Plasticity-induced repression of Irf6 underlies acquired resistance to cancer immunotherapy

Acquired resistance to immune checkpoint immunotherapy remains a critical yet incompletely understood biological mechanism. Here, using a mouse model of pancreatic ductal adenocarcinoma (PDAC) to study tumor relapse following immunotherapy-induced responses, we found that tumors underwent an epithelial-to-mesenchymal transition (EMT) that resulted in reduced sensitivity to T cell-mediated killing. EMT-transcription factors (EMT-TFs) ZEB1 and SNAIL function as master genetic and epigenetic regulators of this tumor-intrinsic effect. Acquired resistance was not due to immunosuppression in the tumor immune microenvironment, disruptions in the antigen presentation machinery, or altered expression of immune checkpoints. Rather, EMT was associated with epigenetic and transcriptional silencing of interferon regulatory factor 6 (Irf6), which renders tumor cells less sensitive to the pro-apoptotic effects of TNF-α. These findings show how resistance to immunotherapy in PDAC can be acquired through plasticity programs that render tumor cells impervious to T cell killing.


Introduction
Immune checkpoint blockade (ICB) has transformed cancer treatment for multiple malignancies 1, 2 , but durable clinical responses remain elusive in many patients. Acquired resistance to immunotherapy represents an important clinical challenge. For example, approximately one third of patients with metastatic melanoma who had objective responses to ICB in a recent clinical trial subsequently relapsed over ~2 years of continuous therapy 3 . Similarly, although PD-1 and PD-L1 therapeutic blockade has revolutionized treatment of patients with non-small cell lung cancer (NSCLC), relapse after initial response is a major challenge 4 .
Unlike primary resistance, which is attributable to factors present when therapy is initiated, acquired resistance emerges over time. In patients treated with targeted therapies (e.g. EGFR inhibition), acquired resistance is often associated with cellular plasticity, a phenomenon that broadly describes changes in cell identity along a phenotypic spectrum 5 . One of the most well studied examples of cellular plasticity is epithelial-to-mesenchymal transition (EMT), during which carcinoma cells lose their epithelial features and acquire the more motile characteristics of broblasts and leukocytes 6,7 . In addition to its phenotypic effects on tumor cells, EMT has also been associated with the development of an immunosuppressive TME in NSCLC, breast cancer, and melanoma 8-11 . These ndings have implications for immunotherapy, as exempli ed by a recent report that the EMT transcription factor SNAIL promotes primary resistance to ICB in breast carcinomas, most notably through the CD73 pathway 12 .
Studies of acquired resistance to immunotherapy have been di cult due to a lack of tractable biological systems with which to model tumor relapses over prolonged periods of time. Here, we present an immunotherapy-sensitive model of pancreatic ductal adenocarcinoma (PDAC) in which recurrent (and therapy-resistant) disease appears weeks-to-months after a complete response to treatment, and delineate underlying cellular and molecular mechanisms.
To test this hypothesis, we established tumor cell lines from cohorts of mice that had different therapeutic outcomes following combination therapy. We rst compared therapy responses by reinoculating each tumor cell line into naïve mice and administering combination therapy. Cell lines derived from non-responsive tumors (termed "Early Progressor (EP) lines", n = 2) gave rise to tumors exhibiting variable responses to combination therapy, with rates of regression and CR that were comparable to mice bearing control cell lines that were never exposed to therapy (termed "Ctrl lines", n = 4) ( Fig. 1e-g). By contrast, cell lines derived from tumors that underwent CR followed by relapse (termed "Escape (Esc) lines", n = 8) gave rise to resistant tumors exhibiting poor survival ( Fig. 1e-g). These results suggest that the mechanisms allowing tumors to grow after therapy-induced CR (i.e. acquired resistance) persist in the Esc lines, whereas the mechanisms that render tumors non-responsive upon rst exposure to immunotherapy (i.e. primary resistance) are not stably preserved in the EP lines.
We considered two models that could explain the emergence of resistant tumors. First, tumor escape might have resulted from the selective expansion of pre-existing resistant clones. Alternatively, cell plasticity, in the setting of immunotherapy, might have fostered the emergence of newly resistant clones.
To distinguish between these models, we generated clonal PDAC lines from the 4662 parental line and examined therapy responses. Consistent with our earlier ndings with the 4662 parental line, individual clones exhibited heterogeneous therapeutic outcomes including escape tumors following CRs (Extended Data Fig. 1e), suggesting that tumor cell plasticity, rather than presence of pre-existing clones, accounts for the resistant phenotype. As expected, clonal tumors that escaped following an immunotherapyinduced CR (C10.e1 and C7.e1) were highly resistant to combination therapy compared to control tumors (Extended Data Fig. 1f).
Next, we set out to understand the mechanisms underlying acquired resistance. We began by performing bulk RNA sequencing on parental 4662 cells, EP cells, and Esc cells. Principle component analysis (PCA) revealed a strong similarity between parental cells and EP cells, while Esc cells diverged from both, suggesting that Esc cells had acquired a unique transcriptional pro le (Fig. 2a). We also observed striking morphological differences; namely, parental cells exhibited epithelial features and gave rise to welldifferentiated tumors, while Esc cells exhibited spindle-like features and gave rise to poorly differentiated tumors ( Fig. 2b and Extended Data Fig. 1g). In accordance with these observations, gene set enrichment analysis (GSEA) showed that Esc cells were highly enriched for the Hallmark EMT gene signature compared to parental tumors (Fig. 2c) and exhibited a decrease in mRNAs associated with the epithelial phenotype and an increase in mRNAs associated with the mesenchymal phenotype (Extended Data Fig. 1h, 2a). In addition to EMT, which was the most signi cantly enriched gene set in Esc tumors, other gene sets enriched in Esc tumors included interferon response, angiogenesis, hypoxia, and in ammation response, while gene sets that were reduced in Esc tumors included androgen/estrogen response and cholesterol homeostasis (Extended Data Fig. 2b). These ndings suggest that Esc tumors exhibit transcriptional hallmarks of EMT.
Given the well-documented role of EMT in various forms of therapy resistance 19 , we hypothesized that EMT was not merely correlated with acquired resistance to immunotherapy but was itself acting as a driver of resistance. To test this, we assessed whether gain-or loss-of-function of the Zeb1 and Snail EMT-TFs alters the response to immunotherapy (CD40 agonist plus ICB). First, we con rmed that overexpression of Zeb1 and Snail in parental tumors, or ablation of both genes in Esc tumors, prompted the expected changes in epithelial-mesenchymal phenotype and morphology (Extended Data Fig. 2c-f).
Next, we performed RNA sequencing of the engineered lines. GSEA revealed that Zeb1 and Snail overexpression in parental tumors (Zeb1/Snail OE) resulted in the enrichment of gene signatures associated with Esc tumors in vitro and in vivo (Fig. 2d, e), indicating that the transcriptional changes induced by these EMT-TFs resemble those associated with acquired resistance to immunotherapy. Consistent with these ndings, Zeb1/Snail OE tumors exhibited reduced responses to immunotherapy compared to EV-transduced controls, resulting in poorer survival (Fig. 2f, g). By contrast, ablation of Zeb1 and Snail (Zeb1 −/− Snail −/− ) in Esc tumors rescued the response to immunotherapy, leading to greater survival in treated mice (Fig. 2h, i). Importantly, none of the tumor cell lines tested above had a defect in cell growth in vitro (Extended Data Fig. 2g-i). Taken together, these data support the hypothesis that EMT promotes acquired resistance to immunotherapy.
EMT drives tumor cell-intrinsic resistance to cytotoxic T cell activity Resistance to immunotherapy in various PDAC models has been associated with an immunosuppressive TME characterized by abundant granulocytic myeloid-derived suppressor cells (gMDSCs) and a paucity of dendritic cells (DCs) and CD8 T cells 16,20,21 . To determine whether EMT fostered the creation of an immunosuppressive TME, we compared the immune pro les associated with parental, Esc, and Zeb1/Snail OE tumors. Contrary to expectations, both Esc and Zeb1/Snail OE tumors exhibited decreased in ltration of gMDSCs and increased in ltration of CD103 + DCs (cDC1) and CD8 + T cells compared to parental tumors (Fig. 3a-c). Similar results were observed in the orthotopic setting (Extended Data Fig. 3a, b), suggesting that EMT has similar effects on the TME regardless of the site of injection.
Likewise, the expression of co-inhibitory molecules such as PD-L1 and CD73 was reduced in both Esc and Zeb1/Snail OE tumors, although the expression of the TIGIT ligand CD155 was slightly increased compared to parental controls (Extended Data Fig. 3c, d). Conversely, ablation of Zeb1 and Snail in Esc tumors resulted in a paradoxical increase in immunosuppressive gMDSCs and a decrease in total T cells (Extended Data Fig. 3e). These results suggest that EMT promotes resistance to immunotherapy by mechanisms other than the creation of an immunosuppressive TME.
To determine whether loss of MHC I and/or Ag presentation accounted for acquired resistance in our model, we transduced parental and Esc tumor cells with chicken ovalbumin (OVA) and assessed the relative intensity of antigen presentation using an antibody (Anti-H-2K b bound to SIINFEKL) that detects OVA peptide in the context of MHC class I. Although Esc lines exhibited mild heterogeneity in antigen presentation (some slightly increased and some slightly decreased compared to parental; Extended Data Fig. 4a), all OVA-expressing Esc lines exhibited marked resistance to killing when co-cultured with OVAspeci c CD8 + T cells (OT-I) (Extended Data Fig. 4b). Importantly, Esc and Zeb1/Snail OE tumors in vivo had comparable MHC I expression compared to parental control tumors (Extended Data Fig. 4c, d). These results suggest that mechanisms other than MHC I downregulation account for EMT-associated acquired resistance in PDAC.
T cells play a crucial role in limiting tumor relapse following tumor clearance [22][23][24] . We con rmed in our model that animals achieving immunotherapy-induced CRs required T cells, but not NK cells, to maintain tumor control (Fig. 3d, e). This suggested two possibilities: (i) EMT induces a tumor cell intrinsic resistance to T cell killing, and/or (ii) EMT in tumor cells induces a defect in T cell function. To explore these possibilities, we transplanted OVA-expressing parental and Esc tumor cells into NOD/SCID mice and measured tumor growth before and after infusion with activated OT-I cells. OVA-expressing parental tumors responded to the OT-I infusion with a slowing of tumor growth (Fig. 3f). However, OVA-expressing escape immune surveillance by evading antigen-speci c T cell immunity.
Next, we established an in vitro co-culture system to determine whether the mechanism of resistance involves a direct interaction between tumor cells and CD8 T cells. Whereas parental cells were killed in the presence of either activated or non-activated OT-I cells (measured as positivity for AnnexinV and 7-AAD), Esc cells were highly resistant to OT-I killing (Fig. 3g). Next, to determine whether resistance to killing was cell-autonomous, we labelled parental and Esc tumors with different uorescent markers and co-cultured them individually or together with OT-I cells. In this mixed co-culture, Esc cells were far more resistant to killing than parental cells (Fig. 3h), suggesting that a tumor cell-intrinsic block to T cell killing drives the resistance phenotype. In addition, we found that naïve OT-I cells were poorly primed when co-cultured with Esc cells compared to parental cells (Extended Data Fig. 4e, f). Consistent with this nding, we noted that parental cells mixed with Esc cells had reduced cell death in OT-I co-culture compared to parental cells co-cultured alone (Fig. 3h, compare "Parental" to "Parental + Esc"). Taken together, these results indicate that Esc cells are intrinsically resistant to T cell killing and have a mild defect in T cell priming ability.
Because resistance arose in the setting of plasticity rather than outgrowth of a pre-existing resistant subclone (Extended Data Fig. 1e), we next studied whether the immune pressure present in the in vitro OT-I co-culture system might induce a similar change. Remarkably, parental cells that survived 2 days of coculture exhibited reduced expression of the epithelial marker E-cadherin, suggestive of an EMT-like process (Extended Data Fig. 5a). To determine whether immunotherapy prompts a similar shift in epithelial-mesenchymal phenotypes in vivo, we implanted parental tumor cells and compared the transcriptional pro les of tumors treated for 2d to those of control (untreated) tumors. GSEA revealed that even this short-term immunotherapy caused tumor cells in vivo to become enriched for signatures associated with the Esc tumors and the Hallmark EMT signature (Extended Data Fig. 5b, c). Together, these ndings strongly suggest that immune pressure selects for PDAC cells with a more mesenchymal phenotype that confers resistance to T cell killing. In line with this idea, we found that Zeb1/Snail OE rendered parental tumor cells resistant to killing by OT-I cells (Fig. 3i), whereas depletion of Zeb1 and Snail made Esc tumor cells more sensitive to T-cell killing (Fig. 3j).
Transcriptional and epigenetic regulation of Irf6 contributes to acquired immunotherapy resistance Given the stability of the EMT-associated resistance phenotype, we reasoned that the underlying mechanism was likely to involve epigenetic remodeling. Consequently, we performed ATAC-seq on parental and Zeb1/Snail OE tumors to identify genes whose chromatin accessibility changed upon EMT induction (in both steady state and in co-culture with OT-I cells) (Fig. 4a). In parallel, we examined the overlap of EMT-associated transcriptional differences across two experimental comparisons -(i) parental cells vs. Esc cells and (ii) EV-vs. Zeb1/Snail-transduced parental cells -and then used GSEA to compile a list of candidate genes whose transcriptional regulation correlated with immune sensitivity across both datasets (Supplementary Table 1). An examination of these epigenetically and transcriptionally regulated gene lists yielded a single gene candidate common to both: interferon regulatory factor 6 (Irf6).
Based on these ndings, we hypothesized that Irf6 plays a role in EMT-associated resistance to T cell killing. As predicted bioinformatically, Zeb1/Snail OE resulted in a loss of chromatin accessibility of the Irf6 locus (Fig. 4b), particularly at the promoter region ( Fig. 4c and Extended Data Fig. 6a, b), leading to a corresponding decrease in Irf6 mRNA (Fig. 4d). In agreement with this nding, Zeb1/Snail OE also resulted in the downregulation of putative Irf6 target genes but not those of unrelated transcription factors such as Six2 ( Fig. 4e and Extended Data Fig. 6c-e). Next, using gene sets generated via ectopic expression of Irf6 in PDAC cells, we found that Irf6 signatures were strongly enriched in therapy-sensitive parental and EV tumors compared to Esc and Zeb1/Snail OE tumors, respectively (Fig. 4f). Interestingly, an analysis of published human scRNA-Seq PDAC datasets revealed that IRF6 expression is largely restricted to epithelial cells, in contrast to other IRF genes [25][26][27] (Extended Data Fig. 6f-h). Furthermore, IRF6 signatures from human PDAC were consistently enriched in therapy-sensitive parental and EV tumors (Fig. 4g). Taken together, these results nominate Irf6 as a candidate EMT-related driver of immune sensitivity whose loss is associated with acquired resistance to immunotherapy.

Irf6 restoration promotes cytotoxic T cell killing and response to immunotherapy in resistant PDAC
To functionally assess the role of Irf6 in acquired immunotherapy resistance, we restored Irf6 expression to Esc tumors and assessed vulnerability to T cell killing in vitro. Esc tumor cells engineered to re-express Irf6 regained sensitivity to T cell killing upon OT-I co-culture ( Fig. 5a), whereas parental cells lacking Irf6 (Irf6 −/− ) became resistant to OT-I cell killing (Fig. 5b). Of note, although Irf6-expressing Esc cells maintained their mesenchymal morphology and had similar growth kinetics in culture (Extended Data Fig. 7a, b), epithelial genes such as Ecad, Ocln, and Cldn7 were upregulated compared to control (EV) cells (Extended Data Fig. 7c). Next, we tested the ability of Irf6 expression to restore immunotherapy responsiveness in vivo. Esc tumors expressing Irf6 partially recovered their response to immunotherapy compared to control (EV) tumors, with some mice exhibiting durable CRs and prolonged survival (Fig. 5c, d). These results suggest that loss of Irf6 in association with EMT promotes resistance to T cell killing in vitro and responses to immunotherapy in vivo, phenotypes that can be rescued by restoration of Irf6 expression.
To determine whether IRF6 might be associated with acquired resistance to immunotherapy in patients, we identi ed a dataset in which pre-and post-treatment transcriptome data were available from lung cancer patients who had initially responded to ICB but later developed resistance 28 . Roughly half of the patients with acquired resistance exhibited decreased expression of IRF6 compared to pre-treatment levels; in those patients, Irf6 signatures were enriched in pre-treatment samples (pre-ICB) compared to those with acquired immunotherapy resistance (IR) (Fig. 5e, f). Importantly, EMT signatures were inversely correlated with IRF6 expression, such that therapy-resistant patients with decreased IRF6 expression were enriched for EMT signatures compared to pre-treatment (Fig. 5g). By contrast, therapyresistant patients with no change or an increase in IRF6 expression showed the opposite result (Fig. 5g). Thus, immunotherapy resistance in a subset of lung cancer patients is associated with loss of IRF6 expression and concomitant acquisition of an EMT signature.

Irf6 facilitates T cell-mediated tumor control via TNF-induced apoptosis
To understand how Irf6 regulates tumor cell-intrinsic resistance to T cell killing, we compared the transcriptomes of control (EV) and Irf6-expressing tumor cells after OT-I co-culture. GSEA identi ed various hallmark gene sets as enriched (cholesterol homeostasis, MYC targets, estrogen response, TNF-α signaling via NFκB, etc.) or depleted (IFN response and EMT) following ectopic expression of Irf6 (Extended Data Fig. 7d). Given the known role of TNF and NF-κB in T cell-mediated killing of target cells 29-31 , we hypothesized that Irf6 loss confers resistance to T cell killing by blocking the pro-apoptotic effects of TNF-α. Consistent with this hypothesis, we found that Esc cells were markedly resistant to TNFα-induced cell death compared to parental cells and Irf6 re-expression restored sensitivity to TNFαinduced killing (Fig. 6a). Re-expression of Irf6 in Esc cells had no detectable effect on NF-κB pathway components (Extended Data Fig. 7e). TNF-α-induced killing was due to apoptosis, as Irf6-expressing cells exhibited greater staining for cleaved caspase-3 compared to control (EV) Esc cells, both in vitro following TNF-α treatment (Fig. 6b) and in vivo following immunotherapy (Fig. 6c). Treatment with the pan-caspase inhibitor z-VAD reversed the enhanced sensitivity to TNF-α of Irf6-expressing cells (Fig. 6d), and deletion of Tradd, Fadd, or Casp8 -genes encoding intracellular mediators of TNF-induced cell death -had similar effects (Fig. 6e).
T cells employ multiple redundant mechanisms to kill their targets. To con rm that death receptor signaling is critical for T cell cytotoxicity of PDAC cells, we assessed the consequences of Tradd, Fadd, or Casp8 deletion on T cell-mediated cytotoxicity. Whereas Irf6 expression restored the ability of OT-I cells to kill OVA-expressing Esc cells, loss of any of these apoptosis mediators signi cantly blunted the effect (Fig. 6f), suggesting that this pathway plays an important role in T cell killing of these PDAC cells. Finally, we used TNF-α neutralizing Ab to determine whether the dependency on Irf6 for e cient T cell killing was speci c for TNF-α. Whereas anti-TNF-α antibodies had no effect on T cell-mediated killing of control Esc cells in OT-I co-culture, antibody treatment reduced the killing of Irf6-expressing cells to the level of control Esc cells (Fig. 6g). Thus, Irf6 sensitizes PDAC cells to T cell-mediated apoptosis by altering the cellular response to TNF.
IRF6-related disorders in humans, including Van Der Woude syndrome, have been linked to point mutations in the IRF6 DNA binding domain 32 . These mutations introduced into the mouse Irf6 gene either partially or completely abrogated Irf6's ability to sensitize cells to TNF-α-induced killing (Extended Data  Fig. 7f), indicating a role for DNA binding in Irf6's effects. Together, these results suggest that Irf6 confers sensitivity to TNF-induced cytotoxicity through a classical TRADD-FADD-CASP8 death receptor signaling pathway.

Discussion
While predictors of primary resistance to immunotherapy in patients have been studied in detail 33-35 , there has been far less investigation of acquired resistance 36 . Our results suggest that distinct mechanisms operate in the two settings: whereas primary resistance is typically associated with a paucity of T cells and an immunosuppressive TME 16 , acquired resistance is paradoxically associated with a hyperimmune TME, where resistance to T cell-mediated cytotoxicity occurs through a change in cell state. Our data suggest that epithelial plasticity (EMT) confers resistance by providing PDAC cells with an intrinsic resistance to T cell killing rather than by assembling an immunosuppressive tumor microenvironment. Mechanistically, EMT drives repression of Irf6, making the tumor cells less susceptible to the pro-apoptotic effects of T cell-derived TNF-α.
Preclinical and clinical studies by us and others have provided encouraging results from combination therapies that include CD40 agonists 13-15, 17, 18, 37 . These studies, and our results here, indicate that therapy responses are heterogenous, re ecting various resistance programs. Oncogenic signaling, defects in Ag presentation, and immune-suppressive elements of the TME are known to induce primary resistance, whereas the mechanisms of acquired resistance are poorly understood and likely involve strategies distinct from those used by primary tumors 38 . This concept is reinforced by our nding that Esc tumors with acquired resistance exhibited a paradoxical increase in CD8 + T cell in ltration, a reduction in gMDSC in ltration, and a decrease in the expression of the co-inhibitory molecules PD-L1 and CD73 -effects that would all be predicted to enhance rather than diminish antitumor immunity. These ndings suggest that once a tumor has responded to immunotherapy, it activates mechanisms other than a recon guration of the TME to evade further immune attack.
An inverse relationship between EMT and immunotherapy response has been well-documented in mice and humans 8, 10-12, 39-41 . Our nding that EMT blunts the tumor cell's response to T cell-derived TNF-α is in line with previous reports implicating TNF-mediated killing as a crucial mechanism of tumor elimination, especially in poor neoantigen-expressing tumors 30,42 . Accordingly, genetic ablation or pharmacological inhibition of the TRAF2/cIAP complex, which facilitates TNF-induced cell death, resulted in improved antitumor responses when combined with ICB in preclinical studies 31,43,44 . In the context of these studies, our work suggests that strategies that re-establish sensitivity to death receptor-mediated killing, as reported recently in the setting of tumor cell intrinsic resistance to CAR T cell killing 45 , may reverse or prevent the emergence of resistance after an initial response to immunotherapy.
Our study revealed acquired resistance to be the product of plasticity rather than the outgrowth of preexisting mutant subclones. This nding is consistent with studies of targeted therapies such as BRAF inhibition, where acquired resistance can result from the outgrowth of rare subclones (i.e. cells carrying with mutations in the drug target) and/or non-genetic cell state transitions that enable the outgrowth of cells that are resistant on the basis of an altered phenotype 46 . Indeed, we found that co-culturing OVAexpressing tumor cells with antigen-speci c T cells for just 2 days resulted in a profound activation of EMT programs in the surviving cells (Extended Data Fig. 5). While we believe this is likely due to the selective outgrowth of cells with a more mesenchymal phenotype (i.e. preferential killing of cells with a more epithelial phenotype), we cannot rule out the possibility that T cells may also possess EMT-inducing properties.
The Irf family of transcription factors have broad activities in immune function that extend beyond their originally described role in type 1 interferon responses 47 . Unlike other Irf family members, Irf6's known roles are limited to the proliferation and function of epithelial (epidermal) cells 32,48 . In humans, heterozygous mutations in IRF6 are associated with van der Woude syndrome, a condition associated with facial malformations due to developmental dysmorphogenesis 49 . Our data show that EMT leads to the repression of Irf6 -either during spontaneous EMT in Esc cells or EMT induced by the expression of Zeb1 and Snail. Irf6 ablation in parental cells protects them from T cell killing, while Irf6 restoration leads to greater T cell killing and immunotherapy responsiveness, effects that may be related to a tumor cell's "TNF cytotoxicity threshold" 31 . Further support for this model comes from our nding that the subset of lung cancer patients whose tumors exhibited decreased IRF6 expression in the setting of immunotherapy resistance also exhibited a strong EMT signature.
While it is unclear how Irf6 loss exerts its protective in uence at a molecular level, a recent report in sh suggests that Irf6 can reduce the activity of IFN and NF-κB reporters in transfected 293T cells 50 . Future studies will be needed to understand how Irf6 regulates TNF cytotoxicity and to determine whether other pathways besides Irf6 and TNF signaling are dysregulated during EMT to contribute to acquired resistance.

Methods
Animals C57BL/6, C57BL/6-Tg(TcraTcrb)1100Mjb/J (OT-I), and NOD/SCID mice were purchased from the Jackson Laboratory and/or bred at the University of Pennsylvania. Mice were housed under a 12h-12h light-dark cycle, temperature of 18-23℃, and pathogen-free conditions. All animal procedures used in this study were performed following the National Institutes of Health guidelines. All mouse procedure protocols used in this study were in accordance with, and with the approval of, the Institutional Animal Care and Use Committee (IACUC) of the University of Pennsylvania (protocols 804643 and 805650).

Cell lines
The 4662 murine PDAC cell line was derived from a spontaneous pancreatic cancer in a female KPC mouse on the C57BL/6J background as previously described 51 . 4662 early progressor (n = 2) and escape cell lines (E1 ~ 8, C7.e1, and C10.e1) were isolated from non-responders and recurrent tumors reaching ≥ 3 mm mean tumor diameter beyond day 75, respectively, following inoculation of the 4662 mouse PDAC cell line and treatment with chemoimmunotherapy or immunotherapy alone. Single cell clones were derived from the 4662 PDAC cell line by limiting dilution. 4662 PDAC cell lines with a full length of OVA and tdTomato as a surrogate have been generated as previously described 51  Tumor implantation and therapy response assessment endpoint criteria included tumor volume exceeding 500 mm 3 , severe cachexia, or weakness and inactivity, as per our mouse protocols. Tumors were harvested 18-21 days following implantation or a week after treatment for ow and tissue analyses. Therapy response assessment was de ned as follows: Durable CR indicated the absence of palpable tumor at the completion of the experiment; CR with relapse indicated CR or near CR (≤ 1x1 mm tumor diameter) followed by progressive tumor growth; partial response denoted tumor regression to ≤ 30% of the maximum tumor diameter followed by progressive tumor growth; and nally, non-responders showed no response to therapy, transient stable disease, or limited response not meeting criteria for partial response.

In vivo treatment and T/NK-cell depletion
The general treatment schedule was previously described 17 . Brie y, mice with mean tumor diameter 5-7 mm (tumor volume around 100 mm 3 ) were enrolled and treated intraperitoneally (i.p.) with anti-PD-1 (clone RMP1-14, 200 µg/dose) for 7 times and anti-CTLA-4 (clone 9H10, 200 µg/dose) for 3 times every 3 days. Gemcitabine (Hospira) and nab-paclitaxel (Abraxane, Celgene) (120 mg/kg each), purchased from the Hospital of the University of Pennsylvania Pharmacy, were administered i.p. one time on the next day of rst ICB treatment. Anti-CD40 agonist (clone FGK45, 100 µg/dose) was co-injected i.p. with second dose of ICB. Control animals were treated with equivalent doses of isotype control antibodies. The pCDH-EF1-FHC vector, a gift from Richard Wood (Addgene plasmid #64874), and pCDH-CMV-EF1α-YFP vector were used for constitutive overexpression. Full-length mouse Zeb1 and Snail were gifts from Thomas Brabletz, University Erlangen, Germany. Murine Irf6 gene was ampli ed based on the cDNA template from parental tumors. Point mutations in Irf6 gene were performed using Q5 Site-Directed Mutagenesis Kit (NEB, E0552S) with mutagenic primers according to manufacturer's instructions and con rmed by sequencing.

RNA-seq
RNA was prepared from tumor cells at steady states, 1 d after co-culture, and 2-3 weeks following implantation as described above. RNA-seq libraries were prepared either by Novogene (California, USA) or with the NEBNext Ultra II RNA Library Prep Kit (NEB, E7770S) according to the manufacturer's protocol.
Libraries were then sequenced on Illumina next generation sequencers, generating either 150 bp paired end or 100 bp single end reads. Salmon v1.8.0 52 was used to generate raw counts in transcripts per million (TPM) through quasi-alignment to the mm39 reference genome using standard settings. The raw count matrix was subsequently imported into R-studio (R v4.1.2) for downstream normalization and differential gene expression analysis using DESeq2 53 . Genes were ranked by their Wald statistic for preranked gene set enrichment analysis (GSEA) 54,55 . Multiple published datasets were analyzed in the same manner. De novo and known motifs were identi ed within 500 bp of promoters of differentially expressed genes using HOMER's (v4.11) ndMotifs.pl command 56 .

Read1 -TCGTCGGCAGCGTCAGATGTGTATAAGAGACAG
Read2 -/5Phos/GTCTCGTGGGCTCGGAGATGTGTATAAGAGACAG Reverse -/5Phos/C*T*G*T*C*T*C*T*T*A*T*A*C*A*/3ddC/ Nuclei were isolated from 50,000 cells, followed immediately by transposition at 37 o C for 30 min. Transposed DNA fragments were puri ed using a Qiagen MinElute Kit, barcoded with primers based on Illumina TruSeq indices, and PCR ampli ed for 5 cycles using NEBNext High Fidelity 2x PCR master mix (NEB). Libraries were column-puri ed with the Qiagen PCR Cleanup kit, followed by 1.0x AMPure bead cleanup. Library quality was assessed on 4200 TapeStation (Agilent), and concentrations were quanti ed by Qubit D1000 assay (ThermoFisher Scienti c). Samples were sequenced using 150-cycle High Output NextSeq kits (Illumina, 20024907) to generate 75 bp paired end reads.
Analysis -Adapters were trimmed with Cutadapt v3.5 and reads were aligned to the mm39 mouse reference genome with Bowtie2 v2.4.4 58 . Sambamba v0.7.1 59 was used to lter out duplicates, while SAMtools v1.9 60 was used to identify and discard reads that aligned to the mitochondrial genome. Peaks were called with Genrich v0.6.1 (https://github.com/jsh58/Genrich) using standard settings in ATAC-seq mode and blacklisted regions were removed with BEDtools v2.30.0 61 . Overlapping peaks were identi ed with BEDtools, then merged. Raw read counts were determined with featureCounts v2.0.1 62 , then imported into R Studio for normalization and differential analysis using DESeq2. Differentially accessible loci were annotated to genomic features using ChIPseeker 63 . Genes and enriched biological pathways associated with differentially accessible loci were determined by GREAT 64 . Motif discovery within differentially accessible loci was done with HOMER's ndMotifsGenome.pl command.
Putative target promoters of transcription factors of interest were determined by overlapping mm39 promoters identi ed through ChIPseeker with known TF motifs. These target promoters were then intersected with the differentially accessible loci and annotated to their nearest genes using a mm39 annotations le and BEDOPS 65 . Genes were de-duplicated and their raw TPM counts were extracted from prior RNA-seq analysis. Genes were then ltered for fold changes of > = 2 or <=-2 and plotted by log2TPM.

scRNA-seq analysis
Published raw single cell RNA-seq data derived from patient samples were analyzed using the 10x and Seurat v4 pipelines 70 . Brie y, raw counts were determined with Cell Ranger using the hg19 reference genome, and imported into R Studio for analysis using Seurat, as previously described. Data were initially ltered to include cells with at least 100 genes and all genes in at least 3 cells. Samples were merged and further ltered by mitochondrial read percentage and total transcripts. Samples were then integrated and normalized, and variable genes were determined. Cell subpopulations were clustered based on the expression of certain gene markers. The epithelial cell cluster was further subsetted by the mean expression level of certain genes of interest, and the top differentially expressed genes were identi ed in the high (> mean) and low (< mean) subsets. These differentially expressed genes were used as gene sets for downstream GSEA.

Immuno uorescence and Hematoxylin & Eosin (H&E) staining
Tissues were xed in Zinc-formalin and embedded in para n for histological analysis and immuno uorescence (IF) staining. For IF staining, sections were depara nized, rehydrated, and prepared by antigen retrieval. They were then blocked in PBS with 0.3% Triton-X and 5% donkey serum for 1 hour, stained with primary and secondary antibodies, and mounted with Aqua Polymount (Polysciences). Primary antibodies used include chicken anti-GFP (Abcam, ab13970) and rabbit anti-cleaved caspase-3 (Cell Signaling Technology, 9661). Slides were visualized using an Olympus IX71 inverted multicolor uorescent microscope equipped with a DP71 camera. ImageJ FIJI software was used for quanti cation, with each data point an average of 2-3 elds per tumor section.
For H&E staining, sections were depara nized, rehydrated, stained with hematoxylin, differentiated with acidic ethanol, stained for eosin, dehydrated, and mounted with Permount. Slides were visualized using an Olympus BX41 microscope equipped with a DP25 camera.

Cell proliferation and viability assay
For checking cell proliferation, 10 4 tumor cells were seeded in each 12-well in triplicates and cell density at indicated time points was measured by staining with Hoechst 33342 Solution (Thermo Scienti c, 62249) and detecting by spectrometry. Data were normalized to background control and calculated by percent of cell growth compared to day 0.
Cell viability in response to TNF was determined as previously described 42 with some modi cation.

Immunoblot analysis
Tumor cells with or without TNF-α treatment or genetic modi cation were lysed in RIPA lysis buffer with protease and phosphatase inhibitor cocktail (Thermo Scienti c, 78444). Equivalent amounts of protein from whole cell lysates were separated by SDS-PAGE and transferred onto PVDF membranes (Bio-Rad).
Membranes were blocked in 5% nonfat milk in PBS plus 0.1% Tween-20 and stained with primary antibodies, followed by probing with horseradish peroxidase-conjugated secondary antibodies (Jackson Immunoresearch Software, statistics, and reproducibility PRISM and R were used for data processing, statistical analysis, and data visualization. The R language and environment for graphics (https://www.r-project.org) was used in this study for the bioinformatics analysis of RNA-seq and ATAC-seq data. The R packages used for all analysis described in this manuscript were from the Bioconductor and CRAN. Statistical comparisons between two groups were performed using unpaired two-tailed Student's t test. For comparisons between multiple groups, one-way ANOVA with Tukey's HSD post-test was used. For survival comparison between two groups, log-rank (Mantel-Cox) P values of Kaplan-Meier curves were determined using GraphPad Prism 9 (GraphPad). On graphs, bars represent either range or standard error of mean (SEM). For all gures, P < 0.05 was considered statistically signi cant, * indicates P < 0.05, ** P < 0.01, and *** P < 0.0001. No statistical method was used to pre-determine the sample size, which was chosen on the basis of previous experimental experience. No data were excluded from the analyses. For measuring tumor phenotypes in vivo, investigators were not blinded to group information, however, the results were reproducible by two independent researchers in different animal facilities. For in vitro experiments, blinding was not required because all the samples were analyzed in a consistent manner.

Declarations
Data availability The RNA-seq and ATAC-seq data generated in this study are available in the NCBI Gene Expression Omnibus (GEO) under accession numbers GSE222223 and GSE222224.

Code availability
All the code used in this study will be available upon request, including RNA-seq and ATAC-seq analyses.

Competing interests
Dr. Stanger receives research funding from Boehringer-Ingelheim and previously served as a consultant to iTeos Therapeutics. Dr. Vonderheide reports receiving consulting fees from BMS, is an inventor on a licensed patents relating to cancer cellular immunotherapy and cancer vaccines and receives royalties from Children's Hospital Boston for a licensed research-only monoclonal antibody.
Correspondence and requests for materials should be addressed to Robert H. Vonderheide or Ben Z. Stanger. and checkpoint blockades αCTLA-4 (C) and αPD-1 (P) Abs. Tumor growth (a) and survival (b) were monitored. c, The proportion of non-responders and mice with partial response (PR), relapsed after complete response (CR), and durable CR upon treatment was depicted. d, Mice with recurrent tumors after CR or near CR were re-treated with GFCP (blue arrows) and tumor sizes were measured. e-g, Tumor cell lines were generated from s.c. tumors with control IgG ('Ctrl' lines, n = 4), early progressor ('EP' lines, n = 2), and the relapsed after CR ('Esc' lines, n = 8) upon therapy, as denoted by blue, green, and red lines in a, respectively. Naïve WT mice challenged s.c. with these cell lines were treated with control IgG (n = 3 or 4 per line) or GAFCP (n = 5 or 6 per line) and tumor growth (e) and survival (f) were monitored. Response rates (g) in mice received each cell line followed by treatment are shown.    Irf6 loss contributes to EMT-induced immunotherapy resistance. a, OVA-tdTomato +