Evolutionary fingerprints of epithelial-to-mesenchymal transition

Abstract

Mesenchymal plasticity has been extensively described in advanced epithelial cancers; however, its functional role in malignant progression is controversial^1–5. The function of epithelial-to-mesenchymal transition (EMT) and cell plasticity in tumor heterogeneity and clonal evolution is poorly understood. Our work clarifies the contribution of EMT to malignant progression in pancreatic cancer. We leveraged somatic mosaic genome engineering technologies to trace and ablate malignant mesenchymal lineages along the EMT continuum. The experimental evidence clarifies the essential contribution of mesenchymal lineages to pancreatic cancer evolution. Spatial genomic analysis, single cell transcriptomic and epigenomic profiling of EMT clarifies its contribution to the emergence of genomic instability, including events of chromothripsis. Genetic ablation of mesenchymal lineages robustly abolished these mutational processes and evolutionary patterns, as confirmed by cross species analysis of pancreatic and other human solid tumors. Mechanistically, we discovered that malignant cells with mesenchymal features display increased chromatin accessibility, particularly in the pericentromeric and centromeric regions, which in turn results in delayed mitosis and catastrophic cell division. Therefore, EMT favors the emergence of genomic-unstable, highly fit tumor cells, which strongly supports the concept of cell-state restricted patterns of evolution, where cancer cell speciation is propagated to progeny within restricted functional compartments. Restraining the evolutionary routes through ablation of clones capable of mesenchymal plasticity, and extinction of the derived lineages, halts the malignant potential of one of the most aggressive forms of human cancer.

Neoplastic cells display virtually unlimited evolutionary potential, which sustains tumor heterogeneity and ultimately drives lethal cancers^6–9. This intrinsic trait of the tumor ecosystem is responsible for unlocking a multiplicity of distinct cell states, a hallmark of cancer defined as phenotypic plasticity¹⁰. Nevertheless, acquisition of phenotypic plasticity and its contribution to tumor progression and patient outcomes remain poorly understood, particularly its role in metastatic dissemination, adaptation to therapy and evolutionary selection of aggressive clones^11–13. As a paradigm of phenotypic plasticity, Epithelial-to-Mesenchymal Transition (EMT), a trans-differentiation program resulting in a continuum of cell-states, allowing the switch from epithelial to a spindle-like morphology, has been frequently described in advanced epithelial cancers and associated to dismal prognosis^14,15. The mechanistic contribution of mesenchymal plasticity to cancer evolution, however, is still a subject of debate, as conflicting evidence from various research groups accumulated over the past two decades^1–3,16,17.

To fill this knowledge gap, we set to clarify the functional role of EMT in cancer evolution focusing on pancreatic ductal adenocarcinoma (PDAC) as a model system, where pathological and transcriptomic analysis of patient derived samples suggests that the emergence of mesenchymal features is associated with disease progression and poor survival^18–20. To study the contribution of mesenchymal plasticity to cancer evolution, we established somatic mosaic genetically engineered mouse model (SM-GEMM) of PDAC leveraging dynamic lineage tracing and lineage ablation systems that allow to trace and selectively remove the tumor lineages derived from cells undergoing EMT without perturbing the tumor microenvironment, which is a substantial technological advancement compared to previous in vivo technologies^1,4. Our findings provide compelling evidence on the functional role of EMT as a crucial phenotypic program for malignant transformation and the expansion of aggressive tumor subpopulations. Detailed annotation of genome-epigenome-phenome associations of SM-GEMM through bulk genomic and epigenomic, coupled with single cell and spatial sequencing technologies reveals that EMT is necessary to unlock the evolutionary potential of epithelial tumors towards aggressive clinical behaviour. Specifically, cross-species analysis of genomic signatures suggests that tumors with mesenchymal features are prone to acquire massive genomic instability characterized, in particular, by complex structural rearrangements and chromotripsis. Our work proves that highly plastic malignant subpopulations contribute to the generation of specific evolutionary routes and display unique patterns of structural genomic signatures. This suggests that diverse mutational processes and mechanisms of selection are at play in phenotypically distinct malignant compartments.

EMT is indispensable for PDAC progression

To study the role of mesenchymal plasticity in PDAC evolution we developed multiple reporter systems that allow for the tissue specific tracing of mesenchymal lineages (defined as malignant clonal expansions derived from an ‘EMT-proficient’ ancestor). We engineered the mouse Vimentin locus for the conditional expression of a synthetic cassette harboring a H2B-GFP and a FLPO recombinase that is actively transcribed upon CRE-mediated recombination to ensure tissue and tumor cell specificity, thus excluding the stromal cellular compartment (Vim^{FLEX(SA-H2B-eGFP-T2A-FlpO-WPRE-pA)}, hereafter Vim^Flpo) (Supplementary information, see ‘Experimental Model Information’). Preliminary characterization of the SM-GEMM confirmed expected patterns of tissue expression and no overt phenotypic effect in both heterozygous and homozygous littermates (Supplementary information). The reporter strain was further crossed into a somatic mosaic model of PDAC (R26^FSF-LSL-TdT-H11^LSL-spCas9-KRas^LSL-G12D) where tumorigenesis is initiated through orthotopic delivery of Adeno-Associated Viral (AAV) particles carrying sgRNAs targeting the most common PDAC tumor suppressor genes, Trp53 and Cdkn2a and Cdkn2b, alone or in combination, along with a CRE-recombinase driving the activation of the conditional Cas9, the KRas^G12D oncogene and the lineage tracing reporter in the edited pancreatic epithelial cells (hereafter PCΦ, see Methods for extended nomenclature and Supplementary information). This model system allows for 2 possible experimental outputs: ‘mesenchymal lineages’ (MS-L: TdT+/H2B-GFP^HIGH, actively in a mesenchymal state, or TdT+/H2B-GFP^LOW, reverted to an epithelial state) and ‘epithelial lineages’ (EPI-L: TdT−/H2B-GFP−) (Figure 1a-b, Supplementary information). Strikingly, gross and histopathological analysis of PCΦ models enginneered for different combinations of driver events (KPCC: KRas^G12D-Trp53^KO-Cdkn2a/b^KO; KP: KRas^G12D-Trp53^KO; KCC: KRas^G12D-Cdkn2a/b^KO) reveal that, advanced tumors and metastatic lesions were universally established by MS-L (Figure 1 c-g, Extended Fig. 1 a-f). Histopathological, immunophenotypic and lineage tracing analysis of EMT in the PCΦ model also demonstrates that metastatic sites are almost universally established by MS-L and their growth is sustained by cells in an active mesenchymal state (TdT+/H2B-GFP^HIGH), irrespective of the engineered driver events (Figure 1 h, Extended Fig. 1 g-h). Furthermore, quantitative histopathological analysis of early lesions shows that EPI-L contribute to pre-neoplastic lesions, but are counter-selected during malignant transformation (Extended Fig. 2 a-b). This evidence suggests that mesenchymal plasticity emerges very early during tumorigenesis, with scattered activation of the TdT reporter in PanIN lesions, thus selecting for malignant populations with high fitness and metastatic competency. To further test the functional advantage of MS-L on EPI-L cells, we performed orthotopic transplantation studies of TdT+ (MS-L) and TdT− (EPI-L) organoid-derived cells engineered with the previously described driver events (Extended Fig. 2 c-d). Indeed, upon transplantation, TdT+ cells display aggressive clinical behavior, tendency to sarcomatoid transdifferentiation and metastatic spread (Extended Fig. 2 e-f). Remarkably, TdT− transplants, while tumorigenic, invariably resulted in TdT+ lesions. These experiments suggest that the activation of an EMT program is indispensable to lineages endowed with malignant potential and that tumoral epithelial cells are highly metastable and prone to mesenchymal reprogramming. To functionally elucidate the contribution of EMT to pancreatic tumorigenesis, we engineered a lineage ablation system leveraging a diphteria toxin A (DTA) suicide cassette. This system selectively eliminate EPI-L cells transitioning to MS-L and activate the Vim^Flpo reporter during tumorigenesis (PCΨ, Methods, Supplemental information) (Figure 1 i). In line with the lineage tracing experiments, complete ablation of MS-L strongly impacted tumor penetrance, survival and metastatic burden resulting in virtually complete abrogation of malignant progression, with the emergence of low-grade cystic lesion lacking invasive potential and characterized by low proliferative index in both primary SM-GEMM and in orthotopic transplants (Figure 1 j-o, Extended Fig. 2 g-j). Overall, these data demonstrate that the ability of malignat cells to undergo EMT is dispensable for tumor initiation, but essential for the emergence and expansion of clones with highly proliferative, invasive and metastatic potential.

Figure 1. EMT is required for malignant progression of pancreatic tumors.

a-b) Schematics of of PCΦ SM-GEMM alleles. c) Representative case of PCΦ SM-GEMM showing presence of TdT positivity at both primary and metastatic sites. Picture is representative of N = 76 mice. d) Histopathological characterization of a PanIN in the PCΦ SM-GEMM. Image is representative of N = 5 independent lesions. e) Scatter plot with individual values displaying total number of TdT+ and TdT− metastasis for the PCΦ model. N = 76 mice. P = 3.93683e-10, 1.71139e-06, 2.2362e-10, 0.0380, 5.4153E-11. f-g) Representative histopathological pictures showing presence of TdT+ cells in a liver metastatic site (f) and a primary tumor site (g) both in low grade (I) and high grade (II) lesions. h) Histopathological quantification of TdT+/GFP− and TdT+/GFP+ cells among primary tumor (PT) and different metastatic sites. N = 25 lesions, 5 lesions/mouse per each anatomical site. i) Experimental schematic of the generation of the PCΨ model allowing for the conditional ablation of tumor cells expressing the synthetic Vim^Flpo cassette. j) Survival curves of PCΨ Vim^+/+ and Vim^Flpo/+ mice; censored events: no tumor event. P = 3.01937e-06. k) Macroscopic pictures of PCΨ Vim^+/+ and Vim^Flpo/+ mice-derived lesions (BF: brightfield and TdT). Vim^+/+: N = 10; Vim^Flpo/+: N = 10. Data are combined from two different experiments. l) Representative MRI axial sections of PCΨ Vim^+/+ and Vim^Flpo/+ mice. m) Scatter plot displaying individual number of metastasis per each PCΨ Vim^+/+ and Vim^Flpo/+ mouse. Vim^+/+: N = 10; Vim^Flpo/+: N = 10. Data are combined from two different experiments. P = 0.002 n-o) Histopathological analysis of PCΨ Vim^+/+ and Vim^Flpo/+ mice-derived tumoral lesions evidencing low proliferation index of Vim^Flpo/+-derived tumors. N = 5 lesions per group. P = 4.89272e-05. Data are presented as mean values +/− SD. P values are calculated as follows: j) Logrank Mentel-Cox test; e,m,o) two-sided t-student test.

EMT sustains tumor growth in PDAC

We further tested whether cells with mesenchymal features are necessary to actively sustain the growth of advanced tumors. To this aim, we generated a second SM-GEMM by knocking-in a synthetic cassette harboring an EGFP and a HSV-TK (Vim^{FLEX(SA-eGFP-T2A-HSV-TK—WPRE-pA)}, hereafter Vim^TK) that is primed for transcription upon CRE-mediated recombination in the pancreatic epithelial compartment to ensure tumor specific ablation of mesenchymal lineages upon ganciclovir (GCV) administration (Supplemental information). Specifically, the system allows for the tissue and time restricted ablation of malignant cells that are actively in a mesenchymal state while sparing the stromal and the epithelial compartments of the tumor (PCΩ, Methods, Supplementary information) (Figure 2 a). This experimental model allowed us to distinguish 2 functional malignant subpopulations: one in an active mesenchymal state, TdT+/GFP^HIGH (MS-L), and one lacking mesenchymal features TdT+/GFP^LOW (EPI-L). Indeed, GCV treatment of pancreatic tumors strongly inhibits tumor growth at primary and metastatic sites in both the short and long term experimental arms, irrespective to the engineered driver events (21 and 90 days of treatment respectively). These effects are completely rescued after GCV pulse-release, suggesting that the malignant epithelial compartment might function as a reservoir for the expansive mesenchymal compartment or that a pool of remnant MS-L fuel tumor recurrence (Figure 2 b-f, Extended Fig. 3 a-d). Histopathological characterization of PCΩ tumors at the end of the treatment shows a significant depletion of malignant cells with mesenchymal features in the treated group, resulting predominantly in well differentiated epithelial structures (Extended Fig. 3 e-g). These evidences led us to the hypothesis that cells in active EMT are critical contributors of tumor growth, sustaining the homeostasis of the proliferative compartment. To further corroborate this hypothesis, we generated tumor explants from PCΩ mice to establish 3D spheroid cultures (Extended Fig. 4a); ex vivo GCV treatment of PCΩ tumor derived cultures resulted in a significant reduction in dimensions, without any significant effect on clonogenic potential upon pulse-release (Extended Fig. 4 b-c). Furthermore, transplantation studies in orthotopic and heterotopic settings demonstrate that MS-L populations promote metastatic outgrowth more efficiently when compared to EPI-L, as previously shown by our group²¹ (P<0.0001)(Figure 2 g-j, Extended Fig. 4 d-l, Extended Fig. 5 a-e).

Figure 2. Tumor cells with mesenchymal features fuel tumor growth.

a) Experimental schematic representing the PCΩ model and treatment schedule. b) Representative T2 weighted axial MRI scans of PCΩ mice. c) Quantification of tumor volume via MRI. N = 4 mice per group. P = 7.27e-13. d-e) Box plot showing quantification of tumor volume and number of metastases Vehicle: N = 20, GCV: N = 10, GCV-lt: N = 10; GCV-rec: N = 10; GCV-lt rec: N = 10 over N=2 experiments. P = 8.922e-12, 0.0008, 0.00011 (d); P = 3.00406e-07, 0.0005, 2.38745e-06 (e). f) Survival curves displaying a significant benefit. Vehicle: N = 20; GCV-rec: N = 10; GCV-rec lt: N = 10. P = 1e-15. g) Representative macroscopic pictures of lungs of mice randomized to Vehicle or GCV. Overlap: Brightfield, TdT and GFP. h-i) Quantification of number of nodules (h) and mean nodule area (i) displaying a significant reduction in the GCV-tv group. N = 5 mice per group. P = 0.027 (h). P = 0.0082, P = 0.0073 (i). j) Survival curves showing significant benefit of GCV treatment in transplantation studies. N = 5 mice/group. P = 0.0026. k) Histopatological analysis of proliferation potential of EPI-L and MS-L cells with representative pictures (left) and quantification (right) in primary tumors. N = 25 lesions from 5 different mice per group. P = 8.02997e-08. l-m) Representative images (l) and quantification (m) of aberrant mitosis in the two subpopulations. N = 5 different tumors. P = 8.75017e-05. Error bars represent SD. Box plots are centered at the 50^th percentile and bounds of box are fixed on the 25^th and 75^th percentile, min and max include all data points. P values are calculated as follows: c,k) two-way ANOVA; d,e,h,i) two-sided student t-test; f,j) Logrank Mentel-Cox test; m) two-sided Fisher’s exact chi-square test.

Histopathological and cytological analysis of PCΩ pancreatic tumors reveals that malignant cells with mesenchymal features MS-L have higher proliferation index (P<0.0001) and a significant increase in the number of aberrant mitosis when compared to EPI-L cells, as assessed by PCNA, pHH3 stainings and visualization of mitotic figures through confocal microscopy. These evidences altogether suggest that mesenchymal plasticity might increase the chance for genomic instability and mitotic errors (Figure 2 k-m). To provide further proof of increased mitotic infidelity in PDAC cells with mesenchymal features, we analyzed a series of histopathological cases from surgically resected patients, annotated for grade and sarcomatoid/anaplastic transdifferentiation²¹. The morphometric analysis confirms that sarcomatoid transformation associates with an increased number of aberrant mitosis (P < 0.0001) (Extended Fig. 5 f-h). Our data collectively provide a counterintuitive model to the established role of EMT during malignant progression^15,22, where mesenchymal plasticity is required for tumor growth and clonal expansion, as demonstrated by GCV pulse-and-release experiments and transplantation studies.

EMT promotes genomic instability

The functional studies described above point towards a potential role of mesenchymal plasticity as a functional driver of genomic instability in PDAC. To test this hypothesis, we performed Slide-DNA-seq²³ on 3 independent tumors established from the PCΦ model 3 weeks after AAV injection, enabling us to capture a spectrum of pre-neoplastic lesions and early maligant nodules. Geospatial characterization of the genomic landscape of early tumors consistently showed the emergence of three clusters where the one overlapping with MS-L areas uniformly displays features of aneuploidy including whole genome duplication (WGD) and chromosome level alterations (gains of 6 and 11, loss of 4 and 15), overall suggestive of chromosomal instability (CIN) (Figure 3 a-c, Extended Fig. 6 a-d). These data prompted us to investigate whether the genetic ablation of mesenchymal lineages restrains the evolutionary patterns of PDAC and the emergence of CIN. Therefore, we performed Whole Genome Sequencing (WGS) of malignant sub-populations isolated from tumors established from PCΨ and PCΩ SM-GEMM, where ablation of mesenchymal lineages is performed at different time points of tumor progression (collectively, EMT-off and EMT-on, please see Methods, Supplemental information and Table 1) aiming to inform on the impact of EMT-proficient lineages on the genomic landscape of these tumors.

Figure 3. EMT promotes genomic instability and acquisition of events of chromotripsis.

a) Slide-DNA-seq of a PCΦ SM-GEMM 3 weeks after AAV transduction showing TdT+ and TdT− regions of pancreatic tumor tissue. b) Normalized copy number scatter plot showing differential copy number aberrations in the three clusters. c) Three-dimensional plot showing geographic distribution of ploidy as calculated from Slide-DNA-seq. d-e) Experimental schematic and circos plot of two exemplary cases of PCΩ models. f) Summary plots showing individual values of mean of ploidy and fraction of copy number altered (fCNA) for EMT on and EMT off groups. N = 42 independent tumors (EMTon: N = 20, EMT off: N = 22) over N = 3 independent experiments. P = 0.0014, P = 0.0129. g) Box and whiskers with data points showing number and type of SV enriched in the EMT-on mice. N = 5 (Vehicle) and N = 7 (GCV). P = 0.0002 h) Copy number signatures for the EMT-on group (top) and EMT-off group (bottom). i-j) Scatter plots and quantification of events of chromotripsis in EMT-on and EMT-off groups showing an increment in the number of chromotripsis in the EMT-on group. N = 42 independent tumors (EMTon: N = 20, EMT off: N = 22) over N = 3 independent experiments. P = 2.15663e-09. k) Schematic and three-dimensional scatter plot of the TCGA cohort. l) Representative histopathological pictures of two TCGA BRCA. Images are representative of the TCGA BRCA dataset, N = 1108 samples. Data are presented as mean values +/− SD. P values are calculated as follows: f) two-sided student t-test; g) two-way ANOVA; j) two-sided student t-test; k) two-tailed spearman correlation.

Strikingly, ablation of EMT lineages results in relatively flat genomic landscapes, while the spontaneous expansion of malignant sub-populations with mesenchymal plasticity results into genomic features of aggressive and advanced PDAC²⁴, specifically frequent copy number aberrations (CNAs, p < 0.01), WGD and a significant increase in the number of complex structural variants (SVs) including deletions, duplications and insertions (p < 0.05). Such phenomena were evident collectively in both models, although more prominently in the PCΨ system, where mesenchymal lineages are genetically removed at the early stages of tumorigenesis, suggesting that context-dependent evolutionary patterns are highly specific for mesenchymal lineages (Figure 3 d-g). These differences were not a result of unequal distribution of driver events, particularly concerning Trp53 alterations, being a gatekeeper of genomic instability, neither a result of single nucleotide variants (SNVs) mutational signatures, which are similarly distributed among groups (Extended data Fig. 6 e-f). This strongly suggests that the activation of EMT is a driver of genomic instability, irrespective to the truncal genomic drivers provided in the model. Next, we interrogated EMT-off and EMT-on tumors on the genetic processes that may be responsible of CIN through analysis of copy number signatures, demonstrating that EMT-on tumors are enriched for signatures related to events of chromotripsis (CN8)²⁵ as compared to EMT-off lesions, where MS-L are genetically depleted (Figure 3 h). Supporting these findings, we recovered the number of chromotriptic events per sample using ShatterSeek and by visual inspection of high-confidence chromotripsis events from each tumor (Methods)²⁶. Strikingly, tumors with unrestrained potential for mesenchymal plasticity (EMT-on) display high burden of chromotripsis when compared to tumors where mesenchymal lineages are ablated (EMT-off, P < 0.0001) (Figure 3 h-j, Extended Fig. 7 a-c). These findings were also recapitulated leveraging WGS data obtained from PCΩ-KP and -KCC models, suggesting that the acquisition of genomic instability in EMT-on tumors is independent from the tumorigenic driver events (Extended Fig. 8 a-e). Cross-species analysis of patients’ derived malignant lesions (TCGA, COMPASS and MDA-PDX-PDAC, please see Table 2)²⁴ annotated for number of events of chromotripsis and signatures of EMT, further confirmed that solid tumors with transcriptomic features of mesenchymality are prone to accumulate chromotripsis, WGD and subclonal genomic evolution (Figure 3 k-l, Extended Fig. 9 a-b). These phenomena are tissue agnostic, as being observed across anatomical disease sites and histological sub-types, strongly supporting the hypothesis that mesenchymal plasticity is permissive for the emergence and selection of complex structural rearrangements and karyotypes. Moreover, after correction for sample purity and TP53 status a significant association between complex chromosomal rearrangment, fraction of copy number altered, ploidy with mesenchymal transcriptomic signatures was detected (TCGA: P<0.0001; COMPASS: P<0.01) (Extended Fig. 9 c-o).

To understand the putative functional impact of chromotriptic events and other structural genomic rearrangments related to EMT during pancreatic cancers progression, we examined chromosomes affected by chromotripsis at high frequency and SVs in the EMT-on group (Figure 4 a-b). In particular, chromosomes 4 and 6 chromotripsis were always present in this cohort. Tumors belonging to the EMT-on group were enriched with SVs affecting the cluster of Interferon ligands genes on chromosome 4C4, synthenic to human 9p21 (P < 0.05), mostly leading to intragenic inversions and deletions through intra-exonic breakpoints, affecting genes integrity (Figure 4 c-d). Furthermore, visualization of the chromotripsis regions on chromosomes 6 revealed copy number amplification of oncogenic drivers, particularly gain of the Kras oncogene (Figure 4 e). These data are strongly suggestive of an active role of EMT in shaping pancreatic cancer genomic evolution through the acquisition of key driver events^27,28. We further leveraged a Zeb1 inducible knockout model of PDAC in a Kras^G12D and Trp53^R172H mutant background (KPCY-Zeb1^fl/fl), which has been previously described to affect the EMT process during tumor progression, as confirmed by histopathological analysis of tumor grade and Vimentin (VIM) positive tumor cells (Figure 4 f-h, Extended Fig. 10 a-b, Supplemental information)^29,30. In line with the previous evidence, WGS analysis of KPCY Zeb1^fl/fl and Zeb1^WT/WT models consistently showed increased CIN (P < 0.05), emergence of recurrent pattern of copy numbers associated with chromotripsis, and a higher magnitude of events of chromotripsis (P < 0.001) in tumors with unaltered Zeb1 (Figure 4 i-k, Extended Fig. 10 c-d). Overall, orthogonal approaches either ablating mesenchymal lineages or impairing the EMT process through conditional deletion of an EMT master transcriptional regulator demonstrate that EMT-on tumors recurrently acquire specific mutational features, converging on complex chromosomal structural rearrangements.

Figure 4. Acquisition of driver events during EMT in pancreatic cancer.

a) Bar plots showing fraction of mice with chromosome-specific events of chromotripsis among the EMT-on and the EMT-off group. b) Venn diagramm showing numbers of private and shared genes affected by chromotripsis among different SM-GEMM as calculated by WGS analysis. c) Bar plot displaying number of SVs per each mouse with breakpoints on the 4C4 cluster of genes of Interferon ligands (Ifn Cluster). N = 26 independent tumors (EMT on: N = 12; EMT off: N = 14) over N = 3 independent experiments. d) Representative inversion (top) and deletion (bottom) involving two genes of the Ifn Cluster. e) Representative case of chromotripsis of chromosome 6 in a mouse belonging to the EMT-on group. f) Representative H&E of KPCY Zeb1^WT/WT and Zeb1^fl/fl primary tumors. g) Representative immunofluorescence of YFP and VIM of primary tumors derived from the KPCY Zeb1^WT/WT and Zeb1^fl/fl GEMM. h) Quantification of YFP+/VIM− and YFP+/VIM+ tumor cells in the KPCY Zeb1^WT/WT and Zeb1^fl/fl GEMM (N = 5 lesions per group). P = 1.16624e-05. i) Representative bar plot showing copy number signatures for the KPCY Zeb1^WT/WT and Zeb1^fl/fl GEMM (right and left respectively. j-k) Box and whiskers plot with data points showing fCNA and number of chromotripsis for the KPCY Zeb1^WT/WT and Zeb1^fl/fl GEMM (N = 3 and 4 respectively). P = 0.0162, P = 0.0006. Scale bars = 100μm if not otherwise specified. Data are presented as mean values +/− SD. P values are calculated as follows: c) two-sided student t-test; h) two-way ANOVA; j-k) paired student t-test.

EMT fuels transcriptomic entropy

To understand the impact of mesenchymal plasticity on functional diversification and heterogeneity in PDAC, we performed single cell RNA and GET sequencing on short term cultures of sorted MS-L and EPI-L subpopulations isolated from the PCΩ SM-GEMM, which allows to infer chromatin velocity and genomic information at the single cell level³¹. Integration of epigenetic and transcriptomic profiles, confirmed the presence of 2 different states defined by the activation of EMT programs (Extended Fig. 11 a-d). Remarkably, cells with mesenchymal features display increased genomic instability, accumulation of genomic SVs and clonotypic diversity along the EMT continuum as displayed by integration of single cell RNA and GET sequencing data (P < 0.0001) (Figure 5 a, Extended data Fig. 11 e-g, Methods). Moreover, detailed inference of genomic complexity from scGET-seq data reveals an enrichment of SVs in chromatin regions with high accessibility, in particular centromeric and peri-centromeric chromosomal regions, overall characterized by a global increase in accessibility (Figure 5 b-f). These evidence suggests that alteration of pericentromeric chromatin compaction might result in delayed mitosis, lagging chromosomes and extranuclear chromatin sequestration which are hallmarks of CIN and permissive of catastrophic genomic events like chromotripsis^32–34.

Figure 5. Chromatin accessibility during EMT drives genomic instability and aggressive behavior of pancreatic tumors.

a) UMAP showing EMT Score, Genomic instability, Normalized SV number and clonotypic diversity as calculated by integration of scRNA and scGET seq data. b-c) Violin plots showing scGET seq accessibility and genomic instability values (b) and single cell correlation values for co-occurrence of SV in Minor and Major repeats and EMT continuum (c). N = 3920 cells over N = 2 independent experiments. P = 1e-15, 6.480e-7. d-e) UMAP showing number of SV-breakpoints (BND-residual) (d) and three-dimensional scatter plot showing single cell distribution, and correlation, among Accessibility, BND-residual and alpha diversity in sc-GET seq data (e). P = 1e-15. f) Pie chart showing correlations among EMT score, pseudotime and genomic instability with top significant transcription listed, based on expression and activity as calculated from scRNA and GET seq data. g) Representative picture (left) and quantification (right) of presence and number of micronuclei/area across EPI-L and MS-L subpopulations. N = 100 independent cells over N = 3 independent experiments. P = 3.71856e-06 h) Time-lapse frames (left) and quantification (right) of mitotic time in EPI-L (top) and MS-L (bottom) subpopulations. N = 32 independent mitosis over N = 3 independent experiments. P = 0.01009. i) Representative pictures and quantification of metaphasic spreads of EPI-L and MS-L subpopulations showing higher median number of chromosomes in MS-L subpopulation. N = 20 metaphases/group over N = 3 independent experiments. P = 1.673e-12. Scale bars = 100μm if not otherwise specified. Error bars represent SD. Box plots are centered at the 50^th percentile and bounds of box are fixed on the 25^th and 75^th percentile, min and max include all data points. P values are calculated as follows: P values are calculated as follows: b) Kolmogorov-Smirnov test; g,i) two-sided paired student t-test; e) two-tailed pearson correlation.

In line with this hypothesis, we demostate that MS-L subpopulations display a significantly greater number of micronuclei, prolonged time of mitosis, increased number of chromosomes per cell, larger nuclear sizes, an increased activation of DNA damage pathway as assessed by γH2AX staining, and sensitivity to the perturbation of the spindle assembly checkpoint, all indicative of mitotic and nuclear abnormalities leading to aneuploidy and complex structural chromosomal aberrations as previously described (Figure 5 g-i, Extended Fig. 11 h, Extended Fig. 12 a-b)^35,36. Complementing our observations, short-term induction of the EMT program in EPI-L models through TGF-beta stimulation leads to a significant increment of mitotic (P < 0.0001), nuclear abnormalities (P < 0.01), and chromatin accessibility at centromeric regions (P < 0.01) (Extended Fig 12. c-h)^37,38. To further prove that the acquisition of genomic instability during the EMT process is a functional driver of aggressive PDAC, we isolated ‘diploid’ and ‘aneuploid’ stable clones from freshly isolated MS-L cells (Extended Fig. 13 a-b). Orthotopic transplantations of diploid and aneuploid clones demonstrate that, indeed, genomic instability promotes aggressive behavior and increases metastatic burden upon transplantation (Extended Fig. 13 c-m). These findings suggest that aneuploidy in the MS-L functional compartment greately affects clinical progression of PDAC.

We further speculated that increased chromatin accessibility and mitotic errors in cancer cells may support the emergence of functional diversification in PDAC. Single cell RNA sequencing data were therefore obtained in the PCΩ model from 14 primary tumors and 6 metastatic sites and 3 experimental conditions: Vehicle, GCV and Recurrence after GCV pulse-chase (Extended Fig. 14 a, Supplementary information, Methods). After sorting of TdT+ populations (malignant compartment), and computational exclusion of TdT− cells and doublets, a total number of 33802 high confidence tumor cells were retrieved from both primary and metastatic lesions (Extended Fig. 14 b, Supplementary information). Deconvolution of cluster distribution, as extrapolated by single cell RNA sequencing data, demonstrates that tumor cells at the metastastic sites are enriched for mesenchymal features (Supplementary information), and that PDAC tumor heterogeneity is mostly due to the contribution of cancer cells from untreated (Vehicle) and recurrent (GCV-rec) tumors (EMT-on), while treated (GCV, EMT-off) tumors display low level of transcriptomic entropy, as calculated with two independent indeces (P<0.01) (Shannon and Simpson, Methods) (Extended Fig. 14 c, Supplementary information). Moreover, lineage trajectory analysis of independent tumors and associated metastasis revealed that EMT-on tumors are characterized by multiple transcriptomic branches (Extended Fig. 14 d, Supplementary information). Bioinformatic analysis of intra-group heterogeneity revealed the emergence of basaloid, stem-like, rhabdoid-like features, and overall increased phenotypic heterogeneity in the EMT-on tumors, while the EMT-off group displayed an enrichment of a ductal-like phenotype, as also confirmed by histopathological analysis and transcriptomic signatures of Classical-like PDAC (Extended Fig. 14 e, Supplementary information)¹⁹. Cross species analysis was performed leveraging a public single cell RNA sequencing dataset comprising of PDAC from a large cohort of patients (N = 24 primary tumors and 11 control pancreas)³⁹. After bioinformatic sorting through infer copy number calling, we retrieved 11854 bona fide cancer cells from 9 primary tumors (Extended Fig. 14 f, Methods). Next, we divided the cells derived from these tumors into two groups (EMT-high and EMT-low) based on the median values of a transcriptomic signature of EMT (Extended Fig. 14 g, Supplementary information)⁴⁰. Strikingly, scRNA sequencing analysis of this cohort recapitulates what observed in SM-GEMM, particularly in terms of increased tumor heterogeneity and transcriptional entropy in the EMT-high group (Extended Fig. 14 h, Supplementary information). In summary, cross-species analysis of transcriptomic trajectories strongly suggests that EMT is required for functional diversification in PDAC, resulting in the emergence of phenotypically different subpopulations that are suppressed during the genetic ablation of cells with mesenchymal competency.

Discussion

Our work clarifies the role of mesenchymal plasticity as a complex biological program observed in several solid tumor types, where it has been generally described as a response to external cues or during natural evolution of malignant tumors. Specifically, EMT has been observed to be a key feature influencing the metastatic process⁴.

While the major paradigm suggests that malignant epithelial cells must activate EMT to become invasive and metastatic, favoring cell motility over proliferation potential^5,41, we demonstrate through selective tracing and ablation of cell-state specific tumor lineages that the acquisition of mesenchymal features is necessary for malignant progression and tumor growth, promoting the emergence of high-fitness populations characterized by complex patterns of genomic instability. This is in line with numerous clinical and histopathological pieces of evidence showing that epithelial cancers, characterized by high grade histologies and/or sarcomatoid dedifferentiation, display increased mitotic rate, proliferative capacity and genomic instability^18,42,43. Indeed, in the context of PDAC, basal-like or quasi-mesenchymal tumors have been associated with increased proliferation rate and aggressive clinical behaviour^19,44.

The emergence of specific copy number signatures in a distinct malignant subpopulation suggests that diverse mutational processes are at play during cell plasticity dynamics, thus contributing to the complexity of genetic and non-genetic heterogeneity in cancers. It is reasonable to speculate that the poor progrosis universally associated with the emergence of mesenchymal features in solid tumors might be a consequence of the selection of phenomena associated with punctuated evolution followed by rapid clonal sweeps in specific transcriptomic/functional cell states^6,7. This is strongly supported by our experimental findings, where the selective ablation of the mesenchymal lineages in vivo results in relatively flat genomes and indolent disease. Our data are recapitulated by genomic and transcriptomic patients’ datasets, where lesions characterized by mesenchymal features are pervasively accumulating chromotripsis and other complex structural aberrations. Remarkably, such events are observed with the highest frequency in primary sarcomas resulting from the malignant transformation of mesenchymal or mesenchymal progenitor cells²⁶. Further studies are required to clarify whether the mesenchymal state (constitutive or acquired through the process of EMT) is universally more permissive or tolerogenic to complex and catastrophic structural rearrangments including chromothripsis, and whether other state-transitions promote the emergence of specific genetic evolutionary patterns.

Mechanistically, we discovered that structural genomic aberrations correlate with a “relaxed” and accessible chromatin state, particularly in the peri-centromeric and centrometic regions that significantly increase the chances of mitotic delays, chromosome lagging and structural rearrangements. Recent evidences, and our previous work, shed light on the critical role of chromatin remodeling during mitosis as a bookmark of lineage fidelity, in particular in the context of SWI/SNF deficient cancer cells^21,45. Of note, loss of markers of heterochromatin in distant sites has been previously associated with the emergence of tumor cells with metastatic potential, likely marking mesenchymal lineages during pancreatic cancer progression^34,46. Altogether, our work points towards an intertwined relationship between chromatin remodelling and the genomic evolution of aggressive neoplasms through acquisition of genomic instability⁴⁷, which concomitantly lead to functional diversification of cancer cells, and an overall increase in tumor heterogeneity.

In conclusion, our findings have important conceptual implications for the field of tumor evolution, particularly introducing the concept of “cell-state specific mutational processes”. Seminal work from Stratton and others has led to the identification of conserved signatures of mutational processes and structural variants in human cancers, informing on the effects of specific oncogenic insults on the genomic landscape of malignant tumors^26,48,49. Here we highlight a novel level of complexity, where different functional subpopulations within a tumor can evolve according to diverse mutational processes that demonstrate a degree of cell state specificity in such processes. These data add another layer of complexity to the concept of intratumoral heterogeneity and tumor cell plasticity as a driver of genomic and functional heterogeneity in PDAC and potentially other aggressive solid tumors.

Methods

Genetic engineered mouse models

The Vim^{FLEX(SA-H2B-eGFP-T2a-FlpO-WPRE-pA)} strain was generated by genOway, targeting exons 2 to 4 of the mouse Vim locus; the design has been carried out to minimize the risk of interference of regulatory elements. The Vim^{FLEX(SA-eGFP-T2A-HSV-TK-WPRE-pA)} strain was generated by genOway, targeting exons 2 to 4 of the mouse Vim locus; the design has been carried out to minimize the risk of interference of regulatory elements. The H11^LSL-Cas9 strain was generated by Dr. Monte M. Winslow and obtained through the Jackson Laboratory, stock No: 027632⁵⁰. The Rosa26^LSL-TdT was generated in Dr. Hongkui Zeng’s laboratory and obtained through the Jackson Laboratory, stock No: 007908⁵¹. The Rosa26^FSF-LSL-TdT was generated in Hongkui Zeng’s lab and obtained through the Jackson Laboratory, stock No: 021875⁵². The Kras^LSL-G12D was generated in Dr. Tyler Jacks lab and obtained through Jackson Laboratory, stock No: 008179⁵³. The Rosa26^LSL-PBT was kindly donated by Dr. Rolan Rad⁵⁴. The Tg(Pdx1-Cre) strain was generated by Dr. David Tuveson and obtained through Jackson Laboratory, stock No: 014647⁵⁵. The Ella^CRE was generated in Dr. Heiner Westphal’s laboratory and obtained through the Jackson Laboratory, stock No: 003724⁵⁶. The genotyping and characterization of the KPC (Tg(Pdx1-Cre), LSL-Kras^G12D/+, Trp53^R172H/+, Rosa-LSL-EYFP) and Zeb1^fl/fl mice have been described earlier^1,29. For details regarding the generation of the mouse models used for this study, refer to Supplementary Information. Strains were kept in a mixed C57BL/6 and 129Sv/Jae background. All animal studies and procedures were approved by the UTMDACC Institutional Animal Care and Use Committee. All experiments conformed to the relevant regulatory standards and were overseen by the institutional review board. Maximal tumor burden was not exceeded according to the institutional review board guidelines: for orthotopic tumors, mice were euthanized upon symptoms of distress. No sex bias was introduced during the generation of experimental cohorts. Littermates of the same sex were assigned randomly to experimental arms.

GEMM nomenclature

PCΦ: Rosa26^FSF-LSL-TdT, H11^LSL-spCas9, KRas^LSL-G12D, Vim^{FLEX(SA-H2B-eGFP-T2a-FlpO-WPRE-pA)}; PCΨ: Tg(Pdx1-Cre), Rosa26^LSL-PBT, H11^LSL-spCas9, KRas^LSL-G12D, Vim^{FLEX(SA-H2B-eGFP-T2a-FlpO-WPRE-pA)}; PCΩ: Rosa26^LSL-TdT, H11^LSL-spCas9, KRas^LSL-G12D, Vim^{FLEX(SA-eGFP-T2A-HSV-TK-WPRE-pA)}. PCΦ, PCΨ and PCΩ models were tested for the following combination of AAV-CRE: KPCC (sgTrp53, sgCdkn2a and sgCdkn2b); KCC (sgCdkn2a and sgCdkn2b); KP (sgTrp53). EMT-on group is composed by PCΨ-Vim^+/+ and PCΩ-Vehicle mice; EMT-off group is composed by PCΨ-Vim^Flpo/+ and PCΩ-GCV mice. For details of the generation of the mouse models used for this study, refer to Supplementary Information.

Animal procedures

Retrograde ductal injection: retrograde ductal injection was performed as previously described⁵⁰. Mice were anesthetized using isofluorane (Henry Schein Animal Health). Analgesia was achieved with buprenorphine ER (0.1 mg/Kg BID) (Par Parmaceutical) via subcutaneous injection, and shaved skin was disinfected with 70% ethanol and betadine (Dynarex). Laparotomy was performed through a midline incision of 1cm starting from the xyphoid process. The appendix was located and pulled out of the abdomen along with about half of the total bowel mass, exposing the duodenum. The sphincter of Oddi was exposed and injection of the AAV or Ad particles was performed with a 30-gauge needle, inserted through the sphincter of Oddi into the common bile duct up to where the cystic duct converges. About 100μL of 10¹⁰ viral particles resuspended in 1X PBS was injected. Then, the needle was slowly removed 30 seconds after completion of the injection. Muscular/peritoneal planes were closed individually by absorbable sutures. The skin/subcutaneous planes were closed using metal clips. Mice were monitored daily for the first three days, and twice/week.

Orthotopic parenchymal pancreatic injection: tumor-explant-derived three-dimensional cultures and genetically engineered pancreatic spheroids were resuspended in single cell suspension using trypsin-based method at a concentration of 1000 cells/μL in 1X PBS. Mice were anesthetized using isofluorane (Henry Schein Animal Health). Shaved skin was disinfected with betadine and 70% ethanol and 1-cm incisions were performed through the skin/subcutaneous and muscular/peritoneal layers. The spleen and tail of the pancreas were identified, exposed, and one injection was performed in the pancreatic tail and body. The muscular/peritoneal planes were closed using continuous absorbable sutures. The skin/subcutaneous planes were closed using interrupted absorbable sutures. Analgesia was achieved with buprenorphine ER (0.1 mg kg-1 BID).

Intrasplenic injections: for intrasplenic injections, tumor-explant-derived three-dimensional cultures were resuspended in single cell suspension using trypsin based method at a concentration of 1000 cells/μL in 1X PBS. Mice were anesthetized using isofluorane (Henry Schein Animal Health). Once the animals were under deep anaesthesia, the area of injection was disinfected with Betadine and 70% ethanol. The spleen was exposed through a small incision. Cells were injected into the spleen with a single injection using an insulin syringe. Cells were given 10 min to travel through the vasculature to the liver, after which the entire spleen was removed to prevent the formation of a large splenic tumour mass. To remove the spleen, a dissolvable 4–0 suture was tied snugly around the base of the spleen including the major splenic vasculature and the spleen was removed; to achieve hemostasis, a cauterizer was used when needed. The muscular/peritoneal planes were closed using continuous absorbable sutures. The skin/subcutaneous planes were closed using interrupted absorbable sutures. Analgesia was achieved with buprenorphine ER (0.1 mg kg-1 BID).

Intravenous injection: tumor-explant-derived three-dimensional cultures were resuspended in single cell suspension using trypsin-based method at a concentration of 1000 cells/μL in 1X PBS. Mice were anesthetized using isofluorane (Henry Schein Animal Health). Cells were intravenously injected through the lateral vein. Mice we monitored three times per week until moribound.

Treatments:

When average tumor size reached ~500mm³ of volume (around day 21 according to MRI imaging) mice received daily intraperitoneal injections with 50mg/kg of ganciclovir (GCV, Invivogen). Treatment schedules are specified for each experiment.

Euthanasia, necropsy, and tissues collection: mice were euthanized by exposure to CO2 followed by cervical dislocation. A necropsy form was filled in with mouse information, tumor size and weight, infiltrated organ annotations, and metastasis number and location. Fluorescent and brightfield images were acquired with Leica. Euthanasia was performed with animals at clinical terminal disease and metastatic tumor burden.

Recombinant DNA

Packages of one or more guide RNAs were designed and synthetized according to the following scheme: MluI restriction site-U6 promoter-gRNA1 sequence-gRNA scaffold-polyA-U6 promoter-gRNAn sequence-gRNA scaffold-polyA-KpnI restriction site. The synthetic sequence was assembled into the AAV:ITR-U6-sgRNA(backbone)-pCBh-Cre-WPRE-hGHpA-ITR vector (Addgene)⁵⁷ into the MluI and KpnI restriction sites. The Ad-PB::reverse complement TdT−(FLEX-FRT)reverse complement CAGG-DTA-polyA-U6 promoter-gRNA1 sequence-gRNA scaffold-polyA-U6 promoter-gRNAn sequence-gRNA scaffold–polyA was constructed as follows: first the PB inverse repeats 5’-reverse complement TdT−(FLEX-FRT)reverse complement CAGG-DTA-polyA-MluI restriction site-KpnI restriction site-PB inverse repeats 3’ was synthetized and clone into the pUC57 into the EcorI and XbaI restriction sites; second the U6 promoter-gRNA1 sequence-gRNA scaffold-polyA-U6 promoter-gRNAn sequence-gRNA scaffold–polyA was derived from the previously described AAV:ITR-U6-sgRNA(backbone)-pCBh-Cre-WPRE-hGHpA-ITR vector and cloned into the pUC57-PB inverse repeats 5’-reverse complement TdT−(FLEX-FRT)reverse complement CAGG-DTA-polyA-MluI restriction site-KpnI restriction site-PB inverse repeats 3’ into the MluI and KpnI restriction sites; the final construct was assembled into the Adeno-X adenoviral system (Takara Bio) via In-Fusion cloning as described by the manufacturer protocol.

Virus production

Plasmid DNA preparations were generated using endotoxin-free MIDI kits (Qiagen). Large-scale AAV and Ad particle production was outsourced to Vector Biolab (10^13 IU/mL).

Lentiviral particles were produced as previously described²¹. Briefly, infectious viral particles were produced using psPAX2 and pMD2G helper plasmids. For transfection, 293T cells were cultured in DMEM containing 10% FBS (Gibco), 100 IU ml⁻¹ penicillin (Gibco), 100 μg ml⁻¹ streptomycin (Gibco) and 4 mM caffeine (Sigma-Aldrich) and transfected using the polyethylenimine method. Virus-containing supernatant was collected 48–72 h after transfection, spun at 3,000 r.p.m. for 10 min and filtered through 0.45-μm low-protein-binding filters (Corning). High-titre preparations were obtained by multiple rounds of ultracentrifugation at 23,000 r.p.m. for 2 h each.

Non-invasive imaging

A 7T Bruker Biospec (BrukerBioSpin), equipped with 35mm inner diameter volume coil and 12 cm inner-diameter gradients, was used for MRI imaging. A fast acquisition with relaxation enhancement sequence with 2,000/39 ms TR/TE, 256x192 matrix size, r156uM resolution, 0.75 mm slice thickness, 0.25 mm slice gap, 40 × 30 cm FOV, 101 kHz bandwidth, and 4 NEX was used for acquired in coronal and axial geometries a multi-slice T2-weighted images. To reduce respiratory motion, the axial scan sequences were respiratory gated. All animal imaging, preparation, and maintenance was carried out in accordance with MD Anderson’s Institutional Animal Care and Use Committee policies and procedures.

Pancreatic organoids

Organoid cultures derived from the PCΦ and PCΨ were maintained as previously described with few modifications⁵⁸. Pancreata were isolated and ductal fragments were isolated by collagenase digestion (C9407, Sigma) for 30 min at 1 mg/ml. Fragments were seeded in growth factor-reduced Matrigel (Corning) and cultured in medium (DMEM/F12 supplemented with 1% penicillin/streptomycin (P/S), HEPES, GlutaMAX), with 2% B27 supplement (Gibco), recombinant mouse noggin (50ng/ml Preprotech), 10% Rspo1 (Millipore-Sigma, SCM104), EGF (50 ng/ml, Peprotech), FGF-10 (100 ng/ml, Preprotech), N-acetylcysteine (1.25 mM, Sigma), A8301 (5 μM, Tocris Bioscience) and primocine (0.1 mg/ml, Invivogen). After 2 weeks, pancreatic spheroids were cultured using DMEM/F12 supplemented with 1% P/S, 10% FBS. After 3 weeks, pancreatic spheroids were dissociated from Matrigel in ice cold PBS, collected and pelleted. Organoids were plated at high confluency in 96 wells with DMEM/F12 supplemented with 1% P/S, 10% FBS and incubated with AAV (10⁷ viral particles) for 8 hours at 37^o C 5% CO2. Single cell suspensions were collected, pelleted and embedded in Matrigel or transplanted, as previously described.

Tumor cell isolation, flow cytometry, cell sorting and culture

Ex vivo cultures from primary tumor explants were generated by mechanical dissociation and incubation for 1 hour at 37°C with a solution of collagenase IV/dispase (2 mg/ml) (Invitrogen), resuspended in DMEM (Lonza) and filtered. Cells derived from tumor dissociation and digestion were resuspended as single cell suspensions in 1X PBS and 5% FBS. Single cell suspensions were processed through fluorescent-activated cell sorting (FACS) using tdTomato fluorescence as a reporter for tumor cells. Single cell suspensions were sorted using a BD Aria II cell sorter. Sorted tdTomato+ cells were centrifuged at 200g per 5 minutes and resuspended in the appropriate medium.

Spheroid formation assay and treatment:

We refer to spheroid as those 3D structures embedded in MethoCult and derived from the PCΩ model. For sphere forming assay, 25–50k cells were carefully resuspended as single cell suspension in 300 mL of Stem Cell medium (SCM). Resuspended cells were then gently mixed with 1.5 mL of reconstituted MethoCult (StemCell Technologies, cat. n. M3134) and aliquot in a single well of an ultra-low attachment 6-well plate (Corning Costar, cat. n. 3471). Cells were kept at 37C and 5% CO2 for 5–10 days to allow sphere formation from single cells. Images of spheres were captured using Nikon microscope, sphere number and size were measured using NiS Element software (Nikon). Second generation spheres were plated as described above and treated with 10 microM GCV.

SCM preparation:

MEBM (Lonza, cat. n. CC-3153) was supplemented with 1% Penicillin/Streptomycin (100 U/ml), 2 mM L-Glutammine (Sigma-Millipore, cat.n. 59202C), 0.5 mM Hydrocortisone (Sigma-Millipore, cat.n. H0135), 1X serum-free B27 Supplement (Gibco, cat.n. 17504044), 20 ng/mL mouse recombinant EGF (Peprotech, cat.n 315–09), 20 ng/mL human recombinant FGF (Peprotech, cat. n. 100–18B), 1X ITS liquid media supplement (Sigma-Millipore, cat.n. I3146), 100 mM 2-mercaptoethanol (Sigma-Millipore, cat.n. M3148).

MethoCult reconstitution:

40 mL of Methocult were diluted to a final volume of 100 mL using MEMB and supplemented as described above for SCM.

EPI-L and MS-L isolation: sorted cells were cultured as two-dimensional cultures in 0.1% gelatin coated plates in DMEM High Glucose, 10% FBS, 1% P/S. After attachment and stabilization, cell cultures were either treated with 10μM GCV or with an equal volume of PBS in DMEM High Glucose, 10% FBS, 1% P/S to isolate EPI-L and MS-L populations respectively. Medium was changed every 24 hours for both EPI-L and MS-L cell cultures. Five passages after treatment, cells were subjected to downstream analysis.

Diploid and aneuploid MS-L isolation:

Short-term cultured (p < 10) MS-L were dissociated and resuspended at single cell suspensions in DMEM High Glucose, 10% FBS, 1% P/S. Cells were sorted in a 96 well plate (Corning) counting ~0.5 cells/well. Two and ten passages after plating, cells were fixed with methanol 100% and stained with DAPI. DNA content was analyzed using a BD FACS Canto II flow cytometer. MS-L clones with stable DNA content at passage 2 and 10 were used for downstream experiments.

Single-guide RNA design and shRNA

Single-guide RNAs (sgRNAs) sequences used in this study were obtained through Genescript and validated elsewhere⁵⁹.

Single-guide RNA sequences: Trp53: CATAAGGTACCACCACGCTG, Cdkn2a: GTGCGATATTTGCGTTCCGC, Cdkn2b: GGCGCCTCCCGAAGCGGTTC.

shRNA plasmids (pLKO.1) targeting Mad2l1 (TCR clone ID: TRCN0000012839) and non-targeting (SHC016) were purchased through Sigma-Aldrich.

Staining

Immunohistochemistry (IHC) and immunofluorescence were performed as previously described⁵³. Antibodies list: RFP (Thermo Fisher, cat. #MA5–15257, 1:200 dilution), GFP (Abcam, cat. #13970, 1:200 dilution), CK19 (Proteintech, cat #10712–1-AP), gamma-H2A.X (Abcam, cat #ab26350, 1:200) phospho-Histone H3 (Cell Signaling, cat#9701, 1:400 dilution), PCNA (Cell Signaling, cat#2586, 1:400 dilution), GATA6 (ProteinTech cat#55435–1AP, 1:500 dilution), KRT17 (ProteinTech cat#17516–1-AP, dilution 1:250), Vimentin (Cell Signaling, cat# 5741, dilution 1:200).

Multispectral imaging using the Vectra. Microwave treatment (MWT) was applied to perform antigen retrieval, quench endogenous peroxidases, and remove antibodies from earlier staining procedures. Akoya Biosciences AR6 antigen retrieval buffer (pH 6) was used for vimentin and RFP staining. The slides were stained with primary antibodies against RFP, GFP, and Vimentin, corresponding HRP conjugated secondary antibodies, and subsequently TSA dyes to generate Opal signal (vimentin, Opal 570; RFP, Opal 620; and Pax8, Opal 690). The slides were scanned with the Vectra 3 image scanning system (Caliper Life Sciences), and signals were unmixed and reconstructed into a composite image with Vectra inForm software 2.4.8.

Metaphase spread and chromosome count

Giemsa staining on metaphasic spread was obtained as previously described with few modifications. Cultures were treated with 100 ng/ml nocodazole for 8 hours overnight, collected by trypsinization, resuspended in 0.2% (w/v) KCl and 0.2% (w/v) trisodium citrate hypotonic buffer at room temperature (20–22°C) for 5 to 10 min, and cytocentrifuged onto SuperFrost Plus glass slides (MenzelGlaser) at 450g for 10 min in a Shandon Cytospin 4. Slides were fixed at room temperature for 10 min in 1× PBS with 4% (v/v) formaldehyde, permeabilized for 10 min at room temperature in KCM buffer (120 mM KCl, 20 mM NaCl, 10 mM Tris (pH 7.5) and 0.1% (v/v) Triton X-100), and blocked with 5% goat serum PBS1X 0.1%Triton 100x BSA 3% for 30 min at room temperature. Slides were incubated with Giemsa solution for 5 minutes at room temperature.

Drug and shRNA response assays

For drug experiments, cells were plated at 2 × 10³ cells per well in a 6-well plate and treated with either vehicle, Reversine or MSP1i (MPI-0479605) for 14 days. Fresh medium with or without drugs was changed daily. MPI-0479605 was purchased from MedChem Express and reversine was purchased from Sigma-Aldrich. Following incubation with the drug, viability was assessed either by crystal violet staining (Sigma). Absorbance was quantified using an Envision Plate Reader (PerkinElmer). Experiments were performed in triplicate. For shRNA response assays, cells were plated 2 × 10³ cells per well in a 6-well plate and transduced with lentivirus harboring shRNA agains Mad2l1 or non-targeting. Fourteen days after incubation viability was assessed by crystal violet staining (Sigma). Absorbance was quantified using an Envision Plate Reader (PerkinElmer). Experiments were performed in triplicate.

For mouse recombinant TGF-beta treatment, EPI-L cells were seeded at 5 × 10⁵ cells per petri dish in a 100 mm petri dish and treated with either vehicle or mouse recombinant TGF-beta at a concentration of 20 ng/mL for 48 h. Medium was changes with fresh mouse recombinant TGF-beta every 24 hours. Mouse recombinant TGF-beta was purchased from R&D Systems, Inc (catalog # 7666-MB).

Next-generation sequencing of murine DNA

Whole genome-libraries were prepared and sequenced using a modified protocol⁶⁰. For murine whole-genome sequencing, after adapter ligation, libraries were only amplified by 2 cycles of PCR. Equimolar quantities of the whole-genome indexed libraries were multiplexed, with 18 libraries per pool. Results from 13 of the 18 libraries were used in our analysis. All pooled libraries were sequenced on an Illumina NovaSeq6000 using the 150 bp paired-end format.

Bioinformatic processing of high-throughput sequencing data

The bioinformatic processing pipeline of raw whole-genome sequencing (WGS) data was adapted for murine data⁶¹. Sequencing fasta files were applied to FastQC for quality controls, adapters were trimmed by Trimmomatic⁶². Reads were aligned to the mouse genome reference (mm10) using Burrows-Wheeler Aligner (BWA, https://github.com/lh3/bwa ) with a seed length of 40 and a maximum edit distance of 3 (allowing for distance % 2 in the seed). BAM files were further processed according to GATK Best Practices, including removal of duplicate reads, realignment around indels, and base recalibration.

Analysis of sgRNA performance

Expected cut sites of sgRNAs were analyzed using CRISPResso2 in “CRISPRessoWGS” mode. BAM files were first filtered with SAMtools to contain reads spanning a 50-bp region centered around the expected sgRNA cut site and passed to CRISPResso2 in “CRISPRessoWGS” mode. The allele frequency of each base position around the cut site window was extracted from the CRISPResso2 results⁶³.

Identification and characterization of somatic mutations

Somatic mutations were detected from murine tumor samples using a combination of MuTect v2⁶⁴ to call somatic single-nucleotide variants (SNVs). Tumor samples from WGS were compared to their respective matched healthy tissue. All mutations were also filtered for depth (tumor sample coverage > 20x, normal sample coverage > 10x) and variant allele frequency (VAF ≥ 0.1). All mutations annotated to genomic regions not targeted by a sgRNA detected in at least one sample were kept. Mutational signatures were called using SigProfilerExtractor⁶⁵.

Identification of somatic copy number, structural variants and chromotripsis.

Sequenza⁶⁶ was used to derive somatic copy number profiles from WGS data using each sample’s matched normal sample. A modified version of the “copynumber” R package (https://github.com/aroneklund/copynumber ) was used to create the necessary support objects and allowed for usage of the mm10 genome in the deployment of Sequenza on the mouse BAM files. Structural variations were called by Manta⁶⁷ applying default settings. Structural variants were filtered for length < 6,000 bp. Chromotripsis events were called using ShatterSeek and following PCAWG guidelines²⁶: high-confidence chromotripsis were called used ShatterSeek with visual inspection of number of structural variants > 6 and number of oscillating copy number > 6. Circos plots were generated with OmicCircos⁶⁸. Copy number signatures were called using SigProfilerExtractor⁶⁵.

TCGA WGS and RNA data analysis

Chromothripsis score was calculated as follows: chromothripsis information of TCGA dataset were downloaded and summarized based on the label information for each chromosome then assign discete values: High confidence = 1; Linked to high confidence = 0.75; Low confidence = 0.5; Linked to low confidence = 0.25; No = 0. For each sample, a final value was calculated by multiplying the sum of these values by the number of chromosomes affected. EMT score was extrapolated from Gibbons and Creighton⁴³.

Slide-DNA-seq

Slide-DNA-seq was performed as previously described²³. Bead arrays were produced and sequenced as described in Stickels et al, 2022. The bead oligonucleotide sequence is specific to slide-DNA-seq, and was produced in-house with the following nucleotide sequence:

5'-TTT_PC_GCCGGTAATACGACTCACTATAGGGCTACACGACGCTCTTCCGATCTJJJJJJJJTCTTCAGCGTTCCCGAGAJJJJJJJNNNNNNNVVGCTCGGACACATGGGCG-3'

Where PC designates a photocleavable linker, J bases represent the spatial barcode (unique to each bead), N bases represent a unique molecular identifier (unique to each oligo), and V represents randomized bases that exclude T. Tissue handling, library preparation, and analysis were as described in Zhao et al., 2022, with the following modification. During the tagmentation stage, the Tn5 transposase solution was refreshed every hour, for a total of four rounds of tagmentation. All other experimental steps proceed exactly as previously described. Analysis steps were also performed as previously described, with scripts available at https://github.com/buenrostrolab/slide_dna_seq_analysis.

scRNA-seq and GET-seq of EPI-L and MS-L

scRNA-seq and scGET-seq were performed as previously described^31,69. Briefly, scRNA-seq was performed on a Chromium platform (10x Genomics) using ‘Chromium Single Cell 3ʹ Reagent Kits v3’ kit manual version CG000183 Rev C (10x Genomics). Final libraries were loaded on a Novaseq6000 platform (Illumina) to obtain 50,000 reads per cell for scRNA-seq and 100,000 reads per nucleus for scGET-seq.

Processing of scRNA-seq data of EPI-L and MS-L

Cell barcodes were identified and quantified using UMI-tools⁷⁰, setting the expected number of cells to 5,000. Read tags from 10x Chromium libraries were aligned to mm10 reference genome using STARsolo v2.7⁷¹. Count matrices were imported into scanpy objects⁷². Doublets were identified with scrublet⁷³ using default parameters. After concatenation of all data samples, predicted doublets, cells with more than 20% counts assigned to mitochondrial genes and cells with less than 2,000 genes by counts were removed from the analysis. Counts were normalized and log1p transformed before cell cycle phase assignment. After that we calculated cell cycle progression value (CC_diff) as G2M_score – S_score. We regressed out CC_diff and %MT before scaling.

We calculated cell kNNgraph using BBKNN⁷⁴ setting the number of neighbors to half the square root of the number of cells. Cell clusters were identified using schist⁷⁵ with default parameters. CytoTrace score⁷⁶ was calculated using cellrank⁷⁷. Simpson and Shannon indeces were calculating on the frequency of cells belonging for each cluster in each group.

To compute the cell-to-cell transition matrix we applied a velocity kernel with dynamical scoring (vk) implemented in cellrank using the experimental time to set the progression. We also applied a connectivity kernel (ck) to stabilize the Schur decomposition. Finally, the transition matrix was calculated using a weighted kernel (0.8vk + 0.2ck). GPCCA method was then used to identify two major cell macrostates, assigned to mesenchymal or epithelial phenotypes. To quantify EMT gene signature we considered the difference of mean expression of genes up and down-regulated⁷⁸.

Activity of transcription factors (TF) was evaluated using decoupleR⁷⁹ with default parameters.

Processing of scGET-seq data of EPI-L and MS-L

Raw reads were preprocessed as previously described⁶⁸ using the scGET workflow⁷⁵ with default parameters. Each sample consists of two matrices representing read counts for the mm10 genome binned at 5 kbp, one for Tn5 and one for TnH. Once count matrices have been generated, we fitted a zero-inflated Poisson distribution on Tn5 and TnH counts per each cell and generated the empirical distribution of the difference between the two transposases. Using such per-cell distribution, we computed the p-value of the difference Tn5-TnH for each genomic bin in each cell. p-values for each tail were log-transformed as previously detailed, so that accessible regions are assigned a positive value and compact regions are assigned a negative value.

We removed regions that were not profiled in at least 90% of the cells, we removed cells with a total coverage lower than 2,000. Reduction of dimensionality was achieved by tensor train decomposition (TTD) using tensorly python library. The first 20 components were used as basis to compute the cell kNN graph, using BBKNN⁷⁴ with cosine distance (‘angular’) and number of neighbors equal to half the square root of the number of cells.

Chromatin Accessibility Score (CAS) was obtained as fraction of regions with enriched Tn5 over TnH, as results of the signed log(p-value) previously calculated.

We applied dot product of the matrix of signed p-values with the adjacency matrix of the kNN graph to obtain a smoothed matrix (Macc) of net accessibility values. This matrix was used to calculate a velocity kernel (vk) using cellrank. We regularized this kernel with a connectivity kernel (ck) (0.8vk + 0.2ck) before applying GPCCA method. Finally, two major cell macrostates were identified, assigned to mesenchymal or epithelial phenotypes.

To calculate the activity of TF from scGET-seq data we selected the top 200 genomic regions that could be considered drivers of each macrostate using cellrank correlation test. For each region we calculated the Total Binding Affinity⁸⁰ of the whole set of transcription factors included in HOCOMOCO v11⁸¹.

To profile genome instability, we applied delly⁸² on aligned TnH reads to retrieve a list of structural variants (SV). As delly returns the name of reads supporting each SV, we matched read names to single cells to build a count matrix of SV. We applied a linear model fitting the log10-coverage (as number of TnH reads per cell) with the log(nSV + 1) (as number of SV per cell). We used the residuals of the fit as measure of genome instability. Positive values indicate an excess of SV at a given coverage (hence high instability).

To characterize the genetic heterogeneity in each group, we characterized Copy Number Alterations in single cells. To this end, we binned genome into 10 Mb windows and computed the per-cell coverage, then we fit ZIP on the coarse count matrix and calculated the hyperbolic arcsine of the ratio between counts and the expected number λ from the Poisson distribution, as a proxy of CNA values. The new matrix was processed do identify cell clusters using schist⁷⁵. The size of intersections of such clusters with the macrostates was used to calculate ²D Hill numbers, α-diversity and β-diversity indices.

Combined analysis of scGET-seq and scRNA-seq data of EPI-L and MS-L

To integrate the scGET-seq and scRNA-seq data we applied MOWGAN⁸³, deep learning framework based on Generative Adversarial Network (WGAN-GP) that uses Wasserstein loss function and gradient penalty. The WGAN-GP training is performed in mini-batches, where cells are sampled from the whole dataset (i.e., from each molecular layer). Input data for WGAN-GP are the single cell embeddings (PCA for scRNA-seq and TTD for scGET-seq), clipped to the same number of dimensions. Cells for each are sorted according to the first component of the Laplacian Eigenmap. Sampled data are combined in a tensor of shape (𝑁, 2, 𝐶), where 𝑁 is the number of cells in the mini-batch (256) and 𝐶 is the number of components in each embedding used for the analysis. Mini-batches are sampled from the same experimental batches to avoid spurious associations. We select N cells from the scRNA-seq layer, then we select n=50 random mini-batches from scGET layer and run Bayesian Ridge regression to fit the best coupling. This step is repeated for every mini-batch that is used during training. The network generator is designed with three convolutional 1D layers and two batch normalization layers. The network critic is designed with two Conv1D layers and a Dense layer with 1 unit. All Conv1D layers are characterized by a kernel size of 𝑀, stride 1 and the ReLU activation function. Finally, different optimizers are used for each component: Adam optimizer for the generator with learning rate = 0.001, beta_1 = 0.5, beta_2 = 0.9, epsilon = 1𝑒 − 07 and the AMSgrad option; RMSprop optimizer with learning rate = 0.0005 for the critic. Once training has been performed, the MOWGAN can generate embeddings for synthetic cells that are artificially paired. For each modality, we created a single scanpy object for including the synthetic and the original data and applied Harmony⁸⁴ to perform the integration. The new embeddings produced by Harmony were used to generate the kNN graphs for each modality. To identify cell states consistent across both modalities, we applied the multilevel Nested Stochastic Block Model, for which a single kNN is created with labeled edges corresponding to each modality.

Processing and analysis scRNA seq data for EMT-on and EMT-off

Raw scRNA-seq fastq files were pre-processed and aligned to mm10 reference genome for mouse scRNA-seq data, and hg38 for human scRNA-seq data using CellRanger Single Cell Software Suite (v. 6.0, 10x Genomics)⁸⁵. Count matrices were imported into Seurat (v.4.0) for quality control and clustering⁸⁶. Doublets were identified with scDblFinder using default parameters⁸⁷. After concatenation of all data samples, removal doublets, cells with more than 25% counts assigned to mitochondrial genes and cells with less than 500 genes by counts were removed from the analysis. Harmony was used for batch effect correction (https://github.com/immunogenomics/harmony?tab=readme-ov-file ). STREAM was applied for single-cell trajectory reconstruction sample wise⁸⁸. Simpson and Shannon indeces were calculated on the frequencing of cells belong for each cluster per each group. To quantify EMT gene signature we considered the difference of mean expression of a previously described signature of EMT⁴⁰. To calculate scClassical and scBasal signatures, we used previously described methodologies⁸⁹. A similar approach was used to analyze human scRNA seq data from Peng et al. To include only tumor cells in our analysis we applied SCEVAN (https://github.com/AntonioDeFalco/SCEVAN ) to retrieve bona fide cancer cells based on infer copy number profiles.

Bulk ATAC seq processing and analysis.

ATAC-Seq library preparation was performed at the MD Anderson Epigenomics Profiling Core following the protocol as previously described^90,91. Briefly, nuclei isolated from EPI-L short-term cultures treated with either vehicle or TGF-beta, as previously described, were incubated with Tagment DNA TDE1 enzyme (Illumina), followed by DNA purification and library preparation. The resulting libraries were sequenced 2 × 50 bp on an Illumina NovaSeq6000.

Raw fastq files were applied to FastQC for quality control, adapters were trimmed by Trimmomatic⁶². Reads were aligned to the mouse genome reference (mm10) using Burrows-Wheeler Aligner (BWA), then sorted and indexed by Samtools. PCR duplicates were marked by Picard. After alignment, peaks were called by macs2⁹², and then annotated by HOMER⁹³.

Statistics and Reproducibility

Data are presented as the mean or medians ± s.d and percentages. Comparisons among biological replicates were performed using two-tailed Student’s t test, two-way ANOVA, Mann-Whitney U test. Results from survival experiments were analyzed with log-rank (Mantel–Cox) test and expressed as Kaplan–Meier survival curves. Results from contingency tables were analyzed using two-tailed Fisher’s exact test or chi-test for multiple comparisons. All the statistical analysis have been performed with GraphPad Prism software. Data distribution was assumed to be normal without formal testing. Group size was determined on the basis of the results of preliminary experiments. No statistical methods were used to determine sample size. No data were excluded from the analysis. Group allocation and analysis of outcome were not performed in a blinded manner. In vitro experiments were repeated three times, while in vivo experiments were performed at least twice.

Extended Data

Extended data Figure 1. Clinico-pathological characterization of the PCΦ SM-GEMM.

a) Representative macroscopic pictures of primary tumors (PT) and liver metastasis (Liver Met) of PCΦ SM-GEMM engineered with different combinations of driver events. b) Survival curves of PCΦ SM-GEMM engineered with different combinations of driver events (KPCC: N = 76; KCC: N = 36; KP: N = 24). P = 1.29752e-07, 6e-15, 3.70032e-08. c) Bar plot showing percentage of mice engineered with different combinations of driver events with either normal pancreas (Normal) or pancreatic tumor (PDAC) at endpoint (KPCC: N = 76; KCC: N = 36; KP: N = 24). P = 2.19655E-06, 5.161e-12. d) Scatter plot showing number of metastasis of PCΦ SM-GEMM engineered with different combinations of driver events (KPCC: N = 76; KCC: N = 36; KP: N = 24). e) Scatter plot showing number of TdT+ and TdT− metastasis of PCΦ SM-GEMM engineered with different combinations of driver events (KPCC: N = 76; KCC: N = 36; KP: N = 24). P = 0.0011, 1e-15, 1e-15. f) Scatter plot showing number of TdT+/GFP− and TdT+/GFP+ metastasis of PCΦ SM-GEMM engineered with different combinations of driver events (KPCC: N = 76; KCC: N = 36; KP: N = 24). P = 0.0383. g) Representative primary tumors histopathology pictures of PCΦ SM-GEMM engineered with different combinations of driver events (H&E: hematoxilyn and eosin; MTS: Masson thrichromic staining). h) Bar plot showing quantification of tumor grade of primary tumors of PCΦ SM-GEMM engineered with different combinations of driver events (KPCC and KCC, N = 10 different primary lesions; KP, N = 9 different primary lesions). i) Bar plot showing quantification of aniline blue positive and negative areas of primary tumors of PCΦ SM-GEMM engineered with different combinations of driver events (KPCC and KCC, N = 10 different primary lesions; KP, N = 9 different primary lesions). j) Bar plot showing quantification of TdT+/GFP− and TdT+/GFP+ tumors cells of primary tumors of PCΦ SM-GEMM engineered with different combinations of driver events (KPCC and KCC, N = 10 different primary lesions; KP, N = 9 different primary lesions). Data are shown as mean with standard deviation. Scale bars = 100μm if not otherwise specified. P values are calculated as follows: b) Logrank Mentel-Cox test; c) Fisher’s exact chi square test; d-f) two-sided paired student t-test; h-j) two-way ANOVA.

Extended data Figure 2. Pancreatic tumors with mesenchymal competency display aggressive clinical behavior.

a) Histopathological characterization of primary tumor site and metastatic sites of PCΦ SM-GEMM showing distribution of TdT+/GFP^HIGH and TdT+/GFP^LOW cells in tumoral areas. Images are representative of N = 10 independent lesions. b) Histopathological characterization of PCΦ SM-GEMM at different clinical stages showing concomitant expression of TdT, GFP and KRT19 tumor cells markers (N = 5 different lesions/group). P = 0.001, 1e-15. c-e) Schematic and representative pictures of the isolation of TdT+ and TdT− cells from PCΦ SM-GEMM-derived organoids (c-d) and representative scatter plot of gating strategies (e). Scatter plot is representative of N = 2 sorting experiments. f-g) Survival curves and metastasis (f) quantification of metastasis (g) of PCΦ SM-GEMM transplanted with either TdT+ or TdT− organoid-derived tumor cells showing reduced survival and metastatic burden in mice bearing TdT+ derived lesions (N = 10 mice/group). P = 4.495e-6, 0.019. h) Representative histopathological pictures of tumors derived by TdT+ or TdT− cells: a tendency to sarcomatoid transformation can be appreciated in lesions derived from TdT+ cells; noticeably, TdT− cells derived tumors evolve to TdT+ lesions. Images are representative of N = 2 experiments. i-j) Representative microscopic pictures of organoids derived from PCΨ Vim^+/+ (right) and Vim^Flpo/+ (left) SM-GEMM (i) and survival curves of these organoids transplanted into the pancreata of recipient mice (j). N = 20 mice per group; data are representative of N = 2 experiments. P = 1.2124E-11. k) Representative immunofluorescence (left) and grading (right) of tumors arose from transplanted organoids derived from PCΨ Vim^+/+ (top) and Vim^Flpo/+ (bottom) SM-GEMM. N = 25 tumor areas. P = 5.32527e-08. l) Histopathological quantification of cells co-expressing VIM and TdT in tumors arose from transplanted organoids derived from PCΨ Vim^+/+ and Vim^Flpo/+SM-GEMM. N = 25 tumor areas. P = 1.74523e-10. Scale bars = 100μm if not otherwise specified. P values are calculated as follows: b) one-way ANOVA with multiple comparisons (Tukey’s multiple comparisons test); f,j) Logrank Mentel-Cox test; g) paired student t-test k) Fisher’s exact chi square test; l) two-way ANOVA.

Extended data Figure 3. Clinico-pathological characterization of the PCW SM-GEMM.

a) Macroscopic pictures of primary tumors and metastatic lesions of a PCΩ SM-GEMM. Pictures representative of N = 20 mice. b) Box plot showing primary tumor dimensions for Vehicle, GCV and GCV-rec cohorts (top, KCC; bottom KP. N = 10 mice/cohort for KCC; N = 5 mice/cohort for KP). KCC, P = 1e-06; KP, P = 0.004, 0.0199, 1e-06. c) Box plot showing number of metastasis for Vehicle, GCV and GCV-rec cohorts (top, KCC; bottom KP. N = 10 mice/cohort for KCC; N = 5 mice/cohort for KP). KCC, P = 0.003, 0.001; KP, P = 0.0089. d) Survival curves of the Vehicle and GCV-rec cohorts, start and end of treatments are highlighted by green and red arrows respectively. (top, KCC; bottom KP. N = 10 mice/cohort for KCC; N = 5 mice/cohort for KP) KCC, P = 0.00096; KP, P = 0.0018. e) Representative histopathological analysis of the different cohorts of the KPCC PCΩ SM-GEMM; pictures are representative of N = 5 mice. f) Bar plot showing quantification of EPI-L and MS-L (TdT+/GFP− and TdT+/GFP+ respectively) of primary tumors for the different cohorts of KPCC PCΩ SM-GEMM (N = 25 lesions/cohort). P = 1e-15, 2e-15, 1e-15, 1e-15. g) Bar plot showing quantification of EPI-L and MS-L (TdT+/GFP− and TdT+/GFP+ respectively) of primary tumors for the different cohorts of KCC PCΩ SM-GEMM (left) and KP PCΩ SM-GEMM (right) (N = 25 lesions cohort). KCC, P = 0.0012, 0.026, 3.10031e-06; KP, P = 1.98129e-06, 0.0406, 0.00045. Scale bars = 100μm if not otherwise specified. Error bars represent SD. Box plots are centered at the 50^th percentile and bounds of box are fixed on the 25^th and 75^th percentile, min and max include all data points. P values are calculated as follows: b,c) one-way ANOVA with multiple comparisons (Tukey’s multiple comparison test); d) Logrank Mentel-Cox test; f-g) two-way ANOVA.

Extended data Figure 4. Cells with mesenchymal features are required for the expansion of tumor engraftments.

a) Representative pictures of pancreatic tumor-derived 3D cultures generated from PCΩ SM-GEMM. Images representative of N = 3 independent primary cell lines. b-c) Scatter plot with individual points (left) and box and whiskers plot showing area and number of pancreatic spheres upon GCV treatment. N = 32 fields per group (b) and N = 10 fields per group. P = 1.4094e-06. d) Schematic showing transplantation experiments and treatment schedule of 3D cultures generated from PCΩ SM-GEMM. e) Representative macroscopic pictures of primary tumors for the three groups. f) Scatter plots with individual values calculating tumor volume and number of metastasis for the same groups. Images are representative of 30 mice; N = 10 per group. P = 8.40064e-07, 1.19482E-08, 0.0071, 0.0002. g) Survival curves of mice bearing transplanted pancreatic tumor-derived 3D cultures upon orthotopic transplantation and GCV treatment. N = 20 independent mice over N = 2 independent experiments. P = 3.51436e-06. h-i) Representative macroscopic and microscopic pictures (left) and quantification of number of nodules (right) of pancreatic tumor-derived 3D cultures transplated via spleen injection and treated or not with GCV. Images are representative of N = 15 mice; N = 5 per group. P = 0.002 j) Survival curves of mice bearing pancreatic tumor-derived 3D cultures upon splenic transplantation and GCV treatment. N = 5 per group. P = 0.002 k-l) Representative histopathological pictures (left) and number of microscopic nodules (right) from mice transplated in the lateral vein with pancreatic tumor-derived 3D cultures. Images are representative of N = 15 mice; N = 5 mice per group. Scale bars = 100μm if not otherwise specified. Error bars represent SD. Box plots are centered at the 50^th percentile and bounds of box are fixed on the 25^th and 75^th percentile, min and max include all data points. P values are calculated as follows: b,c,f,i,l) two-sided student t-test; g, j) Logrank Mentel-Cox test.

Extended data Figure 5. Tumors with sarcomatoid features display increased aberrant mitotic events.

a) Schematic showing experimental workflow. b) Survival curves showing improved survival for mice in the DTR-tr rec and GCV-tr rec group. N = 10 mice/group over N = 2 independent experiments. P = 0.0004, 4.83711e-06, 0.0128. c-d) Box and wiskers with data point showing tumor dimension (c) and number of metastasis (d) in the different groups. N = 10 mice/group over N = 2 independent experiments. c, P = 7.17017e-06, 2.65958e-05, 7.38067E-05; d, P = 4.5863E-05, 0.00011, 8.12171e-05. e) Bar plot showing distribution of EPI-L and MS-L population among the different groups. N = 5 lesions/group. P = 2.72642e-05, 1.92443e-05, 0.0027. f) Representative histopathological pictures showing patient-derived pancreatic tumors pathologically annotated for sarcomatoid features and high magnification of normal and aberrant mitosis. g) Box and wiskers with data points showing number of mitosis in patient derived pancreatic tumors annotated with sarcomatoid features. h) Bar plot showing distribution of normal and aberrant mitosis in patient derived pancreatic tumors annotated with sarcomatoid features. Error bars represent SD. Box plots are centered at the 50^th percentile and bounds of box are fixed on the 25^th and 75^th percentile, min and max include all data points. P values are calculated as follows: c,d, g) two sided paired student t-test; h) Fisher’s exact chi square test.

Extended data Figure 6. Spatial genomics of tumor cells with mesenchymal proficiency.

a-b) Slide-DNA-seq of two PCΦ SM-GEMM 3 weeks after AAV transduction showing TdT+ and TdT− regions of pancreatic tissue, differentially clustering according to genomic analysis. c-d) Normalized copy number scatter plots showing differential copy number aberrations in the three clusters as previously calculated with principal component analysis; particularly evident are gains of 6 and 15 and minor loss of chromosome 4. e) Heatmap showing genomic characteristics of tumors belonging to the EMT-on and EMT-off groups together with analysis of putative driver events as expected by genomic engineering. f) Bar plot showing percentage of distribution of single nucleotide variants in EMT-on and EMT-off groups. It can be readily appreciated that the two groups are characterized by similar mutational signatures belonging to clock-like types of mutagenesis (SBS5 and SBS1).

Extended data Figure 7. Structural variants and chromotripsis in KPCC EMT-on and EMT-off groups.

a) Box and whiskers with individual points showing number and type of SVs in the PCΨ Vim^+/+ and Vim^Flpo/+ SM-GEMMs. N = 7 mice per group. b) Two representative scatter plots showing absolute copy number values of chromosome 4 for two different mice belonging to the EMT-on (top) and EMT-off (bottom groups). Poliploidy and focal copy number gains can be readily appreciated. c) ShatterSeek output plots for 4 different mice belonging to the EMT-on (left) and EMT-off (right) group. Error bars represent SD. Box plots are centered at the 50^th percentile and bounds of box are fixed on the 25^th and 75^th percentile, min and max include all data points. Visual inspection, SVs count and number of oscillating copy number (Oscillating CN) were used to call high-confidence chromotripsis: the two cases on the left met the criteria (Methods).

Extended data Figure 8. Genomic features of KP and KCC EMT-on and EMT-off groups.

a) Representative circos plots of genomic features of mice belonging to the KP and KCC genotype in the EMT-on and EMT-off experimental groups. b) Bar plot with individual point showing fCNA values for tumors belonging to the KP (left) and the KCC (right) genomic backgrounds. N = 27 independent mice (KP: N = 10; KCC: N = 17) over N = 2 independent experiments. P = 0.015, 0.027. c) Bar plot with individual points showing SV load values for tumors belonging to the KP and the KCC genomic backgrounds. N = 27 independent mice (KP: N = 10; KCC: N = 17) over N = 2 independent experiments. P = 0.0063, 1e-05. d) Bar plot with individual points showing number of chromotripsis for tumors belonging to the KP and the KCC genomic backgrounds. P = 0.0354, 0.002. e) Representative ShatterSeek plots for tumors belonging to the KP and the KCC genomic backgrounds. N = 27 independent mice (KP: N = 10; KCC: N = 17) over N = 2 independent experiments. Data are presented as mean values +/− SD. P values are calculated as follows: b) two-sided Mann-Whitney test; c) two-sided Tukey’s multiple comparison test, d)two-sided paired student t-test.

Extended data Figure 9. Transcriptomic signatures of EMT correlate with genomic instability in human cancers.

a) Scatter plot showing values of EMT score per each sample among WGD status (0: no WGD; 1: one WGD event; 2: two WDG events). P = 0.0004. b) Three-dimensional scatter plot showing correlations between EMT score, estimated ploidy and tumor subclonal fraction (Sub. fr.). c) Scatter plot showing correlation between EMT score and chromotripsis is high purity samples from COMPASS cohort. N = 23. P = 0.0160. d) Box and whiskers with all datapoints showing no statistical difference in terms of EMT score in TP53 wild type (WT) and mutant (Mutant) samples. N = 23. e) Schematic showing the analytical worflow used for the TCGA Solid Tumors cohort (Methods). f-g) Scatter plot showing correlation between Chromotripsis score and a transcriptomic signatures of EMT (Methods) of TCGA Solid Tumors filtered by purity 0.75. P = 0.0056, 4.90308E-08. h) Bar plot showing distribution of samples with at least one event of chromotripsis based on TP53 genomic status of TCGA Solid Tumors filtered by purity 0.75. P = 3.61974e-05. i) Box plot with individual values showing Chromotripsis score based on TP53 genomic status of TCGA Solid Tumors filtered by purity 0.75. N = 217 independent tumors. j-k) Violin plots showing EMT scores based on TP53 genomic status of TCGA Solid Tumors filtered by purity 0.75. l) Scatter plot showing correlations between EMT scores and Subclonal fraction of TCGA Solid Tumors filtered by purity 0.75. P = 5.50454e-08. m) Box and whiskers with individual data points displaying Chromotripsis score values between groups divided by median value of EMT scores from the COMPASS dataset filtered by purity 0.75. N = 23 independent tumors. P = 0.0018. n) Scatter plot showing correlation between EMT score and ploidy from the COMPASS dataset filtered by purity 0.75. P = 0.014. o) Scatter plot showing correlation between fCNA and EMT score from the PDAC-PDX MDACC dataset. P = 0.025. Error bars represent SD. Box plots are centered at the 50^th percentile and bounds of box are fixed on the 25^th and 75^th percentile, min and max include all data points. P values are calculated as follows: b-c,f-g, l-o) two-tailed Spearman correlation; h) Fisher’s exact chi square test; a,d,m) two sided Mann Whitney test; i-k) two sided paired student t test.

Extended data Figure 10. Histopathological and genomic characterization of the KPCY Zeb1WT/WT and Zeb1fl/fl GEMM.

a) Representative H&Es of primary tumors derived from KPCY Zeb1^fl/fl and KPCY Zeb1^WT/WT GEMMs. b) Bar plot showing quantification of grade in the two GEMMs (N = 10 areas per group). P = 0.0023. c) Circos plots showing genomic events in one KPCY Zeb1^fl/fl mouse and KPCY Zeb1^WT/WT mouse. d) Representative cases of chromotripsis as calculated via ShatterSeek. Scale bars = 100μm if not otherwise specified. P values are a result of b) Fisher’s exact chi square test.

Extended data Figure 11. scGET and RNA seq of EPI-L and MS-L subpopulations.

a-b) UMAPs showing cluster distribution of MS-L and EPI-L tumor cells and top 10 significant pathways enriched in cluster n.1 according to scGET and RNA data integration. c) UMAPs showing the distribution of EPI-L and MS-L subpopulation in scRNA and GET seq data. d) UMAP showing clonotype distribution as calculated by copy number analysis of scGET seq data. e) Scatter plot displaying correlations between Pseudotime α-Diversity and β-diversity as calculated from scGET seq data. f) UMAP showing distribution of structural variants as calculated from scGET seq data after normalization of sequencing coverage. g) Violin plot showing average nuclear area per cell as calculated by histopathological analysis of PCΩ SM-GEMMs. P = 0.0099. h) Histopathological representative picture (left) and quantification (right) of PCΩ SM-GEMMs with the DNA damage marker γH2AX. N = 88 independent areas over N = 3 independent experiments. P = 6.26675e-10. Scale bars = 100μm if not otherwise specified. Data are presented as mean values +/− SD P values are calculated as follows: g) two-sided student t-test; h) two-way ANOVA.

Extended data Figure 12. TGF-beta drives acquisition of nuclear abnormalities and chromatin remodelling.

a) Bar plot showing Mean Optical Density of MS-L and EPI-L cultures treated with either Reversine, MPS1i (MPI-0479605) or vehicle 14 days after plating. N = 3 biological replicates/group over N = 3 independent experiments. P = 7.36e-12, 1.13e-12, 6.63e-11, 4.803E-12. b) Representative pictures showing colonies forming under the effect of Mad2l1 knockdown in the EPI-L and MS-L subpopulations. c) Schemating showing experimental strategy to induce a TGF-beta driven EMT in the EPI-L subpopulation with analysis of nuclear and chromatin changes. d) Representative high magnification pictures showing normal and aberrant cell-cycle phases (anaphase, metaphase and interphase); red arrow indicates micronucleus in interphase. e) Bar plot with individual data points showing number of micronuclei per HPF in EPI-L subpopulations 48h after incubation with either TGF-beta or vehicle (N = 20 HPF images per group). P = 0.015 f) Bar plot showing incidence of aberrant and normal mitosis per HPF in EPI-L subpopulations 48h after incubation with either TGF-beta or vehicle (N = 50 mitosis per group). P = 3.32908E-05. g) Bar plot with individual data points showing number of peaks as extrapolated from ATAC-seq in EPI-L subpopulations 48h after pulses of either TGF-beta or vehicle (N = 5 independent EPI-L models per group). P = 0.0085. h) Bar plot showing number and type of peaks in Centromeric genes as extrapolated from ATAC-seq in EPI-L subpopulations 48h after incubation with either TGF-beta or vehicle (N = 5 models per group). P = 0.021. Scale bars = 100μm if not otherwise specified. Data are presented as mean values +/− SD P values are a result of: e) two-sided student t test; f) Fisher’s exact chi-square test; g) two-sided Mann-Whitney test; a,h) two-sided two-way ANOVA.

Extended Data Figure 13. Characterization of MS-L ‘diploid’ and ‘aneuploid’ clones.

a) Schematic showing experimental design of the isolation and transplantation of MS-L diploid and aneuploid clones from a stable MS-L ex vivo model. b) Representative pictures of metaphasic spreads of diploid and aneuploid MS-L (top) and quantification of chromosomes (bottom). N = 10 metaphases/group over N = 3 independent experiments. P = 1.82851E-08. c) Representative macroscopic pictures of liver metastasis derived from orthotopically transplanted MS-L diploid and aneuploid cells. d) Survival curves of tumor bearing mice transplanted with either MS-L diploid or aneuploid cells (N = 10 mice/group). P = 3.06476e-06. e) Scatter plot showing primary tumor size and number of metastasis of tumor bearing mice transplanted with either MS-L diploid or aneuploid cells (N = 10 mice/group). P = 0.047. f) Box and whiskers with individual data points showing number of mets in mice transplanted with either MS-L diploid or aneuploid cells (N = 10 mice/group). P = 0.033. g) Circle plots showing incidence of ascites in mice transplanted with either MS-L diploid or aneuploid cells (N = 10 mice/group). P = 0.019. h) Representative H&E of transplanted diploid and aneuploid tumors (top) and quantification of grade (bottom). i) DNA content histogram showing ploidy of MS-L bulk and clone#1 MS-L diploid and aneuploid derived from the bulk population. DNA profiles are similar in the isolated clones 10 passages after single cell seeding. j) Representative pictures of primary tumors derived from MS-L diploid and aneuploid clones transplanted orthotopically in the pancreas of recipient mice. k) Survival curves of mice transplanted with MS-L diploid and aneuploid clone #1 and #2. l-m) Number of metastasis and primary tumor volumes of mice transplanted with MS-L diploid and aneuploid clone #1 and #2. N = 5 mice/group over N = 2 independent experiments. Error bars represent SD. Box plots are centered at the 50^th percentile and bounds of box are fixed on the 25^th and 75^th percentile, min and max include all data points. P values are a result of: d,k) Logrank Mentel-Cox; f,l-m) two-sided Tukey’s multiple comparison test; e) two sided pearson correlation; b,f) two sided paired student t test; g-h) chi square test.

Extended Data Figure 14. EMT unlocks functional heterogeneity of pancreatic tumors.

a-b) Schematic and UMAP showing scRNA sequencing data processing of PCΩ tumors and their distribution across experimental groups and values of a transcriptomic signature of EMT. c) Scatter plot showing distribution of Simpson Index, Shannon Index values and branches as calculated by STREAM (Methods) in samples belonging to the EMT-on and EMT-off groups. Correlation and statistics are reported in figure. d) Representative stream plot showing inferred distribution of a tumor and metastasis derived from one GEMM belonging to the Vehicle and GCV group. A high number of putative transcriptomic branches can be appreciated. e) Representative H&E pictures of primary tumors belonging to Vehicle, GCV and GCV groups showing the phenotypic variability of pancreatic tumors (left) and quantification of observed phenotypes by pathological analysis in the EMT-on and EMT-off groups (N = 5 primary tumors/group). P = 0.0453. f) Schematic of the analytical workflow used for the analysis of patient-derived pancreatic tumor single cell RNA sequencing dataset retrieved from Peng et al. g) UMAP showing distribution of two groups characterized by high and low levels of EMT, as calculated by a transcriptomic signature of EMT (EMT-high and EMT-low respectively, Methods). h) Scatter plot showing distribution of Simpson Index, Shannon Index values and number of cells per each sample included in the analysis of the dataset published by Peng et al. Comparison and P values are reported in figure. Scale bars = 100μm if not otherwise specified. P values are calculated as follows: c,h) two-tailed spearman correlation; e) chi-square test.

Data availability

All data supporting the findings of this study are available within the Article and its Supplementary Information. Murine WGS, single cell RNA, GET seq and bulk ATAC seq raw data have been deposited on Sequence Read Archive (PRJNA1085291) and ArrayExpress under accession code: E-MTAB-13051 and E-MTAB-13052. Mouse and human genomic references (mm10, hg38) were downloaded from https://genome.ucsc.edu/. Human single cell RNA sequencing data were downloaded from: Genome Sequence Archive project PRJCA001063. Human clinico-genomic and transcriptomic information were downloaded from: https://portal.gdc.cancer.gov/ (TCGA) and EGAD00001003584, EGAD00001003585 (COMPASS). Source data for in vitro and in vivo are provided with this paper. Requests for resources and reagent can be directed to the lead contact G.G.

Code and software availability

Codes used for this manuscript have been previously published and adequately referenced in this manuscript. Methodological details on parameters used are available in the Methods section of this manuscript.