employ orthotopic transplantation of a tdTomato⁺ (Rosa26^mTmG/+) PDAC KPF fluorescent reporter cell line into the pancreas of immunocompromised NSG mice (Fig. 1H). We have previously shown that this model reconstitutes the cell state heterogeneity of the KPC GEMM (12). Further, we introduced a lentiviral vector encoding Gaussia princeps luciferase (GLuc), which is secreted into the bloodstream, and enables longitudinal tracking of tumor burden via small-volume, non-terminal sampling of the peripheral blood (27). Using serial GLuc measurements, we observed a significant decrease in tumor burden to ~10% of baseline at 7 days of KRAS-G12D inhibition (Supplementary Fig. 1G). This time point was selected for scRNA-seq analysis because it represents the peak response (nadir) to MRTX1133.

We conducted principal component analysis (PCA) on all malignant cells and found that the first principal component (PC1) was strongly associated with genes from the classical-basal axis. To quantify this association, we developed a similarity metric to measure the strength of the similarity between each principal component with human classical-basal gene signatures (13). The correlation of PC1 to classical-basal signatures was markedly stronger than those of all other principal components (Fig. 1I, J). Comparing KRAS inhibitor-treated tumors with vehicle tumors, we observed a robust shift along PC1 towards classical identity under therapy (Fig. 1I, J; Supplementary Fig. 2A). Analyzing MRTX1133 vs vehicle-treated cells in bulk, we found that signatures of classical PDAC identity across multiple datasets were enriched. Conversely, signatures of basal identity were strongly suppressed under KRAS inhibition (Fig. 1K) (13–16). Importantly, the PDAC cells remained within the phenotypic space of pre-treatment tumors even in the context of short-term KRAS inhibition, indicating that new therapy-adapted states are not induced by KRAS inhibition in this time frame. These data collectively point to two main conclusions: First, the classical-basal axis remains the dominant source of transcriptional variation in the context of KRAS inhibition. Second, residual disease following acute KRAS inhibition is strongly enriched for the classical state.

We hypothesized that the differential response to KRAS inhibitors between classical and basal states may be linked to variations in oncogenic KRAS activity across cell states. To test this, we measured levels of phosphorylated ERK, a key KRAS effector and biomarker of KRAS activity, in unperturbed autochthonous KPC tumors. Notably, classical PDAC cells marked by galectin-4 exhibited significantly lower levels of phosphorylated ERK (p-ERK) compared to KRT17⁺ basal cells within the same tumor (Fig. 1L, M). Further, our analysis revealed that classical cells have lower transcriptional activity of an established set of KRAS downstream targets, including Dusp6, Ccnd1, Etv1, Etv5, and Spry2 (11,28), as well as decreased expression of a recently published KRAS gene signature in PDAC (29) (Supplementary Fig. 2B, C). These results suggest that classical cells are less dependent upon KRAS signaling compared to basal cells, thereby providing a possible explanation for the intrinsic resistance of the classical cells to KRAS inhibition. Independent of classical or basal identity, we noted that resistant cells altered common biological pathways in response to KRAS inhibitors, suggesting that these pathways may be controlled by oncogenic KRAS or may have roles in survival after treatment (Fig. 1N, Supplementary Table 1).

Classical PDAC cells give rise to disease relapse

To study the dynamics of the classical cell state and its functional role in resistance to KRAS inhibition in PDAC cells, we knocked a “MACD” reporter cassette into the Lgals4 locus encoding galectin-4, a highly specific and sensitive marker of classical identity (Fig. 1B, Supplementary Fig. 1A) (13,15,16). The MACD cassette comprises (1) an mScarlet red fluorescent protein allowing detection of classical cells via microscopy or flow cytometry; (2) AkaLuc bioluminescence (30); (3) tamoxifen-inducible Cre recombinase (CreERT2); and (4) the diphtheria toxin receptor (DTR) suicide gene (Fig. 2A; Supplementary Fig. 3A, B). These cells were also transduced with a lentiviral vector delivering ubiquitously expressed GLuc, a far-red fluorescent protein (iRFP670), and a Cre-inducible flex-tagBFP2-FLAG (BFP) fluorescent protein. These components enable longitudinal monitoring of tumor burden, marking of all cancer cells, and lineage-tracing, respectively. Upon orthotopic transplantation of the cell line into the pancreas, the resulting tumors showed morphological heterogeneity, with epithelial cancer cells expressing mScarlet and all cancer cells expressing iRFP (Fig. 2A). To further validate reporter fidelity, we sorted mScarlet⁺ cells from PDAC tumors and demonstrated strong enrichment of classical state marker genes when compared to mScarlet⁻ cells (Fig. 2B).

Figure 2. Classical cells in PDAC give rise to relapsed disease, and classical state cytoablation enhances response to KRAS inhibition.

(A) Genetic Lgals4-MACD reporter enabling visualization (mScarlet), non-invasive longitudinal tracking (AkaLuc), lineage tracing (CreER) and ablation (DTR) of classical KP PDAC cells. Frt-Stop-Frt-mScarlet-Akaluc-CreERT2-DTR reporter construct knocked in in-frame after exon 10 of Lgals4. Endogenous mScarlet signal expressed from Lgals4-MACD reporter labels classical states and aligns closely with galectin-4 immunofluorescence (green, lower right). The lentiviral reporter (described in C) labels all tumor cells with iRFP670 (cyan). Scale bar, 100 μm. (B) Classical marker gene expression in mScarlet cells (high, dim) as compared with mScarlet negative cells, isolated by flow cytometry from orthotopic Lgals4-MACD PDAC tumors. n = 3 biological replicates from independent tumors per cell population. One-way ANOVA was used to test for statistical significance between mScarlet⁻ and mScarlet⁺ (dim, high) populations. (C) PGK-GLuc-miRFP-EFS-lox-BFP-lox lentiviral lineage-tracing vector was introduced into Lgals4 -MACD reporter cells. This vector enables tumor burden measurements via secreted Gaussia luciferase (GLuc). Lineage-tracing of classical Lgals4+ cells occurs following a tamoxifen pulse, which triggers BFP expression. The ratio of AkaLuc/GLuc, tracked over time, indicates changes in the proportion of classical states in the tumor. (lower right) After lineage tracing, cells that retain the BFP mark, but are no longer mScarlet⁺, arise through cell state transitions. (D) (top) Outline of experimental design to investigate Lgals4+ classical states during and following withdrawal of KRAS(G12D) inhibition in orthotopic KP; Lgals4-MACD PDAC tumors using bioluminescence imaging (BLI; for AkaLuc+ classical states) and serum measurements (GLuc; for tumor burden) at the indicated time points. (bottom) In vivo AkaLuc luciferase imaging (classical state) of mice with orthotopic Lgals4-MACD pancreas tumors subjected to vehicle or MRTX1133 before, ON (7, 10 days), and OFF (15, 25 days) treatment. (E) Tumor burden measured by Gaussia princeps luciferase (GLuc) luminescence in response to MRTX1133 therapy over time in orthotopic Lgals4-MACD tumors, treated with MRTX1133 vs vehicle (on drug), followed by relapse (off drug), n = 8–12 mice/group. Error bars indicate SEM. (F) Quantification of bioluminescence from Lgals4: : AkaLuc normalized to tumor burden (GLuc), expressed as a fold-change from baseline, during MRTX1133 treatment and after relapse; n = 8–12 tumors/group. (G) Outline of experimental design to lineage trace Lgals4⁺ classical cells on and after MRTX1133 treatment vs. vehicle control in orthotopic Lgals4-MACD tumors. (H) Representative flow cytometry plots of cancer cells (miRFP760⁺/ CD45⁻/CD31⁻/CD11b⁻/TER119⁻/DAPI⁻); depicting expression of mScarlet (x-axis; classical identity) and BFP (y-axis; lineage-traced cells) from orthotopic tumors (Lgals4-MACD) on treatment with MRTX1133 (Day 7) as compared with time of relapse (Day 21). (I) Quantification of the fraction of classical lineage-traced states (mScarlet⁺/BFP⁺) as compared with all lineage-traced states (BFP⁺; history of classical identity at the time of nadir) from orthotopic tumors (Lgals4-MACD) on treatment with MRTX1133 (Day 7) as compared with time of relapse (Day 21). Note that samples for each group were collected during independent experiments with the same experimental design. (J) Single PDAC cell transcriptomes from tumors at 14 days of relapse, after lineage-tracing classical states at nadir (left) Unsupervised clustering of cells (miRFP760⁺/ CD45⁻/CD31⁻/CD11b⁻/TER119⁻/DAPI⁻), annotated based on (12) (middle) gene expression of Lgals4 projected into UMAP (right) lineage-traced cells (BFP+) projected onto UMAP; arrows indicates the transdifferentiation of classical cells at treatment-nadir into other PDAC cell states at relapse. (K) Representative images of BFP and mScarlet immunofluorescence in KP; Lgals4-MACD reporter tumors on MRTX1133 (left) or at 14 days relapse (right). Note labeling of BFP+/mScarlet+ cells (white arrowheads) on MRTX-1133 and an increase of transdifferentiated classical cells (BFP+/mScarlet-; yellow arrowheads) during relapse. Scale bar, 100 μm. (L) Percentage of lineage-traced cells (BFP⁺) as quantified by flow cytometry in Lgals4-MACD orthotopic tumors that underwent treatment with MRTX1133 followed by relapse, as compared with vehicle. (M) Outline of experimental design to test ablation of Lgals4-MACD⁺ cells via diphtheria toxin (DT) in the context of MRTX1133 treatment vs vehicle control in orthotopic tumors. (N) Tumor burden measured by Gaussia luciferase in MRTX1133-treated, MRTX1133+DT-treated, DT-treated, or vehicle-treated orthotopic Lgals4-MACD tumors, n = 5–12 mice/group. (O) Tumor burden measured by Gaussia luciferase in vehicle-treated, gemcitabine/nab-paclitaxel-treated, MRTX1133-treated, or MRTX1133+ gemcitabine/nab-paclitaxel-treated mice in orthotopic Lgals4-MACD tumors, n = 4–8 mice/group

Our reporter system enables a real-time non-invasive measurement of the classical state via AkaLuc bioluminescence imaging, as well as overall tumor burden via GLuc (Fig. 2C, D). As previously, KRAS inhibition with MRTX1133 led to a ~90% reduction in tumor volume in 7 days (Fig. 2E). The tumors rebounded rapidly upon treatment cessation (Fig. 2E), suggesting PDAC cells within the minimal residual disease harbor the capacity to reignite tumorigenesis upon treatment failure or discontinuation. We tracked the emergence of the classical state during the response to MRTX1133 using AkaLuc bioluminescence imaging (Fig. 2D), normalizing to GLuc (the total tumor burden). We observed a striking, 10-fold increase in the fraction of classical cells during MRTX1133 treatment, indicating an enrichment of classical PDAC cells in residual disease (Fig. 2F). To verify that the observed increase was attributable to the fraction of cells in this state rather than enhanced Lgals4 gene expression, we harvested tumors and measured the proportion of classical (mScarlet⁺) cells after seven days of KRAS inhibition. The intensity of mScarlet reporter fluorescence did not increase with treatment; instead, the proportion of mScarlet⁺ cells increased tenfold, confirming the relative abundance of classical cells rather than an increase in Lgals4 gene expression is the principal driver of the observed increase in AkaLuc bioluminescence (Supplementary Fig. 3C, D). Upon discontinuing MRTX1133 treatment, we noted a rapid decline of the Lgals4⁺ classical fraction back to baseline (Fig. 2F). To confirm that this finding is not specific to the genetic background or the classical marker gene, we utilized a distinct genetic reporter in a different genetic background; here, we used an established KPC4662 (Kras^LSL-G12D/+, Trp53^LSL-R172H/- ) murine PDAC cell line and integrated a similar cassette (DMAC) into the Porcn gene locus, which we found to be restricted to classical cells in a recent study (bioRxiv 2023.12.30.573100) (Supplementary Fig. 4A). We observed similar kinetics of the Porcn⁺ classical state using this system (Supplementary Fig. 4B–E). Together, these results demonstrate that classical cells are strongly enriched in residual disease in response to KRAS inhibition, and that the shift towards a classical identity in PDAC is directly induced by targeting KRAS.

Given that classical cells are enriched upon MRTX1133 treatment, we next investigated the possibility that classical cells serve as a source of relapsed disease. Observing that therapy discontinuation leads to a decrease in the Lgals4⁺ classical state suggests two possibilities: either residual non-classical cells proliferate in the setting of KRAS reactivation, or the classical state transdifferentiates to give rise to non-classical cells. To investigate this, we labeled classical cells at the nadir of residual disease (7 days on MRTX1133 treatment) with a heritable BFP switch via tamoxifen-induced Cre-recombinase activation (Fig. 2G). MRTX1133 was then withdrawn to model the loss of drug activity and reactivation of KRAS (Fig. 2G). We observed that our system traced the Lgals4⁺ state with high fidelity (Fig. 2H). Interestingly, following therapy cessation, we found that ~70% of Lgals4⁺ classical cells lose their identity and enter a non-classical state (Fig. 2H–K).

To definitively profile the fate of the classical state within residual disease upon disease relapse, we performed scRNA-seq of lineage-traced classical states following 14 days of withdrawal of MRTX1133. Consistent with our non-invasive measurements (Fig. 2F), classical Lgals4⁺ cells represented a small fraction of cancer cells after drug withdrawal (Fig 2H, 2J). Notably, a significant fraction of cells that were classical at the time of treatment nadir (BFP⁺) were detected in basal and mesenchymal compartments of these tumors (Fig. 2J, K). Further, we noted that the fraction of traced (BFP⁺) cells within the tumor that emerge from the classical state was significantly higher as compared with vehicle-treated tumors (Fig. 2L). To examine the function of the classical state in spontaneous disease relapse, we lineage-traced the Lgals4+ classical cells at nadir under continuous KRAS inhibitor therapy (Supplementary Fig. 5A). The classical cells showed a strong enrichment at the point of maximal response, followed by a decline at the time of acquired resistance and regrowth (Supplementary Fig. 5B, C). As in the context of drug withdrawal, the classical cells traced at the timepoint of maximal treatment response have the capacity to give rise to non-classical states in resistant tumors (Supplementary Fig. 5D, E). In sum, our data establish residual classical cells that survive KRAS inhibition are a key driver of disease relapse.

Targeting the PDAC classical cell state augments response to KRAS inhibitors

As classical PDAC cells play a critical role in acute KRAS inhibitor resistance, we next interrogated the therapeutic potential of co-targeting classical cells and KRAS to improve treatment response. To perform these studies, we ablated classical PDAC cells through systemic administration of diphtheria toxin (DT). As expected, while MRTX1133 treatment led to increased reporter activity, DT administration eliminated classical cells in both the vehicle- and MRTX1133-treated context (Supplementary Fig. 5F). Importantly, simultaneous administration of DT and MRTX1133 led to a 66% further reduction in tumor size compared to MRTX1133 alone (Fig. 2M, N). Interestingly, elimination of the classical cells alone did not impact tumor growth relative to vehicle, suggesting that the classical state is not functionally important for tumor maintenance at baseline, yet becomes important in the context of KRAS inhibition (Fig. 2N).

Clinical studies have suggested that PDAC tumors with a predominance of classical cells are more sensitive to chemotherapies (31). Thus, we hypothesized that a mainstay chemotherapy combination regimen, gemcitabine/nab-paclitaxel, may direct cytotoxicity to classical cells and thus enhance the efficacy of KRAS inhibition. We found that, alone, gemcitabine/nab-paclitaxel had a modest effect on tumor growth compared to vehicle (Fig. 2O). However, in combination with MRTX1133, we observed a trend towards deeper tumor regressions, with the mean tumor burden decreasing by ~60% when compared to single-agent MRTX-1133 (Supplementary Fig. 5G) —a similar effect to what was observed in our genetic classical state ablation model. Further, the combination regimen strongly delayed relapse (Fig. 2O).

Together, these data demonstrate that residual classical states that survive KRAS inhibitor treatment have the capacity to regrow tumors upon treatment failure. Further, cytoablative or chemotherapies that eliminate classical cells strongly augment response to KRAS inhibition.

The classical state is enriched in residual disease following KRAS inhibition in human PDAC

To explore the relevance of our findings in mouse models for human PDAC, we subjected three KRAS-G12D mutant PDAC PDX models to MRTX1133 therapy (Fig. 3A; Supplementary Fig. 6A). These models exhibited significant pre-treatment state heterogeneity, enabling us to study the selective pressure exerted by MRTX1133 on classical and basal cells in the same tumor. All PDXs responded to MRTX1133 with reduced growth (Fig. 3B). Using established markers of cell states in human PDAC (32), we observed a significant increase in GATA6⁺ (classical) cells and a decrease in S100A2⁺ (basal) cells 10 days following MRTX1133 therapy (Fig. 3C, D). The human PDAC cells underwent morphological changes under KRAS inhibition, showing glandular differentiation and intracellular mucin, suggestive of a phenotypic shift (Fig. 3C). Interestingly, chemotherapy showed a trend towards an opposite finding, with (S100A2⁺) basal cells becoming enriched under treatment and cancer cells showing elongated, fibroblastic morphology and loss of epithelial architecture (Fig. 3C). These findings are consistent with basal subtype cells harboring chemoresistance and sensitivity to KRAS inhibitors, and vice versa for classical cells.

Figure 3. KRAS inhibition enriches for the classical epithelial state in human PDAC.

(A) Experimental design to investigate KRAS(G12D) inhibition in human PDAC PDX models using MRTX1133. (B) Tumor volume in response to MRTX1133 treatment (n = 7–8 mice/group). Error bars indicate SEM. (C) Immunofluorescence of classical (GATA6⁺) and basal (S100A2⁺) states in PDX PC106 after 14 days of MRTX1133 or FOLFIRI treatment. Note glandular features and intracellular mucin in MRTX1133-treated tumors. Scale bar, 100 μm (D) Quantification of classical (GATA6⁺) and basal (S100A2⁺) states in treated tumors. (n = 3–4 tumors/group). One-way ANOVA was used to test for statistical significance. Error bars indicate SD. (E) Representative flow cytometry plots of human PDX cells (mouse H-2Kd⁻/mouse CD45⁻/human CD45⁻/mouse CD31⁻/mouse CD11b⁻/mouse TER119⁻/DAPI⁻); depicting expression of TSPAN8 (x-axis; classical) and CD44 (y-axis; basal) in PDAC tumors from PDX PC69 after MRTX1133 compared with vehicle control. (F) Quantification of the proportion of classical/basal states using flow cytometry after MRTX1133 compared with vehicle control, n = 3 mice/group. Unpaired t tests were used to test significance. Error bars indicate SD. (G) Principal component analysis (PCA) of human PDX cells (PC69) from MRTX1133-treated vs control tumors (n =3 in each group, n = >2000 cells per condition). Unsupervised clustering of PDAC single cell transcriptomes, colored and annotated based on human gene signatures (13). (H) Heatmap depicting the difference in gene signature scores between MRTX1133-treated tumors, as compared with vehicle-treated tumors, for each PDX model. Classical and basal signatures drawn from four independent human data sets (12–15). Difference calculated with the average signature score over all single-cell transcriptomes per condition; n = 3 tumors/condition; ~2000 cells/condition. Statistical significance assessed over all single-cell transcriptomes between treatments by a two-sided Wilcoxon rank-sum test. (I) Immunofluorescence of classical marker genes (galectin-4, GATA6, and TFF-1) in a longitudinal tumor tissue biopsy obtained from a liver metastasis in a human patient harboring KRAS G12C PDAC prior to and on adagrasib therapy. Note increase in TFF-1 (classical) expression on treatment. Scale bar, 200 μm. (J) Schematic summary of key findings.

We developed a flow cytometry-based phenotyping method for human PDAC cells, using TSPAN8 as a marker for classical identity, and CD44 as a marker for basal identity, which we validated in a human PDX (Supplementary Fig. 6B). Using this phenotypic readout, we observed a strong enrichment of classical states under MRTX1133 treatment (Fig. 3E, F), a phenotype that persists at 4 weeks of treatment (Supplementary Fig. 6C).

ScRNA-seq analysis of the three PDXs revealed that a major source of transcriptional variation in the MRTX1133-treated and vehicle treated tumors occurred along the classical-basal axis, similar to our mouse model data (PC3; Fig. 3G; Supplementary Fig. 6D). MRTX1133-treated PDXs exhibited a pronounced shift towards the classical identity, as observed along the principal component axis (Fig. 3G). In both PDXs, we also observed that human gene signatures of classical identity were enriched, and basal signatures diminished upon treatment relative to vehicle (Fig. 3H). Similar findings were observed upon analysis of publicly available bulk RNA-sequencing data obtained from a MRTX1133-treated human PDAC cell line xenograft after 7 days of treatment (4) (Supplementary Fig. 6E). Given the association between KRAS genomic amplification, subtype identity, and KRAS inhibitor resistance (8,33), we further explored whether genomic amplifications of KRAS contribute to acute resistance to KRAS inhibitors. We did not observe differences in KRAS copy number after acute treatment in our PDX models (Supplementary Fig. 6F, G).

Finally, we obtained a pair of clinical samples from a patient with a KRAS-G12C mutant human PDAC, treated with the KRAS-G12C-specific inhibitor adagrasib (Fig. 3I). Biopsies were collected longitudinally from liver metastases before and on-therapy. We conducted multiplex immunohistochemistry on this case to assess malignant state identity (32). At baseline, the specimen showed expression of the classical state markers galectin-4 and GATA6 with no basal marker expression. During treatment, these markers were retained and a significant increase in another classical marker, TFF-1, was observed, indicating a shift toward classical identity from baseline (Fig. 3I).

Discussion

Our study reveals that classical PDAC cells play a critical role in mediating resistance and acute adaptation to KRAS inhibitors. We observe this acute phenotypic shift towards classical identity across a range of in vivo models, including mouse models and human PDAC PDXs with multiple genetic backgrounds. The convergent nature of this response to KRAS inhibition contrasts sharply with the diversity of genetic mutations and transcriptional adaptations that have been observed at the time of acquired resistance to KRAS inhibitors in colorectal and lung cancer (7,8,34,35). Based on our findings, we propose targeting the classical cells may represent a therapeutic entry point with the potential to pre-empt the complex and diverse genetic acquired mechanisms of resistance (Fig. 3J).

There is precedent for exploiting vulnerabilities that emerge in the context of MAPK inhibition. For example, in BRAF-mutated thyroid cancer, which is typically poorly differentiated and resistant to radioactive-iodine therapy (RIT), MAPK inhibition can lead to the restoration of glandular architecture within tumors and re-sensitization to RIT (36). Similarly, we recently demonstrated that targeting KRAS in lung adenocarcinoma drives a shift of the cancer cells towards a highly differentiated alveolar identity (37). In PDAC, classical cells are marked not only by epithelial differentiation, but also by increased levels of mucin secretion and a transcriptional overlap with an endodermal/gastric lineage, which express common markers such as Cldn18.2, MUC1, and TFF1/2 (38). Excitingly, Cldn18.2 and MUC1 are already well-developed targets in oncology, with blocking monoclonal antibodies, antibody-drug conjugates, and CAR-T cells currently in clinical development (39). Our data strongly encourage combining such extant cytoablative therapies targeting classical PDAC states with KRAS inhibitors. Importantly, our findings suggest that classical state-directed therapies may only be beneficial in the context of the selective pressure of KRAS inhibition, rather than as standalone treatment, as we observed that elimination of classical cells alone did not impede tumor growth. Mechanistic studies aimed at elucidating the dependencies of drug-resistant classical cells are an important area for further study and may offer additional therapeutic opportunities. Moreover, exploring how resistant classical cells re-establish the basal/mesenchymal state at the time of tumor relapse may reveal strategies to combat plastic cell state shifts under drug treatment.

Several clinical studies have suggested that mainstay chemotherapy regimens, like FOLFIRINOX and gemcitabine/nab-paclitaxel, are more effective in classical-dominant subtype tumors when compared to the basal-dominant subtype—a finding that is corroborated by our results in PDX models (13,40). Thus, the differential activity of these agents on classical and basal states provides a potential explanation for the synergy between chemotherapy and KRAS inhibition, combinations which are currently under exploration in PDAC. Prior clinical studies have shown that the basal PDAC subtype portends a poor response to chemotherapy (31,40). Importantly, our findings indicate that the basal subtype may serve as a positive biomarker for response to KRAS inhibitors, a hypothesis that will be important to address in ongoing clinical trials.

It is notable that prior in vitro studies have suggested that basal/mesenchymal PDAC cell lines are more resistant to genetic inactivation of KRAS, and induction of mesenchymal-epithelial transformation e.g. through Zeb1 inactivation has the capacity to restore KRAS dependence in PDAC—findings which are inconsistent with our observations (15,41). Our prior work and the work of others has shown that growth conditions (2D vs. 3D culture) and the presence of growth factors dramatically impact PDAC cell states, which are likely to be critical determinants of treatment response (12,13). Our findings here further highlight the importance of investigating PDAC cell states using in vivo models, which faithfully approximate cell state heterogeneity observed in human disease. Our findings indicate that elevated KRAS activity is a characteristic feature of the basal/mesenchymal state, corroborating prior findings and providing a mechanistic explanation for the sensitivity of these states to KRAS inhibition (42).

Despite considerable data confirming the co-existence of classical and basal states in human PDAC, still little is known about the differential sensitivity of these states to treatments or about the cell state transitions induced by treatment. Obtaining longitudinal biopsies is particularly challenging in PDAC patients, where tissue sites are often inaccessible and yield limited amounts of tissue. In this study, we develop new genetic reporter systems that enable non-invasive, quantitative monitoring of cell state dynamics in developing tumors. The genetic tools developed in this study are versatile and provide a platform for extending these findings to other cell states and drug classes. Furthermore, we contribute flow cytometry markers for the identification and isolation of the classical, basal, and mesenchymal cancer cell subsets from mouse and human PDAC tissues. Using these PDAC models and approaches, we report the selective enrichment of the classical state upon KRAS inhibition. Further, we demonstrate the classical state drives tumor relapse and harbors the capacity for plastic cell state change in this context. Collectively, our results cast the classical PDAC state as an attractive target to limit resistance to KRAS inhibition and improve outcomes in patients.

Methods

Animal Studies.

All animal studies were approved by the MSKCC Institutional Animal Care and Use Committee (protocol # 17–11-008). MSKCC guidelines for the proper and humane use of animals in biomedical research were followed. All genetically engineered mice were maintained on C57/BL6 and Sv129 mixed backgrounds. Autochthonous Kras^LSL-G12D/+; Trp53^flox/flox; Pdx1-Cre; Rosa26^{LSL-tdTomato/+} (KPCT) mice were generated by crossing published alleles, as previously described (12). Mouse genotyping was performed at 2 weeks and tumors were harvested at planned endpoints or until they reached a humane endpoint (weight loss, distended abdomen, hunching). Mice were euthanized by CO₂ asphyxiation.

Orthotopic models.

Mouse orthotopic pancreas transplants were performed in NOD.Cg-Prkdc^scid; Il2rg^tm1Wjl/SzJ (NSG) mice. Orthotopic tumors were generated by implanting 50,000 cells from 2D culture into the tail of the pancreas, as previously published (12). A pancreas cancer cell line derived from a Kras^FSF-G12D/+; Trp53^frt/frt; Pdx1-FlpO; Rosa26^mTmG/+ mouse was generated with published alleles, as previously described (12). A Lgals4-MACD cell line from a Kras^FSF-G12D/+; Trp53^frt/frt in the C57BL/6J background was established (described below) and used in this study. Additionally, murine KPC4662 PDAC cells (Kras^G12D/+, Trp53^R172H/+, Pdx1-Cre mouse in the C57BL/6J genetic background) (43) with a Porcn-DMAC reporter, as previously generated (bioRxiv 2023.12.30.573100), were used in this study. Primary 2D mouse cell lines were cultured in RPMI (Gibco) supplemented with 1% GlutaMax (#35050061, Thermo Fisher Scientific), 1% Pen/Strep (#15070063, Thermo Fisher Scientific), and 10% Heat-Inactivated FBS (#SH30910.03, Hyclone). All mouse cell lines were tested for Mycoplasma every 2 months using a PCR assay and confirmed to be negative. Cells were thawed and then passaged for 1–2 months prior to orthotopic transplantation. Animal studies began approximately 3 weeks after implantation. All animals were monitored weekly, and tumors were harvested at a maximum size of 1 cm³ or if the animals reached a humane endpoint for euthanasia.

Patient-derived Xenograft Models.

PDXs were established under MSKCC IRB #06–107 and IRB #12–245. Participating patients signed written informed consent for these clinical trials and biospecimen protocols. This study was conducted in accordance with ethical guidelines in the Declaration of Helsinki. Focused exome profiling (MSK-IMPACT) was performed on clinical samples as a part of routine care at MSKCC (44). Mutational analysis of all PDXs is provided in Supplementary Fig. 6A. Before transplantation, PDXs were thawed and resuspended in S-MEM and mixed with Matrigel at a 1:1 ratio. 100,000 cells were implanted subcutaneously into one flank of an NSG mouse. PDX volume was measured with a digital caliper twice per week.

Administration of KRAS Inhibitors, Chemotherapy, and Diphtheria Toxin.

KPCT mice bearing autochthonous PDAC tumors or NSG mice bearing orthotopic KP cell line allografts were intraperitoneally administered freshly prepared MRTX1133 in captisol (#HY-17031, MedChem Express) at 30 mg/kg b.i.d., as previously described, starting at 7 weeks after birth or three weeks after transplantation of cell lines, or in both cases when tumors were palpable (4). Mice body condition was monitored daily, and no significant cytotoxic side effects were observed during treatment. In genetic ablation experiments, NSG mice bearing KP; Lgals4-MACD and KP; Porcn-DMAC cell line orthotopic allografts were administered 50 μg/kg DT (#D0564, Sigma) resuspended in sterile PBS every other day by intraperitoneal injection at the start point of treatment. In chemotherapy experiments, NSG mice bearing KP cell line transplants were treated with gemcitabine (25 mg/kg) twice weekly (day 1, 4, 8, 11) and nab-paclitaxel (30 mg/kg) (day 1, day 8) once weekly. PDX transplants were treated with MRTX1133, as above. For FOLFIRI experiments, 5-FU (30 mg/kg), irinotecan (73 mg/kg), and leucovorin (90 mg/kg) were administered intraperitoneally 3 times weekly for the duration of treatment. Vehicle controls were used for all drug treatments.

Isolation of Primary Pancreatic Tumor Cells (GEMMs, orthotopic, human PDXs).

Orthotopically transplanted and primary tumors were dissociated into single-cell suspensions and cancer cells were isolated for scRNA-seq. We provide detailed information in the Supplementary Methods.

Generation of Lgals4-MACD Reporter and Lineage-Tracing Cell Line.

Homology arms ∼1,500 bp in length 5′ and 3′ flanking the end of Lgals4 exon 10 were amplified from C57/BL6 genomic DNA using high-fidelity PCR (KME-101, Toyobo). A homology-directed repair template donor vector where frt-PGK-Hygro-pA-frt-T2A-mScarlet-Akaluc-T2A-CreER-P2A-DTR is flanked by the homology arms was cloned into the pUC19 plasmid backbone (#638949, Takara). Donor vector and RNP complex were transfected into mESC line harboring Kras-frt-stop-frt(FSF)-G12D/⁺, Trp53(frt/frt) by electroporation (4D Nucleofector, Lonza). Clones were validated by genotyping and sequencing using primers below specifically detecting DTR, mScarlet, and spanning the homology arm. Chimeric F0 mice were obtained by injecting the donor ESCs into host embryos at the 8-cell stage. Pancreatic epithelial cells were isolated from F0 Kras-frt-stop-frt(FSF)-G12D/⁺; Trp53frt/frt; Lgals4-MACD^+/− wild-type hybrids and cultured in 3D spheres, as previously described (17). After establishment, the cells were transduced by lentivirus PGK-Flpo to remove the frt-TAG-bGlobinpA-(PGK-Hygro-pA)i-frt “STOP” cassette. To select for excision of the frt-stop-frt cassette, Nutlin-3 (#S1061, Selleck Chemicals), which selects for p53 knockout, was dissolved in DMSO (stock concentration of 10 mM) and used at a final concentration of 10 μM. Excision of the STOP cassette was confirmed by genotyping using PCR spanning mScarlet and the left homology arm. (Supplementary Fig. 3A, B; Supplementary Table 2). The lentiviral lineage-tracing vector (PGK-GLuc-miRFP-EFS-lox-BFP-lox) was integrated into the Lgals4-MACD reporter tumor organoids by spin infection at 300 × g for 30 minutes at 37°C. Transduced cells were subsequently expanded in 2D culture prior to orthotopic transplantation.

In vivo lineage-tracing.

NSG mice transplanted with KP cell lines containing Lgals4-MACD and lineage tracing vectors were established and monitored as above. To evaluate Lgals4⁺ classical cells at the time of relapse, tumor-bearing mice were treated with one dose of tamoxifen (200 mg/kg via oral gavage) at day 6 on treatment with MRTX1133, and MRTX1133 was withheld after 48 hours to allow lineage tracing to occur while on treatment and relapse to occur subsequently. Tamoxifen was dissolved in corn oil at 20 mg/mL and dissolved at 60°C for 1 hour.

In vivo luciferase imaging.

NSG mice transplanted with AkaLuc-expressing orthotopic tumors were dosed with 100μl of 30mM of AkaLumine-HCl substrate resuspended in PBS (#808350, Millipore Sigma, TokeOni) and imaged on an IVIS Lumina II (Perkin Elmer). Signal was recorded for 20 15-second exposures to capture peak luminescence. AkaLuc values were calculated through analysis on the Living Image software (Revvity). The total flux (p/s) for each mouse was calculated for the 20 exposures, and the peak value was used for each timepoint in individual animals. A fold-change over time was generated by dividing the calculated value at each time point from the pre-treatment value.

Plasma sampling and Gaussia luciferase measurements.

Whole venous blood was harvested by puncturing the submandibular vein and collecting 100μl in blood collection vials (#02–675-185, Fisher Scientific) (45). Plasma was separated by centrifugation at >8000 g for 10 minutes at 4°C and diluted 1:10 in PBS. Plasma was treated with the Gaussia luciferase substrate Coelenterazine-h (#301, NanoLight) at 200uM and luminescence was immediately measured on an BioTek Cytation 5 (Agilent). When calculating the tumor burden from Gaussia luciferase measurements, two technical replicates from the serum measurement were averaged and background (control serum from a non-tumor bearing mouse) was subtracted. A fold-change from the pre-treatment value was calculated at each time point for each individual mouse within the treatment cohorts. To calculate the fold change in the classical tumor fraction over time, the AkaLuc ratio (as calculated in the prior section) was divided by the Gaussia luciferase ratio (fold-change in tumor burden) at each time point. Reported values represent the mean and SEM across individual mice in each treatment cohort.

Tissue histology and immunofluorescence.

Tissue immunofluorescence was performed as before (37). Detailed information about the protocols and antibodies used can be found in the Supplementary Methods and in Supplementary Table 2.

FISH Analysis.

KRAS FISH analysis was performed using a 2-color KRAS/Cen12 probe mix, developed at MSKCC (8). The probe mix consisted of bacterial artificial chromosome clones containing the full-length KRAS gene (clones RP11–29515 and RP11–707F18; labeled with red dUTP), and a centromeric repeat plasmid for chromosome 12 served as the control (clone pa12H8; labeled with green dUTP). Probe labeling, hybridization, washing, and fluorescence detection were performed according to standard procedures. Representative images were taken at 63x for each treatment group and at least 20 cells were manually reviewed and scored for quantification.

Analysis of PDAC patient undergoing adagrasib therapy.

Tissue from a patient with KRAS -G12C-mutated PDAC undergoing adagrasib therapy were obtained under MSKCC Institutional Review Board approval (IRB #19–408 and IRB #06–107). Participating patients signed written informed consent for these clinical trials and biospecimen protocols. This study was conducted in accordance with ethical guidelines in the Declaration of Helsinki. The human specimen was collected from a 62-year old woman with metastatic PDAC to the liver undergoing treatment with adagrasib. Specimens were analyzed using multiplex IF using Galectin-4, GATA6, TFF1, KRT17, S100A2, and phospho-ERK antibodies. The patient underwent serial liver biopsies 02/2022. MSK-IMPACT sequencing (targeted exome) prior to treatment showed KRAS G12C, TP53 p.N288Rfs*54, KDM6A p.X128_splice, NFE2L2 R34Q.

Image Analysis.

Images were processed using a customized pipeline in HALO A.I. module v.3.6.4134 (Indica Labs). HALO A.I. module was used to obtain nuclear segmentation. Annotations of nuclear staining were manually reviewed for accuracy, and filtered to exclude errors though a DAPI-based criteria using size and intensity thresholds. In mouse tumors, cancer cells were identified through a marker-based (pan-cytokeratin) intensity threshold, and the expression of cell-state based markers (mouse Galectin-4, mouse KRT17) were used to classify states as positive or negative. In patient-derived-xenografts, human cells were morphologically distinguishable from mouse stroma, and a custom A.I. classifier was used to separate tumor cells from stroma. Annotations were manually reviewed for accuracy. Thresholding parameters for each marker were kept consistent across all slides. For quantifying the strength of phospho-ERK expression in mouse tumors, a uniform radius around the nucleus of each cell was used to measure the mean cell intensity, and these values were plotted as a distribution across each cell state.

Quantitative PCR (qPCR).

RNA was isolated from FACS-isolated cell populations and quantitative PCR was performed using standard procedures. More detailed information about the protocols and antibodies used can be found in the Supplementary Methods and in Supplementary Table 2.

Generation and analyses of single-cell RNA-seq data.

Single-cell cDNA libraries were prepared with the 10X Genomics Chromium Single Cell 3’ kit (v2 and v3) according to manufacturer’s instructions, and sequenced on an Illumina NovaSeq S4 platform. Single-cell cDNA library preparation and sequencing have been performed at the Integrated Genomics Operation (IGO) core facilities of Memorial Sloan Kettering Cancer Center. Detailed computational analyses of the datasets are described in the Supplementary Methods.

Statistics and reproducibility

GraphPad Prism software v.9.5.1 was for statistical analyses or in-house scripts in Python, which are available from the corresponding author upon request. Variance was similar between compared groups and p-values were determined by two tailed Student’s t-test for all measurements comparing untreated to treated samples of single time points. One-way analysis of variance (ANOVA) with Sidak’s multiple comparisons correction listed in the figure legends for comparisons across more than two groups. Figure legends denominate statistical analysis used. Investigators were not blinded to allocation during experiments and outcome assessment. Detailed computational and statistical approaches are described in the Supplementary Methods.

Statement of Significance.

KRAS inhibitors hold promise in pancreas cancer, but responses are limited by acquired resistance. We find a classical epithelial cancer cell state is acutely resistant to KRAS inhibition and serves as a reservoir for disease relapse. Targeting the classical state alongside KRAS inhibition deepens responses, revealing a potent therapeutic strategy.

Data Availability.

The data generated in this study are available within the article and its supplementary data files. scRNA-seq raw and processed data reported in this article have been deposited at the NCBI Gene-Expression Omnibus with accession number GSE271300 and are publicly available. Jupyter notebooks documenting the computational pipeline are available on GitHub (https://github.com/dbetel/PDAC_Singhal). All other raw data are available upon request from the corresponding author.