ZNF274 constrains lineage plasticity and drives intrinsic resistance to CDK7 inhibitors in pancreatic cancer

Abstract

Pancreatic Ductal Adenocarcinoma (PDAC) is characterized by two distinct transcriptional subtypes: classical and basal, which may interconvert. We show the KRAB-ZNF protein, ZNF274, correlates with PDAC subtype and regulates sensitivity to CDK7 inhibition by facilitating heterochromatin maintenance and gene suppression. We find ZNF274 loss drives a classical to basal transition, induces invasive protrusions and facilitates invasion capacity. We define two mechanistic arms to this regulation. ZNF274 directly suppresses ZEB1 expression. When ZNF274 is lost, ZEB1 facilitates acquisition of mesenchymal features, keratin gene expression, and susceptibility to CDK7 inhibition. Second, ZNF274 dampens expression of repetitive elements including human endogenous retroviruses (HERVs). HERV expression following ZNF274 loss induces a double stranded RNA response that reinforces the classical to basal subtype transition. Here, we show ZNF274 is an important epigenetic regulator of cellular plasticity and sensitivity to CDK7 inhibition, presenting a therapeutic liability of PDAC subtype transition that is actionable in the clinic.

Pancreatic ductal adenocarcinoma (PDAC) classical and basal subtypes can interconvert. Here, the authors discover that ZNF274 loss drives a classical to basal transition in PDAC via suppression of ZEB1 and HERV-driven interferon signaling, increasing the sensitivity to CDK7 inhibition.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) presents with poor prognosis as well as a rising incidence rate, driving it to become the 2nd-leading cause of cancer-related deaths in the US by 2030¹. The mutational spectrum in PDAC is very similar both between patients and between primary and metastatic lesions^2,3 yet, patients respond differently to standard of care chemotherapy⁴. PDAC is characterized by two major transcriptional subtypes: classical and basal^5–11. These PDAC transcriptional subtypes are determined by specific chromatin states and epigenetic landscapes¹². These epigenetic landscapes are defined by a complex regulatory network of super-enhancers and transcription factors that control cellular pathways and functions in the classical tumors, but which are lost in the basal tumors^12,13. Lineage plasticity exists within the subtypes¹⁰ and is likely governed by epigenetic mechanisms¹³. Loss of epigenetic and transcriptional regulators can increase phenotypic plasticity which is vital for tumorigenesis and cancer progression¹⁴. Plasticity between cell states enables transition from treatment-sensitive to treatment-resistant states¹⁵, which can be exploited to uncover new subtype specific therapeutic liabilities for this devastating disease.

ZNF274 is an epigenetic factor involved in regulating genomic stability and lineage specific gene clusters^16–18. ZNF274 is a member of the Krüppel-associated box (KRAB) zinc finger (ZNF) family, the largest class of sequence-specific DNA binding proteins in the human genome^19,20. KRAB-ZNF genes evolved in competition with repetitive elements^21–25 and cluster on chromosome 19^26,27. KRAB-ZNF proteins form a complex with KAP1 (a scaffold or co-factor), SETDB1 (histone methyltransferase), HDACs (histone deacetylases), DNMTs (DNA methyltransferases), and HP1 (heterochromatin protein 1)^28,29. This KRAB-ZNF complex facilitates silencing of repetitive elements through heterochromatin maintenance by removal of active chromatin marks and installation of repressive chromatin marks, including H3K9me3^28,30. ZNF274 is a master regulator of other KRAB-ZNF genes^16,17 and has been shown to directly bind SETDB1 and KAP1 in the KRAB-ZNF complex^31,32. Other KRAB-ZNFs have been found to regulate tumorigenesis in cancer^33–38. Howevert nothing is known about the role of ZNF274 in pancreatic cancer or how ZNF274 regulation of heterochromatin maintenance effects cellular plasticity and drug sensitivity.

Here, we show ZNF274 functions to suppress the basal PDAC cell state through direct regulation of ZEB1 and human endogenous retroviruses (HERVs). We provide functional evidence of basal state migration in classical PDAC lines and organoids, driven by loss of ZNF274. We show that aberrant expression of HERVs activates the dsRNA and IFN response pathways which in turn facilitate commitment to the basal cell state. We demonstrate that ZNF274’s ability to regulate PDAC cell state has specific therapeutic implications. ZNF274 controls sensitivity of PDAC tumors to CDK7 inhibitors, identifying a subtype specific therapy for this disease. Altogether, our data indicates that ZNF274 is an important epigenetic regulator with a broad impact on cellular plasticity and sensitivity to subtype specific therapies.

Results

ZNF274 controls cell state plasticity

We investigated the biological function of ZNF274 in PDAC plasticity, given its known role in regulating lineage specific gene clusters. Ablation of ZNF274 expression, by knockdown, in classical PDAC cell lines resulted in a dramatic morphological change. Classical PDAC cell lines present a more epithelial phenotype with growth in colonies while basal PDAC cell lines present a more mesenchymal phenotype with growth as individual, spindle shaped cells³⁹. Interestingly, we observed classical PDAC cell lines lose their epithelial phenotype and acquire a more mesenchymal phenotype as seen by spindle shaped cells and a more diffused growth pattern upon ZNF274 loss (Fig. 1A). Following this observation, we assessed ZNF274 expression in our basal and classical PDAC cell lines and organoids. Classical PDAC presented higher expression of ZNF274 than basal PDAC (Fig. 1B, C) suggesting a correlation between ZNF274 expression and PDAC subtypes. We further sought to validate these findings in a subtyped human PDAC tumor dataset. Analysis of ZNF274 expression showed classical tumors trend towards higher, but not statistically significant, RNA levels of ZNF274 (Fig. S1A, p = 0.31, Wilcoxon rank sum test). ZNF274 activity, predicted using transcription factor enrichment analysis with the ChEA3 database, was significantly increased in classical tumors compared to basal tumors (Fig. S1B, p = 6 × 10–6, Wilcoxon rank sum test). Overall, these findings support greater transcription factor activity of ZNF274 in classical tumors.

Fig. 1. ZNF274 control of the keratin program and PDAC subtype specific markers.

A Phase contrast images of live classical PDAC cells and brightfield color images of crystal violet-stained classical PDAC cells treated with non-specific or ZNF274-specific siRNAs. Images have a scale bar of 100 μm. Images are representative of 3 replicates. B qPCR for ZNF274 in a panel of human PDAC cell lines. Bar plots are displayed as mean ± SEM; Biological replicates: n(AsPC1) = 6, n(SUIT2) = 5, n(CAPAN2) = 5, n(CFPAC1) = 4, n(HPAFII) = 4, n(DANG) = 3, n(8988 T) = 5, n(MiaPaCa2) = 4, n(BxPC3) = 5, and n(Panc1) = 4; asterisks, P values after one-way ANOVA. ****P = < 0.0001 Classical vs Basal. C Immunofluorescence images of ZNF274 stained embedded classical (black) and basal (red) PDAC organoids. Images have a scale bar of 100μm. Images are representative of 4 replicates. D WB for ZNF274, TP63ΔN, keratin17, keratin5, and ACTIN in classical PDAC cells lines treated with non-specific or ZNF274-specific siRNAs. The samples derive from the same experiment but different gels for ZNF274 and another for TP63ΔN were processed in parallel. Data are representative of 2 replicates. E Heat map of qPCR data for GATA6, TP63deltaN, KRT5, KRT17, and KRT81 in classical PDAC cell lines treated with non-specific or ZNF274-specific siRNAs (biological replicates, n = 2). **P $\le$ 0.01, ***P $\le$ 0.001, ****P $\le$ 0.0001 (two-tailed unpaired Student’s ttest). F Gene Set Enrichment analysis (GSEA) of RNA-seq data comparing ZNF274 OE and control in basal PDAC cell lines for the computational GSEA set (biological replicates, n = 3). G, qPCR data for GATA6 and KRT5 in a basal PDAC cell line infected with an empty vector or ZNF274 overexpression vector (biological replicates, n = 2). H Immunofluorescence images of KRT81 and GATA6 stained and embedded classical PDAC organoids with doxycycline induced control or ZNF274-specific shRNAs. Images have a scale bar of 100 μm. Quantification was performed in FIJI using gray scale intensity values. All images were taken on the Nikon Live microscope. Biological replicates: n(GATA6 shCTRL) = 17; n(GATA6 shZNF274) = 24; n(KRT81 shCTRL) = 17; n(KRT81 shZNF274) = 24; asterisks, P values after two-tailed unpaired Student’s t-test; *P = 0.0179 GATA6 shCTRL vs shZNF274; ****P = < 0.0001 KRT81 shCTRL vs shZNF274. Source data and exact P values are provided within the Source Data file.

To further explore ZNF274 function in classical PDAC, we investigated the hypothesis that manipulation of ZNF274 expression may induce a state change, specifically loss of ZNF274 in classical PDAC may drive a cell state conversion from classical to more basal PDAC. The deltaN form of TP63 (TP63ΔN), Keratin5 (KRT5), and Keratin17 (KRT17) are all well-known markers of the squamous subtype which is the most aggressive form of basal PDAC^40–42. We assessed protein expression of all three basal markers for any changes induced upon ZNF274 knockdown (KD). We observed clear increases in TP63ΔN and KRT17 protein expression and a dramatic increase in KRT5 protein expression in two classical PDAC cell lines with ZNF274 KD compared to control (Fig. 1D). To further investigate the change in cell state markers when ZNF274 is lost, we assessed RNA expression of basal identity marks, TP63ΔN, KRT5, KRT17 as well as GATA Binding Protein 6 (GATA6), a clinical biomarker of classical PDAC^43,44, and Keratin81 (KRT81), a marker of aggressive PDAC that predicts worse overall patient survival and poor response to chemotherapy³⁹. We observed a significant increase in transcripts of TP63ΔN, KRT5, KRT17, and KRT81 (p < 0.01) and significant reduction in transcripts of GATA6 (p < 0.01) following ZNF274 KD (Fig. 1E). This suggests the loss of ZNF274 is driving a reduction in markers of classical PDAC and an acquisition of basal PDAC markers.

In contrast, RNA-seq analysis of ZNF274 overexpression (OE) in a basal PDAC cell line (Fig. S1C), was sufficient to drive an enrichment of epithelial and classical PDAC GSEA^11,45 gene signatures (Fig. 1F, S1D). To further validate this acquisition of a more classical cell state, we checked validated markers of classical PDAC and basal PDAC. GATA6 mRNA expression increased with ZNF274 OE while KRT5 mRNA expression decreased with ZNF274 OE (Fig. 1G), suggesting that ZNF274 is sufficient to drive a more classical cell state in PDAC.

Next, we analyzed gene sets enriched after ZNF274 depletion using GSEA (Fig. S1E). Extracellular matrix (ECM) and collagen genes were upregulated upon ZNF274 depletion (Fig. S1F). Keratin expression is correlated with cellular migration speed and ECM elasticity^46,47 and is dynamic in epithelial to mesenchymal transition (EMT)^48,49. This data suggests that acquisition of a more basal cell state could lead to phenotypic changes in cell extrinsic factors. We opted to validate that induction of a more basal cell state is achieved in a 3D human PDAC organoid model system. For this purpose, we created stable, doxycycline (dox) inducible shRNA organoids with either a control, non-targeting shRNA, (shCTRL) or a ZNF274 targeting shRNA (shZNF274). We used immunofluorescence to show a clear reduction in ZNF274 signal in the shZNF274 organoids compared to the shCTRL organoids (Fig. S1G). Then, we observed an activation of the basal program through increased expression of KRT81 and mild suppression of the classical PDAC subtype through reduction of GATA6 expression in these organoids following ZNF274 KD (Fig. 1H). This confirms these classical PDAC organoids are undergoing a cell state conversion to become more basal in nature.

Collectively, we have identified a necessary and sufficient dependency on ZNF274 expression and maintenance of the classical PDAC cell state. We uncovered that ZNF274 controls cell state plasticity through suppression of basal state keratin markers and that manipulation of ZNF274 expression can drive subtype switching between classical and basal PDAC.

ZNF274 is necessary to prevent migration and invasion of classical PDAC cells

We performed a scratch assay to evaluate the phenotype sequelae of the cell state conversion imparted by ZNF274 loss with respect to cell mobility and migration. Depletion of ZNF274 in classical PDAC cell lines increased cell migration via wound confluence over time (Fig. 2A). In contrast, overexpression of ZNF274 in a basal PDAC cell line decreased cell migration via wound confluence over time (Fig. S2A). This data suggests that ZNF274 influences migration potential as part of its regulation of cell plasticity in basal and classical PDAC. We next compared the baseline structure of our classical and basal PDAC organoids, since classical PDAC organoids are known to form glandular cysts^50,51. We found the classical organoids form a single layered, ordered structure while the basal PDAC organoids form a disordered, multilayered structure (Fig. S2B). We investigated if ZNF274 depletion affected PDAC organoid morphology, since organoid morphologic changes such as induction of multilayering through delamination are uniformly seen in conjunction with acquisition of mesenchymal features and cell mobility^52–55. ZNF274 depletion, using doxycycline-inducible shRNA, led to an increase in multilayering of the organoid structure in our classical PDAC organoids, as quantified by local thickness (Fig. 2B). Quantification of these organoids showed basal PDAC organoids have a significantly thicker cellular layer compared to classical organoids (Fig. 2C). These data demonstrate that ZNF274 depletion induces loss of classical PDAC morphology and drives acquisition of basal morphology and mesenchymal features including migration ability.

Fig. 2. Classical PDAC cells acquire increased migration and invasion capacities upon removal of ZNF274.

A Scratch assay plot showing percent of wound confluence over 24 h following doxycycline induction of control or ZNF274-specific shRNAs in two classical PDAC cell lines (left) and paired WB for ZNF274, ZEB1 and ACTIN (right) for each line. Line graphs are displayed as mean ± SEM; Biological replicates: n(ASPC1 shCTRL) = 16, n(APSC1 shZNF274) = 16, n(CFPAC1 shCTRL) = 10, and n(CFPAC1 shZNF274) = 10; asterisks, P values after two-tailed paired Student’s ttest; **P = 0.0016 ASPC1 shCTRL vs shZNF274; *P = 0.0439 CPFAC1 shCTRL vs shZNF274. B Brightfield color images of H&E slides from embedded classical PDAC organoids with doxycycline induced of control or ZNF274-specific shRNAs (left) and quantification of relative thickness by FIJI local thickness plug in (right). All images were taken on the Nikon Live microscope, main images have a scale bar of 100 μm, and inset images have a scale of 50μm. Data is representative of 4 independent replicates. C Quantification of relative thickness by FIJI local thickness plug in from brightfield color images of H&E slides from embedded classical and basal PDAC organoids. Bar plots are displayed as mean ± SEM; Biological replicates: n(PPTO69 shCTRL) = 9, n(PPTO69 shZNF274) = 14, n(PPTO97 shCTRL) = 15, n(PPTO97 shZNF274) = 20, n(Classical PPT069) = 9, n(Classical PPTO97) = 15, n(Basal PPTO74) = 8, and n(Basal PPTO154) = 8; asterisks, P values after two-tailed unpaired Student’s t-test. ****P $\le$0.0001 PPTO69 shCTRL vs shZNF274; ****P $\le$0.0001 PPTO97 shCTRL vs shZNF274; ****P $\le$0.0001 Classical vs Basal. D Brightfield images of live classical PDAC organoids with doxycycline induced control or ZNF274-specific shRNAs grown in collagen to assess invasion. Images were taken on the Leica DMi8 inverted microscope and have a scale bar of 200 μm. Quantification of number of spines protruding from shCTRL and shZNF274 classical PDAC organoids (right). Violin plot biological replicates: n(shCTRL) = 26, and n(shZNF274) = 32; asterisks, P values after two-tailed unpaired Student’s t-test. ****P $\le$0.0001 shCTRL vs shZNF274. E Confocal, immunofluorescence images from z-stack showing phalloidin (F-actin) in magenta and DAPI in cyan for doxycycline induced control or ZNF274-specific shRNAs classical PDAC organoids grown in collagen. Images have a scale bar of 50μm. Data are representative of 2 independent experiments. Source data and exact P values are provided within the Source Data file.

Next, we used our dox-inducible shZNF274 classical PDAC organoids to assess migration and invasion in a 3D collagen matrix assay. This assay mimics the invasion requirements of a cell squeezing through the extracellular matrix surrounding a tumor. For this assay, invasion is quantified as an increase in invasive strands or protrusion out of the organoid. Very strikingly, we observed a statistically significant increase in the presence of invasive strands in the ZNF274 depleted organoids as compared to an absence of invasive strands in the control organoids (Fig. 2D). Immunofluorescence staining of these organoids within the collagen matrix clearly revealed the dramatic morphological changes undergone following ZNF274 loss. The ZNF274 depleted organoids have acquired a disordered structure in which cells fill the center of the organoid (Fig. 2E), suggesting an increase in cell mobility. This observation, in conjunction with the invasive strands reaching into the collagen matrix, shows the creation of a dynamic organoid structure (Fig. 2E). ZNF274 KD in these organoids has effectively disrupted the cystic structure that is characteristic of the control organoids and previously reported of classical PDAC^50,51. Combined, these results demonstrate that ZNF274 is necessary to prevent migration and invasion in classical PDAC tumor cells.

Loss of ZNF274 drives the epithelial to mesenchymal transition

Given that loss of ZNF274 induces a more basal-like phenotype with mesenchymal features, we hypothesized that ZNF274 may also regulate EMT pathway components. Previous literature has identified regulators of the mesenchymal cell state, such as ZEB1, in high-grade (basal) PDAC⁵⁶. To investigate this hypothesis, we assessed both protein and RNA expression of markers of EMT to identify changes upon ZNF274 KD in classical PDAC cell lines. ZEB1, a transcriptional driver of EMT and master regulator of the mesenchymal state^57–59, showed a dramatic increase in both chromatin-bound protein and global protein expression when ZNF274 is lost (Fig. 3A, B). In contrast, EpCAM, an epithelial cell marker, was substantially decreased upon ZNF274 loss (Fig. 3A) and both EpCAM and another marker of epithelial cell state, E-cadherin, showed decreased protein expression in the absence of ZNF274 compared to control (Fig. 3A, B). We then assessed mRNA expression in our ZNF274 KD and control cells and observed the same pattern in an expanded panel of EMT markers where those characterizing the mesenchymal state (ZEB1, SNAIL1, SNAIL2) were upregulated and those characterizing the epithelial state (EpCAM and E-cadherin) were downregulated in the absence of ZNF274 (Fig. 3C). To validate this phenotype in our RNA-seq data, we ran GSEA pathway analysis for hallmark signatures and uncovered the epithelial to mesenchymal transition hallmark to be the most enriched pathway upon loss of ZNF274 (Fig. 3D). These data demonstrate that ZNF274 depletion drives a clear induction of the epithelial to mesenchymal transition.

Fig. 3. ZNF274 loss promotes epithelial to mesenchymal transition in Classical PDAC.

A WB for chromatin bound ZNF274, ZEB1, EpCAM, and Total H3 in classical PDAC cells lines treated with non-specific or ZNF274-specific siRNAs. The samples derive from the same experiment but a different gel for ZNF274 and EpCAM was processed in parallel. Data are representative of 2 independent experiments. B WB for ZNF274, ZEB1, EpCAM, E-cadherin, and ACTIN in classical PDAC cells lines treated with non-specific or ZNF274-specific siRNAs. The samples derive from the same experiment but a different gel for ZNF274 and another for E-cadherin was processed in parallel. Data are representative of 2 independent experiments. C Heat map of qPCR for ZNF274, markers of the mesenchymal cell state: ZEB1, SNAIL1 and SNAIL2; and markers of the epithelial cell state: EpCAM and CDH1 (E-cadherin) in classical PDAC cells lines treated with non-specific or ZNF274-specific siRNAs (biological replicates, n = 2). D Gene Set Enrichment analysis (GSEA) of RNA-seq data comparing ZNF274 KD and control classical PDAC cell lines (biological replicates, n = 3). Nominal Pvalue$\le$0.0001 ASPC1 ZNF274 KD vs Control; Nominal P-value$\le$0.0001 CFPAC1 ZNF274 KD vs Control. ChIP-qPCR for ZEB1 3 (promoter), ZEB1 5 (3’ exon), ZEB1 6 (3’ exon) with pulldown for ZNF274 (E), H3K9me3 (F), and SETDB1 (G) in a classical PDAC cell line with doxycycline induced control or ZNF274-specific shRNAs (biological replicates, n = 3). Asterisks, P values after two-tailed unpaired Student’s t-test. Figure E: *P = 0.029212 ZEB1 3 shCTRL vs shZNF274; **P = 0.003988 ZEB1 5 shCTRL vs shZNF274; *P = 0.012494 ZEB1 6 shCTRL vs shZNF274. Figure F: **P = 0.005997 ZEB1 3 shCTRL vs shZNF274; ns P = 0.051843 ZEB1 5 shCTRL vs shZNF274; ns P = 0.050052 ZEB1 6 shCTRL vs shZNF274. Figure G: **P = 0.006646 ZEB1 3 shCTRL vs shZNF274; **P = 0.003646 ZEB1 5 shCTRL vs shZNF274; **P = 0.001914 ZEB1 6 shCTRL vs shZNF274. Source data and exact P values are provided within the Source Data file.

To determine if ZNF274 directly regulates a driver of EMT, we explored ZEB1 as a potential target given that ZEB1 shows the most consistent expression change of all the EMT markers upon ZNF274 KD (Fig. 3A–C). ZEB1 is a master regulator of the EMT pathway, it is a known repressor of epithelial genes and an activator of mesenchymal genes^57–59. Specifically in pancreatic cancer, ZEB1 is a very important EMT transcription factor, known for promoting stemness, invasion and metastasis in pancreatic cancer mouse models^60,61. We hypothesized that ZNF274 may regulate H3K9me3 at the ZEB1 gene locus, thereby controlling ZEB1 expression and regulation of EMT. We first assessed changes in H3K9me3 as well as binding of ZNF274 and SETDB1 at the ZEB1 locus by using ChIP-qPCR on ZNF274 KD and control cells. We used the dox-inducible shZNF274 system in classical PDAC cell lines to enable prolonged depletion of ZNF274 and induction of ZEB1 (Fig. S2C). Since ZNF274 has been shown to bind the 3’ exons of other ZNF genes¹⁷, we designed ChIP primers to both the promoter and 3’ exons of ZEB1. Interestingly, ChIP pulldown with ZNF274 showed a significant reduction in ZNF274 binding at the ZEB1 locus (Fig. 3E, p < 0.05). Known ZNF274 bound genes (positive control genes: ZNF180 and ZNF554)¹⁷ were also significantly reduced (p < 0.01) comparing control and ZNF274 KD while genes not bound by ZNF274 (negative control genes: ZNF555 and ZNF556)¹⁷ showed little to no binding in either condition (Fig. S2D). H3K9me3 ChIP showed a decrease in H3K9me3 at positive control genes (ZNF180 and ZNF554, p < 0.05) and a significant decrease in H3K9me3 at the ZEB1 locus (p < 0.01) upon loss of ZNF274 (Fig. 3F, S2D). ChIP pulldown with SETDB1 showed a reduction in SETDB1 binding at the ZEB1 locus (p < 0.01) and positive control genes (ZNF180 and ZNF554, p < 0.05), and little to no change at the negative control genes (ZNF555 and ZNF556) in ZNF274 KD cells compared to control (Fig. 3G, S2D). In summary, we observed significant reduction in H3K9me3, ZNF274 and SETDB1 binding at the ZEB1 loci and positive control genes. Collectively, we have identified an additional function of ZNF274—the ability to directly regulate gene expression of ZEB1 in PDAC.

SETDB1 regulation and complex formation in classical PDAC

Given that SETDB1 functions as the histone methyltransferase component of the ZNF274 KRAB-ZNF complex^17,29, SETDB1 should be a main factor driving ZNF274 regulation of ZEB1 and EMT. Furthermore, SETDB1 has been implicated as a regulator of EMT in multiple cancer types^62,63. Therefore, we sought to investigate the role of SETDB1 regulation of EMT in PDAC. We reasoned that, if ZNF274 recruitment of SETDB1 and deposition of H3K9me3 is the mechanism by which ZNF274 is controlling ZEB1 transcription (Fig. 3E–G), then SETDB1 KD should induce a similar change in ZEB1 expression and downstream EMT targets. We knocked down SETDB1 in our classical PDAC cell lines and, surprisingly, observed no dramatic changes in ZEB1 expression. However, SETDB1 KD did cause a clear increase in chromatin bound protein expression of the epithelial genes, EpCAM and E-cadherin (Fig. S3A). This was unexpected given that ZNF274 depletion increases ZEB1 expression and decreases EpCAM expression. RNA expression of the EMT pathway was assessed by qPCR and a slight decrease in ZEB1 expression was observed along with a clear increase in E-cadherin RNA expression upon SETDB1 KD (Fig. S3B), validating the protein expression data. This unexpected data prompted us to hypothesize that SETDB1 may function to regulate epithelial genes independent of ZNF274 binding. SETDB1 has previously been shown to regulate epithelial genes, such as EpCAM⁶⁴. We theorized that ZNF274 sequesters SETDB1 at ZNF274 bound genes, while loss of ZNF274 may releases SETDB1. This in turn allows SETDB1 to increase suppression of epithelial genes independent of ZNF274 binding. Under normal conditions, SETDB1 could be bound to chromatin both dependent and independent of ZNF274. SETDB1 KD may predominantly impact expression of genes bound by SETDB1 independent of ZNF274 and therefore have little to no impact on ZNF274 dependent SETDB1 bound genes. To test this hypothesis, we performed co-immunoprecipitation of SETDB1 to validate direct binding with ZNF274 and other complex components, KAP1 and DNMT1. We observed cooperation of SETDB1 with all three complex components validating that SETDB1 binds to ZNF274, KAP1 and DNMT1 (Fig. S3C). To confirm that these components cooperate to form a complex and that complex formation is dependent on ZNF274, we performed a sucrose gradient experiment and assessed ZNF274, SETDB1, KAP1, DNMT1, and H3K9me3 in a classical PDAC cell line with and without ZNF274. In the presence of ZNF274, we observed fractionation of ZNF274 with SETDB1, KAP1, and DNMT1 (siCTRL, lanes #5 & 6) and ZNF274 with H3K9me3 (siCTRL, lane #10) demonstrating the formation of this complex in classical PDAC. In the absence of ZNF274, we observed a substantial decrease in fractionation of all complex components indicating that ZNF274 is necessary for complete complex formation and regulation of ZNF274 dependent H3K9me3 (Fig. S3D, siZNF274, lanes #5, 6, & 10). We also observed weak, residual presence of SETDB1 with KAP1, and DNMT1 (siZNF274, lanes #5 & 6) and SETDB1 with H3K9me3 (siZNF274, lane #10) in the absence of ZNF274 (Fig. S3D). This suggests that low levels of SETDB1 may remain bound to KAP1 and DNMT1 but the majority of SETDB1 bound in the ZNF274 complex has been released from complex formation and is free to function independently in the cell. To validate that SETDB1 suppresses epithelial gene expression independent of ZNF274, we performed ChIP on ZNF274 KD and control cells for H3K9me3, SETDB1, and H3K4me3. H3K4me3 marks transcriptional start sites and is positively correlated with active transcription⁶⁵. We observed slight increases in H3K9me3 and SETDB1 binding at epithelial genes, E-cadherin (CDH1) and EpCAM, with ZNF274 KD as compared to control (Fig. S3E). Additionally, we observed a decrease in H3K4me3 at E-cadherin (CDH1) and EpCAM which indicates a reduction in active transcription at those loci (Fig. S3F). Together, this data validates our previous finding that ZNF274 KD decreases expression of epithelial genes in the EMT pathway and confirms the role of SETDB1 in this regulation.

Overall, we discovered that, in the presence of ZNF274, SETDB1 complexes directly with ZNF274, KAP1, & DNMT1 to suppress gene expression through H3K9me3 at target loci (Fig. S3G). Contrastingly, in the absence of ZNF274, SETDB1 is released from the complex and thereby free to regulate epithelial gene expression (Fig. S3G). We find that SETDB1 functions in both ZNF274-dependent and ZNF274-independent manners in classical PDAC.

ZNF274 regulates global H3K9me3 and ZEB1 target genes

To expand our understanding of ZNF274 gene regulation, we performed CUT&RUN for H3K9me3 and ZNF274. We observed a global reduction in H3K9me3 following loss of ZNF274 (Fig. S4A). ZNF274 is a known regulator of ZNF genes and repeat elements localized to chromosome 19^16,17. We observed the greatest reduction in H3K9me3 at ZNF genes (Fig. S4B) and localized to chromosome 19 (Fig. S4C-D) which is consistent with known ZNF274 functions. Next, we correlated our gene expression data with the binding data and identified that genes were upregulated when ZNF274 binding was decreased (Fig. S4E). A motif analysis identified the ZNF274 binding motif is significantly enriched at locations with differentially enriched H3K9me3 after loss of ZNF274 (Fig. S4F, Fisher’s exact test, median p = 5.62 × 10−²⁸, median p = 2.43 ×10−⁴, respectively). Across H3K9me3 peaks that changed after ZNF274 loss, the ZNF274 motif was strongly and reproducibly enriched in the H3K9me3 DOWN set, but not in the UP set (Fig. S4F). In the DOWN peaks, the enrichment was consistent across background resampling (median OR = 1.498, empirical 95% OR 1.485–1.511), with a higher motif-hit fraction in the selected peaks (31.5%) compared with the matched background (23.5%); correspondingly, the signal was robustly significant (median p = 2.22 ×10⁻⁴, Fisher’s exact test, 200/200 resamples with p < 0.05) (Fig. S4F, H3K9me3 DOWN changes). In contrast, H3K9me3 UP peaks showed only a modest shift in motif-hit frequency (26.9% vs 23.5% background; median OR = 1.199, empirical 95% OR 1.188–1.210) that was not statistically supported under the same resampling framework (median p = 0.227, Fisher’s exact test, 0/200 resamples with p < 0.05) (Fig. S4F, H3K9me3 UP changes). Overall, these results indicate that loss of H3K9me3 is enriched in ZNF274 motif–containing sequences, consistent with a stronger association of ZNF274-linked sequence features with H3K9me3 loss over gain.

Across H3K9me3 peaks that change after ZNF274 loss, the ZEB1 binding motif is significantly enriched in the H3K9me3 UP set, but not in the DOWN set (Fig. S4G). In the UP peaks, the enrichment is consistent across background resampling (median OR = 1.328, empirical 95% OR 1.318–1.339) and the signal is significant (median p = 0.028, Fisher’s exact test, 200/200 resamples with p < 0.05) (Fig. S4G, H3K9me3 UP changes). In contrast, H3K9me3 DOWN peaks show little shift in motif-hit frequency (median OR = 1.169, empirical 95% OR 1.160–1.178) that is not statistically supported under the same resampling framework (median p = 0.121, Fisher’s exact test, 0/200 resamples with p < 0.05) (Fig. S4G, H3K9me3 DOWN changes). These results indicate that sites gaining H3K9me3 are the subset most enriched for ZEB1 motif–containing sequences. ZEB1 is a known repressor of epithelial gene expression, and we see increased SETDB1 deposition of H3K9me3 at ZEB1 regulated epithelial genes (Fig. S3E), suggesting ZEB1 may cooperate with H3K9me3 to induce repression of ZEB1 target genes following ZNF274 loss.

ZEB1 regulates a subset of basal cell state genes

Given that ZEB1 can function as both a transcriptional repressor and a transcriptional activator, we wanted to further explore how ZNF274 regulation of ZEB1 drives a subtype transition and how ZEB1 could act as an activator of basal state genes. We first performed a ZNF274 KD time course experiment to validate that the state change, via induced keratin expression, occurs after ZEB1 expression. We observed an early increase in ZEB1 while the basal state markers induced after or in tandem with ZEB1 and CDH1 expression decreased last (a proxy for full EMT induction) (Fig. 4A, S5A). To show sufficiency of ZEB1 activation of basal markers, we overexpressed ZEB1 in classical PDAC cells. We performed lentiviral transduction with a dox-inducible ZEB1 overexpression (OE) or empty vector (EV). ZEB1 OE significantly upregulated TP63ΔN and KRT5 while KRT17 and KRT81 remained unchanged (Fig. 4B), suggesting that ZEB1 regulates a subset of basal state genes.

Fig. 4. ZEB1 regulates a subset of basal cell state genes.

A qPCR for ZEB1, CDH1 (E-cadherin), TP63ΔN, KRT5, KRT17, and KRT81 in classical PDAC cell lines treated with non-specific or ZNF274-specific siRNAs for 24, 48, 72, 96 h (biological replicates, n = 2). B qPCR for ZEB1, CDH1 (E-cadherin), TP63ΔN, KRT5, KRT17, and KRT81 in classical PDAC cell lines infected with an empty vector or ZEB1 overexpression vector (biological replicates, n = 2). C ChIP-qPCR for KRT5_ZB1 and KRT5_ZB2 with pulldown for ZEB1 in a classical PDAC cell line with doxycycline induced control or ZNF274-specific shRNAs (biological replicates, n = 3). Error bars represent +/- SEM; asterisks, P values after two-tailed unpaired Student’s ttest. ***P = 0.000633 KRT5_ZB1 shCTRL vs shZNF274; ns P = 0.092422 KRT5_ZB2 shCTRL vs shZNF274. Source data and exact P values are provided within the Source Data file.

ZEB1 is known to function as a transcriptional repressor on its own but forms a trans-activating complex with AP-1 and TEAD family proteins to function as a transcriptional activator⁶⁶. To further explore ZEB1 activation of these basal state genes, we performed motif scanning across the KRT5 and TP63ΔN promoter regions. We identified abundant ZEB1 sites and detected multiple TEAD-family and AP-1 (FOS/JUN) motif matches within the same promoters, indicating that both loci contain sequence features compatible with potential TEAD and AP-1 regulatory input (Fig. S5B). We tested whether TEAD or AP-1 motifs are preferentially positioned in close proximity to ZEB1 sites and identified that TEAD and AP-1 motifs are on average 500 base pairs from the closest ZEB1 site (Fig. S5C). A within-promoter circular-shift permutation test, which preserves motif density and spacing while randomizing alignment, showed the observed number of ZEB1 sites with nearby TEAD motifs within ±100 to ±200 bp and the median nearest ZEB1 to TEAD distances were not more extreme than expected by chance in either promoter (Fig. S5D). While the ZEB1 and TEAD/AP-1 motifs are not significantly close in proximity, this does not preclude the ability of ZEB1 to form an activating complex with TEAD/AP-1 or other co-activating factors.

To confirm the presence of ZEB1 at the KRT5 genomic locus, we performed ChIP with a ZEB1 antibody followed by qPCR upstream of the KRT5 transcriptional start site (TSS). KRT5 gene expression is known to be regulated distal to the TSS⁴⁰. Therefore, we selected a distal ZEB1 binding (ZB) site with a nearby TEAD motif from our motif analysis that also corresponded with known ZEB1 binding from a previous ZEB1 ChIP-seq in PDAC⁵⁶. Our ChIP data showed binding of ZEB1 at the KRT5 ZB sites following ZNF274 KD (Fig. 4C). ChIP investigation at the TP63ΔN genomic locus did not yield consistent binding, suggesting that the ZEB1 activating complex may be difficult to detect at some genomic loci. When ZEB1 functions in its activating complex, ZEB1 is not required to be bound to a high-affinity motif and can be recruited indirectly by other complex components⁶⁶.

Altogether, these results demonstrate that ZEB1 regulates not only suppression of epithelial genes but also activation of KRT5 and TP63ΔN basal state genes through potential formation of a ZEB1 containing activating complex. This highlights ZEB1 as an important regulator of cell plasticity in PDAC.

Regulation of basal cell state by HERVs and the dsRNA response pathway

Given that other basal state markers such as, KRT17 and KRT81, remained unchanged following ZEB1 OE (Fig. 4B), we hypothesized that another player may function in conjunction with ZEB1 to drive the basal cell state. We turned to the evolution of KRAB-ZNFs, including ZNF274, and their known function of suppressing repeat elements^22–30. Expression of repeat elements was recently identified to correlate with PDAC methylation subtypes⁶⁷, induce cellular plasticity, increase mesenchymal gene expression, and drive an IRF3 dependent interferon (IFN) response⁶⁸. Therefore, we sought to uncover if ZNF274 depletion increased expression of repeat elements which could facilitate induction of the basal cell state in PDAC.

First, we investigated which class of repeat elements are expressed in the basal PDAC subtype. Our broad analysis of repeat elements uncovered that the predominate class of repeat elements expressed in basal PDAC were ERVs (Fig. 5A). We then analyzed human ERV (HERV) expression between classical and basal PDAC cell lines. Differential expression analysis of HERVs showed that classical PDAC cell lines have lower expression of HERVs at baseline while many HERVs are significantly upregulated in basal PDAC (Fig. 5B, C). Analysis of HERV subgroups identified that HERVH, HERVL, HERVE, and HERVK (aka HMLs) were top HERV families expressed in basal PDAC (Fig. 5D). HERVH and HERVK were previously identified as being upregulated in PDAC⁶⁸ confirming the validity of these results across studies. Together, this data identifies the importance of HERV regulation in basal PDAC.

Fig. 5. Basal state regulation by HERVs, dsRNA, and IFN pathways.

A Pie chart of top 30 up-regulated repeat elements in basal PDAC grouped by repeat element class (biological replicates, n = 3). B Volcano plot showing differential expression analysis of HERVs between basal and classical PDAC cell lines (biological replicates, n = 3). C Heat map of HERV expression in normalized transcripts per million (TPM) between basal line, Panc3.27, and classical lines, ASPC1 and CFPAC1, at baseline (biological replicates, n = 3). D Plotted up-regulated HERVs in basal PDAC grouped by HERV families (biological replicates, n = 3). E Pie chart of top 20 up-regulated repeat elements with ZNF274 KD in classical PDAC cell lines grouped by repeat element class (biological replicates, n = 3). F Plotted up-regulated HERVs with ZNF274 KD in classical PDAC cell lines colored by HERV family (biological replicates, n = 3). G Gene Set Enrichment analysis (GSEA) RNA-seq data comparing control and ZNF274 KD classical PDAC cell lines for the Hallmarks GSEA set (biological replicates, n = 3). H qPCR for KRT81 in classical PDAC cell lines treated with non-specific, ZNF274-specific, or MAVS-specific siRNAs for 72hrs (biological replicates, n = 2). Source data and exact P values are provided within the Source Data file.

Next, we investigated which class of repeat elements are expressed upon ZNF274 KD (Fig. S1E). Of all repeat elements, ERVs were the most upregulated in the context of ZNF274 KD (Fig. 5E). Further analysis of HERV subgroups identified significant upregulation of elements from the HERVH, HERVE, HERVL, and HERVK (aka HMLs) family groups (Fig. 5F). In our binding data, we observed that loss of ZNF274 binding peaks showed significant enrichment at HERV loci relative to all other ZNF274 binding peaks (Fig. S5E, Fisher’s exact test, p = 5.49 × 10^−8), indicating that loss of ZNF274-associated signal is disproportionately linked to HERV-proximal regions. These data reveal that ZNF274 directly regulates HERV expression in PDAC. Furthermore, gene expression pathway analysis identified an enrichment of interferon and inflammatory response genes in addition to EMT following ZNF274 loss (Fig. 5G, S5F). Together, these data suggest that induction of HERV expression upon loss of ZNF274 could induce an IFN response and lead to possible regulation of the basal cell state.

Aberrant expression of repeat elements, including HSATII and HERVs, can lead to production of endogenous dsRNA and activate IFN and inflammation response pathways in the more aggressive form of PDAC^68–70 and other model systems^71,72. The presence of dsRNA in a cell induces a viral infection response mechanism. RIGI (retinoic acid-inducible gene I) and MDA5 (melanoma differentiation-associated gene 5) detect the presence of dsRNA in the cell and activate MAVS (mitochondrial antiviral signaling) which initiates a type I IFN response^73–75. To investigate if ZNF274 loss induces dsRNA expression, we performed immunofluorescence with antibody for dsRNA in control and ZNF274 KD classical PDAC cell lines (Fig. S6A). dsRNA staining was significantly increased with ZNF274 KD suggesting an increase in dsRNA expression (Fig. S6A, p < 0.001). To define the role of dsRNA induction in regulation basal state keratin genes, we performed KD of MAVS (activated by the sensing of dsRNA) in the absence of ZNF274 to assess the necessity of MAVS activation for induction of the basal state. We observed a partial rescue of KRT81 induction by ZNF274 loss when MAVS is knocked down (Fig. 5H, S6B). This partial rescue confirms our previous hypothesis that ZEB1 is not the only regulator inducing the basal cell state. Additionally, we see an increase in IFN expression with ZNF274 KD and this increase is partially recused by the addition of MAVS KD (Fig. S6C, D). Collectively, our data demonstrates that ZNF274 is necessary to suppress HERV expression, which in turn prevents dsRNA production, IFN activation, and acquisition of the basal cell state.

ZNF274 promotes resistance to CDK7 inhibition in classical PDAC

In our previous work, we discovered basal PDAC is exquisitely sensitive to transcriptional inhibitors, most notably cyclin-dependent kinase 7 inhibition (CDK7i). This sensitivity was typified by THZ1 (a covalent inhibitor of CDK7/12/13) and YKL-5-124 (a covalent inhibitor of CDK7 alone), but not to the non-transcriptional inhibitors gemcitabine or the CDK4/6 inhibitor Palbociclib⁷⁶. CDK7 regulates transcription by phosphorylating the C-terminal domain (CTD) of RNA Pol II which enables transcription initiation and productive elongation^77,78. Inhibition of CDK7 blocks phosphorylation of the CTD of RNA Pol II and prevents transcription in many different cancer types^79–84. SY5609 is a clinically approved, non-covalent inhibitor of CDK7 which has cleared phase I clinical trials and shows exciting promise in PDAC patients^85–87.

To elucidate the impact of ZNF274 regulation of cellular plasticity on sensitivity to CDK7 inhibitors, we performed knockdown of ZNF274 by siRNA and shRNA followed by treatment with increasing doses of THZ1 (Fig. 6A) or YKL-5-124 (Fig. 6B and S7A). ZNF274 knockdown increased ZEB1 expression between the siCTRL DMSO (dimethyl sulfoxide) and the siZNF274 DMSO validating induction of EMT. Additionally, ZNF274 knockdown increased sensitivity to both CDK7 inhibitors, THZ1 and YKL-5-124, through an increase in cleaved caspase-3 (CC3) activation in multiple independent classical PDAC cell lines (Fig. 6A, B, S7A). To validate this phenotype in a 3D model system, we treated our dox-inducible control and ZNF274 knockdown classical PDAC organoids with increasing doses of YKL-5-124. Upon completion of treatment, the organoids were fixed and embedded in a cone of agarose, paraffin processed and embedded, and then stained for H&E. The shCTRL organoids show intact organoid structure as indicated by their round and cystic morphology (Fig. 6C). This intact structure is then maintained throughout treatment with increasing doses of YKL-5-124. In contrast, shZNF274 organoids show intact organoid structure by their thick, multicellular-layer morphology prior to treatment, and then upon treatment, increasing doses of YKL-5-124 causes disintegration and loss of organoid structure in the shZNF274 organoids (Fig. 6C). This loss of integrity of the shZNF274 organoid structure indicates increased sensitivity to CDK7 inhibition upon loss of ZNF274 in classical PDAC organoids.

Fig. 6. ZNF274 is necessary to maintain resistance to CDK7 inhibitors.

A WB for ZNF274, ZEB1, PARP, CC3 and ACTIN in classical PDAC cells lines transfected with non-specific or ZNF274-specific siRNAs and treated with 500 nM or 2.5uM THZ1 for 40 h. Data are representative of 2 replicates. B WB for ZNF274, ZEB1, PARP, CC3 and ACTIN from doxycycline inducible control or ZNF274-specific shRNAs classical PDAC cell lines treated with 500 nM or 1uM YKL-5-124 for 72 h. Data are representative of 2 replicates. C Brightfield color images of H&E slides from embedded classical PDAC organoids with doxycycline induced of control or ZNF274-specific shRNAs and then treated with 500 nM, 1 μM, or 2.5 μM YKL-5-124 for 72 h (left). Images were taken on the Nikon Live microscope and have a scale bar of 100 μm. Paired WB for ZNF274, ZEB1 and ACTIN from classical PDAC organoids with doxycycline induced of control or ZNF274-specific shRNAs (right). The samples derive from the same experiment but a different gel for ZEB1 was processed in parallel. Data are representative of 2 independent experiments. WB for ZNF274, CC3 and ACTIN in a basal PDAC cell line infected with an empty vector or ZNF274 overexpression vector and treated with 500 nM or 1uM THZ1 for 24 h (D) and 500 nM or 1uM YKL-5-124 for 72 h (E). Data are representative of 2 replicates. Normalized tumor growth (F) and mouse weights (G) of subcutaneously implanted ASPC1 and CFPAC1 dox-inducible shCTRL or shZNF274 cell lines treated with once-daily SY5609 (0.5 mg/kg) compared with Vehicle control (5% captisol) for 21 days (each arm n = 3, biological replicates). Error bars represent ± SEM; asterisks, P values after two-tailed unpaired Student’s ttest. *P = 0.0362 ASPC1 shZNF274 Vehicle vs shZNF274 SY5609; *P = 0.0483 CFPAC1 shZNF274 Vehicle vs shZNF274 SY5609. Source data and exact P values are provided within the Source Data file.

To investigate if ZNF274 is sufficient to promote resistance to CDK7 inhibition, we overexpressed ZNF274 in basal (CDK7i sensitive) PDAC cell lines and treated with increasing doses of CDK7 inhibitors. As previously shown, ZNF274 OE in a basal PDAC cell line drives an enrichment of the classical cell state as seen by epithelial and classical PDAC GSEA gene signatures (Fig. 1F). We then treated ZNF274 OE cells with increasing doses of CDK7 inhibition and saw a substantial reduction in sensitivity to THZ1 and YKL-5-124 as indicated by reduced activation of cleaved caspase-3 compared to the empty vector control (Fig. 6D, E and S7B). Thus, our results suggest that ZNF274 is sufficient to promote resistance to CDK7 inhibition while the absence of ZNF274 enables sensitivity to CDK7 inhibition. Since the absence of ZNF274 drives an increase in ZEB1, we next sought to identify if ZEB1 is also sufficient to drive CDK7i sensitivity. We overexpressed ZEB1 in a classical PDAC cell line that has low baseline expression of ZEB1 and high baseline expression of ZNF274. We found that ZEB1 OE does not substantially change ZNF274 mRNA expression but does increase sensitivity of a classical PDAC cell line to THZ1 as shown by an increase in cleaved PARP (Fig. S7C, D). Together this data indicates that sensitivity to CDK7 inhibition is at least partially regulated by ZNF274 control of ZEB1 expression and that modulations to either ZNF274 or ZEB1 is sufficient to alter sensitivity to CDK7i in either subtype of PDAC.

Finally, to show that the absence of ZNF274 confers sensitivity to CDK7 inhibition in vivo, we implanted the dox-inducible shCTRL and shZNF274 classical PDAC cell lines subcutaneously into immune-compromised nude female mice and treated them daily with SY-5609 (a clinically approved oral CDK7 inhibitor) for 3 weeks. SY-5609 significantly reduced growth of the ZNF274 knockdown tumors but not the control tumors (Fig. 6F). Mice did not experience any toxicity from the drug as indicated by stable mouse weights (Fig. 6G). Combined, our results identify ZNF274 regulation of cellular plasticity as a potential therapeutic option for sensitizing classical PDAC to CDK7 inhibitors.

DNMT inhibition sensitizes classical PDAC to CDK7 inhibition

Unfortunately, ZNF274 is not targeted by any clinically approved drugs, but other components of the KRAB-ZNF complex do have clinically approved inhibitors. Clinically approved drugs exist to target the DNMT and HDAC components of the complex. We already identified that DNMT1 cooperates with SETDB1 (Fig. S3C) and is part of the complex formed in the presence of ZNF274 (Fig. S3D). In the absence of ZNF274, the presence of DNMT1 in the complex also decreases (Fig. S3D) suggesting that DNMT1 is a main component of the complex and could influence its regulatory capabilities. To further validate this observation, we assessed DNMT1 expression in ZNF274 KD cells and observed a slight decrease in DNMT1 expression and chromatin binding in ZNF274 KD cells compared to control (Fig. S8A). This aligns with our complex formation assay and identifies that ZNF274 influences DNMT1 expression and localization. This suggests that DNMT1 may function as part of the ZNF274/SETDB1 complex to facilitate regulation of ZEB1 and HERVs expression. Next, we performed DNMT1 KD in a classical PDAC cell line to determine if loss of DNMT1 alone could induce ZEB1 and EMT. We observed no change in ZNF274 protein expression but did observe a slight increase in ZEB1 as well as decreases in EpCAM and E-cadherin protein and RNA expression in DNMT1 KD cells compared to control (Fig. S8B, C, D). This data indicates that DNMT1 loss can induce changes in EMT genes and that DNMT1 functions to help promote the ZNF274 driven EMT phenotype. In conjunction with this data, we would expect a DNMT inhibitor to disrupt the regulatory function of the complex promoting similar gene expression changes to ZNF274 KD.

There are two clinically approved DNMT inhibitors: decitabine (DAC) and azacytidine (AZA). Given that our goal is to mimic ZNF274 loss, we treated classical PDAC with either a DAC or AZA and assessed changes in ZEB1 expression as a proxy for ZNF274 regulation. We observed that AZA induced a robust increase in ZEB1 protein expression while we observed no change in ZEB1 expression with DAC treatment (Fig. 7A and S8E). This data suggests that DAC is not a suitable DNMTi for further experiments given that it does not recapitulate the increased ZEB1 expression phenotype that is characteristic of ZNF274 loss. Therefore, AZA was selected as the appropriate drug to proceed into the next experiments. We then expanded our investigation to assess three classical PDAC cell lines and showed that RNA expression of ZEB1 increased with AZA treatment across all three cell lines and continued to increase with higher doses of AZA (Fig. 7B) while RNA expression of epithelial genes did not change with AZA treatment (Fig. S8F). This suggests that an AZA induced state change could be driven by ZEB1 without full induction of EMT. To identify if AZA treatment increases expression of HERVs in classical PDAC, as seen with ZNF274 KD, we performed RNA-seq with ribosome depletion in two classical PDAC cell lines treated with 5uM AZA. Differential HERVs expression analysis showed significant upregulation of a subset of HERVs in the AZA treated versus DMSO control samples including the HERVH, and HERVL families we saw previously in the ZNF274 KD data set (Fig. 7C). Additionally, we observed the same cellular morphological change, from a colony-based growth pattern to a more spindle shaped and individual growth pattern, with AZA treatment as described previously with ZNF274 KD (Fig. S8G). The ability of AZA treatment to mimic the cellular changes induced by ZNF274 KD is further apparent in classical cells where blocking dsRNA signaling prevents full induction of basal state keratin genes. AZA treatment of the siCTRL caused increased expression of KRT17, KRT81, ZEB1, and IFNβ (Fig. S8H). KD of MAVS rendered the classical cells incapable of fully inducing ZEB1 and the basal state keratin genes in response to AZA (Fig. S8H). Together these data validate that AZA induces cellular plasticity in classical PDAC enabling activation of a mechanistic cell state change.

Fig. 7. DNMT inhibition induces ZEB1 and enhances sensitivity to CDK7 inhibition.

A Representative WB for ZEB1 and Actin in a classical PDAC cell line treated with 1uM Azacytidine for 48 h (left). WB quantification of relative ZEB1 expression normalized to ACTIN for classical PDAC cells treated with 1 μM Azacytidine for 48 h replicates (right) (biological replicates, n = 4). Error bars represent ± SEM; asterisks, P values after two-tailed unpaired Student’s ttest. **P = 0.0052 DMSO vs AZA 1 μM. B qPCR for ZEB1 in classical PDAC cells lines treated with increasing doses of Azacytidine for 48 h (biological replicates, n = 2). C Plotted up-regulated HERVs with 5 μM AZA treatment in classical PDAC cell lines colored by HERV family (biological replicates, n = 3). D WB for ZEB1, PARP, CC3, and ACTIN in classical PDAC cells treated with 5 μM Azacytidine plus either 500 nM or 1 μM SY-5609 for 48 h. Data are representative of 2 replicates. E Tumor volume of basal and classical PDAC xenografts after treatment with once-daily SY5609 (0.5 mg/kg) (red) compared with Vehicle control, 5% captisol, (black) for 21 days (Vehicle n = 8, SY-5609 n = 8, biological replicates). Error bars represent ±  SEM; asterisks, P values after two-tailed unpaired Student’s ttest. **P = 0.005584 PAAD141b (Basal) Vehicle vs SY5609; **P = 0.006391 PAAD169b (Basal) Vehicle vs SY5609; ns P = 0.109041 PAAD154b (Classical) Vehicle vs SY5609; ns P = 0.218005 PAAD200b (Classical) Vehicle vs SY5609. F Tumor volume of a classical PDAC xenograft after treatment with once-daily combo (SY-5609 0.5 mg/kg and azacytidine 2.5 mg/kg) (dotted blue), azacytidine alone (2.5 mg/kg) (solid red), SY-5609 alone (0.5 mg/kg) (dotted red), or Vehicle control (5% captisol) (solid black) for 21 days (n = 7 for each treatment group, biological replicates) and mouse weights. Error bars represent ± SEM; asterisks, P values after two-tailed unpaired Student’s ttest. ***P = 0.000793 PAAD200 (Classical) SY5609 vs SY5609 + AZA. G Endpoint survival study for a subset of AZA + SY combination study mice. Treatment was stopped at 22 days and mice were monitored for a total of 80 days post initiation of treatment (Vehicle and SY5609 only n = 3, biological replicates; AZA only and AZA + SY n = 5, biological replicates). A log rank (Mantel-Cox) test was performed to determine significance between the AZA + SY arm or the AZA alone arm and the Vehicle control arm. *P = 0.0136 PAAD200 (Classical) Vehicle vs AZA; **P = 0.0032 PAAD200 (Classical) Vehicle vs SY5609 + AZA. Source data and exact P values are provided within the Source Data file.

Given that AZA induces cellular plasticity, we hypothesize that this could sensitize classical PDAC to CDK7 inhibition. We identified that basal PDAC presents similar sensitivity to the clinical CDK7 inhibitor, SY5609, as we had previously observed with THZ1 and YKL-5-124⁷⁶. We observed clear increases in the apoptosis marker, cleaved PARP, and lower IC50 in basal PDAC while classical PDAC presented with resistance to apoptosis and higher IC50 (Fig. S9A, B). Given that classical PDAC is resistant to CDK7i by SY5609, we performed a combination treatment of AZA with SY5609 in classical PDAC cell lines to assess if the addition of AZA could overcome resistance to SY5609 in classical PDAC. Cells treated with the combination of AZA and SY5609 showed a substantial increase in apoptosis as seen by increased cleaved PARP and cleaved caspase 3 (Fig. 7D). This data indicates that AZA treatment can sensitize classical PDAC cells to CDK7 inhibition, broadening the therapeutic application of CDK7 inhibitors in PDAC. To verify that this sensitization is through ZNF274 controlled pathways, we performed the same combination treatment with AZA and SY5609 in the presence and absence of ZEB1. We observed the expected sensitization of classical PDAC cell lines in the presence of ZEB1 with treatment of both AZA and SY5609, while the reduction of ZEB1 by siRNA rescued the sensitization phenotype as seen by a decrease in cleaved caspase 3 in the combination treatment (Fig. S9C). Combined, our results demonstrate the importance of ZNF274 and ZEB1 epigenetic regulatory networks in determining drug sensitivity in PDAC.

To further investigate this potential treatment in vivo, we first treated four patient-derived xenograft (PDX) models of PDAC (2 basal and 2 classical) with SY5609 dissolved in 5% Captisol and administered once daily by oral gavage to nude female mice for 3 weeks. Basal PDAC PDX mice showed marked sensitivity to SY5609 treatment through dramatic reduction of tumor growth (Fig. 7E, red) while the classical PDAC PDX mice showed no change in tumor growth with SY5609 treatment (Fig. 7E, black). This data shows clear differential sensitivity of basal and classical PDAC tumors to the clinically approved CDK7 inhibitor, SY5609, and potentially efficacy of SY5609 to treat PDAC patients with more basal-like tumors. However, since PDAC patient tumors are intrinsically heterogeneous, the best treatment course would be able to target both basal and classical PDAC cells within the tumor. To determine if a combination treatment of AZA and CDK7 inhibition would be able to treat more classical PDAC patient tumors, we utilized one of our PDX mouse models of classical PDAC to test the efficacy of the combination treatment. We split the trial into four arms of treatment, vehicle (5% Captisol), SY5609 only, AZA only, and AZA plus SY5609 combination, which were administered by oral gavage to nude mice for 3 weeks. The combination treatment of AZA plus SY5609 showed dramatic reduction of tumor growth over all other treatment arms in our classical PDAC PDX model (Fig. 7F). Additionally, the combination treatment, as well as the solo arms, showed no toxicity to the mice as seen by continued stable weight of the mice throughout the trial (Fig. 7F). A subset of the combination trial mice were enrolled in an endpoint survival study in which the treatment was stopped after 22 days and the mice were monitored for 80 days post initiation of treatment. Endpoint was called when a tumor reached ~2 cm in diameter indicating a notch on the survival curve. Interestingly, mice treated with AZA alone showed a slight but significant increase in survival even though no decrease in tumor growth while on treatment was observed (Fig. 7F, G). This indicates that AZA alone may have a small impact on tumor growth over an extended period of time, suggesting a time-dependent effect of EMT on tumor growth with AZA treatment. Mice from the combination treatment arm (AZA + SY) survived significantly longer than any of the single arm treatments or the vehicle control arm, indicating that the combination treatment has a long-term effect on tumor growth even after removal of the treatment doses (Fig. 7G). Together, we have shown that combination treatment of AZA plus the clinical CDK7 inhibitor, SY5609, would be an effective treatment for classical PDAC patient tumors as well as any mixed classical and basal tumors, thereby dramatically increasing the application of CDK7 inhibition in treatment of PDAC patient tumors.

Discussion

Here, we report our discovery that the zinc finger protein, ZNF274, has the ability to control cell state in PDAC. With the identification of subtype specific therapies for pancreatic cancer, it is vital to understand the biology of how cell state and lineage plasticity are controlled in PDAC. Our data demonstrates that loss of ZNF274 induces the basal PDAC cell state through two arms of regulation. First, loss of ZNF274 increases expression of ZEB1, a known driver of EMT. EMT signatures are a known part of the basal PDAC subtype, but it previously remained unclear whether it is a necessary step in the state conversion. We show that EMT is an early event and indeed part of the conversion to a basal cell state as typified by KRT5, KRT17 & KRT81. Second, loss of ZNF274 leads to de-repression of repeat elements and production of dsRNA, which activates the dsRNA response and IFN pathway. When ZNF274 is lost, both ZEB1 and repeat RNAs drive cellular plasticity and a basal state conversion.

Not only does loss of ZNF274 induce a more basal cell state but this loss also induces a morphology change. PDAC cells that lose ZNF274 acquire the morphology of a more basal cell. Tumors with basal/mesenchymal features tend to have an increased ability to migrate and invade nearby tissues^88,89. Indeed, we found that cells with ZNF274 loss gain the ability to migrate and invade as is characteristic of a basal cancer cell (Fig. 2). Additionally, our 3D pancreatic cancer organoid models undergo a multilayering organogenesis following loss of ZNF274 which enhances the ability of the tumor cells to invade (Fig. 2B-D). Our organoid phenotype may be the result of a process called delamination which refers to the formation of multiple layers⁵². The delamination process can involve a partial or complete EMT as seen in neural crest cells, which enables them to migrate and colonize different organs and tissues^52,53. Another study showed that organoids from human mammary tissues can have distinct morphologies based on their cell lineage. It was shown that mature luminal organoids formed a structured acini whereas their basal organoids formed large, disordered spheres with branching outgrowth⁵⁴. Breast cancer organoids have a multilayering morphology in which a layer of basal cells forms over the luminal structures thus enabling branched outgrowth and increased invasion⁵⁵. PDAC organoids are known to form glandular cysts, and a glandular histology has been shown to correlate with the classical PDAC subtype^50,51. Therefore, our ZNF274 KD causes a loss in glandular identity of the classical PDAC organoids and increasing multilayering thereby driving them towards a more basal cell state (Fig. 2B). The ability of ZNF274 regulation to induce such dramatic changes in cellular plasticity truly highlights the importance of understanding epigenetic regulation of cell state and the new therapeutic targets that can be discovered.

ZEB1 is a well-known master regulator of EMT; it functions as a repressor of epithelial genes and an activator of mesenchymal genes^57–59. ZEB1 is also known for promoting stemness, invasion and metastasis in pancreatic cancer mouse models^60,61. Our data supports the published literature and identifies ZEB1 as an important component of the mechanism regulating cell plasticity in PDAC. However, the regulation of ZEB1 itself is not as well studied as ZEB1 regulation of other genes, such as EMT pathway genes. In this study, we determined that ZNF274 is a regulator of ZEB1 thereby controlling ZEB1’s influence over EMT. Interestingly, ZNF274 expression is inversely correlated with ZEB1 expression in classical PDAC cell lines (Fig. 3A). This points to a complex, homeostatic relationship between ZNF274 and ZEB1 in which when one goes up the other goes down and vice versa to maintain a specific cellular balance. This connection between ZNF274 and ZEB1 is one of many connections which builds the complex regulatory network that defines the chromatin landscape of classical PDAC.

A few limitations of this study include the heterogeneous nature of patient tumors in their subtype classification. Here we provide a better understanding of PDAC tumor subtype vulnerabilities and conversion capacity which enable us to identify a combination therapy for classical and mixed PDAC tumors. In addition, we recognize that our PDX models lack an immune component. There is continued value in exploring the immunogenic effects of HERVs activation and the potential activation of anti-tumor immunity. Investigation of the immunological consequences of HERV depression through immunocompetent and humanized models are important next steps to understanding the contribution of viral mimicry and immune activation in response to ZNF274 loss.

Furthering the clinical application of this work, we uncovered that the absence of ZNF274 or inhibition of DNMT by AZA sensitizes classical PDAC models to inhibitors of CDK7. These manipulations lead to disruption of the complex chromatin landscape that enables classical PDAC to maintain its epithelial cell state and resistance to CDK7 inhibition. A previous study found their KRAS-dependent (classical) PDAC models to be more sensitive to another DNMT inhibitor, decitabine, compared to their KRAS-independent (basal) PDAC models⁹⁰. We found that treatment with Decitabine did not induce ZEB1 expression (Fig. S8E) and therefore was not the appropriate inhibitor to disrupt function of the ZNF274 complex. We did not observe any sensitivity to AZA alone in classical PDAC and the previous study did not provide a mechanism to explain their sensitivity, therefore their mechanism of action may be independent of the ZNF274 complex. Collectively, this suggests DNMTs may function with a multifaceted role in maintenance of classical PDAC cellular integrity.

Treating PDAC patients with AZA was attempted in the clinic using an oral azacytidine (CC-486) followed by adjuvant therapy. This trial showed no improvement in patient disease response and no increase in time-to-relapse⁹¹. This is not surprising given that our study as well as others have shown that AZA treatment alone does not inhibit growth until very high doses are reached, such as 5uM or more in PDAC cell lines in culture⁹². The equivalent treatment of these higher doses in humans would likely be too toxic, therefore a better approach is to pair AZA with another drug that can improve its anti-tumor effect and enable lower treatment doses. Recent work has identified the importance of using epigenomic inhibitors to alter gene expression patterns in basal and classical PDAC subtypes to better understand vulnerabilities and develop new therapies⁹³. In our study, we show that treatment of classical PDAC with AZA increases cellular plasticity and forces classical PDAC towards a more treatment sensitive state. The addition of AZA to CDK7 inhibition accelerates the apoptotic effect compared to treatment with each drug alone (Fig. 7F). Leveraging cellular plasticity to develop new therapies and enhance current ones will substantially increase therapeutic options for patients and help combat treatment resistant disease¹⁵. Here we showed that KRAB-ZNF complex members including ZNF274 and DNMT can manipulate cellular plasticity and targeting this strategy is therapeutically beneficial. Our combination treatment strategy utilize AZA to induce cellular plasticity and control sensitivity to CDK7 inhibitors which is an ideal option to treat heterogeneous PDAC tumors in the clinic.

Methods

Research ethics

All animal experiments were conducted under protocol PROTO201900024 in accordance with the Fred Hutchinson Cancer Center (FHCC) Institutional Animal Care and Use Committee (IACUC) guidelines and the ARRIVE guidelines. All cell lines, human data, and biological material were publicly sourced or commercially available.

Cell lines

PDAC cell lines ASPC-1 (CRL-1682), Capan2 (HTB-80), CFPAC1 (CRL-1918), HPAFII (CRL-1997), MiaPaCa2 (CRL-1420), BxPc3 (CRL-1687), and Panc-1 (CRL-1469) were obtained from the American Type Culture Collection (ATCC) and grown in their required growth medium per the ATCC description. The DanG (ACC 249) and PaTu-8988T (ACC 162) PDAC cells were obtained from DSMZ while SUIT2 was obtained from Japanese Collection of Research Biosources (JCRB) cell bank and grown in their required growth medium per the DSMZ and JCRB descriptions. Cells were passaged by trypsinization. All studies were done on cells cultivated for less than ten passages.

Organoids

Basal and classical subtyped human PDAC organoids were generously gifted from the Notta Lab at the University of Toronto and cultured according to the methods detailed in Tuveson Laboratory Murine and Human Organoid Procotols. Human PDAC organoids were grown by embedding in Matrigel (Corning) and cultured in human feeding medium (AdDMEM/F12 medium supplemented with HEPES [1×, Invitrogen], Glutamax [1×, Invitrogen], penicillin/streptomycin [1×, Invitrogen], B27 [1×, Invitrogen], Primocin [1 mg/mL, InvivoGen], N-acetyl-L-cysteine [200 μg/mL, Sigma], Wnt3a-conditioned medium [20% v/v], RSPO1-conditioned medium [30% v/v], Noggin [0.1 μg/mL, Peprotech], Human EGF, [50 ng/mL, Peprotech], Gastrin [0.21 μg/mL, Tocris], Human FGF10 [0.1 μg/mL, Preprotech], Nicotinamide [0.045 mg/mL, Sigma], and A83-01 [0.21 μg/mL, Tocris]). To passage, organoids were dissociated into single cells using TrypLE (GIBCO) and then re-seeded into Matrigel (Corning) and topped with human feeding media. All studies were done on organoids cultivated for less than ten passages.

Real-time quantitative PCR (qPCR)

Total RNA extraction, cDNA synthesis and real-time quantitative PCR were performed as previously described⁷⁶. Data were expressed as relative mRNA abundance normalized to the β-ACTIN expression level in each sample or as fold change over control and error bars are represented as mean ± s.e.m. between two independent experiments unless otherwise indicated in the figure legend. The primer sequences are listed in Table S1.

Protein isolation and Western blot

Chromatin fractions were prepared by resuspending the cell pellet in lysis buffer containing 10 mM HEPES pH 7.4, 10 mM KCl, 0.05% NP-40 supplemented with a protease inhibitor cocktail (Complete EDTA-free, Roche Applied Science), 5 μM TSA, 5 mM sodium butyrate, 1 mM DTT, and phosphatase inhibitors (Phosphatase Inhibitor Cocktail Sets I, II, and III Calbiochem) and incubated on ice for 20 min. The lysate was then centrifuged at 18,000 x g for 10 min at 4 °C. The supernatant was removed (cytosolic fraction) and the pellet (nuclei) was acid-extracted using 0.2 N HCl and incubated on ice for 20 min. The lysate was then centrifuged at 18,000 x g for 10 min at 4 °C. The supernatant (contains acid soluble proteins) was neutralized using 1 M Tris-HCl pH 8.

Whole cell lysate (WCL) was prepared by resuspending the cell pellets in RIPA lysis buffer supplemented with a protease inhibitor cocktail (Complete EDTA-free, Roche Applied Science), 5 μM TSA, 5 mM sodium butyrate, 1 mM DTT, and phosphatase inhibitors (Phosphatase Inhibitor Cocktail Sets I, II, and III Calbiochem) and incubated on ice for 20 min. The lysate was then centrifuged at 18,000 x g for 10 min at 4 °C and the supernatant was harvested. Protein concentration was measured by using a BCA protein assay kit (Pierce). 50 μg of the cell lysate (from lysis with buffer) electrophoresed on a 4–20% gradient polyacrylamide gel with SDS (Genscript) and electroblotted onto polyvinylidene difluoride membranes (PVDF) (Millipore). Membranes were blocked in TBS with 5% non-fat milk and 0.1% Tween and probed with antibodies. Bound proteins were detected with horseradish-peroxidase-conjugated secondary antibodies (Vector Biolaboratories) and Clarity Max Western ECL Blotting Substrate (Biorad). Antibodies used were anti-ZNF274 (1:500, Novus Biologicals #H00010782-A01), anti-H3K9me3 (1:1000, CST #13969), anti-ZEB1 (1:500, abcam #ab203829), anti-EpCAM (1:1000, abcam #ab32392), anti-E-cadherin (1:1000, CST #3195), anti-SETDB1 (1:1000, Proteintech #11231-1-AP), anti-DNMT1 (1:1000, abcam #ab13537), anti-TP63 (1:1000, CST #39692), anti-Cytokeratin17 (1:100, Invitrogen #MA1-06325), anti-Cytokeratin5 (1:500, Leica #CK5-L-CE-H), anti-Cleaved Caspase-3 (1:1000, CST #9664), anti-PARP (1:1000, CST #9542), anti-RNA polymerase II CTD phospho S2 (1:1000, abcam, #ab5095), anti-RNA polymerase II CTD phospho S7 (1:1000, abcam, #ab126537), anti-Total Histone H3 (1:1000, abcam #ab1791), and anti-betaACTIN (1:5000, Sigma #A5316) as a loading control. Uncropped and unprocessed blots are included in the associated Source Data file.

For Western blots performed on organoids, organoids were cultured in human feeding media. Organoids were pelleted using Cell Recovery Solution (Corning) to dissolve Matrigel, lysed in RIPA buffer, and processed as detailed above.

Organoid embedding

Human PDAC organoids were cultured as described above and then fixed in their Matrigel domes by removing the culture media and replacing with 4% PFA in PBS, then incubated for 1–2 h at room temperature. Organoids were collected in a 50 ml conical tube and centrifuge at 350 x g for 5 min at room temperature then the supernatant was removed. Washes were done consecutively by adding ddH2O, followed by centrifugation, then washed with 0.2% BSA in PBS and centrifuged. Organoids are suspended with 2% agarose and centrifuge at 200 x g, for 3 min, at room temperature and then cooled on ice. The solid gel from the conical tube is transferred to a histology cassette using disposable plastic scoop, stored in 70% ethanol, and submitted to the experimental histology core for FFPE processing and H&E staining.

Immunofluorescence

Following organoid embedding, unstained slides are deparaffinize by placing in the following liquids for the specified amount of time each: xylene for 15 min, xylene for 15 min, 100% ETOH for 5 min, 100% ETOH for 2 min, 100% ETOH for 2 min, 95% ETOH for 2 min, 70% ETOH for 2 min, 50% ETOH for 2 min, H2O for 2 min, H2O for 2 min. Slides are then placed in 1x antigen retrieval buffer (BRAND) and heated up using a pressure cooker. Slides are then cooled for 30 min to about 50 °C and then washed with 1x PBS twice for 2 min each, then in PBS for 10 min, and then in permeabilization solution (PBS/gelatin/Triton 0.25%) twice for 10 min each. Slides are then blocked in 5% BSA in permeabilization solution in a moist chamber for 1 h at room temperature. Primary antibodies are diluted with 1% BSA in permeabilization, added to the slides and placed in a moist chamber at 4 °C overnight. The next day, the slides are washed twice with PBS for 10 min each and then placed in permeabilization solution (PBS/gelatin/Triton 0.25%) once for 10 min. Secondary antibodies are diluted with 1% BSA in permeabilization solution (Alexa Fluor 555, Invitrogen, 1:400 dilution), added to the slides and incubated for 1 h at room temperature. Next, perform 3x washes in TBST 0.1% Tween for 5 min each. Mount using mounting media containing DAPI (H-1500, Vectashield Vector Laboratories), top with coverslip and seal. Antibodies used were: anti-ZNF274 (1:250, Novus Biologicals #H00010782-A01), anti-TP63 (1:200, CST #39692), GATA6 (1:200, CST #5851), Phalloidin (1:200, Proteintech #PF00003), dsRNA (1:500, Exalpha #10010200), and Keratin81 (1:200, Santa Cruz #sc-100929).

Crystal violet staining

Grow cells in a 6 cm or 10 cm plate, then place cells on ice and wash twice with cold PBS. Fix for 10 min with ice-cold 100% methanol and then aspirate. Move the cells off ice to room temperature and cover them with 0.5% crystal violet solution in 25% methanol. Incubate for 10 min and then remove the crystal violet. Wash the cells in water several times, until the dye stops coming off. Allow the cells to dry at room temperature (possibly overnight) and then store at room temperature.

siRNA transfection

Cells were plated and then forward transfected the next day with siRNAs in a 6 cm or 10 cm culture dishes by replacing the media with 2 mL or 4 mL of OptiMEM low-serum media (Gibco) and then adding 500 μL or 1 mL OptiMEM low-serum media (Gibco) containing 10 μL or 20 μL of Lipofectamine RNAiMAX (Invitrogen) and either non-targeting (Control) or targeting siRNAs (0.25 μM). The plates were supplemented with 2 mL or 4 mL of complete media after 8 h of incubation at 37 °C and then completely replaced with fresh media 24 h after transfection. Cells were cultured for an additional 40 h (72 h total) unless otherwise stated in the figure legend, after which cell pellets were extracted for downstream assays including Western blotting and qPCR. siRNAs were obtained from Dharmacon: ZNF274 (L-013359-02), SETDB1 (L-020070-00), MAVS (L-024237-00), DNMT1 (L-004605-00), ZEB1 (L-006564-01), Control (D-001810-10).

Human PDAC patient tumor data analysis

Patient tumor specimens

All patients provided written informed consent allowing the molecular characterization of their tumor samples and follow-up on their clinical information under the International Cancer Genome Consortium (ICGC) protocol. Patient samples were mostly accrued at Princess Margaret Cancer Center at the University Health Network (Toronto, Canada). Resectable tumors were obtained from the UHN Biospecimens Program, and advanced tumors were obtained from the COMPASS trial (no. NCT02750657). As part of previous studies^94–97, some resectable tumors were also obtained from Sunnybrook Health Sciences Center (Toronto), Kingston General Hospital (Kingston), McGill University (Montreal), Mayo Clinic (Rochester), Massachusetts General Hospital (Boston) and Sheba Medical Center (Tel Aviv). Approval for the study was obtained through the University Health Network Research Ethics Board (nos. 15-9596, 13-6377, 18-5116, 20-5594, 21-5648 and 32517). This study complies with all relevant ethical regulations. Tumor samples were obtained and processed as described previously^96,97. Briefly, resectable tumors were obtained from surgical specimens and advanced tumor cores were obtained by image-guided percutaneous core needle biopsy.

Cell enrichment using laser capture microdissection for bulk sequencing

Fresh tumors were embedded in optimal-cutting-temperature (OCT) compound and snap-frozen in liquid nitrogen. Frozen biospecimens for bulk RNA-sequencing and whole genome sequencing (n = 490 from 464 patients) underwent laser capture microdissection (LCM) to enrich for tumor cells as described previously^96,97.

Bulk RNA sequencing

RNA was extracted from LCM tissue using the PicoPure RNA Isolation Kit (ThermoFisher Scientific) and quantified using the Qubit dsRNA High Sensitivity Kit (Invitrogen). Quality was assessed using the RNA ScreenTape Assay on the 2200 TapeStation Nucleic Acid System (Agilent Technologies). RNA with RIN (RNA integrity number) > 7 was used to prepare sequencing libraries using the TruSeq RNA Access Library Sample prep kit (Illumina) or the TruSeq Stranded Total RNA Library Prep Gold kit (Illumina) according to the manufacturer’s instructions. Libraries were then quantified using the KAPA Illumina Library Quantification Kit (Roche) according to the manufacturer’s protocol. Paired-end sequencing was carried out on the Illumina HiSeq 2500 platform (2×126 cycles) or the NovaSeq 6000 (2×151 or 2×101 cycles).

RNA sequencing (RNA-Seq) was conducted at the Ontario Institute of Cancer Research following the protocol outlined by O’Kane et al.⁹⁸. Sequencing reads were aligned to the human reference genome (hg38) and transcriptome (Ensembl v100) using STAR v.2.7.4a. Duplicate reads were marked with Picard v.2.21.4. Gene expression levels were quantified in transcripts per million (TPM) using the stringtie package v.2.0.6, and log2-transformed expression values were used for subsequent analyses.

Transcriptional subtyping and transcription factor enrichment analysis

To determine the transcriptional subtypes of the bulk RNA-seq cohort using various schemes, genesets were obtained from Moffitt et al.⁶, Collisson et al.⁵ and Bailey et al.⁷ and individual samples were assigned the subtype with the highest median gene expression. The Moffitt subtype was also determined using PurIST⁹⁹ run with default parameters.

The ZNF274 transcription factor enrichment gene set “ZNF274 21170338 ChIP-Seq K562 Hela” was obtained from the ChEA3 database¹⁰⁰. Briefly, the ChEA3 database integrates multiple source modalities, including ChIP-seq and RNA-seq, to predict genes affected by the transcription factor of interest. The enrichment gene set was scored in individual samples of the bulk cohort with gene set variation analysis¹⁰¹ using the GSVA library (v1.50.0) in R (v4.3.0). Correlation analysis in R was used to confirm that enrichment scores were positively associated with ZNF274 expression.

Constructs and viral infection

Plasmids for viral infection: SMART vector shZNF274 and shCTRL, pTRE-ZNF274-3xHA, pTR-ZEB1-3xHA, pSCALP-ZEB1.

SMARTvector Inducible Lentiviral shRNA plasmids containing ZNF274 or control short hairpin RNAs were purchased from Dharmacon (#V3SH11255-01EG10782). These plasmids are used for dox-inducible knockdown of ZNF274 or control and were used to create stable lines. pTRE-ZNF274-3xHA was kindly gifted from Dr. Martina Begnis in the Trono lab at the Swiss Federal Institute of Technology in Lausanne. This plasmid is a dox-inducible ZNF274 overexpression construct. Gibson cloning was used to replace ZNF274 with ZEB1 to create pTRE-ZEB1-3xHA, a dox-inducible ZEB1 overexpression construct. pSCALP-ZEB1 (ZEB1 overexpression plasmid) was kindly gifted from Dr. Giuseppe Diaferia at Humanitas University in Italy.

Viral particles were synthesized and infected as described in ref. ⁷⁶. Cells infected with the SMART vector shZNF274 and shCTRL were then selected for single clones by dilution for single cells in a 96 well plate and grown out to form stable cell lines.

Chromatin immunoprecipitation

Cells were cross-linked with 1% paraformaldehyde for 15 min at room temperature. The reaction was quenched for 5 min at room temperature by adding 0.125 M glycine. After three washes with 1X PBS, cells were lysed with lysis buffer (1% SDS, 10 mM EDTA pH 8, 50 mM Tris-HCl pH 8) supplemented with a protease inhibitor cocktail (Complete EDTA-free, Roche Applied Science), 5 μM TSA, 5 mM sodium butyrate, 1 mM DTT, and phosphatase inhibitors (Phosphatase Inhibitor Cocktail Sets I, II, and III Calbiochem). Lysates were incubated on ice for 20 min and then sonicated using a QSONICA Q800R3 sonicator (10 seconds ON, 20 s OFF, 30% AMP; samples in 4 °C water bath throughout sonication cycles). Size of fragments obtained (between 200 and 1200 bp) was confirmed by electrophoresis. Soluble chromatin was collected after centrifugation at 18,000 x g at 4 °C for 10 min and diluted to 1/5 in dilution buffer (1% Triton X-100, 2 mM EDTA, 150 mM NaCl, 20 mM Tris-HCl pH 8.1) supplemented with protease, deacetylase, and phosphatase inhibitors. Soluble chromatin (1%) was kept as input control. Soluble chromatin was precleared with 100 μg/ml of salmon sperm (Amersham Biosciences), 2.5 μg/ml of unspecific IgGs, and 20 μL of protein-PLUS A/G Sepharose beads overnight at 4 °C in continued rotation. The next day samples were centrifuged, supernatants were collected, and specific antibodies were added. Mixtures were incubated at 4 °C overnight in rotation. The following day protein-PLUS A/G Sepharose beads (sc-2003) were added to the samples and incubated for 2–4 h at 4 °C in rotation. Beads were collected and washed sequentially at 4 °C for 10 min with TSE I (150 mM NaCl, 0.1% SDS, 1% Triton X-100, 2 mM EDTA, and 20 mM Tris-HCl (pH 8.1)), TSE II (500 mM NaCl, 0.1% SDS, 1% Triton X-100, 2 mM EDTA, and 20 mM Tris-HCl (pH 8.1)), buffer III (0.25 LiCl, 1% Nonidet P-40, 1% deoxycholate, 1 mM EDTA, and 10 mM Tris-HCl (pH 8.1)) and then 1x PBS. Immunoprecipitates were eluted with 100 μL of with elution buffer (0.1 M NaHCO3 and 1% SDS) and rocked for 40 min. Reversion of cross-linking was performed overnight by adding 0.2 M NaCl and heating and shaking samples and input controls at 65 °C. Samples were then treated with 0.2 mg/mL RNAse A (Qiagen) and incubated for 1 hr at 37 °C followed by addition of 0.01 M EDTA, 0.04 M Tris-HCl pH 6.5 and 4 U/mL of Proteinase K (Promega) and samples were incubated at 45 °C for 1 hr. DNA was then purified using the Monarch DNA Clean up kit (NEB). 1 μL of purified DNA was then used in real-time quantitative PCR as described above. The primer sequences are listed in Table S2. Antibodies used for ChIP were anti-ZNF274 (1:200, Novus Biologicals, catalog # H00010782-A01), anti-H3K9me3 (3 μg, CST #13969), anti-SETDB1 (1:200, Proteintech Group #11231-1-AP), anti-H3K4me3 (1 μg Millipore #07-473), anti-ZEB1 (1:200, GeneTex # GTX105278), and anti-IgG XPA Isotype Control (5 μg, CST #3900S).

Scratch assay

Plate 30,000 cells per well in a 96-well plate. The next day perform scratch, wash twice with PBS, and replace each well with 200 μL of media. Place plate in IncuCyte and program to take images every 4 h for 24 h. After imaging completes, remove plate, export images, and analyze using Sartorius.

Invasion assay

Human PDAC organoids were cultured as described above and then were recovered from Matrigel by resuspension in cold cell recovery solution and incubated on ice. Organoids were centrifuged at 200 x g and resuspended in cold DMEM and kept on ice.

Organoids were embedded in 1.5 mg/mL collagen gels using protocols previously published, which were prepared from the 4.5 mg/mL rat-tail collagen made by the Cheung lab^102,103. 4.5 mg/mL collagen was diluted to 1.8 mg/mL in 0.1% acetic acid. This was then mixed with 10X DMEM, 1 N NaOH, and ultrapure water at the following ratio:

1 mL collagen (1.8 mg/ml): 100 μL 10X DMEM: 32 μL NaOH: 66.66 μL H2O

This mixture was allowed to polymerize at 4 °C and was then mixed with organoids at a concentration of ~180 organoids/100 μL. Glass-bottom 24 well plates were prepared by plating 40 μL of collagen (without organoids) in the center of each well to be used. The plate was placed at 37 °C for 5–10 min to allow the collagen disks to polymerize, after which 100 μL of the collagen/organoid mixture was plated on top as a dome. This was allowed to polymerize for an additional 45–60 min at 37 °C. After complete polymerization, 1 mL of human organoid feeding media was added to each well.

At the end of the experiment, the media was removed, and PBS was added for 5–10 min. The organoids in collagen were fixed with 4% PFA for 20 min at room temperature in the 24 well plate. Wells were washed with PBS and then stored at 4 °C. Plates were imaged using an Andor CSU-W confocal spinning disk on a Leica DMi8 inverted microscope. All organoids per well were imaged and counted for invasive protrusions.

RNA-seq data analysis

RNA-seq libraries were prepared with extracted total RNA from cell lines using the Illumina Stranded Total RNA Prep with Ribo-Zero Plus kit having an ERCC spike-in. Libraries were barcoded and pooled, with 30 samples in a single pool, and sequenced using 50 bp paired-end reads in Illumina NextSeq P4 flow cell. Image analysis and base calling was performed using Illumina’s Real Time Analysis v4.12.2 software, followed by demultiplexing of indexed reads and generation of FASTQ files using Illumina’s bcl2fastq Conversion Software v2.20. An average of 61 M 50 bp paired-end reads per sample passed Illumina’s default quality filters and were retained for further analysis. Alignment was performed against the iGenomes hg38 reference genome using STAR v-2.7.10b in the two-pass alignment mode¹⁰⁴. FastQC¹⁰⁵ and RSeQC¹⁰⁶ were used to perform pre-alignment and post-alignment quality checks respectively. STAR2’s –quantMode flag was used with Gencode v38 gene definitions to perform gene expression quantification and generate counts. GSEA was used to identify enriched pathways across phenotypes of interest¹⁰⁷.

Broad repeat analysis

To enable identification of transposable elements in the data, samples were realigned to the reference genome to allow for inclusion of reads that map to multiple loci (up to 100). All other alignment parameters were unchanged. The TEtranscripts software package from the Hammell lab at CSHL was then used to identify and quantify TEs present¹⁰⁸. Differential expression analysis across groups of interest was performed using Bioconductor edgeR¹⁰⁹. A significance threshold of log2FC >= 1 or log2FC <= -1 at 5% FDR was used to define the genes and TEs of interest in each comparison.

HERVs quantification analysis

The DIGS tool¹¹⁰ and a reference dataset of consensus HERV probes¹¹¹ were used to screen the human genome (assembly version hg19) for HERV elements; “hits” were filtered by bitscore (LTR hits >= 200, RT hits >= 80, TM hits >= 70, ORF hits >= 100, ORF “pro” hits >= 180). For each scaffold, the distance between hits was computed, revealing a bimodal distribution on a log₁₀ scale (Fig. S10A). Most contiguous hits were separated by 4.5 to 7.5 kb, corresponding to the trailing end of the first peak.

Following the typical provirus gene order: LTR-gag-pro-pol/RT-env/TM-LTR, hits that were within 4.5-7.5kbps (see “allowed distance per chromosome” in Table S3) were consolidated into proviruses, allowing for repetition of same-gene hits. This led to identification of 15,354 DIGS-consolidated HERV loci. DIGS-consolidated HERVs genomic coordinates were overlapped with HERV loci in the RetroTector dataset¹¹² and Repbase LTR loci (obtained from UCSC Genome Browser), using bedtools merge -d 0 -s version v2.31.1. This compilation resulted in a list of 135,450 HERV loci. Source database, name and gene information were preserved for each overlapped locus. Distance between HERVs was re-analyzed and found to be in a log2 distribution skewed right (Fig. S10B). To allow for cases where said databases did not capture full-length HERV loci, elements within 533 bp (5% right tail of aforementioned distribution) were merged using bedtools merge with options -d 533 -s. This led to a total of 128,825 merged HERV loci.

To identify high confidence non-solo LTR HERVs, only loci supported by DIGS-consolidated and/or RetroTector (with or without Repbase) were considered (15,030 loci). Of these 6891 loci contained hits to internal regions (non-LTRs) in DIGS (i.e., ORFs, RT or TMs). To account for internal regions that were not detected via DIGS, the size of merged HERV loci that contained only hits to LTRs in DIGS was computed (Fig. S10C) and those with size >= 2418 bp (the 3rd QT size) were regarded as non-solo LTRs (433 loci). Finally, 129 loci were composed of Retrotector elements without DIGS support (with or without Repbase support). In combination, these strategies resulted in 7,453 high confidence non-solo LTR HERV loci (Supplementary Data 1)

RNAseq data was processed using a NextFlow pipeline. For the HERV analysis, reads were trimmed using bbduk from the BBMap package (v38.97). Trimmed reads were aligned and quantified using SALMON v1.9.0 against a non-solo LTR HERV-inclusive hg19 GTF, described above. Differential expression analysis was performed on normalized TPM values for all genes using DESeq2. DEGs were defined as |log2FC | >=1 and padj <= 0.05. HERV classifications for differentially expressed HERVs were annotated using table 4 from PMID: 29308093¹¹³. Annotations for newly defined HERVs were assigned first at the Group level and then consolidated to the Superfamily level if multiple Group annotations were observed. For broad repetitive elements quantification, trimmed reads were aligned using STAR v2.7.11 using a transposon-inclusive hg19 transcriptome annotation, including a subset of transposable elements derived from RepeatMasker, including LINEs, SINEs, LTRs, and DNA transposons. Features were quantified using featureCounts v2.0.3. Differential expression analysis was performed on normalized TPM values for all genes and repeat elements using DESeq2. TPMs from transposable elements were summarized into larger classes, then normalized by dividing by the total number of elements within each repClass. Data was analyzed and visualized in R v4.4.2 and ggplot2 v3.5.

Coimmunoprecipitation

AsPC1 cells grown to ~80% confluence in a 15-cm dish were scraped into 5 ml nuclear lysis buffer (10 mM HEPES pH 7.4, 10 mM KCl, 0.05% NP-40, Complete EDTA-free protease inhibitor cocktail [Roche Applied Science], 5 μM TSA, 5 mM sodium butyrate, 1 mM DTT, and Phosphatase Inhibitor Cocktails I, II, and III [Calbiochem]). Cells were incubated on ice for 20 min and centrifuged at 300 × g for 10 min at 4 °C. Nuclear pellets were resuspended in 500 μl Co-IP buffer (50 mM Tris-HCl pH 7.4, 170 mM NaCl, 20% glycerol, 15 mM EDTA, 0.1% Triton X-100, and inhibitors as above) and sonicated three times for 10 s on/20 s off using a Qsonica sonicator at 20% amplitude. Soluble chromatin and proteins were collected by centrifugation at 300 × g for 10 min at 4 °C. Nuclear protein lysates were split in half, and 3 μg of SETDB1 antibody or rabbit IgG were added to each half and incubated overnight. The next day, 50 μl Dynabeads Protein A were added and incubated for an additional 2 h with rotation. Beads were collected using a magnetic rack and washed four times with 500 μl ice-cold wash buffer (25 mM Tris-HCl pH 7.4, 135 mM NaCl, 2.5 mM KCl, plus protease and phosphatase inhibitors). Proteins were eluted in 25 μl 2× SDS sample buffer by boiling at 95 °C for 5 min. After bead removal, protein eluates were resolved by SDS–PAGE and analyzed by western blotting using standard methods with antibodies against ZNF274 (1:200, Novus Biologicals #H00010782-A01), SETDB1 (1:1,0000, Proteintech #11231-1-AP), KAP1 (1:1000, Proteintech #15202-1-AP), and DNMT1 (1:1,000, Abcam #ab13537).

Native chromatin fractionation and sucrose gradient ultracentrifugation

AsPC1 cells were transfected with either siCtrl or siZNF274 for 72 h. One 15-cm plate per condition was scraped in culture medium, and cell pellets were washed once in 1× DPBS and incubated in 500 μl nuclear lysis buffer (10 mM HEPES pH 7.4, 10 mM KCl, 0.05% NP-40, Complete EDTA-free protease inhibitor cocktail [Roche Applied Science], 5 μM TSA, 5 mM sodium butyrate, 1 mM DTT, and Phosphatase Inhibitor Cocktails I, II, and III [Calbiochem]) for 20 min on ice. Cells were centrifuged at 300 × g for 10 min at 4 °C, and nuclei were washed once in nuclear lysis buffer and resuspended in 1 ml native lysis buffer (0.2% IGEPAL, 100 mM NaCl, 50 mM Tris-HCl pH 7.0, 1 mM EDTA, 0.2 mM DTT, protease and phosphatase inhibitors, 5 μM TSA, 5 mM sodium butyrate). Samples were sonicated three times for 10 s on/20 s off using a Qsonica sonicator at 20% amplitude. Lysates were cleared by centrifugation at 300  × g for 5 min at 4 °C, and 1 mg total protein was used for ultracentrifugation.

Linear 5–40% (w/v) sucrose gradients were prepared in native lysis buffer in polyallomer centrifuge tubes (Beckman). Cleared lysates were layered on top of 9 ml gradients and centrifuged for 16.5 h at 240,486.6 × g in an SW41 Ti rotor (Beckman Coulter). Ten fractions (1000 μl each) were collected manually from the bottom of each tube. Proteins were precipitated with 20% (w/v) TCA, washed with cold acetone, and resuspended in 1× Laemmli buffer. Proteins were resolved on 4–20% Criterion TGX Precast Midi Protein Gels (Bio-Rad) and transferred for immunoblotting using standard procedures with antibodies against ZNF274 (1:200, Novus Biologicals #H00010782-A01), SETDB1 (1:1,0000, Proteintech #11231-1-AP), KAP1 (1:1000, Proteintech #15202-1-AP), and DNMT1 (1:1,000, Abcam #ab13537), H3K9me3 (1:1,000, Active Motif #39062), and H3 (1:1,000, Abcam #ab4729).

AutoCut&Run and data analysis

AsPC1 cells expressing shCtrl or shZNF274 were cultured for 9 days, and 1 × 10⁶ cells per condition were submitted to the Fred Hutch Genomics Core for AutoCUT&RUN with a 1:50 antibody dilution. The following antibodies were used: ZNF274 (Novus Biologicals #H00010782-A01), H3K9me3 (Active Motif #39062) and IgG (Antibodies Online #ABIN101961). Libraries were sequenced on an Illumina HiSeq platform. Read quality assessment and adapter trimming were performed with Trim Galore (https://github.com/FelixKrueger/TrimGalore), and reads were aligned to the human reference genome using Bowtie2. Genome coverage tracks were generated with bedtools genomecov. Peaks were called with both MACS2 and SEACR using IgG controls; IgG reads were deduplicated with GATK MarkDuplicates, and MACS2 was run in target-only and IgG-controlled modes. For broad histone marks (e.g. H3K9me3), target-only peaks overlapping IgG-controlled peaks and passing stringent enrichment and q-value thresholds were retained. For each target, differential peaks between shCtrl and shZNF274 were identified with edgeR. Replicated peaks were first defined by merging peaks present in ≥2 replicates, then consensus peak sets were generated across all samples. After filtering with edgeR’s filterByExpr, count data were TMM-normalized and tested using generalized linear models with quasi-likelihood; significance thresholds are described in the Differential Peaks section. Consensus peaks were annotated to genomic features and nearest genes with HOMER.

Cut&RUN differential peak analysis and tiering

All analyses were performed in R v4.5.2 (2025-10-31) on macOS (Sequoia 15.7.2; time zone America/Los_Angeles) using Bioconductor and CRAN packages as recorded by sessionInfo(). Cut&RUN datasets for ZNF274 and H3K9me3 were analyzed separately as two marks under shRNA conditions (shZNF274 vs shCTRL), and then integrated with RNA-seq differential expression measured under siRNA conditions (siZNF274 vs siCTRL). Differential peak tables for each mark were imported into R and standardized to common variables including a unique peak identifier (peak_id), genomic coordinates (chrom, start, end), effect size (log2FC/logFC), nominal significance (pvalue/PValue) and multiple-testing corrected significance (padj/FDR). Homer peak annotation outputs were imported and merged to associate peaks with nearby genes (gene fields such as SYMBOL/GeneName). For downstream analyses H3K9me3 peaks were retained if they met FDR < 0.05 and |log2FC | ≥ 1 (both gains and losses were included unless otherwise stated), whereas ZNF274 peaks were retained using a more inclusive filter of FDR < 0.10 and |log2FC | ≥ 0.5, and for ZNF274-focused analyses we further restricted to ZNF274 loss peaks defined as log2FC < 0 passing the same thresholds.

Cytoband enrichment analysis

Peaks were imported from the H3K9me3 differential peak table and split by direction; only H3K9me3 DOWN peaks (log2FC < 0 according to the script’s direction == “DOWN” labeling) were carried forward as the interest set, while the complete set of tested/kept H3K9me3 peaks constituted the universe/background. Peak coordinates were represented as genomic intervals and handled with GenomicRanges v1.62.1, IRanges v2.44.0, S4Vectors v0.48.0, and GenomeInfoDb v1.46.2, with chromosome naming harmonized and analysis restricted to primary chromosomes to avoid alternative contigs. Cytoband annotations for the same genome assembly as the peak coordinates were converted to genomic intervals (GRanges) and overlaps between peaks and cytobands were computed. For each cytoband, enrichment of H3K9me3 DOWN peaks was quantified using a 2×2 contingency table comparing (i) counts of DOWN peaks overlapping that cytoband versus all other cytobands and (ii) counts of universe peaks overlapping that cytoband versus all other cytobands; significance was tested using Fisher’s exact test (one-sided enrichment) and reported as odds ratios and p-values, with Benjamini–Hochberg correction applied across cytobands to control the false discovery rate (FDR). Results were exported as CSV tables and visualized as bar plots of enriched cytobands using dplyr v1.1.4, tibble v3.3.0, readr v2.1.6, and ggplot2 v4.0.1.

Chromosome-level enrichment analysis

Chromosome-level enrichment of H3K9me3 DOWN and ZNF274 DOWN peak sets was assessed in R v4.5.2 using dplyr v1.1.4, tibble v3.3.0, readr v2.1.6, and ggplot2 v4.0.1. Differential peak tables were filtered to retain DOWN peaks for each mark, and each peak was assigned to a chromosome using the coordinates contained in the corresponding “universe/all tested peaks” set to ensure consistent coordinate provenance across analyses. Only primary human chromosomes were considered (chr1–chr22 and chrX/chrY; filtering implemented by a regular expression on chromosome names to exclude alternative contigs). For each chromosome and each mark, enrichment of DOWN peaks was quantified by comparing the fraction of interest peaks on that chromosome to the fraction of all tested (universe) peaks on that chromosome, and statistical significance was evaluated using Fisher’s exact test (one-sided, alternative = “greater”) on a 2×2 contingency table of (interest vs non-interest) × (on chromosome vs not on chromosome). P-values were corrected for multiple testing across chromosomes using the Benjamini–Hochberg procedure to obtain FDR-adjusted p-values, and chromosomes with FDR < 0.05 were reported as significantly enriched. For visualization, log2 fold-enrichment was displayed in bar plots, and to avoid undefined values when chromosome counts were zero, a small pseudocount (0.5) was applied to compute a plot-safe log2 fold-enrichment for display only; statistical tests were performed on the original integer counts.

RNA-seq expression analysis and integration with Cut&RUN

RNA-seq counts were provided as per-sample STARcounts files and were merged into a single gene-by-sample count matrix in R using the gene identifier as the join key (gene_id). A minimal sample table was defined in-script specifying sample and condition (siCTRL vs siZNF274), and differential expression was computed using DESeq2 with the design formula ~ condition. For heatmap visualization, normalized expression values were obtained using DESeq2’s variance-stabilizing transformation (VST), and expression was scaled gene-wise (row z-scores) to emphasize relative differences across samples.

To generate an integrated heatmap that combines RNA expression with chromatin changes, we defined a “top gene” set from the RNA-seq results using script variables (e.g., top_n, rna_fdr_thr). Genes were ranked primarily by adjusted p-value (padj, ascending) and secondarily by absolute RNA effect size ( | log2FoldChange | , descending); the heatmap gene list was then defined as the union of the top upregulated and top downregulated genes passing the selected RNA threshold (e.g., padj <rna_fdr_thr). Cut&RUN changes were mapped onto this gene list by collapsing peak-level results to gene-level summaries using the Homer-associated gene field (e.g., SYMBOL/GeneName). When multiple peaks mapped to the same gene, a single representative peak per gene and mark was selected by prioritizing smallest adjusted p-value (padj/FDR) and, if tied, largest absolute log2 fold-change ( | log2FC | ). The representative peak’s log2FC and direction were retained as the gene-level Cut&RUN summary. For the ZNF274-loss view, only ZNF274 loss peaks (log2FC < 0 passing the stated cutoffs) contributed to the ZNF274 annotation track, while H3K9me3 annotations could include both gain and loss peaks passing the H3K9me3 cutoffs. Heatmaps were generated using pheatmap, displaying the RNA expression matrix (row-z-scored VST values) with adjacent annotation tracks capturing mark-specific direction/state and/or continuous Cut&RUN log2FC, and gene label sizes were adjusted through plotting parameters to maintain readability for larger gene sets.

Motif enrichment analysis for sequence features

Motif enrichment analyses in H3K9me3 peak subsets (UP vs DOWN, as defined by the differential direction labels in the script) were performed using motifs from JASPAR2022 v0.99.8 accessed through TFBSTools v1.48.0, with sequence retrieval and motif matching implemented using Biostrings v2.78.0 and BSgenome v1.78.0 where applicable, motif occurrence was evaluated over peak-centered windows (e.g., 200 bp) and motif hits were called using a relative minimum score threshold (e.g., 85% of maximal PWM score), and enrichment was summarized by odds ratios and Fisher’s exact tests comparing motif-positive fractions in the interest set versus a defined background/universe set, with resampling-based background evaluation performed over 200 iterations to obtain distributions of effect sizes and p-values and to support violin/boxplot visualizations and comprehensive table exports. General data manipulation and plotting used dplyr v1.1.4, tibble v3.3.0, readr v2.1.6, ggplot2 v4.0.1, and related tidyverse dependencies.

HERV overlap and enrichment testing

For HERV analyses, curated HERV interval coordinates were imported from a BED resource (see Methods section HERVs quantification analysis for source of 135,450 HERV loci) and represented as GRanges. ZNF274 loss peaks were defined as above (log2FC < 0 with the ZNF274 cutoffs) and converted to GRanges using coordinates from the same differential peak universe to avoid coordinate convention mismatches. Overlaps between selected peaks and HERV intervals were computed with findOverlaps(ignore.strand = TRUE). Enrichment of HERV overlap among ZNF274 loss peaks was quantified using Fisher’s exact test comparing the overlap fraction in the selected set versus the full ZNF274 peak universe (background), reporting odds ratios and 95% confidence intervals along with p-values. Counts of overlaps per HERV identifier were summarized to rank the most frequently overlapped HERV elements.

Motif analysis in promoter and regulatory regions

Motif analyses were performed in R (version 4.5.2) on macOS (Sequoia 15.7.2; aarch64-apple-darwin20) using Bioconductor packages. Promoter sequences were extracted from the human reference genome (hg38) using BSgenome and transcript annotations from the UCSC knownGene gene model (TxDb.Hsapiens.UCSC.hg38.knownGene). For each target gene, a strand-aware promoter window was defined relative to a representative transcription start site (TSS) as 20 kb upstream and 2 kb downstream. For KRT5, a representative promoter was defined from annotated transcripts by selecting a single representative TSS and constructing the corresponding promoter window. For TP63, a proxy ΔNp63 promoter was derived from annotated TP63 transcripts by computing TSS positions for all TP63 transcripts, clustering TSS coordinates within 5 kb, and selecting the most downstream TSS cluster in a strand-aware manner; the median TSS coordinate of the selected cluster was used as the representative ΔNp63 TSS. Promoter coordinates were clamped to chromosome bounds using hg38 seqlengths to avoid out-of-range sequence extraction.

Transcription factor motifs were retrieved from JASPAR CORE using TFBSTools, and motif collections for ZEB1, TEAD-family factors, and AP-1-family factors were assembled by regex-based filtering of motif names restricted to Homo sapiens (NCBI taxonomy ID 9606). AP-1-family motifs were defined by matching FOS, JUN, AP1, FOSL, JUND, or JUNB, and TEAD-family motifs were defined by matching TEAD or TEF. Motif position frequency matrices were converted to position weight matrices prior to scanning to ensure compatibility with the motif search routines. Promoter sequences were scanned for motif occurrences using TFBSTools with a minimum relative score threshold of 85 percent. For each motif hit, coordinates were recorded in promoter coordinates and converted to genomic coordinates, enabling visualization of motif tracks across the promoter regions.

To quantify motif proximity, distances were computed from each ZEB1 motif midpoint to the nearest TEAD-family and AP-1-family motif midpoint within the same promoter. ZEB1-centered co-occurrence was summarized as the number of ZEB1 sites with at least one TEAD-family motif or AP-1-family motif within a symmetric proximity window around the ZEB1 site, as well as the number of ZEB1 sites with both motif classes within the same window. Primary analyses were performed using a ± 200 bp window, and proximity sensitivity analyses evaluated additional windows including ±100 bp and larger thresholds. Distributions of nearest distances and region-level co-occurrence counts were visualized using ggplot2.

Statistical significance was evaluated using empirical null frameworks. A within-promoter circular-shift permutation test was used to test preferential motif alignment independent of motif density and clustering: TEAD-family motif coordinates were circularly shifted by a random offset along the promoter length with wrap-around while ZEB1 coordinates were held fixed, generating a null distribution for the number of ZEB1 sites with nearby TEAD motifs and for the median nearest ZEB1-to-TEAD distance. Empirical permutation p-values were derived from these null distributions for enrichment in close co-occurrence or depletion of nearest-distance statistics as appropriate. All plots and summary tables were generated in R and exported as PDF/PNG figures and CSV files for downstream reporting.

Software versions used included magrittr 2.0.4, dplyr 1.1.4, ggplot2 4.0.1, readr 2.1.6, purrr 1.2.0, Biostrings 2.78.0, GenomicRanges 1.62.1, IRanges 2.44.0, BSgenome 1.78.0, GenomicFeatures 1.62.0, AnnotationDbi 1.72.0, rtracklayer 1.70.0, and TFBSTools 1.48.0.

Proliferation IC50 assay

Assay was performed as described in previous work⁷⁶ using increasing doses of SY-5609 or DMSO control (BP231-100; ThermoFisher Scientific). The proliferation assay was performed in duplicate, and data are represented as mean ± SEM among three independent experiments unless otherwise indicated in the figure legend.

Subcutaneous Cell line Xenografts

All mouse procedures were conducted under protocol PROTO201900024 in accordance with the Fred Hutchinson Cancer Center (FHCC) Institutional Animal Care and Use Committee (IACUC) guidelines and the ARRIVE guidelines. Athymic nude (NU/J) female mice (002019, JAX) were purchased from the Jackson Laboratory. Mice were housed in pathogen-free facilities under controlled conditions: 12 hr light/dark cycle, ambient temperature of 75^oF, and a humidity level of 30-70%. Injections of cell lines for the generation of subcutaneously grafted cell line tumors were prepared by resuspending 1.5×10^6 cells in a 100 μL suspension of 50% Matrigel (Corning) in phosphate-buffered saline (PBS). Cell suspensions were injected into each flank of the mouse. 5 days after injection the mice were put on water with 200 μg/mL of doxycycline. 3 days after the initiation of doxycycline water, initial tumor volumes were measured, and mice started on once daily oral gavage treatment of vehicle (5% captisol) or SY5609 (0.5 mg/kg, Syros) dissolved in 5% captisol for 22 days. Caliper measurements of tumors and body weights were recorded three times per week. Endpoint is called when the first vehicle tumor diameter reaches ~2 cm. The tumor diameter size represents the maximal tumor size/burden permitted by our institutional review board and was not exceed for this study. Tumor volumes were normalized to the initial tumor volume for each mouse and plotted as normalized tumor growth.

Patient-derived Xenograft

PDX models were acquired from the NIC PDMR and JAX (PAAD154b - NCI 885724; PAAD141b - NCI 463931; PAAD169b - JAX TM 00176; PAAD200—JAX TM01212). PDX implantations were administered as previously described in ref. ⁷⁶ followed by post-surgical monitoring. Tumors were measured with electronic calipers and upon reaching an average tumor volume of ~50-70 mm^3, mice were treated with a once daily oral gavage of vehicle (5% captisol), SY5609 alone (0.5 mg/kg, Syros), Azacytidine alone (2.5 mg/kg, S1782, Selleck), or combination of SY5609 (0.5 mg/kg) plus AZA (2.5 mg/kg) for 22 days. Caliper measurements of tumors and body weights were recorded three times a week. A subset of the combination drug trial mice were taken out to endpoint following termination of treatment at 22 days. Endpoint is called when the tumor diameter reaches ~2 cm and this indicates a notch in the survival curve. The tumor diameter size represents the maximal tumor size/burden permitted by our institutional review board and was not exceed for this study. Additionally, mice were euthanized if animals exhibited one or multiple predefined conditions: tumor ulceration, impaired mobility, inability to reach food/water, deterioration of vital physiological functions. The survival study concluded at 80 days post initiation of treatment.

Statistics and reproducibility

Statistical significance was determined by specific tests that are indicated in the figure legends. Statistical analyses were performed using GraphPad Prism v10.4.2. Unpaired Student’s two-tailed t test was used when comparing data from two groups: *P   ≤  0.05; **P   ≤   0.01, ***P   ≤   0.001, and ****P   ≤   0.0001. In Fig. 1b, P value was determined using one-way ANOVA. In Fig. 2a, P value was determined using paired Student’s t test. For RNA-seq, a Wilcoxon Rank sum test was applied and specified. For binding motif analysis and HERV enrichment, a Fisher’s exact test (one-sided) was applied and specified. No statistical method was used to predetermine sample size of represented experiments. The experiments were not randomized, and the investigators were not blind to allocation during experiments and outcome assessment.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

JEG and SK conceptualized and designed the project. JEG, AS, SD, AWP, AD, LM, JH, NR collected data. SEK, PC, SG performed RNA-seq analysis, MP merged the dataset, and RJG performed the DIGS screen. NY performed murine surgical procedures. LDB performed RNA-seq library preps. FN provided human PDAC organoids. JEG and SK analyzed the data. JEG, DBM, KC and SK contributed to the experimental design. JEG wrote the paper. All authors discussed the results and commented on the manuscript.

Peer review

Peer review information

Nature Communications thanks Jing Xue and the other anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

S.K. discloses support for the research of this work from National Institutes of Health (R37CA241472) and the V Foundation: Women Scientists Innovation Award for Translational Cancer Research (T2023-008). J.E.G discloses support for the research of this work from Fred Hutchinson Cancer Center Human Biology Pilot Grant and the ARCS Foundation Fellowship. S.D discloses support for the research of this work form American Cancer Society fellowship (PF-24-1196662-01-RMC). A.W.P discloses support for the research of this work from F32 fellowship (F32CA284723). K.J.C discloses support for the research of this work from National Institutes of Health (R37CA234488) and DOD CDMRP (BC240512). S.E.K discloses support for the research of this work from V Foundation for Cancer Research (V2022-033), the Searle Scholars Program (SSP-2023-103), and an F31 Fellowship F31CA306243.

Data availability

All data associated or generated with this study are present in paper or the supplementary materials, Source Data file, or used publicly available datasets. The RNA-seq and CUT&RUN data generated in this study are publicly available in GEO at the National Center for Biotechnology Information under the accession number: GSE285406. The human patient PDAC tumor RNA-seq data used for analysis in this study is publicly available in European Genome-Phenome Archive under the accession number: EGAS00001002543. Source data are provided with this paper.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-73380-x.