Hypoxia-induced metastatic heterogeneity in pancreatic cancer Authors: Pradeep Moon Gunasekaran, Qianqian Wang, Polina Guseva, Yoke-Chen Chang, Rajika Chauhan, Alexander Kley, Mengdie Xu, Gene Lee, Gregory Marshall, Siddharth Ghosh Roy, Yousef Masoudpoor, Oscar Li, Arthur Roberts, Kelly Watkins Walton, Lucyann Franciosa, Shafiq Bhat, Emmanuel Zachariah, Kishan Patel, Zhongren Zhou, Jian Cao, Wenjin Chen, Julie Zhouli Ni, Sam Guoping Gu, Cristina Montagna, Shin-Heng Chiou Published: 2026 ABSTRACT AbstractIn solid tumors, hypoxia is a key driver of metastasis by promoting cellular plasticity and chromosomal instability (CIN). Despite this, the mechanisms by which malignant cells concurrently co-opt these elements of hypoxic adaptation to promote metastasis remains unclear. Here we report that hypoxia promotes metastasis by suppressing the JmjC-containing histone lysine demethylase KDM8. Kdm8 targeting in a Kras;Trp53-driven mouse model of pancreatic ductal adenocarcinoma induces a profound loss of the epithelial morphology and widespread metastatic disease. Mechanistically, Kdm8 suppression in normoxia recapitulates major aspects of the global epigenetic changes, transcriptomic rewiring, and mitotic spindle defects induced by hypoxia. Of note, disruption of Kdm8’s demethylase function phenocopies the effects of Kdm8 loss, whereas expression of hypermorphic Kdm8 variants that are resistant to hypoxic suppression reduces metastasis beyond the levels achieved by the wildtype counterpart. Through the suppression of Kdm8 demethylase activity, hypoxia unleashes a potent metastatic program by simultaneously advancing cellular plasticity and CIN.Subject terms: Pancreatic cancer, Metastasis Introduction Pancreatic ductal adenocarcinoma (PDA) remains one of the deadliest malignancies, largely due to its highly aggressive metastatic behavior1. As with other carcinomas, PDA metastasis requires the acquisition of diverse and often opposing cellular programs that enable dissemination and colonization at distant sites2. In line with this, paired assessments of genomic alterations in primary and metastatic tumors provided limited insights into the molecular drivers of metastasis3–5. In contrast, recent studies highlight transcriptomic and epigenetic plasticity as key malignant features underlying treatment resistance, subtype differentiation, and metastatic progression in PDA6–17. Despite these advances, the cancer cell-extrinsic factors within the tumor microenvironment that drive metastatic plasticity remain elusive. Intratumoral hypoxia is a potent inducer of metastasis18. Under hypoxic conditions, the stabilization of the hypoxia-inducible factors (HIFs) promotes epithelial-to-mesenchymal transition (EMT), a gene expression program broadly linked to metastasis across multiple cancer types19. In PDA specifically, EMT has been shown to facilitate the metastatic dissemination6,12,20,21. However, genetic depletion of Hif1a, which encodes the hypoxia-stabilized subunit of the Hif1a/Arnt heterodimer, unexpectedly accelerated KrasG12D-driven pancreatic tumor progression and metastasis22,23, suggesting that hypoxia may act through additional, Hif1-independent mechanisms. Beyond HIFs signaling, hypoxia triggers extensive epigenetic reprogramming, in part by inhibiting the activity of JmjC-domain-containing histone lysine demethylases (KDMs), which require oxygen (O₂) as a cofactor24,25. In human PDA, genetic loss of KDM6A ( ~ 6%), an X-linked histone 3 lysine 27 (H3K27)-specific demethylase, is associated with the basal-like or squamous-like state and poor prognosis26–28. Yet, studies using genetically engineered mouse (GEM) models indicate that Kdm6a suppresses PDA progression through a demethylase-independent mechanism, raising the possibility that other KDM members may mediate hypoxia-driven reprogramming of the malignant state26. In addition to its epigenetic effects, hypoxia is known to promote metastasis through DNA overreplication29. Hypoxia-induced DNA damage and genomic alterations have traditionally been attributed to the suppression of multiple DNA damage response pathways30. Large genomic studies utilizing bulk whole genome sequencing (WGS) data from patients suggest a clear link between hypoxia in solid tumors and distinct mutational and copy number alteration (CNA) signatures31,32. Despite this, the mechanism by which hypoxia promotes chromosomal instability (CIN) remains unclear30. CIN in cancer enables the selective amplification of genomic loci that enhance cancer cell fitness under stress conditions, including hypoxia33–37. One such example is KRAS copy number gain, which becomes increasingly prevalent during PDA progression and is most pronounced in metastatic lesions when compared to the primary site38–40. Mounting evidence suggests that oncogenic KRAS drives PDA progression through diverse mechanisms, including metabolic re-adaptations41–44, immune suppression45–47, and subtype differentiation39,48. However, whether and how KRAS copy number in PDA arises as a direct consequence of hypoxia remains unresolved. KDM8 is a histone lysine demethylase that targets H3K36 di-methylation (H3K36me2) and regulates cell proliferation in mammalian cells49,50. In C. elegans, disruption of the KDM8 ortholog jmjd-5 results in the loss of germ cell identity and reduced fertility51. Genetic depletion of Kdm8 in mice leads to embryonic lethality, attributed to dysregulation of Cdkn1a via increased H3K36me2 levels50,52. In human embryonic stem cells, KDM8 (a.k.a. JMJD5) is required to maintain pluripotency53, and germline mutations are associated with severe developmental disorders including intellectual disability and craniofacial abnormalities54. Furthermore, jmjd-5/KDM8 has also been implicated in regulating chromosome segregation and maintaining genomic integrity, presumably through its JmjC domain-dependent activity55,56. Despite its established roles in development, pluripotency, and genome maintenance, the role of KDM8 in regulating cellular plasticity and metastatic progression in PDA remains unknown. In this study, we present compelling evidence that hypoxia promotes PDA metastasis through the suppression of KDM8. Using a custom PDA GEM model that employs CRISPR/Cas9 genome editing and barcoded KrasG12D alleles57,58, we show that Kdm8 depletion reprograms the epithelial state toward EMT, driving widespread metastasis. Notably, the metastatic potential of PDA can be modulated by toggling the activity of KDM8’s JmjC domain, the inactivation of which heightens CIN to fuel metastatic progression. These findings offer further mechanistic insights into how hypoxia enhances the metastatic proclivity of malignant PDA cells. Results Knockdown of Kdm8 drives metastasis in multiple transplanted tumor models Unlike the HIF pathway, select KDM family members respond to hypoxia by inducing immediate epigenetic changes24. Our initial effort focused on identifying functionally important KDM family members that (i) induce embryonic lethality when genetically deleted and (ii) play a crucial role in regulating PDA metastasis. We performed subcutaneous tumor studies using the immunocompromised NSG (NOD-scid;Il2rgnull) mice to functionally interrogate five prioritized Kdm genes for their effect on lung metastasis using the 688 M KrasG12D;Trp53R172H murine PDA cell line (Figs. 1A, S1A, and S1B)11. Compared to the other Kdm family members, Kdm8 knockdown with two independent short hairpin RNAs (shRNAs) reproducibly promoted lung metastasis without affecting subcutaneous primary tumor growth (Figs. 1B and S1C–S1E). KDM8 is a PDA metastasis suppressor gene. A Schematic of the subcutaneous tumor model using Tomato-positive 688 M PDA cells to evaluate lung metastasis. B Lung metastasis counts in NSG mice transplanted with 688 M cells expressing control (shControl, n = 8) or shRNAs targeting Kdm8 (shKdm8-1, n = 6; shKdm8-2, n = 6). The experiments were repeated three times with similar results (rep 1, total n = 20 mice; rep 2, total n = 20 mice; rep 3, total n = 11 mice). C Schematic of the pancreatic orthotopic tumor model to evaluate lung and liver metastasis. Representative H&E staining of livers (D) and corresponding liver metastasis counts per NSG mouse orthotopically transplanted with shControl- or shKDM8-expressing human Capan-1 (E) or MIA PaCa-2 (F) cells. p-values were derived using Wilcoxon test. The experiments were repeated twice with similar results. Representative light (left) and fluorescent (right) images of the lungs (G) and H&E staining of the livers (H), tallies of mice with ascites (I), and counts of the liver (left, J) and lung (right, J) metastases in the NSG mice orthotopically transplanted with 688 M PDA cells expressing shControl (n = 8 mice) or two shKdm8s (Kdm8-1: n = 6 mice; Kdm8-2: n = 6 mice) as in (B). The experiments were repeated three times with similar results. Representative H&E of the livers (K), tallies of mice with peritoneal metastases (L), primary PDA tumor weight (left, M), and counts of liver metastases in the C57BL/6 J (B6) mice (right, M) orthotopically transplanted with the Tomato-negative 703 T PDA cells expressing shControl (n = 10 mice) or shKdm8 (n = 10 mice). NS, not significant. Representative light (left) and fluorescent (right) images of the lungs (N), tallies of mice with ascites (O), and counts of the lung (left, P) and peritoneal (right, P) metastases in the B6 mice orthotopically transplanted with the Tomato-positive 606 T PDA cells expressing shControl (n = 10 mice) or shKdm8 (n = 10 mice). The experiments were repeated twice with similar results. Arrowheads indicate liver metastases. Boxes represent median values and 25th to 75th percentile with whiskers extending to 1.5x IQR (B, E, F, J, M, P). P-values were determined by two-sided (E, M, P) and one-side (F) Wilcoxon signed-rank test, and Kruskal-Wallis test followed by Dunn’s multiple comparison test (B, J). To further investigate the role of KDM8, we orthotopically transplanted control and shKDM8-expressing human PDA cells, Capan-1 and MIA PaCa2, which express high levels of endogenous KDM8 protein, into the NSG mice (Figs. 1C and S1F). Consistent with the subcutaneous tumor experiments, KDM8 knockdown promoted liver metastasis in these orthotopic PDA tumor models (Figs. S1G–S1J, and 1D-1F). Furthermore, compared to the control counterpart, orthotopically transplanted Kdm8-deficient 688 M tumors seeded between 12 and 133-fold more metastases in the livers and lungs, respectively (Figs. 1G-J, S1K, and S1L). Notably, the increased metastases were accompanied by more frequent ascites formation in mice transplanted with Kdm8-deficient 688 M cells (Fig. 1I). Similarly, Kdm8 knockdown in two other murine PDA cell lines (703 T and 606 T, both KrasG12D;Trp53flox/flox) greatly promoted ascites formation and metastatic burdens in the lungs and livers when transplanted orthotopically into the immunocompetent C57BL/6 J mice (Fig. 1K–P and S1M-S1Q). Comparable to the findings in NSG mice, Kdm8-deficient 606 T orthotopic tumors seeded 25-fold more pulmonary micrometastases compared to the control group, with many metastatic lesions appearing as individual cancer cells (Figs. 1P and S1O). Notably, the increase in metastases following Kdm8 depletion appeared to be independent of any effect on the growth of established metastatic tumors (Figs. S1D, S1L, and S1Q). In summary, we screened five KDM family members and identified Kdm8 as a potential metastasis suppressor in PDA. Rigorous transplanted tumor studies employing multiple in vivo models demonstrated that Kdm8 and KDM8 knockdown in both murine and human PDA cell lines, respectively, increased PDA metastasis. These results were consistent across various transplantation routes, the Trp53 status of the murine PDA cell lines, short hairpin sequences for Kdm8 targeting, and the presence or absence of anti-tumor immune surveillance. Establishing an autochthonous PDA mouse model that enables Kdm8 genetic depletion using somatic genome editing To functionally interrogate Kdm8 in neoplasms initiated by oncogenic Kras and Trp53 inactivation, we built upon a previous approach58 that enables the induction of spontaneous PDA in the Trp53flox/flox;R26LSL-Tomato;H11LSL-Cas9 (PTC) mice (Fig. 2A) by delivering an engineered adeno-associated virus (AAV) into the adult murine pancreas through a retrograde pancreatic ductal injection procedure57. The engineered AAV for PDA initiation includes a U6 promoter-driven expression of an sgRNA targeting exon2 of Kras (U6-sgKras), a 2.1Kb Kras genomic template with the codon 12 glycine-to-aspartic acid mutation (KrasG12D) flanked by diverse combinations of synonymous mutations that serve as molecular barcodes for lineage tracing, a set of sgKras-resistant synonymous mutations, and PGK-Cre (Fig. 2B). Pancreatic infusion of the engineered AAV into adult PTC mice leads to the expression of sgKras and Cre, which in turn drives the expression of Cas9, deletion of Trp53, and introduction of the engineered Kras variant into the Kras locus through homology-directed repair, ultimately facilitating the growth of PDA that is histologically indistinguishable from those derived from conventional KPC (KrasLSL-G12D;Trp53LSL-R172H or Trp53flox/flox;Pdx1-Cre) mice57. Following the surgical procedure, an overall 74.2% of the mice developed histologically confirmed PDA by the time endpoints were reached (n = 23/31, Supplementary Data 1). Two cohorts received AAV engineered to carry either an additional U6-sgKdm8 (PTCKdm8KO, n = 18) or non-targeting U6-sgNT (PTCNT, n = 13) cassette. To assess the metastatic burden in these mice, 23 PTC mice with locally advanced PDA were carried forward to subsequent analyses (PTCNT, n = 13; PTCKdm8KO, n = 10). Similar to the control cohort, PTCKdm8KO mice developed solid tumors 4–6 months post-surgery with comparable overall survival (Fig. 2C and D). For each PTCKdm8KO tumor (total n = 7), we detected a minimum of 80% indel (insertion and deletion) formation at the Kdm8 targeted site estimated by the Surveyor nuclease assay, thus validating the effectiveness of the current method in deleting Kdm8 (Fig. 2E). CRISPR-mediated Kdm8 mutation induces profound loss of differentiation and widespread metastatic lesions in autochthonous PDA. Schematic of the allelic configurations of PTC mice (A) and the AAV library (B). Barcodes represent codons 2–31 of murine Kras engineered with synonymous mutations (purple, yellow, blue) and the KrasG12D mutation (red) flanked by ~2.1Kb homology arms. sgKras targets endogenous Kras; an additional U6-sgRNA serves as non-targeting control (sgNT) or Kdm8 targeting (sgKdm8). ITR, inverted terminal repeats. C Kaplan–Meier survival analysis of PTC mice receiving AAV-sgNT (PTCNT, n = 13 mice) or AAV-sgKdm8 (PTCKdm8KO, n = 10 mice). p = 0.16 by log rank test. D Representative light and fluorescent images of a PTCKdm8KO tumor. E Agarose gel electrophoresis of genomic DNA from control and PTCKdm8KO malignant cells using Surveyor nuclease assay. F Representative light and fluorescent images of the lungs from a control and PTCKdm8KO mouse. G Representative H&E staining of the livers from a PTCNT and PTCKdm8KO mouse. H Counts of lung (left, PTCNT, n = 13 mice; PTCKdm8KO, n = 10 mice) and liver (right, PTCNT, n = 13 mice; PTCKdm8KO, n = 9 mice) metastases in each PTC mouse of the indicated group. Boxes represent median values and 25th to 75th percentile with whiskers extending to 1.5x IQR. P-values were determined by two-sided Wilcoxon signed-rank test. Representative immunofluorescent images (top, I) with select regions (bottom, I) and the quantification of Kdm8 and Galectin-4 in 168,567 Tomato+ segments (J). Dotted line demarcates differentiated (Galectin-4 + ) region (bottom, I). The result is representative of tumors from n = 3 PTCKdm8KO mice. K Representative H&E and immunohistochemistry (IHC) of Kdm8, and Vimentin of serial sections from a liver (top, K) and a lung (bottom, K) metastasis. L Zoomed-in images of the selected region of the liver metastasis highlighted by the red rectangle in K. Representative H&E and IHC of Kdm8, Vimentin, and Tomato of serial sections from a primary PDA region (M) and the quantification of Kdm8 and Vimentin in 1415 and 3372 Tomato+ regions from 4 PTCNT (top) and 4 PTCKdm8KO (bottom) mice, respectively (N). The Spearman correlation efficient ρ is shown. Dotted lines demarcate metastases. The experiments were repeated at least three times with similar results (D, n = 12; E n = 3; F n = 12; G n = 6; K n = 3; L n = 3; M n = 3). CRISPR-mediated Kdm8 targeting induces a profound loss of differentiation and widespread metastatic lesions in autochthonous PDA Compared to the PTCNT cohort, PTCKdm8KO mice developed dramatically higher metastatic disease burden, including numerous micrometastases in the lungs (Fig. 2F–H and S2A-S2C). Of note, no mice in the PTCNT cohort developed any histologically discernable liver metastasis, consistent with previous reports that KrasG12D;Trp53-null PDA tumors are rarely metastatic (Figs. 2G, H, and S2A)59,60. Conversely, two-third of the PTCKdm8KO mice had macroscopically evident metastases in the liver (PTCNT, n = 0/13 vs PTCKdm8KO, n = 6/9, p = 0.001 by two-sided Fisher’s exact test). To gain further insights into the mechanism underlying this profound difference in metastasis, we examined the histological features of the primary PDA using Hematoxylin and Eosin staining (H&E), immunohistochemistry (IHC), and immunofluorescence (IF). While the PTCNT tumors exhibited homogeneous Kdm8 staining across histologically distinct regions, Kdm8 expression mostly restricted to fully differentiated glandular areas in the PTCKdm8KO PDA (Fig. S2D). Indeed, quantitative analysis revealed a strong positive correlation between the classical subtype marker Galectin-4 and Kdm8 (Figs. 2I, J, and S2E). This suggests that Kdm8 is required to maintain fully differentiated morphology in PDA, prompting us to test the hypothesis that the Kdm8-negative, less differentiated cancer cells might have transitioned into a mesenchymal state through EMT. To do this, we performed IHC across multiple primary and metastatic PTCKdm8KO tumors. Indeed, PTCKdm8KO metastases often lost Kdm8 expression while upregulating the EMT markers Vimentin and Zeb1 (Figs. 2K, L, M, S2F, and S2G). Furthermore, quantification across thousands of histologic regions in primary tumors revealed a striking anti-correlation between Vimentin and Kdm8 IHC (Fig. 2N). These results document a general enrichment of Kdm8-depleted malignant cells in the metastatic tumors of PTCKdm8KO mice, suggesting enhanced metastatic fitness upon Kdm8 loss. In summary, Kdm8 restrains phenotypic reprogramming by limiting the loss of differentiation and subsequently EMT, thereby potently suppressing metastatic disease. The demethylase function of Kdm8 is critical for hypoxia-induced transcriptomic changes and metastatic progression To examine the comprehensive gene expression changes induced by Kdm8 depletion in vitro, we comparatively profiled the full transcriptomics of control and shKdm8 688 M cells cultured in normoxia (ambient O2) and hypoxia (0.5% O2) by performing bulk RNA-Seq analysis. Consistent with the IHC results, Kdm8 knockdown in normoxia led to an upregulation of several hundred genes significantly enriched in EMT (Fig. 3A and S3A). In contrast, Kdm8 knockdown significantly decreased the classical-like signature38, aligning with the loss of the differentiated state61,62 (Figs. 2I and J, 3A). Of note, Kdm8 or KDM8 knockdown in murine and human PDA cell lines, respectively, reduced the expression of the rate-limiting enzyme Nsdhl or NSDHL for cholesterol biosynthesis63, which may explain the loss of glandular morphology in Kdm8-depleted PDA cells (Fig. S3B). Interestingly, Kdm8 knockdown in normoxia largely recapitulated the hypoxia-induced gene expression program, suggesting that a significant aspect of the hypoxic response in PDA cells is likely mediated through the suppression of Kdm8 function (Figs. 3A, B, and S3C). Consistent with this notion, Kdm8 knockdown in hypoxia failed to induce any Geneset Enrichment Analysis (GSEA) Hallmark transcriptomic programs, presumably due to the already diminished Kdm8 demethylase activity under hypoxic conditions (Figs. 3A and S3D). Kdm8 demethylase function is critical to hypoxia-induced transcriptomic reprogramming and metastatic progression. A, B GSEA of the classical, basal, and 688 M hypoxia gene signature (B), and hallmark signatures enriched in Kdm8-knockdown 688 M cells under normoxia (left) or hypoxia (right) (A). NES and FDR-adjusted p-values are shown; dotted line represents adjusted p = 0.05. C Densitometric quantification of normalized H3K36me2 immunoblots in Kdm8 knockout (Kdm8KO) 688 M cells re-expressing GFP (control) or indicated Kdm8 variants under normoxia (left) or hypoxia (right) (n = 3 technical replicates). D GSEA of the indicated hallmark signatures enriched in hypoxic vs. normoxic Kdm8KO 688 M cells expressing Kdm8WT, Kdm8H319A, Kdm8W308Y, or Kdm8W412N. NES and FDR-adjusted q-values are shown. Lung metastasis counts in NSG mice orthotopically transplanted with Kdm8KO 606 T (E, GFP, n = 19 mice; Kdm8WT, n = 10 mice; Kdm8H319A, n = 10 mice; Kdm8W308Y, n = 10 mice; Kdm8W412N, n = 10 mice) or 688 M (F GFP, n = 12 mice; Kdm8WT, n = 8 mice; Kdm8H319A, n = 8 mice; Kdm8W308Y, n = 9 mice; Kdm8W412N, n = 8 mice) cells expressing GFP (controls) or specified Kdm8 variants. P-values were determined by Kruskal–Wallis test (E, p = 5.6×10-7) and one-way ANOVA (F, p = 6.1 × 10–11), followed by Dunn’s and Dunnett’s multiple comparisons tests, respectively. GSEA of KDM8-induced (KDM8-ind), KDM8-suppressed (KDM8-sup), and hallmark signatures in basal vs. classical bulk tumors from TCGA (G) and PanCuRx (H) datasets. I Pearson correlation (ρ) between KDM8-induced gene signature score and classical subtype score across indicated PDA bulk tumor datasets. The solid line represents the linear regression fit and the shaded area denotes the 95% confidence interval. J Kaplan–Meier survival analysis of PDA patients from the PanCuRx dataset stratified by high (hi, >2nd IQR; n = 97 patients) vs. low (lo, n = 98 patients) KDM8-induced signature scores (p = 3.6 × 10-4 by log rank test). K Univariate logistic regression assessing associations between liver metastasis at diagnosis and indicated clinical or molecular variables in the PanCuRx cohort (n = 195 patient samples). Error bars indicate 95% confidence interval for the odds ratios; dotted line represents odds ratio=1. Boxes represent median values and 25th to 75th percentile with whiskers extending to 1.5x IQR (C, E, F). FDR-adjusted p-values for all GSEA were derived using empirical permutation tests using default setting (A, B, D, G, H). To show that Kdm8 activity is suppressed by hypoxia in vivo, we established orthotopically transplanted PDA tumors using a hypoxia-responsive elements (HRE)-driven UnaG64 hypoxia reporter PDA cell line derived from 688 M cells. This strategy enabled the isolation of pure Tomato-positive malignant cells stratified as UnaGneg (normoxic), UnaGdim (moderately hypoxic), and UnaGhigh (highly hypoxic) populations (Fig. S3E). To assess the level of Kdm8 functional suppression in these cells, we nominated 6 hypoxia-induced, HIF-independent Kdm8-suppressed genes as a surrogate readout of Kdm8 enzymatic activity. First, we generated Arnt (HIF1β)-knockout PDA cell clones and profiled their full transcriptomes under normoxic and hypoxic conditions through bulk RNA-Seq analysis. In contrast to Kdm8 depletion, Arnt loss abolished hypoxia-induced glycolytic gene upregulation, while only a small subset of Kdm8-suppressed genes lost hypoxia inducibility (Fig. S3F–S3J). Based on these analyses, we prioritized 30 hypoxia-induced Kdm8-suppressed genes whose hypoxic induction is independent of HIF signaling (Fig. S3J). From these, we selected 6 genes whose expression is sufficiently abundant for qPCR-based assessment of Kdm8 functional suppression in vivo. Notably, UnaGhigh malignant cells consistently exhibited the most pronounced induction of the indicated HIF-independent, Kdm8-suppressed genes relative to UnaGneg cells. Furthermore, UnaGdim cells displayed intermediate levels of upregulation, suggesting a transitional hypoxic state characterized by the activation of all tested Kdm8 targets (Fig. S3K). Together, these results support our model that hypoxia-mediated suppression of Kdm8 activity in vivo drives extensive transcriptomic reprogramming, with the magnitude of suppression directly reflecting the gradient of hypoxia exposure. To further test this hypothesis, we created Kdm8 hypermorphs by overlaying the JmjC domains of human KDM8 with two other KDM family members known for their high O2 binding affinities (KDM4A and KDM6B)65–67. We identified four non-conserved residues, Y177/1379 (KDM4A/KDM6B) -> W310 (KDM8), E190/1392 (KDM4A/KDM6B) -> D323 (KDM8), S196/1398 (KDM4A/KDM6B) -> N327 (KDM8), and N290/1484 (KDM4A/KDM6B) -> W414 (KDM8), which engage either α-ketoglutarate or Fe+2 according to previous reports and may enhance KDM8 demethylase function upon replacement68,69 (Fig. S4A). Compared to the wildtype counterpart (Kdm8WT), which is functional only in normoxia, expression of Kdm8W308Y or Kdm8W412N (murine orthologs to human KDM8W310Y and KDM8W414N, respectively) in the Kdm8 knockout 688 M cells reduced H3K36me2 level in both normoxia and hypoxia, indicating heightened demethylase activities (Figs. 3C, S4B, and S4C). In support of this, Kdm8-deficient 688 M cells expressing either Kdm8 hypermorph demonstrated reduced sensitivity to hypoxia-induced cell cycle arrest compared to the Kdm8WT-expressing counterparts (Fig. S4D). Additionally, expression of the hypermorphic Kdm8 variants abolished hypoxia-induced changes in multiple transcriptional programs, including the upregulation of EMT and downregulation of cell cycle (Fig. 3D). In contrast, hypoxia-induced suppression of the classical-like program and fatty acid metabolism was sensitive to the expression of the demethylase inactive variant Kdm8H319A (Fig. 3D). Consistent with the major role of the HIF pathway in facilitating anaerobic glycolysis, hypoxia-induced glycolysis appeared to be unaltered by changes in Kdm8 activity (Fig. 3D)70. We next examined the changes upon expressing these Kdm8 variants in the context of metastasis. To see whether the hypermorphic Kdm8 variants have an even stronger anti-metastatic effect, we performed orthotopic tumor studies using Kdm8 knockout 688 M and 606 T PDA cells. In both models, expression of Kdm8H319A did not alter the level of lung metastasis contrasted by the Kdm8WT counterpart, supporting the role of the Kdm8 JmjC domain in restoring a low metastatic state (Figs. 3E, F, S4E, and S4F). Notably, expression of the hypermorphic Kdm8W308Y variant led to significantly reduced lung metastatic burdens compared to the Kdm8WT counterparts in both models (Fig. 3E and F). In summary, these observations suggest that hypoxia promotes metastasis through the suppression of Kdm8, and that Kdm8 activity is a major rheostat that controls PDA metastatic ability. KDM8-induced gene signature predicts survival and metastatic disease in human PDA Our results establish that Kdm8 demethylase function suppresses metastasis and maintains a differentiated state in murine PDA. We next sought to determine whether the KDM8-induced gene signature predicts survival and disease behavior in human PDA. KDM8 is relatively homogeneously expressed, thus we performed correlative analyses using KDM8-regulated gene signatures. Comparative profiling of bulk RNA-Seq data from The Cancer Genome Atlas (TCGA)71 and the Ontario Institute for Cancer Research (PanCuRx Translational Research Initiative)38 reveals a consistent enrichment of the KDM8-induced gene signature in the classical subtype (Fig. 3G and H). Accordingly, classical-like and KDM8-induced signature scores were well correlated across four independent PDA cohorts72,73 (Fig. 3I). Notably, despite minimal overall gene overlap between these signatures (Fig. S4G), the overlapping genes alone preserved a strong correlation with the classical subtype score (Pearson correlation coefficient ρ: PanCuRx, 0.59; ICGC, 0.90; TCGA, 0.82; Dhara et al., 0.71), indicating that this association is driven by a limited subset of key classical-like genes likely maintained by KDM8 function. Using single nuclei transcriptomes isolated from a separate PDA cohort74, we found a striking correlation between the classical-like and KDM8-induced gene signatures in malignant cells, thus eliminating the possibility that the correlations in the bulk RNA-Seq datasets were driven by non-malignant cell types in tumors (Fig. S4H and S4I). In the PanCuRx cohort, the KDM8-induced gene signature is among the top predictors for prognosis, with higher scores correlating with better outcomes (Fig. 3J). This prognostic value was consistently observed across independent PDA patient cohorts (Fig. S4J and S4K). Importantly, patients exhibiting a high KDM8-induced gene signature or a classical-like score showed significantly reduced odds of liver metastasis at diagnosis (Fig. 3K). As anticipated, the hypoxia signature is a potent risk factor for liver metastasis4 (Fig. 3K). These results suggest that in human PDA, KDM8 function maintains the classical program and is associated with reduced metastasis prior to therapeutic intervention. Malignant state plasticity is regulated by Kdm8 in vivo In PDA, transcriptomic programs define distinct disease subtypes that are clinically relevant27,38,75,76. The basal-like subtype (a.k.a. squamous-like or quasi-mesenchymal) is associated with advanced stage38, resistance to treatment74,77–79, and serves as an independent predictor for poor prognosis in multiple studies4,27,75,76,80. More recently, the neural-like progenitor malignant cell program has been identified as one of the most resistant subtypes in chemotherapy and radiotherapy, correlating with poor outcomes in independent cohorts74. Despite this, the role of the hypoxia-Kdm8 axis in the plasticity of malignant programs remains undefined. To investigate the transcriptomic state plasticity in the PTCKdm8KO mice, we performed single cell RNA-Seq (scRNA-Seq) on ex vivo malignant cells isolated using fluorescence-activated cell sorting (FACS) for Tomato-positive and lineage marker (Cd45, Ter119, F4/80, Cd31) negative population (Figs. 4A and S5A)11. The scRNA-Seq dataset comprises 32,878 cells with a median of 17,107 unique molecular identifiers (UMIs) and 4061 genes detected across 11 tumors and metastatic lesions (5 primary PDA tumors, 3 disseminated cancer cell samples from the ascites, and 3 lung metastases). Utilizing previously reported signatures, we uncovered 9 malignant programs that closely resembled human PDA, including classical-like, EMT, classical/basal hybrid, basal/hypoxia hybrid, hypoxia, Myc, neuroendocrine, and transitional states (Figs. 4B and S5B)74. Furthermore, by employing morphologically associated gene signatures that define distinct PDA histological patterns61, we confirmed that the classical, classical/basal hybrid, and transitional states predominantly resembled morpho-biotype A (ductal or glandular-like structure), while the EMT state aligned with morpho-biotype B (ill-defined glands, Fig. S5C). Malignant cell state plasticity is regulated by Kdm8 in vivo. A Schematic of the single-cell sequencing experiment using dissociated malignant cells from the PTCKdm8KO mice. UMAP embeddings of single-cell profiles of all Tomato+ malignant cells derived from 11 tumors and metastatic lesions colored by post hoc cell cluster annotations using Seurat 4.1.3 (B), indicated malignant transcriptomic programs (C), and latent time (D) or pseudotime trajectory (E). F Distributions of the identified malignant programs as in (B) across the 11 tumors and metastatic lesions. Stacked bar plots showing the proportions (Y-axis) of malignant programs across samples isolated from 5 PTCKdm8KO mice. Fisher’s exact test p-values for the association between the advanced malignant state (EMT, neuroendocrine, and basal/hypoxia) and metastasis are shown (top). *, association between the malignant state and ascites is shown. G Log KrasG12D and KrasWT transcript abundance in Tomato+ malignant cells isolated from PTCKdm8KO tumors (n = 7194 cells) that are colored according to their KrasG12D/ KrasWT allelic conformation. Differentially expressed genes (H) and GSEA for the hallmark gene signatures enriched in malignant cells with KrasG12D/ KrasWT allelic imbalance compared to counterparts without the allelic imbalance (I). Somatic cells shown in Fig. G were excluded from this analysis. ES, enrichment score. J Increase in Kdm8-suppressed gene signature score along the latent time trajectory defined in (D). Cells are colored by the Kdm8 status using Kdm8-Seq. Kdm8-sup sig, Kdm8-suppressed gene signature. GSEA of the hallmark gene signatures independently identified transcriptomic programs, including EMT and hypoxia (L), that are enriched in Kdm8 negative (Kdm8neg) malignant cells (K). M Differentially expressed genes identified by comparing Kdm8neg over Kdm8pos cancer cells. N Abundance of select malignant programs (Hwang et al. 2022 Nat Genet), GSEA hallmark gene signatures, Kras expression, and Kdm8-suppressed gene signature in Kdm8 positive (n = 797 cells) and negative cancer cells (n = 6202 cells). P-values were derived using two-sided Wilcoxon signed-rank test (H, M, N). Boxes represent median values and 25th to 75th percentile with whiskers extending to 1.5x IQR. FDR-adjusted p-values for all GSEA were derived using empirical permutation tests using default setting (I, K, L). Consistent with human PDA, a striking positive correlation between the states defined by KDM8-induced gene signature and the classical-like signature was observed (Spearman correlation coefficient ρ = 0.73, Figs. 4C and S5D). Surprisingly, the expression of KDM8-suppressed genes showed a high correlation with the neuro-like progenitor signature74, even though the two gene lists minimally overlapped, implicating the role of hypoxia in suppressing Kdm8 and the subsequent induction of this aggressive malignant cell program (Spearman correlation coefficient ρ = 0.67, Figs. 4C, S5D, and S4G). Of note, a low Myc score, which corresponds to a low proliferative state, coincided with a high abundance of genes suppressed by Kdm8, suggesting a cell cycle-antagonizing effect following Kdm8 inactivation (Fig. S5D). We further validated this by demonstrating reduced cell proliferation upon Kdm8 knockdown in multiple PDA cell lines in vitro as well as in vivo (Figs. S5E–S5H). Further, we examined differentiation trajectories across the malignant cell states by performing RNA velocity analysis to infer latent time and differentiation directionality (Fig. 4D and E)81,82. Consistent with previous findings6,83, our data revealed a transcriptomic continuum spanning Myc, classical-like, and classical/basal hybrid states through to hypoxia and transitional states, and ultimately the basal/hypoxia hybrid, neuroendocrine, and EMT programs (Figs. 4D, E, and S5I). Along the latent time trajectory, EMT, neuro-like progenitor, and Kdm8-suppressed gene expression programs increased, whereas classical-like and Myc gene signatures declined (Fig. S5I). As anticipated, disseminated and metastatic lesions show significant enrichment of the basal/hypoxia hybrid, neuroendocrine, and EMT programs compared to their autologous primary PDA (Fig. 4F). In addition to scRNA-Seq for cell state profiling, the full-length 10X cDNA libraries were subjected to paired analysis measuring the abundance of the barcoded KrasG12D alleles (KrasBC-Seq, Fig. 4A and S6A). This paired measurement of barcoded KrasG12D and wildtype Kras alleles narrowed down the original scRNA-Seq dataset to 7,194 cells with at least one detected read count. As expected, the overall Kras transcript abundance quantified with KrasBC-Seq correlated with those in scRNA-Seq, albeit with much higher resolution (left, Fig. S6B). Consistent with a previous report58, we found that a vast majority of the metastatic lesions derived from the largest primary PDA clonotypes. Notably, multiple malignant programs - including the classical-like, Myc, hypoxia, and basal/hypoxia states - were detected within individual KrasG12D clonotypes, indicating substantial transcriptional heterogeneity within a single clonal lineage (Fig. S6C). Given that KRAS allelic imbalance correlates with PDA progression and is linked to advanced PDA subtypes and metastasis38,39,84, we sought to examine the level of Kras allelic imbalance relative to distinct malignant cell states in the PTCKdm8KO mice. Leveraging the barcode strategy that enabled unambiguous calling of KrasG12D and wildtype Kras alleles, we clearly distinguished malignant cells with balanced versus imbalanced KrasG12D/KrasWT expression (Fig. 4G). A small fraction (n = 621/7194) of the cells exhibited higher KrasWT expression compared to the barcoded KrasG12D alleles, presumably comprised primarily nonmalignant cells in the current PDA model. Interestingly, comparative profiling using GSEA Hallmark signatures uncovered hypoxia and Kras signaling as the top enriched programs in malignant cells with Kras allelic imbalance (Fig. 4H and I). Furthermore, Kras allelic imbalance was consistently enriched in hypoxic cells across all PTCKdm8KO mice, indicating that hypoxia might exert strong selective pressure favoring increased KrasG12D/KrasWT ratios (Fig. S6D). To directly assess the impact of Kdm8 depletion on transcriptomic state plasticity, we performed paired Kdm8 targeted sequencing with the scRNA-Seq cDNA libraries (Kdm8-Seq, Fig. 4A). Consistent with the low Kdm8 expression by IHC observed in metastases (Fig. 2K and L), metastatic lesions yielded minimal PCR amplicon with poor library quality. The paired analysis successfully generated targeted Kdm8 sequencing libraries from a total of 5 primary PTCKdm8KO PDA tumors that also had scRNA-Seq and KrasBC-Seq datasets available (n = 6999 cells). As expected, we noted an exceedingly high correlation between Kdm8 measured with Kdm8-Seq and scRNA-Seq (right, Fig. S6B) and a bimodal distribution of Kdm8 abundance with 11.4% (n = 797/6999) of the cells expressing Kdm8, comparable to results obtained using the Surveyor nuclease assay (Fig. S6E). Strikingly, Kdm8pos cancer cells were non-uniformly distributed, with the majority localized near the early regions of the trajectory and progressively fewer cells observed toward later transcriptional states, suggesting that Kdm8 loss drove cancer cells into advanced transcriptomic states (Fig. 4J). Indeed, comparison of Kdm8neg versus Kdm8pos cells using GSEA hallmark signatures and differential gene analysis revealed significant enrichment of hypoxia, EMT, and Kras-upregulated genes, while E2F targets, G2M checkpoint, and Kras-downregulated gene signatures were depleted (Fig. 4K–N and S6F). These transcriptomic shifts align with the increased cell state entropy and advanced latent time following in vivo Kdm8 depletion (Fig. S6G–S6I). Although RNA velocity-based latent time inference alone does not permit definitive conclusions regarding transition directionality, integration of inferred Kdm8 status with trajectory relationships provides additional biological constraints. Given the irreversible CRISPR-mediated Kdm8 targeting, these observations support that state transitions in the current GEM model are biased toward a predominantly unidirectional continuum, with Myc and classical-like programs preferentially occupying earlier regions and advanced transcriptional states - including hypoxia-associated, transitional, basal/hypoxia hybrid, neuroendocrine, and EMT states - representing later stages along a continuous transcriptional landscape. Together, these results support an in vivo role for Kdm8 in suppressing advanced malignant programs and Kras signaling. Kdm8 depletion recapitulates hypoxia-induced epigenetic reprogramming To define the global epigenetic changes resulted from hypoxia-induced Kdm8 suppression, we performed chromatin immunoprecipitation sequencing (ChIP-Seq) to quantify various histone modifications, including H3K36me2 and H3K27 trimethylation (H3K27me3), which are likely impacted by Kdm8 inactivation based on previous findings49–51. We quantitatively measured H3K36me2 and H3K27me3 global changes in control and shKdm8 688 M cells cultured in normoxia and hypoxia. In the same experiment, we also performed H3K27 acetylation (H3K27ac) ChIP-Seq and assay for transposase-accessible chromatin with sequencing (ATAC-Seq) to simultaneously assess these chromatin features as potential secondary changes that might explain transcriptomic reprogramming following Kdm8 knockdown or hypoxia. Kdm8 knockdown induced a marked increase in H3K36me2 levels across 7952 and a decrease in 7140 genomic regions (Fig. 5A). Similarly, global changes in H3K27me3 were observed upon Kdm8 knockdown (Fig. 5A). Of note, short-term exposure to hypoxia resulted in substantial global changes in both H3K36me2 and H3K27me3 (Fig. S7A). In addition, the global changes in H3K27ac upon Kdm8 knockdown were only partially mirrored by hypoxia (Figs. 5A and S7A). Finally, Kdm8 depletion has a minimal impact on global chromatin accessibility, whereas hypoxia induced robust genome-wide changes (Figs. 5A and S7A). Kdm8 depletion recapitulates hypoxia-induced epigenetic reprogramming. A Kdm8 knockdown (Kdm8KD)-induced differential H3K36me2, H3K27me3, and H3K27ac ChIP-Seq signals, alongside ATAC-Seq chromatin accessibility (left to right), plotted against log mean read counts. Dotted lines represent zero log2 fold change (FC); numbers of significantly altered regions (FDR < 0.05) are shown. B Scatter plots showing genome-wide relationships between Kdm8KD-induced (x-axis) and hypoxia-induced (y-axis) FC in H3K36me2 (left) and H3K27me3 (right). Significantly altered regions are highlighted with quadrant counts and Pearson correlation coefficients (ρ) indicated. C Homer transcription factor (TF) motif enrichment analysis in Kdm8KD-induced (x-axis) and hypoxia-induced (y-axis) regions for H3K36me2 (left) and H3K27me3 (right). Highlighted TF motifs relate to high-altitude adaptation, stress response, and PDA progression. P-values were derived using one-sided Fisher’s exact tests with Benjamini-Hochberg correction for multiple comparisons. D Densitometry of immunoblots quantifying hypoxia-induced FC of H3K36me2/H3 (left) and H3K27me3/H3 (right) ratios in control (parental) and Kdm8 knockout (Kdm8-KO) 688 M cells (n = 5 technical replicates). Boxes represent median values and 25th to 75th percentile with whiskers extending to 1.5x IQR. E Principal component (PC) analysis of genome-wide variance in H3K36me2 across GFP, Kdm8WT, and Kdm8H319A groups along PC1 and PC2, with percentage variance indicated. F Scatter plots showing genome-wide relationships between FC in H3K36me2 (left) and H3K27me3 (right), comparing GFP vs Kdm8WT (x-axis) and Kdm8H319A vs Kdm8WT (y-axis). Highlighted regions include quadrant counts and Pearson correlation coefficients (ρ). G H3K36me2, H3K27me3, H3K27ac, and ATAC-Seq signals at the Lif locus in control and Kdm8KD 688 M cells under normoxia or hypoxia. H Contingency tables showing the number of genes co-induced by Kdm8KD and hypoxia that overlap with regions induced by either hypoxia or Kdm8KD (left), and genes co-suppressed by Kdm8KD and hypoxia that overlap with regions suppressed by either condition (right). Control (Ctrl) genes are defined by p > 0.9 for both conditions. P-values derived by two-sided Fisher’s exact test. I Scatter plot correlating Kdm8KD-induced H3K36me2 and H3K27me3 log₂ FC with Pearson correlation coefficient (ρ). Solid lines represent linear regression fits; shaded areas denote the 95% confidence intervals (B, F, I). To gain further insights, we examined the distribution of regions with changes in H3K36me2 and H3K27me3. Interestingly, hypoxia suppressed 27.4% of all H3K36me2 regions and 10.1% of all H3K27me3 regions that are at least 1 kb away from the nearest transcription start site (TSS) or gene distal, which was more than double the suppression observed in their gene-proximal counterparts (bottom, Fig. S7B; top, Fig. S7C). Similar patterns of the reduced H3K36me2 and H3K27me3 regions were observed following Kdm8 knockdown (top, Fig. S7B; top, Fig. S7C). Notably, suppression of these H3K36me2 gene-distal regions may be a secondary effect of Kdm8 knockdown and could be attributed to increased Kdm2a expression following Kdm8 depletion or hypoxia (Fig. S3A and S3C, Source Data), consistent with a prior report of Kdm2a’s demethylase activity at gene-distal H3K36me2 sites85. In contrast, an opposite pattern was observed for hypoxia-promoted H3K36me2 and H3K27me3 regions, as well as for the differential H3K27me3 regions induced by Kdm8 knockdown, with these regions being predominantly gene-proximal rather than gene-distal (Fig. S7B; bottom, Fig. S7C). Remarkably, a striking correlation between the fold changes of all regions induced by hypoxia and Kdm8 knockdown was noted, indicating that Kdm8 knockdown itself is sufficient to recapitulate most hypoxia-induced changes in these histone marks (Fig. 5B). Furthermore, assessment of enriched transcription factor motifs within H3K36me2 and H3K27me3 regions induced by hypoxia or Kdm8 knockdown revealed co-enrichment of multiple binding sites for factors critical in high-altitude adaptation, stress response, and PDA progression (Fig. 5C)13,86–88. Consistent with Kdm8’s function in hypoxia-induced epigenetic changes, Kdm8 depletion through CRISPR knockout completely and partially abolished hypoxia-induced accumulation of H3K36me2 and H3K27me3, respectively (Figs. 5D and S7D). We validated the requirement for Kdm8 JmjC domain activity in regulating H3K36me2 and H3K27me3 across the vast majority of genomic regions by reintroducing the Kdm8WT or enzymatically inactive Kdm8H319A variant into Kdm8-deficient cells. Principal component analysis of genome-wide ChIP-Seq signals revealed that restoration of Kdm8 JmjC domain activity re-established a global H3K36me2 chromatin state—and, to a lesser extent, an H3K27me3 chromatin state – that was markedly distinct from those observed in Kdm8-knockout or Kdm8H319A-expressing cells (Figs. 5E and S7E). Consistently, we observed substantial correlations between genome-wide fold changes in both histone marks comparing Kdm8-null (GFP-expressing) cells to Kdm8WT-rescued cells, and those comparing JmjC-inactive Kdm8H319A-expressing cells to Kdm8WT-rescued cells (Fig. 5F). Together, the impact of Kdm8 inactivation mirrors a large part of hypoxia-induced chromatin changes, strongly supporting the model of the hypoxia-Kdm8 axis in epigenetic reprogramming. To gain additional insights into the regulation of gene expression during the hypoxic response, we focused on differential gene proximal regions induced by Kdm8 depletion and hypoxia (Fig. S7B). We identified numerous genes, including Lif, where the expression changes induced by hypoxia or Kdm8 knockdown aligned with changes in H3K36me2 (Figs. S3A and 5G). Consistent with H3K36me2’s role in actively transcribed loci, both hypoxia and Kdm8 knockdown resulted in increased Lif expression, coinciding with elevated H3K36me2 signals at multiple peaks near the Lif locus (Figs. S3A and 5G). Overall, 86 of the 341 (25.2%) genes that were induced by both hypoxia and Kdm8 knockdown were associated with hypoxia- or Kdm8 knockdown-induced H3K36me2 proximal regions, whereas 41 of the 265 (15.5%) suppressed genes were linked to hypoxia- or Kdm8 knockdown-repressed H3K36me2 proximal regions (Fig. 5H). Comparative profiling using GSEA c2 curated gene sets uncovered an H3K27me3-regulated gene signature89 as the top depleted program upon Kdm8 knockdown in normoxia, suggesting that the hypoxia-Kdm8 axis might regulate the transcriptomes through coordinated epigenetic changes at multiple histone marks and genomic contexts beyond H3K36me2 (Fig. S7F). Consistent with this, we noted a coordinated increase in H3K27ac and a decrease in H3K27me3 within the H3K36me2 regions induced by both Kdm8 knockdown and hypoxia (n = 3973 regions). Conversely, the suppressed H3K36me2 regions (n = 3838) were associated with reduced H3K27ac and enhanced H3K27me3 signals (Fig. S7G and S7H). Overall, we observed a robust and genome-wide anti-correlation between Kdm8 knockdown-induced changes in H3K27me3 and H3K36me2, consistent with the opposing roles of these marks in transcriptional regulation (Fig. 5I). Collectively, these results support a mechanism by which the hypoxia-Kdm8 axis induces major transcriptomic reprogramming by coordinating global epigenetic changes. The hypoxia-Kdm8 axis promotes Kras copy-number gain through accelerated CIN CIN is a well-established driver of metastasis in multiple solid tumor models90. Given the critical role of jmjd-5 in preserving chromosome integrity in C. elegans55 and the association of JMJD5 germline mutations with DNA replication stress in humans54, we hypothesize that hypoxia promotes CIN by suppressing Kdm8 demethylase activity. To assess the role of Kdm8 in regulating CIN, we applied complementary assays: (i) WGS to measure focal CNA and aneuploidy, (ii) DAPI staining of genomic DNA to quantify micronuclei formation, and (iii) immunofluorescence to analyze mitotic spindle organization during anaphase. We first assessed micronuclei formation in PDA cells following KDM8 knockdown. In both murine and human PDA cell lines, knockdown of Kdm8 or KDM8, respectively, increased micronuclei under normoxic conditions (Fig. 6A). In 688 M PDA cells, Kdm8 knockdown alone recapitulated hypoxia-induced preponderance of micronuclei, indicating that hypoxia may promote CIN through Kdm8 suppression (Fig. 6A). Whole chromosome painting in 688 M cells revealed a tetraploid karyotype with increased double minutes derived from chromosome 6, whereas chromosome 11 remained unaffected, indicating that Kdm8 knockdown drives chromosome-specific structural alterations induced by the hypoxia-Kdm8 axis (Fig. 6B and C). Sporadic chromothripsis events were observed in Kdm8 knockdown cells, absent in controls, albeit without statistical significance. These results indicate that Kdm8 inactivation promotes CIN in PDA cells. The hypoxia-Kdm8 axis promotes Kras copy-number gain through accelerated chromosomal instability. A Quantification of micronuclei formation, shown as normalized ratios of micronuclei counts per nucleus per field of view (FOV; number of FOVs indicated), in control vs. shKDM8-expressing human (Capan-1 and MIA PaCa-2) PDA cells cultured in normoxia, and murine 688 M PDA cells cultured in normoxia or hypoxia. Boxes represent median values and 25th to 75th percentile with whiskers extending to 1.5x IQR. Representative images (B) and quantifications (C) of whole chromosome 6 and 11 painting in control and shKdm8-expressing 688 M cells. The proportions of metaphase cells in control and the shKdm8 group associated with the indicated numbers of chromosome 6- and 11-derived double minutes are shown, with p-values calculated using one-sided Wilcoxon signed-rank test on double minute tallies (C). Representative images of anaphase cells in normoxia (D) and the proportions of control and Kdm8 deficient 688 M PDA cells cultured in normoxia or hypoxia with the indicated anaphase polarities quantified using γ-Tubulin immunofluorescence (IF, E). F Proportions of KDM8 knockout Capan-1 cultured in normoxia (left) or hypoxia (right) with anaphase polarities quantified by Pericentrin IF expressing the indicated KDM8 variants. Representative images (G) and quantifications (H) of fluorescent in-situ hybridization (FISH) using probes for Kras (red), chromosome 6 centromeric (green) and distal (cyan) regions, and chromosome 18 (yellow) in control (normoxia, n = 49 metaphases; hypoxia, n = 50 metaphases) and shKdm8-expressing 688 M cells cultured in normoxia (shKdm8-1, n = 53 metaphases; shKdm8-2, n = 51 metaphases) or hypoxia (shKdm8-1, n = 51 metaphases; shKdm8-2, n = 50 metaphases). The proportions of control and the shKdm8 metaphase 688 M cells cultured in normoxia or hypoxia with the indicated FISH probe tallies are shown (H). Numbers in the plot indicate the number of cells quantified in each group (E, F, H); n = 2 (B–E, G, H) and 3 (F) independent biological repeats. Unless otherwise stated, p-values were derived using two-sided Wilcoxon test for comparisons of fluorescent quantification between two groups or a Kruskal–Wallis test for comparisons among three or more groups, followed by Dunn’s test for sub-comparisons with Benjamini-Hochberg adjustment. Next, to determine whether hypoxia compromises mitosis through the suppression of Kdm8 demethylase activity, we quantified spindle organization defects by immunofluorescence. Consistent with previous reports56,91, Kdm8 knockdown in both 688 M and 606 T murine PDA cell lines increased mitotic abnormalities, specifically multipolar spindle formation, a known cause of chromosome mis-segregation92 (Figs. 6D, E, and S8A). Similar spindle defects were observed in human Capan-1 cells and non-malignant MCF10A cells upon KDM8 depletion (Fig. S8A). Hypoxia promoted multipolar spindles during anaphase in control 688 M cells, while Kdm8 knockdown under normoxia produced the same phenotype, concordant with the micronuclei observations (Fig. 6E). Importantly, expression of KDM8WT, but not the demethylase-inactive KDM8H321A variant, in KDM8 knockout Capan-1 cells restored spindle organization in normoxia (Fig. 6F). Similarly, expression of the demethylase-inactive Kdm8H319A variant in the murine PDA 688 M cells had minimal effect on spindle polarity (Fig. S8B). In hypoxia, expression of the hypermorphic KDM8W310Y variant in human PDA cells or the corresponding Kdm8W308Y variant in murine PDA cells—but not their WT counterparts—significantly reduced spindle defects, demonstrating that KDM8 demethylase activity is essential for maintaining mitotic fidelity and limiting CIN (Figs. 6F and S8B). Notably, the mitotic errors and subsequent CIN induced by the hypoxia-Kdm8 axis appeared independent of a disrupted DNA damage response or alterations in H3K36me2 levels (Figs. S9A–S9F and S10A–S10F). Instead, we found that the exogenous expression of the non-motile microtubule depolymerases Kif2b or Kif2c90—factors known to promote kinetochore-microtubule turnover—significantly rescued the multipolar spindle phenotype triggered by Kdm8 depletion and attenuated the highly metastatic phenotype (Fig. S11A and S11B). These results identify the regulation of microtubule-kinetochore attachment dynamics as a potential mechanism through which Kdm8 restrains CIN in PDA. Given the increased mitotic errors, we performed bulk and single-cell WGS (scWGS) to further examine karyotypic heterogeneity upon Kdm8 knockdown. Interestingly, Kdm8 depletion had distinct impacts on the ploidies of different chromosomes, suggesting that CNAs in select chromosomes may serve as preferred genomic features for in vitro growth (Fig. S12A and S12B). Notably, chromosome 6, which contains Kras, was recurrently amplified following Kdm8 knockdown using different hairpins (Fig. S12B and S12C). In line with these observations and the increased Kras signaling in PTCKdm8KO cancer cells (Fig. 4K, N, and S6F), we hypothesize that hypoxia promotes Kras amplification through the suppression of Kdm8 demethylase function. In patients with advanced PDA, the KDM8-induced gene signature is associated with reduced KRAS allelic imbalance, a feature largely driven by oncogenic KRAS amplification39 (Fig. S12D). To test this hypothesis, we assessed the impact of Kdm8 knockdown or hypoxia on Kras CNAs using fluorescent in situ hybridization (FISH). Consistent with the WGS data, Kdm8 knockdown with two independent hairpins resulted in significant chromosome 6 copy-number gains, relative to chromosome 6 copies quantified using a centromeric probe (Fig. 6G and H). In contrast, reference chromosome 18 exhibited minimal ploidy changes following Kdm8 knockdown (Fig. S12A and S12E). Hypoxia also increased chromosome 6 ploidy in control 688 M cells, though to a lesser degree. Both hypoxia and Kdm8 knockdown increased Kras copy number along with a distal reference region on chromosome 6, consistent with chromosome 6 polyploidy. These data indicate that Kras copy-number gain arises through chromosome 6 amplification (Fig. 6G and H). In agreement with this observation, Kdm8 and KDM8 knockdown in multiple murine and human PDA cell lines significantly increased sensitivity to the KRASG12D inhibitor MRTX1133 (Fig. S12F and S12G). In summary, the hypoxia-Kdm8 axis accelerates CIN, leading to chromosome 6 polyploidy and subsequent Kras amplification. Discussion This work reveals a mechanism by which PDA malignant cells co-opt select elements of the hypoxic response that drive metastatic spread. Our data suggest that by suppressing KDM8 demethylase function, hypoxia unleashes an adaptive program characterized by reduced cell division, increased CIN, loss of differentiation, and enhanced EMT. As a hallmark of cancer93, CIN, enabled by ongoing mitotic errors, generates a pool of malignant cells with diverse genomic and phenotypic traits – some of which are readily selected by and survive environmental constraint. In the context of hypoxia, this model explains the adaptability of genomically unstable malignant cells to cope with the metabolic challenges within the tumor microenvironment. In support of this, our results show that Kdm8 suppression in cancer cells induces the acquisition of higher Kras copy numbers, a heritable trait that facilitates the removal of hypoxia-induced reactive oxygen species through increased glutathione biosynthesis94 and is preserved in metastases38. Furthermore, hypoxia-mediated Kdm8 suppression promotes a metastatic program by facilitating EMT and loss of differentiation through (i) H3K36me2 upregulation and (ii) induction of CIN, both of which have been previously reported to reinforce an EMT-like phenotype85,90. Restoration of the hypoxia-resistant, hypermorphic Kdm8 variant almost completely blocked metastasis and could provide a viable therapeutic strategy to reverse the highly metastatic behavior in PDA. Consistent with previous reports49,50, Kdm8 depletion reduced the proliferation of PDA cells. In the PTC mice, Kdm8 knockout in PDA cancer cells diminished the Myc gene signature, which was enriched near the root of the latent time trajectory. In human PDA, the MYC transcriptomic state shows strong negative correlations with the KDM8-suppressed gene signature, EMT, and hypoxic response (Fig. S5D). This finding aligns with an earlier report analyzing bulk RNA-Seq data from 500 patients across more than two dozen carcinoma types, which revealed a strong negative correlation between EMT and proliferation signatures95. Under normal circumstances, cells undergoing morphogenesis (e.g., EMT) are evolutionarily hardwired for cell cycle arrest to prevent severe developmental defects caused by mitotic checkpoint failure96,97. Despite this, MYC is a PDA oncogene98 that is amplified in metastases99,100, which appears paradoxical given that a suppressed MYC gene signature is part of the hypoxia-induced metastatic reprogramming. As a hallmark of cancer93, unrestrained cell proliferation induces genomic instability in epithelial cancer cells undergoing EMT101. Consistent with this, a mesenchymal state defined as Vimentin-positive is strictly required for CIN in PDA, irrespective of the genomic drivers used in the GEM models21. Thus, in the context of hypoxia-induced EMT, it is conceivable that increased MYC dosage would be a highly favorable trait for promoting CIN and metastasis102,103. Although beyond the scope of the study, it would be of interest to elucidate a potential mechanism by which EMT, regardless of its induction by hypoxia or other factors, fuels MYC amplification in promoting PDA metastasis. High prevalence of KRAS and TRP53 genetic mutations in human PDA underscores the development of PDA GEM models104. Due to their similarity to human disease, KPC mice continue to play a pivotal role in advancing our understanding of PDA biology. However, these GEM models are labor-intensive and highly variable. To address these challenges, we pioneered a CRISPR/Cas9-based approach that eliminates the need to create a floxed Kdm8 allele while allowing head-to-head comparisons between Kdm8-sufficient and knockout malignant cells within the same mouse. Thus, several unique conclusions can be drawn from the features of our GEM model. First, our data confirm that cellular plasticity is likely shaped by extrinsic factors, such as the hypoxia-Kdm8 axis. Second, hypoxia appears to be a strong predictor of Kras imbalance, which can be attributed to the loss of the KrasWT allele and/or the amplification of the oncogenic KrasG12D allele. Similar to human PDA84, our FISH analysis supports arm-level Kras copy number gain as a primary cause of the allelic imbalance observed in this GEM model. Third, stochastic depletion of Kdm8 in malignant cells within a tumor enabled major transcriptomic and histological changes that are associated with Kdm8 loss. These changes include an almost invariant loss of differentiation and Vimentin expression upon Kdm8 ablation in malignant cells juxtaposed with Kdm8-expressing neighbor cells, a feature simply not available in conventional KPC mice. The understanding of how epigenetic dysregulation contribute to CIN in cancer remains incomplete. Nevertheless, accumulating evidence has established histone demethylases as key regulators of genomic instability, particularly under hypoxic conditions. Our findings provide a crucial piece to this emerging framework, demonstrating that hypoxia promotes CIN through the regulation of epigenetic modifiers. Previously, Whetstine and colleagues showed that hypoxia induces transient copy-number gains at specific genomic sites by stabilizing KDM4A37. Mechanistically, hypoxia facilitates the recruitment of KDM4A to H3K4-methylated genomic sites, including EGFR, by modulating multiple histone lysine methyltransferases and demethylases105. In this context, hypoxia promotes locus-specific, transient copy-number gains by enhancing KDM4A demethylase activity, thereby driving re-replication of the amplified regions during S phase106. By comparison, our data support a distinct model in which hypoxia-induced suppression of KDM8 demethylase activity promotes global epigenetic re-wiring, CIN, and nontransient polyploidy affecting select chromosomes. However, the mechanisms through which Kdm8 deficiency leads to unique mitotic defects, such as multipolar spindle formation, remain poorly understood. One possibility is that dysregulated H3K36me2 near centromeres results in reduced chromatin compaction during mitosis, a process previously implicated in genomic instability107. Alternatively, Kdm8 may regulate mitotic fidelity through non-histone substrate(s), such as tubulin, where post-translational modifications including methylation are known to play crucial roles108. Given the complexity of cellular adaptations to hypoxia, it remains to be determined whether additional O2-sensing epigenetic regulators contribute to CIN. Methods All in vivo experiments were performed in accordance with the Rutgers University Institutional Animal Care and Use Committee (IACUC) guidelines (PROTO201900174). For all tumor studies, the humane endpoint is defined as tumors interfering with normal activity, reaching a maximum diameter of 1.75 cm, exceeding 10% of the animal’s pre-inoculation body weight, or surpassing a total subcutaneous tumor burden of 1 cm3. In all tumor studies, maximum tumor burden allowed was not exceeded. All recombinant DNA, lentiviral/retroviral work, and biohazardous materials were handled in strict accordance with biosafety guidelines approved by the Rutgers Institutional Biosafety Committee. Autochthonous tumor model The PTC mice (Trp53flox/flox;R26LSL-Tomato;H11LSL-Cas9, courtesy of Monte M. Winslow)58 were outbred to the C57BL/6 J background. To initiate PDA in the PTC mice, retrograde pancreatic ductal injection of engineered AAV was performed as reported previously57. Briefly, mice were anesthetized prior to laparotomy and bowel displacement. The duodenum and a portion of the pancreas were gently repositioned to visualize the ampulla of Vater - the injection site where the common bile duct enters the duodenum. A microclip (Micro Vascular Clip RS-6472, Roboz) was applied to the common bile duct near the gallbladder base to increase intraductal pressure. A 30-gauge needle was then inserted at a shallow angle into the bile duct, while a serrated curved forceps stabilized the duodenum. Under maintained pressure, 50–100 µL of engineered AAV was slowly infused into the pancreas over 2–3 min. Following injection, the pancreas and intestines were returned to the abdominal cavity. The abdominal muscle was sutured, and the skin was closed using surgical staples. Postoperative analgesia was provided using Ethiqa XR (Fidelis Animal Health). Humane endpoints criteria included signs of pain, distress, or moribund status due to PDA tumor burden. Mice that survived beyond one year post-surgery without evidence of PDA development were excluded from the study. Transplanted tumor models PDA cells were washed three times in cold PBS before resuspension in 100% Matrigel (356231, Corning) and kept on ice. For subcutaneous tumor studies, 5 × 105 cells were injected into the dorsal flank of the NSG mice. For orthotopic pancreatic tumor studies, cells were resuspended in 100% Matrigel at a concentration of 4 × 106 cells/mL prior to the procedure. Briefly, mice were anesthetized before a small incision was made to expose the pancreas tail and the spleen. The resuspended PDA cells were injected into the pancreas tail to form a blister (~25 μL) and allowed to reach body temperature and solidify. The pancreas was then placed back into the abdominal cavity before suturing the abdominal muscle and closing the incised skin with surgical staples. For all in vivo studies, mice were sacrificed upon reaching the endpoint tumor volume (~1 cm3) or morbidity and dissected for fluorescence microscope imaging and histology. Metastases were quantified using FIJI macro. Inferred relative metastatic tumor size for each individual mouse was defined as the 90th percentile of all detected Tomato-defined areas, following application of a threshold to exclude background signal. Both male and female NOD/SCID/γc (NSG, 005557, Jackson Lab) mice were used. In the orthotopic tumor studies using the 703 T and 606 T cell lines, male C57BL/6 J (000664, Jackson Lab) mice were used to match the PDA cell background. Immunoblot analysis Cell lysates were prepared using 2x Laemmli buffer (histone 3) or RIPA buffer supplemented with protease inhibitors (P8340, Millipore Sigma). PAGE separated proteins and subsequently transferred onto an Immobilon PVDF membrane (1620177, Millipore Sigma). Membranes were incubated with primary antibodies (Kdm8: NBP1-77074, Novus Biologicals; β-Actin, 4967S, Cell Signaling; total Histone 3, 4499 T, Cell Signaling; H3K27me3, 9733 T, Cell Signaling; H3K36me2, ab9049, abcam; HIF-1β/ARNT, 5537 T, Cell Signaling; FLAG, F3165, Sigma Millipore) in 5% skimmed milk overnight at 4 °C, followed by three washes with TBST (1% Tween20), and stained with horseradish-peroxidase-conjugated secondary antibodies. After additional washes, chemiluminescence was performed to visualize proteins of interest. Protein densitometry was performed using FIJI. Construction of the AAV-Barcode Library A previously generated AAV-KrasHR2.1Kb-U6-sgKras-PGK-Cre vector58 and pAAV-CAG-GFP (37825, Addgene) were used to generate the backbone of the AAV library (sgKras: 5’ GACTGAGTATAAACTTGTGG). An additional U6-sgRNA cassette targeting Kdm8 or non-targeting control (sgKdm8: 5’ tcaggtgctgatcatgtcag; sgNT: 5’ gcgaggtattcggctccgcg) was subcloned into the AAV vector from the Lenti-CRISPRv2 vector (52961, Addgene) using the NEBuilder HiFi assembly kit (E2621, NEB). The resulting AAV-KrasHR2.1Kb-hU6-sgKdm8-U6-sgKras-PGK-Cre vector was then validated for ITR integrity before HiFi assembly using the de novo AvrII and BsiWI sites to build the barcoded KrasG12D library as reported58. Two separate oligo pools with a predicted diversity of 2,359,296 synonymous mutations (mostly at wobble bases) on both sides of the oncogenic KrasG12D mutation were generated from IDT (oPool, IDT) and PCR amplified using the flanking sequences that are homologous to the linearized AAV vector. In addition, we incorporated 4 additional synonymous mutations at Kras codon 3, 4, 5, 6, 7, and 8 to confer sgKras resistance (GAGTATAAACTTGTGGTG to GAaTAcAAgCTaGTaGTc). The pooled oligos and the linearized AAV vector were assembled overnight, ethanol precipitated, and electroporated into Stbl4 electrocompetent cells (11635018, Thermo Fisher Scientific) to maintain a representation of 20.84 colonies per KrasG12D barcode. The assembled AAV vectors were then validated through Illumina sequencing and packaged into the AAV library (AAV8, SignaGen). Genomic cleavage detection assay Targeted indel formation was determined using the GeneArt Genomic Cleavage Detection Kit (A24372, Thermo Fisher Scientific). Briefly, PCR was performed using the genomic DNA isolated with the DNeasy Blood and Tissue Kit (Qiagen) and primers flanking the Kdm8 target region. The PCR amplicon was purified using the NucleoSpin PCR clean-up kit (Takara) and subjected to slow re-annealing using a thermocycler. The heteroduplex DNA containing indels was subsequently detected using an endonuclease and resolved on a 2% agarose gel. Quantitative PCR Cells were homogenized using TRIzol (Invitrogen) to release the total RNA and quantified by a Nanodrop spectrophotometer before conversion to cDNA with the High-Capacity cDNA Reverse Transcription kit (43-688-14, Thermo Fisher Scientific). Transcript quantification was performed using PowerUp SYBR™ Green Master Mix (Thermo Fisher Scientific). Histopathology and IHC Primary PDA tumors, lungs, and livers were dissected from tumor-bearing mice at the endpoint and fixed in 4% formalin and paraffin embedded. Hematoxylin-eosin staining was performed with a Sakura Prisma Hematoxylin and Eosin stainer according to the manufacturer’s protocol (Sakura Finetek USA). Antibodies for Kdm8 (101664-T40, SinoBiological), Vimentin (5741, Cell Signaling), ZEB1 (70512, Cell Signaling), Phospho-Histone H3 (9701, Cell Signaling), Tomato (600-401-379, Rockland), and Cytokeratin 19 (TROMA-III, Developmental Studies Hybridoma Bank) were used in IHC on a Roche Discover Ultra Immunostainer (Roche). Immunofluorescence and the analysis of Kdm8 and Galectin-4 in PDA tumors Sections were first deparaffinized on the instrument and subjected to buffer washes. Antigen retrieval was performed using a pH 9 retrieval buffer for 76 min at 95 °C. An inhibitor block was then applied to reduce nonspecific binding and inhibit endogenous enzymatic activity. The primary Kdm8 antibody (1:700) was manually applied for 60 min at 37 °C. After a buffer wash, OmniMap anti-Rabbit secondary antibody (760-4311, Roche) was applied at room temperature, followed by the addition of Rhodamine 6 G (760-244, Roche). Antibody denaturation was performed followed by manual application of the primary Galectin-4 antibody (1:4000, PA5-34913, Fisher Scientific) for 60 min at 37 °C. The OmniMap Rabbit secondary was re-applied, followed by the addition of Cy5 (760-238, Roche). After denaturation, the anti-Tomato antibody (1:3500) was then manually applied for 60 minutes at 37 °C, followed by the addition of OmniMap Rabbit and FAM (760-243, Roche). DAPI (760-4196, Roche) was used for nuclear staining. Four-channel fluorescence images were acquired using an Olympus® VS120 whole-slide fluorescence scanner equipped with a 20x objective. Image analysis was performed using the Visiopharm® digital pathology analysis platform. Regions of interest (ROIs) were manually annotated, and automated cell segmentation was performed based on DAPI nuclear staining. Quantification of fluorescence signal intensity was conducted on a per-cell basis within the annotated ROIs. Quantification of Kdm8 and vimentin in PDA tumors Kdm8, Vimentin, and Tomato IHCs were performed on serially sectioned slides and imaged at 20x magnification on an Olympus VS-120(R) whole slide scanner. Each set of whole slide image stack was digitally aligned with the VisioPharm(R) digital pathology analysis suite. The aligned images were divided into thousands of rectangular regions per sample with 354 microns in width and height, and the regions of interest were defined as having more than 40% of the normalized Tomato signal. Normalized expression and the correlation between Kdm8 and Vimentin across all regions of interest were quantified. Human PDA data and analyses Bulk RNA-Seq data from the PanCuRx study38, ICGC Controlled Data109, TCGA (unrestricted access through cBioPortal)71, and Dhara et al.72 were acquired through granted access to the respective data portals. Processed snRNA-Seq data from Hwang et al. was accessed through the Broad Institute Single Cell portal74 (https://singlecell.broadinstitute.org/single_cell). The PanCuRx study was conducted with support of the Ontario Institute for Cancer Research (PanCuRx Translational Research Initiative) through funding provided by the Government of Ontario, the Wallace McCain Center for Pancreatic Cancer supported by the Princess Margaret Cancer Foundation, the Terry Fox Research Institute, the Canadian Cancer Society Research Institute, and the Pancreatic Cancer Canada Foundation. For GSEA, genes were ranked by DESeq2-derived log2 fold change between all basal-like and classical-like samples using raw read counts from the PanCuRx and TCGA data sets and the enrichments of hallmark and custom gene signatures were quantified. Kaplan-Meier survival of the KDM8-induced gene signature score high (>2nd interquartile range or IQR value) and low PDA patients from the PanCuRx data set was performed using ‘rms’ and ‘survival’ packages in r. To quantify the significance of association between the status of liver metastasis at diagnosis and the following independent variables, including tumor size (scaled not centered), classical and basal subtype gene signatures, hallmark hypoxia, KDM8-induced gene signature, and Eastern Cooperative Oncology Group performance status (ECOG) in the PanCuRx cohort, we performed univariate logistic regression using the ‘pROC’ package. To quantify the association between the KRAS allelic imbalance (ordinal variables 0-3) and independent variables including the anatomic location (liver metastasis=1, primary PDA = 0), classical and basal subtype gene signatures and state (classical=1, basal=0), hallmark gene signatures (n = 50), and KDM8-induced and suppressed gene signatures, we performed univariate ordinal logistic regression to derive the coefficients, confidence intervals, and nominal p-values using the ‘ordinal’ package in r. All gene signatures were scaled using scale in r (center = T, scale = T). FDR-adjusted p-values were derived for the correction of multiple comparisons. Malignant cell isolation from the PTC Mice Tumors and metastases harvested from the PTC mice were dissected, minced, and digested to facilitate tissue dissociation as previously described11. Briefly, dissected tumors were dissociated in a digestion buffer (HBSS-free with collagenase IV, dispase, protease inhibitor cocktail) at 37 °C for 30 min followed by quenching with pre-chilled quench solution (L-15 medium, FBS, DNase). Dissociated cells were passed through a 40 μm strainer and resuspended in FACS buffer. Malignant cells in the ascites were prepared through ACK lysing buffer incubation at room temperature for 1–2 min prior to straining and FACS buffer resuspension. In most cases, a small aliquot of dissociated cells was dispensed on a 10 cm plate for the establishment of parental cell lines. The ex vivo isolated cells were subsequently stained with DAPI for viability and antibodies to the lineage markers including CD45 (103111, Biolegend), CD31 (102409, Biolegend), Ter119 (116211, Biolegend), and F4/80 (123115, Biolegend) for the exclusion of most non-malignant cells. Pure, viable, Tomato-positive, DAPI-negative, and lineage-negative malignant cells were isolated using a BD Biosciences Influx High Speed Cell Sorter at the Flow Cytometry Core Facility of Rutgers Cancer Institute. Single-cell RNA-Seq library preparation Using the ex vivo sorted cells, single-cell RNA-seq libraries were prepared using the 10X 5’ v2 dual indexed kit according to the user guide. In 3 metastatic lesions, we spiked in different numbers of Jurkat cells due to low cell counts of sorted Tomato-pos malignant cells (A252, A147, and L254). Following cDNA amplification, the cDNA pool was split and used for scRNA-Seq library construction, Kdm8-Seq, and KrasBC-Seq. ChIP-seq preparation and sequencing Chromatin immunoprecipitation (ChIP) sequencing was performed as reported previously110. Briefly, 2.5 × 107-1 × 108 688 M cells were cultured for each ChIP. To crosslink, formaldehyde solution was directly added to cell culture media to a final concentration of 1% for 10 min, followed by quenching with 125 mM glycine solution. Fixed cells were rinsed twice in ice-cold PBS and lysed in ice-cold lysis buffer (50 mM HEPES-KOH, pH 7.5, 140 mM NaCl, 1 mM EDTA, 10% Glycerol, 0.5% NP-40, 0.25% Triton X-100, with protease inhibitor cocktail). Lysates were transferred to a 1.5 ml TPX tube (Diagenode) and sonicated using Bioruptor (Diagenode) three times of 10 min (output level: high, interval: 0.5). An aliquot (50 μl) of the lysates was reserved for QC and input library. For immunoprecipitation, 4μg of ChIP-grade antibodies against H3K36me2 (ab9049, abcam), H3K27me3 (ab192985, abcam), or H3K27ac (ab4729, abcam) was added to the lysate (containing ~50 μg of sheared gDNA) and rotated overnight at 4 °C, followed by an incubation with 25 μl of Protein A Dynabeads (Life Technology) for 2 h at 4 °C. The beads were washed three times with 800 μl ice-cold LiCl washing buffer (100 mM Tris-HCl, pH7.5, 500 mM LiCl, 1% NP-40 and 1% sodim deoxycholate) prior to elution and reverse crosslink through an incubation with 400 μl of digestion buffer (0.1 M Tris-HCl, pH 7.5, 100 mM NaCl, 50 mM EDTA, 1% SDS, and 200 μg/ml of proteinase K) at 65 °C for 2 h with agitation every 30 min. The Input lysate was treated the same way to reverse crosslink. Eluted DNA samples were purified by organic extraction and ethanol precipitation. DNA libraries were prepared using KAPA Hyperprep Kit (Roche) and sequenced on the Illumina Nova-seq X platform. ATAC-seq library preparation and sequencing Libraries were generated using an updated method111. Briefly, viable 688 M PDA cells were first treated with DNase in HBSS for 30 min at 37 °C prior to nuclei release in pre-chilled ATAC resuspension buffer (Tris-HCL pH 7.4, digitonin, Tween-20, NP40). Transposition was carried out in transposition mixture (1xTD buffer, 100 nM Tn5 transposase, digitonin, Tween-20) for 30 min at 37 °C in a thermomixer with 1000 rpm. Transposed fragments were cleaned up with a Zymo DNA Clean and Concentrator-5 Kit. Partially amplified libraries were first quantified by qPCR to determine additional cycles needed and fully amplified using the NEBNext 2x MasterMix. Libraries were sequenced on an Illumina NextSeq 2000 sequencer. Library preparation for targeted sequencing The preparation for targeted KrasBC-Seq and Kdm8-Seq was based on a modified 10X VDJ immune profiling protocol. Briefly, amplified 10X cDNA libraries were used as templates for nested PCR reactions with non-adapted primers for PCR I (12 cycles for KrasBC-Seq and 13 cycles for Kdm8-Seq) and adapted primers with sample indices (denoted as “N”) and Illumina adapter sequences for PCR II (10 cycles for KrasBC-Seq and 11 cycles for Kdm8-Seq; KrasBC-Seq primers, forward: 5’ aatgatacggcgaccaccgagatctacactctttccctacacgacgctc; reverse: 5’ caagcagaagacggcatacgagatNNNNNNNNgtgactggagttcagacgtgtgctcttccgatctgtaggagtcctctatcgtagggtc; Kdm8-Seq, forward primer is the same as KrasBC-Seq, reverse: 5’ caagcagaagacggcatacgagatNNNNNNNNgtgactggagttcagacgtgtgctcttccgatctagctcctcctctttgtcaggc). Both PCR reactions were performed using ExTaq polymerase (Takara) and the amplicons were purified and size-selected using SPRIselect Beads (Beckman Coulter). Beads purified PCR amplicons from PCR II were quantified on a Qubit HS DNA quantification kit and a BioAnalyzer (Agilent) prior to sequencing (Novogene). Target gene knockdown, knockout, and re-expression in PDA cell lines To knockdown a gene in PDA cells, shRNA sequences were designed and subcloned into the TRC pLKO.1 vector (10878, Addgene; shGFP(Control): 5’ gcaagctgaccctgaagttcat; shScramble(Control): 5’ cctaaggttaagtcgccctcg; shKdm8-1: 5’ cgcacattcttcataacacca; shKdm8-2: 5’ gctcctgatgtcatgttagag; shKdm2a: 5’ gctccaaaccaacaaatataa; shKDM8-1: 5’ cctgttcatcccggtgaaata; shKDM8-2: 5’ tcagcaaatacatcgtgaatg; shKDM8-3: 5’ cgaggtacacagatgaggaat; shNsd2: 5’ ccagaaagagcttggatattt; shSmad4: 5’ gccagctacttaccatcataa). To knockout Kdm8 or Arnt (sgKdm8-1: 5’ caaggtacacagatgaagac, sgKdm8-2: 5’ atgtactgcagactacgaga; sgKdm8-3: 5’ tcgtagtctgcagtacatcc; sgArnt-1: 5’ gacatcagatgtaccatcgc; sgArnt-2: 5’ agggtttccagaagcaatgg), we subcloned sgRNA sequences into the non-viral PX458 vector (48138, Addgene) and transiently transfect PDA cells using TransIT-293 transfection reagent (MIR2704, Mirus Bio). Single GFP+ cells were sorted onto 96-well plates 48 hours post transfection to establish knockout clones and validated using western blot. To re-express Kdm8 cDNA, murine Kdm8 coding sequence was subcloned into the Lv241 lentiviral vector (GeneCopoeia) using TOPO TA cloning kit (450641, Thermo Fisher Scientific). Kdm8 variants were subsequently generated using Q5 site-directed mutagenesis kit (E0554, NEB). Kdm8 knockout PDA clones were infected with lentiviral vectors carrying Kdm8 variants and selected with puromycin or blasticidin. Lentivirus production and infection Lentivirus were generated by co-transfecting the viral shuttle vector with Delta8.2 and VSV-G packaging plasmids into the 293 T cells using polyethylenimine (23966-1, Polysciences). Medium was replenished 24 h post transfection and supernatants were collected at 48 h, frozen, thawed, and spun down at 2000 rpm before infecting PDA cells. Bulk RNA-seq library preparation and analysis Total RNA was isolated from PDA cells using RNeasy Mini kit (Qiagen) prior to library construction using the NEBNext Ultra II RNA Library Prep Kit for Illumina (NEB). Libraries were sequenced on an Illumina NovaSeq 6000 sequencer (paired-end 2x150bp, Novogene). Reads were mapped to the mm39 genome build using the STAR aligner (v2.7.5a). Reads counts at both transcript and gene levels were derived using htseq-count (--stranded=no --type=exon --mode=intersection-nonempty) and transcripts per kilobase of exon and scaled in million or TPM was derived using a mm39 reference gtf file (UCSC genome browser) and the ‘GenomicFeatures’ package (v1.58.0) in r. Differentially expressed genes, fold change, and adjusted p-values were derived from read counts using DESeq2 (v1.46.0)112. Genes with low abundance were removed from further analysis. For GSEA (Broad Institute), log2 fold change ranked gene lists were assessed for enrichment of hallmark (v7.5.1), curated gene sets (c2.all.v7.5.1.symbols, n = 4,133), or custom gene signatures. In vitro cell count and viability assays Cells were seeded in triplicate into a 6-well plate at a density of 5 × 104 cells/well on day 0. The plates were then cultured under normoxia or hypoxia for 96 h. On day 2 and day 4, the cells were trypsinized and total cell count was determined using a Vi-CELL XR Cell Viability Analyzer (Beckman Coulter). For viability, cells were seeded onto a 96-well plate at a density of 2.5–10 × 103/100 μL/well and allowed to attach overnight. Cells were then treated with the Kras inhibitor for 48 h before equal volumes of CellTiter-Glo (Promega) were added and mixed on an orbital shaker for 3 min. Luminescence was quantified using a plate reader (Tecan). Single-cell transcriptomic pipeline Fastq files were mapped to a custom mm39 genome build (GRCm39 release 108 with the Tomato transcript) using the 10X CellRanger pipeline (v6.0.2; custom scripts are available via Zenodo repository under https://doi.org/10.5281/zenodo.20684364). To exclude Jurkat cells in samples A252, A147, and L254, reads were also mapped to the GRCh38 genome reference (GRCh38 v84 build). We used the ‘Seurat’ (v4) R package for the identification of cell clusters, differential genes, and all the downstream cell state analyses with default settings113. Briefly, CellRanger output sparse matrices were merged and read counts were normalized after removing potential doublets, dying cells, and Jurkat cells based on GRCm39 transcript count, mitochondria percent, and GRCh38 transcriptomic features, respectively. Distinct clusters of malignant cells were uncovered using SNN modularity optimization-based approach following linear transformation of data and dimensionality reduction. To calculate gene signature scores, we performed single sample GSEA using the ‘GSVA’ r package (v4.4.2). For RNA velocity analysis, we inferred the latent time and estimated differential kinetics using scVelo82 (v0.3.3). Briefly, mapped bam files from the CellRanger pipeline were used to generate spliced and non-spliced expression matrices using default velocyto run10x command. Preprocessing was performed to compute first- and second-order moments for velocity estimation and subsequently the full splicing kinetics using the scv.tl.velocity() function. We then projected the velocity or latent time estimates onto the UMAPs for visualization. KrasG12D barcode calling for clonal identity To quantify the expression of wildtype and barcoded KrasG12D transcripts in each cell, KrasBC-Seq 3’ reads that passed QC were first trimmed using cutadapt (v3.4, -g “ctctatcgtagggtcata;max_errors=0.2;optional…cattttcagcaggcct;max_errors=0.2”). Trimmed reads were subsequently mapped to a custom reference containing wildtype Kras and the engineered KrasG12D BC sequences using bowtie2 (v2.4.1). Additionally, the paired reads were subjected to the CellRanger pipeline to resolve UMI and the 10X barcode (cellBC) identities. Mapped reads were collapsed with an identical cellBC, KrasG12D BC (or wildtype Kras), and UMI combination into a unique transcript in each cell. Owing to the extremely low incidence of homology-directed repair that is anticipated to occur twice in a cell58, a vast majority of malignant cells should only carry one KrasG12D barcode. Based on this assumption, we quantified the UMI counts of all KrasG12D BCs with edit distance less than 3 (ED ≤ 2) and assigned the dominant KrasG12D BC to each unique cellBC. This ED threshold was optimized via sensitivity analysis to the point of structural saturation; beyond this value, the KrasG12D BC cluster architecture stabilized, and further increments yielded marginal shifts in lineage assignments across the dataset. Non-malignant (somatic) cells were defined as having more than one copy of wildtype Kras and no detectable KrasG12D BC. Cells without a dominant KrasG12D BC and wildtype Kras were excluded from downstream lineage analyses. The abundance of Kras in a cell quantified by KrasBC-Seq was defined as the sum of unique UMI counts of KrasG12D BC and wildtype Kras normalized against the count of the entire transcriptome and scaled according to sequencing depths. We used the ‘networkD3’ (v0.4) package in r to create the Sankey diagrams using default parameters. Kdm8-seq to infer Kdm8 genome editing Similar to the KrasBC-Seq pipeline, Kdm8-Seq 3’ reads were first trimmed using cutadapt to remove the first 26 bp. Trimmed 3’ reads were subsequently mapped to the Kdm8 transcript (NM_029842) using the bwa aligner (v0.7.17-r1188). To include all malignant cells from the primary PDA tumors, cells without detectable Kdm8-Seq signal were imputed as having zero read count. UMI and cellBC identities in the paired 5’ reads were uncovered using the CellRanger pipeline. Reads were then collapsed into unique UMI and cellBC combinations. The abundance of Kdm8 in a cell or cellBC quantified by Kdm8-Seq was defined as the unique UMI count normalized against the count of the entire transcriptome and scaled according to sequencing depths. ChIP-Seq and ATAC-Seq data analysis Both ChIP-Seq and ATAC-Seq reads that passed the QC process were aligned to a mm39 genome build using bowtie2 (v2.4.1). Bam files were filtered for reads mapped to blacklisted regions and duplicated reads. For the ATAC-Seq pipeline, insertion bed files were made with bedtools (v2.25.0) accounting for the +4/-5 bp offsets by Tn5114. Merged peaks were called using MACS3 (v3.0.0a6) to derive broad and narrow peak files by comparing ChIP pulldown signals with input samples. Only reads with low false discovery rate (qValue > 5) were included in the analysis. Counts of the pulldown or Tn5 insertions within merged peaks were derived using bedtools and data from technical replicates were used to derive differential change and adjusted p-values using DESeq2, normalized against size factors that reflected mapped read depths. Peaks were annotated using Homer annotatePeaks.pl (v5.1) with the mm39 reference. To quantify the enrichments of transcription factor binding motifs within differential regions, we performed Homer findMotifsGenome.pl (-size given) using curated collection of mouse motifs from the CisBP database (n = 7000 motifs) and comparable numbers of constitutive regions as background. To compare ChIP profiles across samples, mapped reads from replicates were first merged with samtools (v1.3.1) and normalized against comparable input samples using bamCompare (v3.5.4.post1). Aggregate ChIP profiles within differentially regulated H3K36me regions were created using deeptools (v3.5.4.post1). Bulk and single-cell whole genome sequencing and analysis For bulk WGS, genomic DNA were extracted using the Qiagen DNeasy Blood and Tissue Kit (69504, Qiagen). The libraries were prepared using the Rapid Plus DNA Lib Prep Kit (ABclonal) prior to sequencing on a NovaSeq X Plus sequencer (Novogene). For single-cell WGS, 688 M cells expressing control or a Kdm8-targeting shRNA were sorted into a 96-well plate prior to library preparation using the PicoPLEX® Gold Single Cell DNA-seq Kit according to the manufacturer’s protocol (R300670, Takara). To assess the copy number variations, we adopted a previously reported pipeline with some modifications115. Briefly, fastq files from both the bulk and single-cell WGS studies were mapped to an mm39 reference prior to alignments using bowtie2 (v2.4.1). To compute bin boundaries, mm39 fasta files for individual chromosomes were downloaded from UCSC (hgdownload.cse.ucsc.edu/goldenPath/mm39/chromosomes) and used to generate reads from the reference genome. The reads were mapped back to the mm39 reference using bowtie2 and the result was used to generate genomic positions that defined 50,000 bins with equal numbers of mappable positions. Counts of reads within the bins were quantified for each sample with the GC content accounted for using a LOWESS regression. The log2-transformed ratios were then subjected to Circular Binary Segmentation via the ‘DNAcopy’ r package to define genomic segments. Absolute copy numbers were inferred by evaluating the density distribution of pairwise segment mean differences to determine the baseline distance of a single-copy change, followed by scaling and rounding segment means to the nearest integer. Cell synchronization and immunofluorescence microscopy Cells were seeded in 6-well plates with coverslips and treated with cell culture media containing thymidine (2 mM) for 16 h, followed by fresh media for 9 h, then thymidine (2 mM) for another 16 h. Cells were subsequently released into fresh media for 8–12 h at anaphase (688 cell lines: 8 h; MCF10A cell lines: 12 h; 606 cell lines: 8 h; Capan-1 cell lines: 10 h). Synchronized cells were fixed with 4% paraformaldehyde for 15 min and permeabilized with 5% Triton X-100 for 15 minutes. Blocking was performed with 5% goat serum/PBS at room temperature for 1 h, followed by overnight incubation at 4 °C with primary antibodies diluted in 0.5% goat serum/PBS targeting α-tubulin (Sigma, T9026, 1:1000), γ-tubulin (Sigma, 1:1000), and Pericentrin (Abcam, ab4448, 1:400). Cells were incubated in fluorophore- conjugated secondary antibodies (1:1000 in PBS, anti-rabbit IgG-Alexa Fluor® 488, 4412S; anti-rabbit IgG-Alexa Fluor® 647, 4414; anti-mouse IgG-Alexa Fluor® 488, 4408, all from Cell Signaling) at room temperature for 1 h and counterstained with DAPI (Fisher, P36935). Images were captured using ZEISS Apotome3 and Nikon A1Rsi Confocal Microscope. Metaphase chromosome preparation and Micronuclei quantification Cells in flasks were made 60–80% confluent and split a day before harvesting for metaphase chromosomes preparation. Cells were then treated with colcemid (0.1 µg/mL, Invitrogen, 15210-016) at 37 °C for 30 min. Harvested cells were incubated in a hypotonic solution (0.075 M KCl) at 37 °C for 25 min, followed by fixation (methyl alcohol/ glacial acetic acid, 3:1 in volume). For micronuclei analysis, fixed cells were dropped onto slides and stained with DAPI. Images of micronuclei were captured and analyzed using a BioView Duet System (Bioview) on an Olympus BX63 fluorescence microscope (Olympus) equipped with a ×60 oil immersion objective (NA 1.3). Chromosome painting and FISH analysis Chromosome painting was performed on fixed cells from metaphase chromosome preparation. Fixed cells were dropped onto slides and baked at 95 °C for 5 minutes. Slides were rehydrated in 2 x saline-sodium citrate (SSC, 175.3 g NaCl and 88.2 g of sodium citrate in 800 mL of ddH2O) at 72 °C for 2 min and then dehydrated in an ethanol series: 70% ethanol, 85% ethanol and 100% ethanol for 2 min each. The custom locus-specific mouse FISH probes for Kras were prepared using BAC clone RP23-300B22 (chr6:145163176-145374812, spanning 0.21 Mb), obtained from the BACPAC Resources Center (Children’s Hospital Oakland Research Institute). All FISH probes were generated following a previously described protocol116. Briefly, the BAC DNA was isolated and labeled by nick translation using modified dUTPs (DY-590-aadUTP) from Dyomics (Jena, Germany). Similarly, 3 additional custom locus-specific mouse FISH probes targeting chromosome 6 centromere (Cep 6 using BAC clone RP23-341C22 and DY-495-dUTP), chromosome 6 distal (distal chr6 using BAC clone RP24-77K22 and DY-431-dUTP), and chromosome 18 centromere (Cep 18 using RP23-337K12 and DY-530-dUTP) were generated by nick translation and used as reference probes in this study. The 4 color custom FISH probes were mixed at equal ratios and hybridized to metaphases as previously described. Briefly, probes and slides were codenatured at 75 °C for 3 min on a slide warmer (ThermoBrite System, Molecular Abbott) and hybridized overnight at 37 °C in a humidified chamber. Post-hybridization washes were performed in prewarmed 0.4xSSC at 74 °C, followed by 2xSSC plus 0.1% Tween at room temperature, serial ethanol dehydration, air drying, and mounting with DAPI antifade VECTASHIELD Antifade Mounting Medium with DAPI (Vector Laboratories) for counterstaining. Images were captured and analyzed by a BioView Duet System (Bioview). Statistics & reproducibility Data normality was assessed prior to statistical analysis using the Shapiro-Wilk test. Sample sizes were estimated via power calculations based on effect sizes derived from initial experimental replicates. For non-parametric analyses, two-group comparisons were performed using the two-sided Wilcoxon signed-rank test, and multi-group comparisons were evaluated using the Kruskal-Wallis test followed by Dunn’s post-hoc test for multiple comparisons, unless otherwise indicated. For parametric analyses, two-group comparisons were performed using the two-sided Student’s t-test, and multi-group comparisons were evaluated using a one-way ANOVA followed by Dunnett’s post-hoc test. For correlation studies, Pearson correlation coefficients (ρ) and their associated p-values were calculated. Statistical significance was defined as an adjusted p < 0.05. For animal experiments, mice were randomly assigned to experimental groups by reshuffling prior to the transplantation of different PDA cell lines carrying distinct Kdm8 variants or knockdown constructs. Investigators were blinded to group allocation during outcome assessment. No data points or animals were excluded from the analyses. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Reporting Summary Transparent Peer Review file Source data Source Data