Heterocellular OSM-OSMR signalling reprograms fibroblasts to promote pancreatic cancer growth and metastasis

Pancreatic ductal adenocarcinoma (PDA) is a lethal malignancy with a complex microenvironment. Dichotomous tumour-promoting and -restrictive roles have been ascribed to the tumour microenvironment, however the effects of individual stromal subsets remain incompletely characterised. Here, we describe how heterocellular Oncostatin M (OSM) - Oncostatin M Receptor (OSMR) signalling reprograms fibroblasts, regulates tumour growth and metastasis. Macrophage-secreted OSM stimulates inflammatory gene expression in cancer-associated fibroblasts (CAFs), which in turn induce a pro-tumourigenic environment and engage tumour cell survival and migratory signalling pathways. Tumour cells implanted in Osm-deficient (Osm−/−) mice display an epithelial-dominated morphology, reduced tumour growth and do not metastasise. Moreover, the tumour microenvironment of Osm−/− animals exhibit increased abundance of α smooth muscle actin positive myofibroblasts and a shift in myeloid and T cell phenotypes, consistent with a more immunogenic environment. Taken together, these data demonstrate how OSM-OSMR signalling coordinates heterocellular interactions to drive a pro-tumourigenic environment in PDA.

P ancreatic cancer is characterised by an extensive desmoplastic reaction, which constitutes on average 85% of the tumour volume 1,2 . Here, mutated tumour cells are embedded in a stiff, remodelled extracellular matrix (ECM), which is abundantly co-inhabited by stromal host cells 3 . Dysregulated immune cells and cancer-associated fibroblasts (CAFs) are a perpetual and dynamically regulated feature throughout PDA development 2,4 . CD4 + regulatory T cells, myeloid derived suppressor cells and macrophages (MØs) infiltrate early during PDA development, and directly promote tumour progression through matrix metalloproteinase (MMP) and IL6 signalling, while simultaneously suppressing anti-tumour immunity [4][5][6][7][8] . Consequently, targeting MØs and granulocytes, such as through CSF1R or CXCR2 inhibitors, increases T cell infiltration and sensitises tumours to immune checkpoint inhibitors (ICIs) 9,10 . In addition, tumour cells conscript resident pancreatic fibroblasts to support tumour cell survival, metabolic adaptation and exclusion of CD8 + T cells [11][12][13][14] . Targeting CAFs and the stiff, remodelled ECM such as through vitamin D receptor agonists, lysyl oxidase and leukaemia inhibitory factor (LIF) blocking antibodies or CXCR4 inhibitors improves response to chemotherapy and sensitises to ICIs 11,[15][16][17] . Tumour-restrictive stromal cell populations have also been described in PDA, where genetic and pharmacologic ablation of CAFs accelerates disease progression and metastatic dissemination [18][19][20][21] . Yet, the mechanisms by which discrete stromal populations coordinate to regulate specific tumour functions remain incompletely characterised 2,22 .
Single-cell analysis has been instrumental in annotating individual stromal cell populations 21,[23][24][25][26][27] . Recently, several distinct CAF populations have been identified, including myofibroblastic CAFs (myCAF), defined by elevated αSMA expression and ECM modulating features; inflammatory CAFs (iCAFs), defined by expression of several inflammatory genes; and antigen-presenting CAFs (apCAFs), defined by expression of MHCII and CD74 21,[27][28][29][30] . The transcriptional state of these CAF populations is dynamically regulated, and pharmacological intervention can therefore be used to manipulate these CAF subsets in vitro and in vivo 29,30 . Several signals have been demonstrated to induce specific CAF subtypes, where tissue stiffness, TGFβ signalling and spatial proximity to tumour nests induce myCAF features, whilst tumour cell-secreted IL1α drives the iCAF phenotype through the JAK-STAT pathway 29,30 . Importantly, CAF and immune cell regulation of PDA pathobiology is highly interdependent, where targeting individual populations is accompanied by a concomitant adaptation of the tumour microenvironment (TME). For example, ablation of αSMA pos or FAP pos CAFs sensitises tumours to ICIs, targeting the stiff ECM alters immune infiltration, and pharmacologic and genetic manipulation of individual immune subsets drastically alters CAF functions 10,11,15,19,20,23,31 . Consequently, determining the functional interdependencies between dynamically regulated CAF subsets, immune infiltrates and tumour cell function remains critical to ascertain the benefits of stromal targeting strategies.
In this work, we explore the impact of heterocellular interactions across tumour cells, fibroblasts and MØs on PDA pathobiology. We observe that tumour cell-derived granulocyte-MØ colony-stimulating factor (GM-CSF) induces expression of Oncostatin M (OSM) in MØs, which in turn induces expression and secretion of pro-tumourigenic inflammatory mediators, such as IL6 family members (e.g. IL6 and LIF) and CXC-and CCchemokines (CXCL1, CCL2 and CCL7), in stromal fibroblasts. This aberrant signalling milieu engages survival, motility and cell migration signalling pathways in tumour cells to accelerate metastasis. Pancreatic cancer cells (PCCs) orthotopically transplanted into immune-competent mice lacking Osm (Osm −/− ) generate significantly smaller tumours which exhibit an altered microenvironment with a reduction in tumour-promoting cytokines and chemokines (IL6, CXCL1, TNFα and GM-CSF) and increased abundance of αSMA pos myofibroblasts. The cognate OSM receptor, OSMR, is expressed in mesenchymal stromal cells, and in particular in fibroblasts and perivascular cells. Notably, OSMR expression level is correlated with inflammatory gene expression, a tumour-permissive immune infiltrate and a poor overall survival in human PDA. Taken together, the data presented establishes a role of OSM-OSMR signalling in driving an inflammatory microenvironment and supporting tumour growth.

Results
OSMR expression is associated with tumour-promoting inflammation and poor outcome in PDA. PDA progression is accelerated by an extensive fibroblast expansion and a concomitant myeloid cell infiltrate 2,4 . Several tumour cell-derived signals have been shown to regulate fibroblast function 29,30,32 , however, it is less clear whether other cell populations in the tumour ecosystem influence CAF function. To identify and characterise novel stromal receptor and ligand pair(s) that may regulate CAF function, we re-analysed available RNA-seq data of murine and human PDA. Differential gene expression analysis of iCAF and myCAF subsets 29 identified several membrane receptors, such as Osmr, Antxr1 and Il1r1, which were expressed at a higher level in iCAFs (Fig. 1a). Single-cell RNA-sequencing confirmed expression of Osmr, Antxr1 and Il1r1 in CAFs as well as in endothelial and perivascular cells in both human and murine PDA 27,33 . Moreover, these receptors were expressed at higher levels in iCAFs compared to myCAFs and apCAFs (Fig. 1b, c and Supplementary Fig. 1a, b). To further analyse OSMR expression in human resected PDA we used multiplexed in situ mRNA hybridisation for OSMR with immunofluorescence for Vimentin (to broadly demark stroma) and pan-Cytokeratin (to demark epithelial cells) ( Fig. 1d and Supplementary Fig. 1c). Consistent with the scRNA-seq data, OSMR was broadly expressed within the VIM pos tumour stroma, however, OSMRpositive cells were also observed within the epithelium (PanCKpos ) cells.
To then determine whether expression of OSMR, ANTX1 and IL1R1 was correlated with patient prognosis, we studied their association with patient survival. Analysis of normal tissue samples (GTEx) and human PDA tumours (TCGA PanCancer PAAD) revealed a significantly higher expression level of OSMR and its cognate ligand, OSM, in human PDA compared to normal pancreatic tissue (Fig. 1e). Moreover, of the three iCAF-enriched receptors analysed, OSMR was the only receptor where expression was associated with poor outcome in PDA patients, where PDA patients with higher mRNA expression levels of OSMR exhibited worse overall survival than patients with lower OSMR expression ( Fig. 1f and Supplementary Fig. 1d). Further investigating the association between Osmr expression and the iCAF phenotype in vivo, we compared the expression of inflammatory mediators in Osmr pos and Osmr neg CAFs using publicly available scRNAseq datasets of KPC 27 and KPP 33 tumours, which revealed elevated expression of Il6, Ccl2, Ccl7 and Cxcl1 in Osmr pos CAFs (Fig. 1g).
To investigate how the OSM-OSMR ligand-receptor interaction may be regulated in PDA, we analysed the expression pattern of cognate ligands for Osmr, Il1ra and Antxr1 receptors in both human and murine scRNA-seq datasets. OSM was expressed in MØs and myeloid cells in both murine and human PDA, as well as in murine neutrophils and dendritic cells (Fig. 1h), contrasting a broader expression of IL1R1 and ANTXR1 ligands (Supplementary Fig. 1e). This data indicated that myeloid/MØ-secreted OSM may act in a paracrine manner on OSMR-expressing mesenchymal cells, such as fibroblasts. To test this hypothesis, we subjected publicly available scRNA-seq datasets of human and mouse PDA to CellPhoneDB analysis and evaluated MØ-fibroblast ligand-receptor interactions 27,34 . This analysis further highlighted OSM-OSMR as a stromal paracrine signal conserved across both human and mouse PDA ( Fig. 1i and Supplementary Fig. 1f-h). Given an established role of CAFs as fibroinflammatory regulators, we hypothesised that heterocellular OSM-OSMR signalling may induce inflammatory fibroblasts in PDA.
To subsequently determine the cellular source(s) of individual signals, we FACS-isolated PCCs, PSCs and MØs from mono-and co-cultures and analysed the expression level of selected signals by RT-qPCR ( Supplementary Fig. 2b). Most transcripts were expressed and regulated in a cell-type specific manner (Fig. 2b). For example, MØs expressed Tnf and Osm, PSCs expressed Il6, Cxcl9 and Cxcl10, and PCCs expressed Csf2 as well as Cxcl5.
Other signals, such as Ccl2 and Ccl7, were expressed in MØs cocultured with either PCC or PCC-PSC. However, Ccl2 and Ccl7 were also expressed in PSCs isolated from 3 W co-cultures with PCCs and MØs (Fig. 2b), demonstrating context-dependent control of inflammatory gene expression. Unsupervised hierarchical clustering and principle component analysis (PCA) of mRNA levels highlight that PSCs engaged discrete secretory programmes depending on specific heterocellular interactions ( Fig. 2b and Supplementary Fig. 2c). In particular, PSCs isolated from PCC-PSC-MØ co-cultures (herein called 3W-PSCs) clustered separately from PSCs isolated from mono-and 2 W co-cultures ( Fig. 2b and Supplementary Fig. 2c left), and displayed increased expression of interleukins such as Il6, as well as CC-/CXC-ligands such as Cxcl1, Ccl2 and Ccl7 ( Supplementary Fig. 2c right). In contrast, PSCs from PCC-PSC 2 W co-culture (herein called 2W-PSCs) exhibited a distinct secretory pattern with increased expression of Cxcl12, Tgfb2 and Ccl11 ( Fig. 2b and Supplementary  Fig. 2c right). The analysed ligands were not differentially expressed in PCCs or MØs isolated from mono-and co-cultures (Fig. 2b), suggesting that PSCs are particularly responsive to changes in the local cellular environment.
To further investigate whether interactions across MØs, PCCs and PSCs induce a transcriptional programme in PSCs that is consistent with iCAFs, we FACS-isolated PSCs from mono-, PCC-PSC and PCC-PSC-MØ co-cultures, and analysed mRNA expression by RNA-seq ( Supplementary Fig. 2d). Gene expression analysis identified 1758 differentially regulated genes in PSCs from 2 and 3 W co-cultures. 2W-PSCs displayed increased expression of genes encoding ECM proteins including collagens Col6a1, Col11a1, Col4a5, Col4a6 as well as Acta2, (α-smooth muscle actin, αSMA) a well-documented marker of myofibroblasts ( Fig. 2c) 22 . Similarly, inflammatory signals, many of which were identified in the conditioned medium from 3 W co-cultures ( Fig. 2a and Supplementary Fig. 2a), were abundantly expressed in 3W-PSCs, confirming fibroblasts as a major cellular source of inflammatory signals ( Fig. 2c and Supplementary Fig. 2e). For example, upregulated genes in 3W-PSCs include CC-chemokines (Ccl2, Ccl7 and Ccl5), JAK-STAT (Osmr, Jak2 and Stat3) and NFκB (Chuk, Nfkb1 and Relb) pathway components, as well as NFκB target genes such as C1s, Vcan, Has2, Mt2 and Il4ra ( Fig. 2c and Supplementary Fig. 2e, f). Indeed, gene set enrichment analysis (GSEA) revealed that 3 W co-culture is associated with an enrichment of "Inflammatory Signalling", "TNFα Signalling via NFκB pathway" and "IL6-JAK-STAT pathway" in PSCs, whereas "Epithelial-Mesenchymal Transition" (EMT) and "Oxidative Phosphorylation" pathways are enriched in 2W-PSCs ( Fig. 2d and Supplementary Fig. 2f). These results are consistent with the previously described iCAF subsets in PDA 27,29,30,33 , and suggest that heterocellular interactions influence fibroblast subtypes. To determine whether the PSC subtype switch is consistent across 2 and 3 W co-cultures with different PCCs, we collected conditioned medium from 2 W PCC-PSC and 3 W PCC-PSC-MØ co-cultures using three different PCCs, and analysed the expression of selected inflammatory (Il6, Il4ra and Cxcl1) and myofibroblastic (Acta2) markers in PSCs stimulated with the conditioned medium (Fig. 2e). Indeed, 2 and 3 W PSCs displayed a transcriptional change consistent with subtype switch across all PCC (Fig. 2e, left). Whereas Acta2 was more highly expressed across all 2 W conditions, 3 W conditioned medium increased inflammatory gene expression and decreased Acta2 (Fig. 2e, right). Moreover, to assess whether the PSC subtype switch was specific to 3 W co-culture conditions, and not due to direct interactions between PSCs and MØs, we compared gene expression in PSCs treated with conditioned medium from PCCs, PSCs and MØs in mono-2W-and 3W-co-cultures. This data confirmed that only in the presence of all three cell types did PSCs decrease Acta2 expression and increase Il6, Il4ra and Cxcl1 expression (Fig. 2f).
To then confirm the co-culture-dependent switch in fibroblast function using an orthogonal readout, we profiled the expression of 26 mesenchymal markers by mass cytometry (MC) in PSCs grown as mono-culture or in 2W-or 3W-co-cultures Fig. 1 OSMR expression is associated with tumour-promoting inflammation and poor outcome in PDA. a Heatmap displaying relative expression levels of membrane receptors of stellate cells adopting a quiescent, inflammatory cancer-associated fibroblasts (iCAF), or myofibroblastic CAF (myCAF) phenotype (dataset: GSE93313). b Violin plot of normalised Osmr expression levels in myCAF, iCAF, and apCAFs in murine PDA (dataset: GSE129455) with mean expression z-scores shown below. c Violin plots of normalised OSMR/Osmr expression levels for individual cell types in human (left) and murine (right) PDA (top) and mean expression z-scores shown below (datasets: GSE129455, phs001840.v1.p1). d Representative in situ mRNA hybridisation (ISH) of OSMR multiplexed with immunofluorescence of VIM and PanCK in human PDA (n = 3). 20× magnification. Upper left quadrant, full overlay; upper right, OSMR; lower left, OSMR with VIM; lower right, OSMR with PanCK. Scale bar = 100 μm. e OSM (left) and OSMR (right) expression in human PDA (TCGA PanCancer PAAD dataset, n = 178) and normal pancreatic tissue (GTEx, n = 171). Mantel-Cox test. T tumour, N normal, TPM transcripts per million. f Kaplan-Meier survival curves showing overall survival of 179 PDA patients from TCGA PanCancer PAAD dataset. Patients were stratified by high (Top 50%, n = 89) and low (bottom 50%, n = 90) OSMR expression levels. Analysis by log-rank test and cox-proportional hazard regression. HR hazard ratio. g Violin plots of inflammatory gene expression in Osmr pos and Osmr neg CAFs in murine PDA tumours (datasets: GSE129455, MTAB-8483) with mean expression z-scores shown below. h Violin plots of normalised OSM/Osm expression across cell types of human (top) and murine (bottom) PDA (datasets: GSE129455, phs001840.v1.p1). Mean expression z-scores shown. i Top 14 predicted interactions between macrophage-derived ligands and fibroblast expressed receptors using CellPhoneDB receptor-ligand interaction statistical analysis (p-values < 0.05). Also see Supplementary Fig. 1. Source data are provided as a Source Data file.
( Supplementary Fig. 2g). This analysis demonstrated a shift in marker levels between the culture conditions (Fig. 2g, Supplementary Fig. 2h), supporting a PSC phenotypic switch between PCC-PSC and PCC-PSC-MØ co-cultures. To confirm the role of MØs, we set up a MØ-limiting titration experiment, where we decreased the number of MØs in 3 W co-culture, and quantified mesenchymal marker expression by MC (Fig. 2h). This demonstrated that markers preferentially expressed in 3W-PSCs, such as PDPN, CD73 and Integrin α 5 , decreased as the number of MØs was reduced in the 3 W co-culture (Fig. 2h).
Heterocellular OSM-OSMR signalling induces inflammatory fibroblasts. Consistent with the in vivo gene expression analysis (Fig. 1c,  where MØs express Osm and PSCs express Osmr ( Fig. 3a and Supplementary Fig. 3a). We therefore hypothesised that heterocellular OSM-OSMR signalling induces expression of inflammatory mediators in PSCs. To test this hypothesis, we analysed the expression of selected myofibroblastic and inflammatory markers in PSCs treated with recombinant OSM or conditioned medium from mono-and co-cultured PCCs, PSCs and MØs (Fig. 3b). Recombinant OSM and 3 W co-culture conditioned medium induced a transcriptional shift in PSCs distinct from mono-and 2 W PCC-PSC co-cultures, as visualised by PCA (Fig. 3b left). Specifically, recombinant OSM increased expression of inflammatory mediators, including Il6, Il4ra Cxcl1, Mt2, Nos2 and Vcan, (Fig. 3b right). Similarly, markers predominantly expressed in PSCs co-cultured with PCCs, such as Thbs1, Col4a5, Col4a6, Acta2 and Ccnd1, are decreased following OSM stimulation ( Fig. 3b right). Several of these genes have previously been associated with inflammatory (iCAF) and myofibroblastic (myCAF) phenotypes, confirming an inverse relationship between iCAF and myCAF transcriptional programmes 27,29,30 .
To further elucidate the molecular dependencies of MØinduced inflammatory gene expression in PSCs, we added OSM blocking or control antibodies to conditioned medium from PCC-PSC-MØ 3 W co-culture and assessed the impact on PSC gene expression ( Fig. 3c and Supplementary Fig. 3b). As inflammatory PSCs isolated from 3 W co-cultures were associated with increased JAK-STAT and NFκB pathway activation (Fig. 2d), we also included respective pharmacological inhibitors ( Fig. 3c and Supplementary Fig. 3b). As anticipated, and in agreement with previous observations 30 , inhibition of JAK2 and IKKβ reduced the expression of inflammatory mediators while simultaneously increasing expression of myofibroblastic genes. This resulted in a transcriptional state more similar to 2W-PSCs ( Fig. 3c and Supplementary Fig. 3b). Similarly, inclusion of an OSM blocking antibody in 3 W conditioned medium also induced a transcriptional shift in the PSCs, marked by decreased expression of inflammatory mediators and increased expression of myofibroblastic genes ( Fig. 3c and Supplementary Fig. 3b). The effect of the OSM blocking antibody as well as the JAK-STAT and NFκB pathway inhibitors was replicated using CRISPR-Cas9 mediated knockout of Osmr in PSCs ( Fig. 3c and Supplementary  Fig. 3b), confirming a role for OSM-OSMR signalling in inducing inflammatory gene expression in PSCs. Notably, inclusion of neutralising antibodies against previously described regulators of inflammatory fibroblasts, such as TNFα, IL1α and IL1β 30 , differentially influences fibroinflammatory gene expression when compared to OSM-OSMR-targeting modalities ( Supplementary  Fig. 3c, d). For example, the OSM blocking antibody decreased expression of inflammatory genes such as Il6, Il4ra, Jak2, Has2, Vcan and Nos2, and increased myofibroblastic genes such as Thbs1 and Ccnd1 ( Supplementary Fig. 3c). However, neutralising antibodies for OSM, TNFα and IL1β were needed in combination ("OTI") to efficiently block Mmp3, Pdpn and Nt5e (Supplementary Fig. 3c). In addition, an IL1α neutralising antibody did not alter 3W-PSC gene expression differentially to an IL1β neutralising antibody ( Supplementary Fig. 3d). This suggests that MØinduced inflammatory genes in fibroblasts relies on multiple combinatorial signals. Together, this demonstrates a requirement of heterocellular OSM-OSMR signalling as a potent inducer of inflammatory fibroblasts.

Tumour cell-secreted GM-CSF increases MØ secretion of OSM.
To identify putative Osm inducing signals, we treated MØs with PCC conditioned medium and determined OSM levels in conditioned medium by ELISA (Fig. 4a). None of the tested PCC cell lines produced OSM in mono-culture, however, conditioned medium from all 13 PCC lines increased OSM levels in MØs (Fig. 4a). To subsequently determine which PCC-derived signals induce OSM, we treated MØs with recombinant signals identified in the LUMINEX analysis of PCC conditioned medium (Fig. 2a), and measured OSM in the conditioned medium by ELISA ( Supplementary Fig. 4). Strikingly, GM-CSF was the only tested soluble factor that induced OSM secretion ( Supplementary  Fig. 4). Moreover, increased levels of GM-CSF were generally observed in the conditioned medium of PCCs that strongly induced OSM (Fig. 4b). To functionally confirm a role of PCCsecreted GM-CSF, we collected PCC conditioned medium and included control or GM-CSF neutralising antibodies before stimulating MØs and measuring OSM by ELISA. GM-CSF neutralising antibodies blocked OSM secretion from MØs when added to PCC conditioned medium and recombinant GM-CSF (Fig. 4c), confirming that PCC-produced GM-CSF is a key regulator of OSM secretion. We then sought to determine whether GM-CSF could substitute PCCs in 3 W co-cultures to induce expression of inflammatory signals in PSC and added recombinant GM-CSF to PSC-MØ co-cultures and stimulated PSCs with the conditioned medium. Recombinant GM-CSF-PSC-MØ cultures induced both Osmr and inflammatory gene expression in PSCs (Fig. 4d). Together this data supports a role of GM-CSF in inducing Osm expression in MØs, thereby enabling MØ-induced inflammatory gene expression in fibroblasts. Finally, analysis of TCGA PanCancer PAAD dataset identified a positive correlation of OSM expression levels and CSF2 (encoding GM-CSF) expression in PDA patients (Fig. 4e), highlighting a clinical association between GM-CSF and OSM expression in PDA. Collectively, these findings imply GM-CSF as a positive regulator of MØ OSM secretion. FC fold change. d Gene set enrichment analysis (GSEA) of RNAseq data from (c). NES normalised enrichment score, FDR false discovery rate. e RT-qPCR assay of selected myofibroblastic and inflammatory genes of PSCs stimulated with conditioned medium from 2 or 3 W co-cultures using different isolations of PCCs (PCC, PCC4 and PCC11). Left: Changes in expression state visualised by principle component analysis (PCA) across experimental conditions (n = 4). Right: Representative mRNA expression of inflammatory (Il6, Il4ra and Cxcl1) and myofibroblastic (Acta2) genes. Results displayed as mean ± SEM, n = 4, one-way ANOVA Tukey test. f RT-qPCR of selected myofibroblastic and inflammatory genes in PSCs stimulated with conditioned medium from 1, 2 or 3 W co-cultures of PCC, PSC and bone marrow-derived MØs. Left: Changes in expression state visualised by PCA plot across experimental conditions (n = 4). Right: Representative mRNA expression of inflammatory (Il6, Il4ra and Cxcl1) and myofibroblastic (Acta2) genes. Results displayed as mean ± SEM, n = 4, one-way ANOVA Tukey test. g Mass cytometry analysis of PSCs in mono-(n = 3) and co-culture with PCC (n = 3) or PCC and MØs (n = 3). Selforganising map clustering (FlowSOM) of PSCs in mono-and co-cultures displayed as t-Distributed Stochastic Neighbour Embedding (tSNE). h Mass cytometry analysis of PSCs from 3 W PCC-PSC-MØ co-culture. Number of macrophages plated in 3 W co-culture was titrated down and PDPN, CD73 and Integrin α 5 expression levels shown as biaxial plots. Also see Supplementary Fig. 2. Source data are provided as a Source Data file.
OSM supports the formation of an immunosuppressive microenvironment. We next sought to determine the role of OSM in vivo and analysed the microenvironment of tumours from PCCs 38 orthotopically implanted into syngeneic wildtype or Osm-deficient C57BL/6 mice (Osm −/− ) (Fig. 5a). First, we quantified a number of inflammatory cytokines with known tumour-promoting effects in PDA. Tumours from Osm −/− animals exhibited markedly reduced levels of several cytokines, including IL6, CXCL1, GM-CSF and TNFα, highlighting a profound change in the tumour environment (Fig. 5b). Immunohistochemical (IHC) analysis of the prototypical myCAF marker, αSMA (Acta2), revealed an increase in the αSMA pos area in Osm −/− animals compared to wildtype animals (Fig. 5c).
CD44 and CD127, which demark T memory cells 45 , were increased across several populations of CD8 pos and CD4 pos T cells in Osm −/− tumours (Fig. 5g). We therefore further characterised memory T cell populations in these animals. Central memory (Tcm) and effector memory (Tem) CD4 pos and CD8 pos T cells displayed reduced levels of T cell exhaustion and inhibitory markers such as LAG3 (Fig. 5h, i), CTLA4 and PD-1 (Fig. 5i), consistent with improved function of memory CD4 pos and CD8 pos T cell populations in Osm −/− tumours (Fig. 5h, i and Supplementary  Fig. 6i). Analysis of immune infiltrates in human PDA patients from TCGA PanCancer 46 by CIBERSORT 47 was consistent with the MC analysis (Fig. 5j). Specifically, we observed that OSMR expression levels, as a marker of iCAFs, correlated with infiltration of M2 MØs and neutrophils and resting or inactivated dendritic cells, mast cells, NK cells and CD4 pos memory T cells in patients with PDA (Fig. 5j). In contrast, infiltration of activated NK cells was negatively correlated with patient OSMR expression (Fig. 5j). Together, this emphasises the relationship between OSMR expression and the establishment of an immunosuppressive microenvironment in patients with PDA, and underscores a functional role of heterocellular OSM-OSMR signalling in mediating an immunosuppressive, tumour-promoting environment.

PCC-PSC-MØ interactions reshape tumour cell signalling.
Given the extensive shift in the cellular signalling milieu between PCC-PSC and PCC-PSC-MØ co-cultures (Fig. 2a), we hypothesised that tumour cell signalling and function is differentially regulated under these two conditions. We, therefore, undertook a quantitative phosphoproteomics analysis of stroma-regulated tumour cell signalling, treating tumour cells with conditioned medium from mono-cultured PCC and 2 W PCC-PSC or 3 W PCC-PSC-MØ co-cultures (Fig. 6a). In agreement with previous observations, tumour-stroma interactions rewired tumour cell signalling to engage additional signalling pathways beyond the tumour cell-autonomous signalling state ( Supplementary  Fig. 7a) 14,48 with additional differences between 2 and 3 W coculture conditions. Specifically, we quantified 2523 phosphorylation sites that were differentially engaged in tumour cells treated with co-culture conditioned medium, of which, 3 W signals specifically regulated 508 unique tumour cell phosphosites, and 2 W signals specifically regulated 433 tumour cell phosphosites ( Supplementary Fig. 7a . 6b and Supplementary Fig. 7b). Indeed, enrichment analysis of phosphorylation motifs revealed that 3W-regulated tumour cell signalling was predominantly regulated by the MAPK-and PKA/AKT-signalling pathways ( Supplementary Fig. 7c). Concordantly, a kinase-substrate directed phospho-network analysis further highlighted MAPK1/3 and AKT as central kinases regulating PCC-PSC-MØ-induced tumour cell signalling (Fig. 6b). For example, PCC-PSC-MØactivated MAPK1/3 can phosphorylate a number of downstream phosphoproteins, such as RPS6KA1 (RSK1), in turn phosphorylating GSK-3β, and 4E-BP1, which are also regulated by AKT 49,50 . Notably, both SHC1-engaged SRC and gp130-activated JAK2 signals increased the phosphorylation of STAT3 [Y705] (Fig. 6b), a critical step in pancreatic cancer progression 36,51 . Furthermore, PCC-PSC-MØ signalling uniquely results in phosphorylation of p38 and JNK pathway components including TAOK1 and MAP2K7 (MKK7) (Fig. 6b, Supplementary Fig. 7b), implicated in proliferation and invasion in multiple cancers 52,53 . To corroborate the global phosphoproteomics analysis we undertook a targeted phosphorylation analysis, which confirmed that PCC-PSC-MØ signals accentuate STAT3, MAPK1/3, AKT, p38 and JNK activation beyond mono-and 2 W co-culture conditions ( Supplementary Fig. 7d). In addition, immunoblotting for phosphorylated STAT3 [Y705] in PCCs treated with conditioned medium across all mono-and co-culture conditions revealed a specific increase only with PCC-PSC-MØ conditioned medium (Fig. 6c). This effect is unlikely to be a direct effect of OSM on PCCs as recombinant OSM only engaged STAT3 phosphorylation in PSCs ( Supplementary Fig. 7e). Furthermore, 3 W co-culture induced STAT3 phosphorylation in PCCs was dependent on OSMR expression in PSCs where CRISPR-Cas9mediated knockout of Osmr in PSCs ablate 3 W co-culture induced STAT3 phosphorylation in PCCs ( Fig. 6d and Supplementary Fig. 7f). These data underscore the importance of interactions between tumour cells, fibroblasts and macrophages in shaping tumour cell signalling.
Single sample GSEA (ssGSEA) of human PDA tumours adjusted for tumour purity ( Fig. 6e and Supplementary Fig. 7g), confirmed differential signalling pathway engagement where OSMR lhigh tumours display increased KRAS pathway activity, PI3K-AKT-mTOR and IL6-JAK-STAT3 signalling. Moreover,  7) and Osm −/− (n = 7) tumours, measured by ELISA. Results displayed as mean ± SEM, two-tailed student t test. c Immunohistochemical analysis of αSMA in wildtype and Osm −/− tumour sections. Left: Representative αSMA staining. Right: Quantification of αSMA pos staining of tumour sections of wildtype (n = 6) and Osm −/− (n = 6) animals. Results displayed as mean ± SEM, two-tailed student t test. d RT-qPCR gene expression analysis of FACS-isolated CAFs from wildtype (n = 5) and Osm −/− (n = 5) tumours displayed as a heatmap (left). Number of isolated CAFs shown as bar plot (right). Results displayed as mean ± SEM, two-tailed student t test. e Immunohistochemical analysis of CD8 in wildtype (n = 6) and Osm −/− (n = 6) animals. Upper: Total CD8 pos staining. Lower: Ratio of peripheral/central CD8 pos staining using the mean CD8 pos stain from four representative peripheral regions and four representative central regions per tumour. Results displayed as mean ± SEM, two-tailed student t test. f Mass cytometry analysis quantifying abundance of antigen presenting cells (APCs) expressing co-stimulatory markers in wildtype (n = 5) and Osm −/− (n = 5) tumours. Results displayed as mean ± SEM. Unpaired two-tail student t test. g Mass cytometry analysis of CD45 + /CD3 + T cells expressing indicated markers in wildtype (n = 5) and Osm −/− (n = 5) tumours. Results displayed as mean ± SEM, two-tailed student t test. h, i Mass cytometry analysis of memory T cells expressing indicated exhaustion/dysfunction markers in wildtype (n = 5) and Osm −/− (n = 5) tumours. Results displayed as mean ± SEM, two-tailed student t test. Tcm central memory, Tem effector memory, Trm resident memory. j Spearman correlation analysis between infiltration of indicated immune cells (determined by CIBERSORT) and OSMR expression levels in PDA patients from TCGA PanCancer PAAD dataset (n = 179). 95% confidence interval as shaded area. Also see Supplementary Figs. S5 and S6. Source data are provided as a Source Data file.
consistent with the observation that PCCs in 3 W co-culture regulated cell-cell adhesion, motility and actin cytoskeleton ( Supplementary Fig. 7h), ssGSEA also identified an enrichment of EMT signature genes in OSMR lhigh tumours ( Fig. 6e and Supplementary Fig. 7i). To further validate this observation, we analysed PCCs treated with conditioned medium from 2 and 3 W co-cultures, which revealed a decrease in E-Cadherin levels concomitant with an increase in VIM and SNAIL/SLUG specifically in PCCs treated with conditioned medium from 3W-cultures (Fig. 6f, and Supplementary 7j). Thus, interactions between tumour cells, fibroblasts and MØs, through OSM-OSMR signalling, induce an inflammatory fibroblast phenotype, which in turn engage migratory and mesenchymal tumour cell signalling pathways.
OSM drives tumour growth and metastasis in vivo. Lastly, we sought to determine the pathobiological role of OSM-OSMR signalling on tumour cell function in vivo. We analysed tumour growth and metastasis in the wildtype and Osm −/− orthotopic mouse model of PDA used in Fig. 5. Due to the desmoplastic reaction of pancreatic tumours, we transduced PCCs with infrared RFP (iRFP) before implantation, thereby enabling quantitative analysis of tumour cell-specific growth 54 . As expected, tumour cells injected into the pancreas of wildtype animals efficiently established primary tumours, as quantified by iRFP signal intensity, primary tumour volume and tumour weight (Fig. 7a). In contrast, injection of the same cells into Osm −/− mice results in a significantly reduced tumour burden, shown by iRFP intensity, tumour volume and tumour weight (Fig. 7a), demonstrating a systemic role for OSM in primary tumour growth. IHC analysis confirms a reduced PanCK pos tissue area in Osm −/− tumours, corresponding to restrained tumour growth (Fig. 7b). No significant differences were observed in staining for the proliferation marker Ki67, however cleaved caspase 3 levels were increased in tumours from wild-type animals ( Fig. 7c and Supplementary Fig. 8a, e). Quantitative morphology analysis of PanCK pos tumour cells highlights a morphological shift from a mixed mesenchymal and epithelial phenotype in wildtype animals towards an epithelial-dominated phenotype in the Osm −/− animals (Fig. 7b). Consistently, IHC staining for the EMT marker SLUG revealed decreased levels in tumours from Osm −/− animals but a non-significant difference in SNAIL levels ( Fig. 7c and Supplementary Fig. 8c, d), supportive of an Osm-dependent regulation of tumour cell EMT.
Liver is the most common metastatic site for PDA, and metastasis is the major contributor of cancer mortality 55 . Due to the observed difference in EMT signature between 2 and 3 W coculture conditions in vitro and OSMR high and OSMR low human PDA tumours (Fig. 6e, f), the morphological shift of tumour cells from a more migratory and mesenchymal to an epithelialdominated phenotype (Fig. 7b), together with reduced levels of pro-metastatic inflammatory signals, such as IL6 (Fig. 5b), we hypothesised that loss of Osm would also be associated with reduced metastasis. Strikingly, whereas more than 50% of wildtype animals develops iRFP + liver metastasis, none of the Osm −/− animals shows any signs of liver metastasis (Fig. 7d) demonstrating a role of OSM in regulating metastatic spread. Notably, staining for CD31 did not reveal any overt differences in the blood vessel abundance between tumours in wildtype or Osm −/− animals ( Fig. 7c and Supplementary Fig. 8b). Taken together, these findings highlight that OSM is a key MØ-derived signal in establishing an inflammatory microenvironment promoting tumour growth and metastasis.

Discussion
Tumours are evolving multicellular ecosystems. Pancreatic ductal adenocarcinoma (PDA) is characterised by an abundant desmoplastic reaction, where CAFs have emerged as coordinators of both tumour-promoting and -restricting functions of the microenvironment 2,22 . Single cell RNA expression and de novo inference analyses have been instrumental in annotating transcriptionally distinct CAF subsets 21,24,26,27,33 , however, further characterisation of individual populations and their heterocellular interdependencies is essential in elucidating pathological functions of CAFs. CAFs are highly plastic and readily adopt two interchangeable transcriptional subtypes, where myofibroblastic CAFs (myCAFs) are characterised by high contractility and expression of αSMA (ACTA2), whereas inflammatory CAFs (iCAFs) produce pro-tumourigenic cytokines such as IL6, CXCL1 and CCL2 27,29,30 . Whereas genetic and pharmacologic ablation of CAFs, results in rapid progression and metastatic spread 19,20 , functional CAF reprogramming induce a microenvironment favourable to chemotherapy and ICI 11,16 , highlighting functional opposing roles of CAFs in the microenvironment. Tumour cells have been described to directly modulate CAF function though secretion of SHH, LIF, TGFβ and IL1 14,17,29,30 , however, it is less clear whether other cell populations in the TME, beyond the tumour cells, can regulate CAF function.
Here, we describe a role for OSM and its cognate receptor (OSMR) in regulating pancreatic fibroblasts and the inflammatory TME (Fig. 7e). In PDA, Osmr is expressed in mesenchymal cells such as CAFs, endothelial cells and pericytes, and Osm is expressed in immune cells such as MØs, myeloid cells, dendritic cells and neutrophils. Notably, Osmr high CAFs express higher levels of inflammatory signals such as Il6, Ccl7, Ccl2 and Cxcl1 in animal models of PDA, OSMR is expressed at increased levels in human PDA and is associated with a tumour-promoting immune infiltrate and worse patient outcome. In vitro, co-cultures of cancer cells, MØs and fibroblasts as well as recombinant OSM increase Osmr expression and production of inflammatory signals in fibroblasts. However, while the presented data indicate a role of heterocellular OSM-OSMR signalling as a regulator of inflammatory fibroblasts, future studies will be critical to establish whether OSM-OSMR signalling directly involving tumour cells and other stromal cell populations play additional roles in PDA.
Reconstituted in vitro co-cultures of tumour cells, stellate cells and MØs activate tumour cell PI3K-AKT-mTOR, MEK-MAPK and EMT pathways. This is consistent with patient data, where high levels of OSMR expression is associated with enrichment of pro-tumourigenic pathways including KRAS signalling, IL-JAK-STAT signalling, PI3K-mTOR signalling and EMT. Tumours from orthotopically implanted PCCs are smaller and appear more epithelial in Osm deficient (Osm −/− ) animals. Moreover, no overt metastasis is observed in Osm −/− animals, but further studies are needed to establish a mechanistic link between OSM signalling and metastasis-in particular given an established relationship between primary tumour size and metastatic dissemination. In Fig. 6 PCC-PSC-MØ interactions reshape tumour cell signalling. a Experimental workflow outlining SILAC-based phosphoproteomics analysis of stromaregulated tumour cell signalling. "Light" (L), "Medium" (M) and "Heavy" (H) labelled PCCs were stimulated (5 min) with PCC or PCC-PSC or PCC-PSC-MØ conditioned medium, respectively (n = 5). Pooled L/M/H peptides (in a 1:1:1 ratio) were fractionated, phospho-peptides enriched and analysed by LC-MS/MS. b PhosphoPath phospho-network analysis summarising PCC-PSC-MØ-regulated tumour cell signalling, relative to PCC-PSCregulated signals (p < 0.05, Fisher exact test with Benjamini-Hochberg multiple testing correction). Arrows indicate kinase-substrate relationships and blunt-ended edges indicate protein-protein interactions. c Western blots of phosho-STAT3 levels in PCCs treated with conditioned medium from 1, 2 or 3 W co-cultures (n = 2). d Western blots of phosho-STAT3 levels in PCCs treated with conditioned medium from 2 W or 3 W co-cultures containing either Osmr-KO or Scr-PSCs (n = 4). e Single-sample gene set enrichment analysis (ssGSEA) of PDA patients from TCGA PanCancer PAAD dataset (n = 179). Patients were stratified into OSMR high (n = 89) and OSMR low (n = 90) groups. f Representative western blot of EMT markers in PCCs treated with conditioned medium from 1, 2, or 3 W co-cultures; E-cadherin, Vimentin, SNAIL, and SLUG (n = 3). Also see Supplementary Fig. 7. Source data are provided as a Source Data file.
other solid cancers such as breast cancer and squamous cell carcinoma, OSMR is directly expressed by the tumour cells and regulates cancer stem cell properties, EMT and metastasis [56][57][58] . Importantly, while the analysed scRNAseq data did not identify OSMR expression in epithelial cells, we were unable to apply a consistent cell-type annotation across other available data sets to confirm whether OSMR was expressed in epithelial cells, thus we cannot conclusively exclude that distinct epithelial populations may express OSMR. In situ hybridisation of human PDA demonstrated that OSMR was expressed both in stromal and epithelial cells, however response to recombinant OSM was restricted to pancreatic stellate in vitro. Moreover, OSMR expression in stellate cells is required for co-culture dependent activation of phospho-STAT3 signalling in tumour cells. Thus, although future studies are required to fully establish whether OSM may directly regulate tumour cell function in human PDA, OSM signalling has been associated with pro-inflammatory and anti-fibrotic effects in cardiac fibroblasts 59 and in inflammatory bowel disease where OSM-OSMR signalling and stromal reprogramming drives colitis and mediates resistance to anti-TNFα treatment 60 . Moreover, a correlation between inflammatory fibroblasts and T cell exhaustion and dysfunction has been found in a de novo study of triple negative breast cancer 61 . Loss of OSM function in vivo is associated with a decrease in the level of several inflammatory signals with known pro-tumorigenic function, a shift in the myeloid to MØ infiltration of orthotopic tumours, as well as improved function of APCs and T cells. This suggests that OSM signalling may act to coordinate the inflammatory microenvironment through functional fibroblast reprogramming, and that the role may be conserved across multiple tissues and inflammatory diseases.
Together, the data presented here demonstrates a role for heterocellular OSM-OSMR signalling in establishing a protumourigenic environment in PDA and highlights OSM-OSMR signalling as a putative therapeutic target in PDA.

Mice and in vivo experiments.
To establish an immune-competent mouse line with whole-body depletion of Osm (Osm −/− ), the cryopreserved sperm from a male C57BL/6 Osm −/− mouse (MMRRC_048921-UCD) was purchased from The KOMP Repository (University of California, Davis, US) and fertilised in vitro with eggs from a female wildtype donor C57BL/6 mouse (purchased from Envigo). The fertilised embryos were implanted into pseudo-pregnant surrogate females. The litters were crossed with age-matched wildtype C57BL/6 mice to produce heterozygotes, which were then crossed further to produce wildtype and Osm −/− breeders. All animals were genotyped and maintained under pathogen-free and ventilated cages with environment enrichment in the Biological Resources Unit at the CRUK Manchester Institute (CRUK MI), and allowed free access to irradiated food and autoclaved water ad libitum in a 12 h light/dark cycle, with room temperature at 21 ± 2°C and humidity between 45 and 65%. All animal experiments were reviewed and approved by the Animal Welfare and Ethical Review Body ) and 1% v/v Hyclone Antibiotic/Antimycotic (Cytiva, SV30079.01). PCC (iKras mouse PDA cancer cell line) were cultured in the presence of 1 µg/mL doxycycline (Sigma, D9891). All cell lines, except PSC, were cultured for no more than ten passages to minimise culture-induced phenotypic drift. PSC cells were cultured for no more than five passages and CRISPR edited PSCs no more than nine passages. All cells used in experiments were routinely tested free of Mycoplasma. In addition, cells used in vivo were further tested for Mouse Hepatitis Virus (MHV) and Mycoplasma 5 days prior to the experiment. To ensure cell surface proteins engaged in signalling are preserved during in vitro experiments, cells were detached using Accutase Solution (Sigma, A6964). All in vitro experiments were performed in Minimal Media (DMEM with 0.5% v/v dialysed FBS (dFBS) (ThermoFisher, 26400044)) to minimise serum-induced signalling. All cell lines used and origin are listed in Supplementary Table 8.
Conditioned medium generation. To generate conditioned medium of mono-and co-cultures of PCC, PSC and MØ, a total of 2.5 × 10 6 or 7.5 × 10 6 cells was plated in either a T25 or T75 flask, respectively, and cultured in Minimal Media for 72 h. Conditioned medium from these mono-and co-cultures were passed through 0.22 µm syringe filters (VWR), and were either used fresh for stimulation experiments or stored at −80°C until used for ELISA/LUMINEX experiments.
Stimulation of target cells with conditioned medium or recombinant protein or inhibitors. Target cells were plated and serum starved in Minimal Media 18 h prior to stimulation. On the day of stimulation, media was aspirated and the target cells were stimulated with fresh Minimal Media containing either conditioned media in a 1:1 ratio (v/v) or indicated concentrations of recombinant protein or inhibitors, and cultured for 72 h. A list of recombinant proteins, inhibitors, and neutralising antibodies can be found in Supplementary Table 9.
Bone marrow-derived MØ generation. Bone marrow-derived MØs (BMDMs) were generated as follows. Femurs and tibias were harvested from 8-to 10-weekold C57BL/6 mice (Envigo). Bones were washed in 70% ethanol, followed by two subsequent washes in ice-cold PBS. Epiphyses of the femurs and tibias were cut off with a disposable scalpel, and bone marrow was flushed out with 10 mL NCM using a 25 gauge needle, and was filtered through a 70 µm cell strainer. Bone marrow cells were washed and cultured in BMDM media (NCM supplemented with 20 ng/mL M-CSF) for 3 days. Media was replaced after 3 days with fresh BMDM media and cells were further cultured for 3 additional days. The resulting cells are bone-marrow-derived MØs and used for experiments.  63 . In brief, tumours were homogenised in PBS containing Protease Inhibitor Cocktail (Sigma, P8340) and PMSF (1 mM). Tumour homogenates were then centrifuged at 14,000 × g for 30 min at 4°C and supernatants were stored at −80°C until required for ELISA assay. Following quantification of  19) and Osm −/− (n = 13) mice. Right: Quantification of iRFP signal intensity, tumour volume and tumour weight. Results displayed as mean ± SEM (Wt, n = 19 and Osm −/− n = 13), two-tailed student t test. b Left: Representative Pan-Cytokeratin (PanCK) staining of wildtype and Osm −/− tumour sections. Right: Quantification of total PanCK pos staining, mesenchymal PanCK pos staining, and epithelial PanCK pos staining following morphology analysis, and the epithelial-mesenchymal PanCK pos ratio in tumour sections from wildtype (n = 6) and Osm −/− (n = 6) mice. A shift of PanCK pos cells from a mixed mesenchymal and epithelial morphology in wildtype tumours to a more epithelial-dominant morphology in Osm −/− animals. Results displayed as mean ± SEM, two-tailed student t test. c Quantification of total Ki67, pos CD31 pos , and SLUG pos immunohistochemical staining of tumour sections from wildtype (n = 6) and Osm −/− (n = 6) mice. Results displayed as mean ± SEM, two-tailed student t test. d Upper: Representative iRFP in vivo imaging (top) and haematoxylin and eosin staining (bottom) of resected livers from wildtype and Osm −/− mice. Lower: Quantification of iRFP pos liver metastasis incidents in wildtype (n = 11/19) and Osm −/− (n = 0/13) animals. e Model outlining heterocellular OSM-OSMR signalling between pancreatic cancer cells, fibroblasts and macrophages. Pancreatic cancer cell secretion of GM-CSF induces macrophage secretion of OSM, which reprograms fibroblasts to an inflammatory phenotype, in turn accentuating tumour growth and metastasis. Also see Supplementary Fig. 8. Source data are provided as a Source Data file.
total protein concentration in each tumour sample using Pierce BCA Assay Kit (ThermoFisher, 23225), tumour sample containing 100 mg of total protein was then used for ELISA assay.
Lentivirus transduction and EGFP-, mCherry-, TagBFP-, iRFP-labelling of cell lines. To generate EGFP-, mCherry-and TagBFP-expressing cell lines, a secondgeneration lentiviral system was used. 293FT cells (ThermoFisher) were plated in a 10 cm culture dish and cultured overnight in NCM. Next day, 4 µg of SFFV-EGFP or SFFV-mCherry or SFFV-TagBFP transfer plasmid, 2 µg of pCMV delta R8.2 packaging plasmid (Addgene, 12263), 1 µg pMD2.G envelope plasmid (Addgene, 12259) and 21 µL Polyethylenimine (PEI) (Sigma, 764647) were combined in 3:1 (PEI:plasmids) v/v ratio and mixed in 1 mL of serum-free DMEM and incubated at RT for 30 min. Plasmid/PEI mixture was added directly to 293FT cells, and incubated overnight. Virus-containing supernatant was collected and stored at 4°C and replaced with fresh NCM and cells were cultured for another 24 h. Virus aliquots from both 24 and 48 h collections were combined, passed through a 0.22 µm syringe filter, and stored at −80°C until required. Target cells were plated at 50% confluence and cultured overnight in NCM. Target cells were incubated in neat virus-containing conditioned media, supplemented with Polybrene (8 µg/mL final concentration) for 24 h. Media was replaced with fresh NCM after several washes with PBS. Cells were cultured until confluence and FACS-isolated for positivity for fluorescent protein expression. Transduced cells were further isolated on FACSAria III Cell Sorter (BD Biosciences, BD FACSDiva Software) to ensure high expression of fluorescent proteins. For iRFP-labelling of PCC12 cell line, lentiviral particles expressing iRFP were generated in 293FT cells using the PGK_iRFP720-WPRE transfer, pMD2.G envelope and psPAX2 packaging plasmids by following the steps shown above. PCC12 cells were labelled with iRFPexpressing viral aliquots as shown above. High iRFP-expressing PCC12 cells were isolated by FACSAria III Cell Sorter. PCC12-iRFP cells were passaged once more and multiple vials were cryopreserved until required for in vivo experiment.
Biomark HD multiplex qPCR. Assay primers and probes were designed using the Roche Universal Probe Library Assay Design Centre Tool (https://lifescience.roche.com/ en_gb/brands/universal-probe-library.html). Refer to Supplementary Table 7 for a full list of primer sequences and respective TaqMan probe numbers of genes used in this study. Where possible, primer pairs spanning two exons were selected to avoid amplification of genomic DNA. cDNA was synthesised from 500 ng of RNA of each sample using Multiscribe RT enzyme (ThermoFisher) according to manufacturer's instructions. Reverse transcription was performed in a cycle of 25°C for 10 min, 37°C for 60 min, 95°C for 5 min and 4°C indefinitely until stored at −20°C. CRISPR-Cas9-guided Osmr knockout of fibroblasts. Osmr knockout of fibroblasts was achieved by using a multi-gRNA approach (Synthego), where PSC cells were transfected with recombinant Cas9 and three individual gRNAs against Osmr, which were designed on https://www.synthego.com/. Specifically, prior to transfection, PSC cells were washed in PBS, detached using Accutase Solution, and washed twice in PBS. Subsequently, 2.5 × 10 5 cells were resuspended in 5 µl of P3 Primary Cell Nucleofector TM Solution (Lonza). In parallel, Ribonucleoprotein (RNP) complexes were prepared by combining 4 μL of 30 μM of multi-gRNA mix (Synthego) and 1 μL of 20 μM Alt-R ® S.p. HiFi Cas9 Nuclease V3 (Integrated DNA Technologies) in 10 μL of Primary Cell P3 Nucleofector solution (Lonza) and incubated at RT for 20 min. Finally, the cell suspension, RNP mixture and 0.8 μL of 100 µM Electroporation Enhancer Solution (Integrated DNA Technologies) were combined together and mixed gently by pipetting. The RNP-cell mixture was transferred into the Nucleocuvette Strip (Lonza) and transfected using the 4D-Nucleofector Core Unit (Lonza) with the pulse "CM-137". Cells were then transferred into a T75 flask filled with pre-warmed NCM and cultured at 37°C, 5% CO 2 . Double-strand cleavage of Osmr gene was confirmed by gel electrophoresis of PCR product using primers flanking the gene-edited region. Osmr knockout was confirmed by Sanger sequencing. All the reagents used for nucleofection were from P3 Primary Cell 4D-Nucleofector X Kit S (Lonza, V4XP-3032). Target sequences of gRNAs and Sanger sequencing confirmation are included in the Supplementary  Information file (Supplementary Table 4, Supplementary Fig. 13).
Bulk RNA sequencing of fibroblasts from mono-and co-cultures Cell Culture. mCherry-labelled fibroblasts (PSC-mCherry) were mono-(n = 4) and co-cultured with PCC-EGFP (n = 4) or PCC-EGFP and MØ-TagBFP (n = 4) in T75 flasks containing Minimal Medium with a total cell number of 7.5 × 10 6 cells per condition. After 72 h of culture, cells were gently washed in PBS, detached using Accutase Solution and resuspended in FACS Buffer (1% FBS in PBS) after filtered through 70 µm cell strainer. PSC-mCherry cells from each culture condition were sorted based on mCherry positivity using FACSAria III Cell Sorter (BD Biosciences), and RNA was immediately extracted using RNeasy Plus Mini Kit (Qiagen, 74134).
RNA sequencing library preparation and analysis. Indexed PolyA libraries were prepared using the QuantSeq 3′ mRNA-Seq Library Prep Kit FWD for Illumina (Lexogen, 015.96). An input of 100 ng of total RNA was used with 14 cycles of amplification. Libraries were then quantified by qPCR using a Kapa Library Quantification Kit for Illumina (Roche, 07960336001). Single read 75 base-pair (bp) sequencing was performed by clustering 1.9 pM of the equimolar pooled libraries on a NextSeq 500 sequencer (Illumina Inc.). RNA-Seq reads were quality checked using FastQC (http://www.bioinformatics.bbsrc.ac.uk/projects/fastqc/) software and trimming of bases was performed. Raw sequencing reads were then aligned to the mouse genome assembly (GRCm38.75) in single-end mode using the STAR 2.7 RNAseq aligner 64 with default parameters. Mapped data were converted to normalised read counts (expression) using featureCounts 65 and the Ensembl GTF annotation (Mus_musculus.GRCm38.75.gtf). Differential expression (DE) was evaluated using the DESeq2 Bioconductor package 65 . Genes with relative expression of Benjamini-Hochberg-adjusted p-value < 0.05, and Log 2 Fold-Change > 1 were considered to be differentially expressed.
Orthotopic transplantation. KPC cells (PCC12, 8513) were prepared for orthotopic injection as follows. Twenty-four hour prior to the transplantation, cells were plated at 70% confluence in normal culture medium (NCM). On the day of injection, cells were washed with ice-cold PBS and dislodged using Accutase Solution. After spinning down at 300 × g, 4°C for 5 min, cells were washed two more times with ice-cold PBS before cell counting. Cell suspension was prepared at 5000 cells in 20 µL ice-cold Matrigel, and was kept on ice until injection. Mice were anaesthetised using isoflurane and subcutaneous injection of Buprenophine (Royal Veterinary College, RVC) (20 µL/10 g body weight, Buprecare) and Carprofen (RVC) (60 µL/30 g body weight, Rimadyl). Prior to surgery, fur was clipped, the exposed skin was cleaned using Chlorhexidine surgical scrub, and the animal was draped for surgery to maintain aseptic conditions. A median laparotomy was performed with a 5 mm incision made to the left abdominal flank. Mouse pancreas was carefully externalised using surgical forceps, and 20 µL of Matrigel cell suspension was injected into the pancreatic tail using an insulin syringe and needle (Braun Omnican 50, VWR). To prevent any leakage at the injection site, the needle was slowly withdrawn and the pancreas was left for 30 s. A successful injection without any leakage was confirmed by a formation of a bubble at the injection site. Muscle and skin layers were sutured with absorbable Vicryl suture (Ethicon). Prior to suturing the skin, Bupivacaine Hydrochloride (1:20 in sterile saline, Marcain Polyamp 0.5%) were administered locally to aid local anaesthesia, and animals were injected subcutaneously with 0.2 mL sterile saline. Mice were weighed twice a week, and were closely monitored for exhibiting any moderate signs of PDA such as swollen abdomen, impaired body conditioning (e.g. reduced mobility), hyperventilation and loss of body weight by more than 10%. Five weeks after the transplantation, mice were euthanised by cervical dislocation and tissues were collected accordingly for further analyses. Resected tumours were weighed, and tumour volumes were measured by external calliper using the formula, tumour volume = ½(length × width 2 ).
iRFP in vivo imaging on LI-COR Pearl imager. Prior to imaging in vivo, mice were sacrificed via cervical dislocation and pancreatic tumours and livers were resected. Tissues were imaged for iRFP excitation (emission wavelength of 685 nm, 700 channel) on Pearl imager using Image Studio V5 software with the default setting for animal imaging as per manufacturer's instructions. All quantifications were performed on Image Studio V5 as per manufacturer's instructions.
Tumour disaggregation for flow and MC. Tumours were freshly isolated from wildtype or Osm −/− animals and immediately placed into ice-cold PBS (Fisher Scientific, 10091403), and kept on ice. All non-tumour tissues attached to tumours were thoroughly removed before proceeding further. Tumours were washed once with ice-cold RPMI media and minced into~1 mm pieces using disposable scalpels in 2 mL of Disaggregation Buffer (DB; 2 mg/mL Collagenase Type IV (Thermo Fisher, 17104019), 1 mg/mL DNase1 (Sigma Aldrich, 10104159001) and 0.5 mg/ mL Hyaluronidase (Sigma Aldrich, H3757) in RPMI). Tumour fragments were transferred into C-Tubes and individually disaggregated using GentleMACS Octo Dissociator (Miltenyi Biotech) on automated "37C_m_TDK1" programme. Once complete, the C-tubes were centrifuged at 100 g for 2 min, and the cell pellets were resuspended with fresh DB, and filtered through a 70 μm strainer. Cells were then washed with 10 mL ice-cold RPMI to quench the digestion. The single-cell suspension was pelleted at 300g for 6 min and used for flow and MC. Antibodies used are can be found in Supplementary Tables 1-3 and 6. CAF isolation from mouse tumours. Single-cell dissociation of freshly isolated tumours was performed as described above. Cells were counted for each sample, and CD45 + cells were depleted using mouse CD45 MicroBeads (Miltenyi Biotec, 130-052-301) and LS columns (130-042-401) following manufacturer's recommendation. To discern viable cells, CD45 + -depleted samples were stained with LIVE/DEAD Near-IR Dead Cell Stain Kit (LD) (ThermoFisher, L10119) following manufacturer's recommendation. Cells were then washed and pelleted and stained with Fc block (BD Biosciences, 558636) for 5 min. A master mix of lineage and fibroblast marker antibodies was then added to each sample and incubated on ice for 30 min. Stained cells were washed in PBS, pelleted and resuspended in FACS Buffer (1% FBS in PBS) and kept on ice until FACS-isolation. Viable CAFs were sorted on BD FACSAria III Cell Sorter based on LD − CD45 − EpCAM − CD31 − PDPN + CD90 + staining profile, and data analysed using Flowjo (BD Biosciences). Sorted CAFs were rapidly lysed and their RNA was extracted using RNeasy Plus Micro Kit (Qiagen, 74034) for RT-qPCR assay.
MC antibodies and antibody conjugation. Fibroblast and immune cell (myeloid/ NK/B cell (M/N/B) and T cell) antibody panels used for MC analyses in this study were generated as follows 21 . The dedicated panel builder (https://www.dvssciences.com) was used to estimate isotope and oxide spill-over and guide channel selection. Indicated antibodies were purchased pre-conjugated from Fluidigm, whereas all other antibodies were purchased in a carrier protein-free form and labelled with the indicated heavymetal tag using Maxpar X8 Antibody Conjugation Kit (Fluidigm) as per manufacturer's recommendations.

MC barcoding and antibody staining
Fibroblast MC analysis. Prior to setting up in vitro cultures, PCCs and MØs were labelled with EGFP using a lentivirus approach, as described above. PSC monoculture (n = 3), PCC-PSC (n = 3) and PCC-PSC-MØ (n = 3) co-cultures were prepared as described above. After 72 h of in vitro cultures, 10 μL of 10 mM 5-iodo-2′-deoxyuridine (IdU) (Sigma Aldrich, 17125) solution in 0.2 M NaOH/water was added directly to the media, gently swirled and the cells incubated at 37°C for 20 min. Cells were then washed in PBS, dislodged using Accutase Solution and quenched with CSM (Cell Staining Buffer), consisting of 5 mg/ml bovine serum albumin (BSA) (Sigma Aldrich, A3294) in PBS. Cells were filtered through a 70 μm cell strainer (BD Biosciences, 340633). Subsequently, cells were resuspended and counted.
Barcoding. Totally, 3 × 10 6 cells from each culture condition were aliquoted into individual 5 mL polypropylene FACS tubes, pelleted and gentle mixed. A unique pair of lanthanide-conjugated anti-ITGB1 antibody pairs in 50 μL of CSM was then added to each sample. This live-cell barcoding method leveraged the consistently high expression of ITGB1 in fibroblasts, and in this instance a 5-choose-2 scheme (ITGB1-115In, -163Dy, -166Er, -169Tm, -172Yb) was used. Each sample was gently mixed and incubated on ice in the dark for 30 min. Cells from each sample were washed twice with CSM and combined into a 5 mL polypropylene FACS tube.
Live/dead cell staining. The pooled cells were resuspended in 300 μL of ice-cold PBS and vortexed well. Totally, 300 μL of 1 μM cisplatin (Fluidigm, 201064) in PBS was added into the cells, and incubated for 1 min. The staining was quenched with CSM, and cells were pelleted and aspirated. Subsequently, 60 μL of 100 U/mL heparin sodium salt (Sigma Aldrich, H3393) solution in PBS and 3 μL Fc block (BD Biosciences, 558636) was added to the cells. The cell mixture was gently mixed and incubated on ice for 5 min.
Fluorophore antibody staining. A master mix of fluorophore-conjugated antibodies in 150 μL CSM was added, gently mixed and incubated on ice in the dark for 45 min. Cells were washed with CSM and pelleted. Extracellular antibody staining. A master mix of 6 equivalents of extracellular targeting, metal-conjugated antibodies in 150 μL of CSM was added, gently mixed and incubated on ice in the dark for 45 min. Cells were washed twice with CSM.
Fixation and permeabilisation. The cell pellet was resuspended in 100 μL of PBS, vortexed and 2 mL of 1× FOXP3 Fixation Buffer from FOXP3 Fixation/Permeabilization Kit (Thermo Fisher, 00-5523-00) was added, followed by thorough vortexing. After 30 min incubation at RT, 2 mL of 1× FOXP3 Permeabilization Buffer was added and the cells were pelleted. Additional 2 mL of 1× FOXP3 Permeabilization Buffer was added and the cells were pelleted. Sixty microlitre of 100 U/mL heparin sodium salt in PBS and 3 μL of Fc block was added and the sample was mixed by gently rocking.
Intracellular antibody staining. Intracellular antibody staining. After incubating for 5 min at RT in the dark, a master mix of 6 intracellular targeting, metal-conjugated antibodies in 150 μL CSM was added. The sample was gently mixed and incubated on ice in the dark for 45 min. Cells were washed twice with CSM. The cell pellet was resuspended in 1 mL of PBS and vortexed well. Two mililtre of 4% paraformaldehyde (PFA) (Thermo Fisher, 28908) in PBS was added. The sample was and stored overnight at 4°C in the dark until the MC analysis following day.

MC analysis of immune cells in mouse tumours.
Single-cell dissociation of freshly isolated tumours was performed as described above and in ref. 21 . MC data processing. FCS files were normalised for signal-drift during the acquisition run using the in-built Helios normalisation tool (Fluidigm). For in vitro fibroblast analysis, individual sample events were deconvoluted using manual gating based on the pre-determined ITGB1 barcode after removing non-fibroblast cell types based on expression of canonical markers and GFP. For immune cell analysis of mouse tumours, individual sample events were deconvoluted using the a stand-alone debarcoder 67 , with a Mahalanobis distance of 15 and a minimum barcode separation of 0.3. Individual sample FCS files were uploaded to the cloudbased cytometry platform Cytobank (https://www.cytobank.org, Beckmann Coulter) and checked for consistent signal across the entire acquisition period and correct deconvolution. As per standard methods, live cell events were selected based on DNA-191Ir positivity and cisplatin-198Pt negativity. 191Ir+ debris, remaining non-fibroblast cell types, cell doublets and aggregates were removed. M/ N/B and T cells were selected by biaxial gating as CD45 + CD3εand CD45 + CD3ε, + respectively, and exported from Cytobank as FCS files, which were subsequently loaded on Cytofkit2 (version 2.0.1). To analyse the different cell populations in an unbiased way, cells were clustered using FlowSOM 40 and visualised using UMAP projections 39 , exporting cell data with annotated clusters for further downstream analysis. Clusters were assigned to specific immune cell populations based on the expression of respective markers, and clusters belonging to the same cell population were merged. For memory T cells, which were not allocated to a specific cell cluster by FlowSOM, were gated manually on biaxial plots, as shown in Supplementary Fig. 6i.
Bioinformatic processing of publicly available murine and human PDA scRNAseq datasets. Two publicly available single-cell RNAseq datasets of murine PDA and one dataset of human PDA were re-analysed in this study. The murine PDA data from KPC GEMM 27 was acquired from Gene Expression Omnibus (GEO) under the accession number GSE129455, whereas the human PDA data 27 was obtained from NCBI dbGaP under the accession number phs001840.v1.p1. Additionally, the mouse PDA data from KPP GEMM 33 was obtained from the ArrayExpress database under accession number E-MTAB-8483. All the subsequent processing and analyses of scRNAseq data described herein were performed using Seurat package (v3.0.2) in R 4.0.0. From the KPC mouse PDA data 11,260 cells expressing 13,813 genes, 14,915 cells expressing 24,392 genes from the KPP mouse PDA, and 21,200 cells expressing 16,267 genes from the human PDA data were selected for downstream analyses. Subsequently, the built-in ScaleData function was used to normalise and scale the data. Using the variance stabilising transformation (vst regime) on the built-in FindVariableFeatures, we identified 2000 most variable genes and applied PCA by RunPCA function. For dimensionality reduction, 50 principal components were subjected to elbow test (ElbowPlot) and "JackStraw" estimation (JackStraw, ScoreJackStraw and JackStrawPlot) for each cell. Subsequently, Seurat's graph-based kNN approach (k-Nearest Neighbours) was applied for cell clustering of top 20 principal components (Dimension 20 and resolution 0.5). Annotation of cell types for each cell was made based on the expression of cell type markers indicated in the original articles. UMAP projection was used for visualisation of the data (RunUMAP). Scatter plot visualisations were done by Seurat's functions DimPlot and FeaturePlot. For violin plot visualisation, Seurat's function VlnPlot was used.
Ligand-receptor interaction analysis. Publicly available single-cell RNAseq datasets of murine and human PDA 27 were subjected to ligand-receptor interaction analysis using CellPhoneDB in Python 3.8.5 34 . Only MØ-fibroblast ligand-receptor interactions were tested in this study. To this end, 700 MØs and 700 fibroblasts were extracted from respective datasets and used as input for CellPhoneDB analysis using default settings. For mouse interaction analysis, gene names were replaced with human orthologues using biomaRt package in R 4.0.0. Only the MØ-fibroblast ligand-receptor interactions with p-value < 0.05, which simultaneously exhibit MØ-macrophage and fibroblast-fibroblast interactions with pvalue > 0.05, were selected for further analysis. In addition, only the ligands with "secreted" function were selected for further analyses. Putative interaction pairs were visualised using ggplot2 package in R 4.0.0.
In-solution protein digestion. The conditioned medium-stimulated PCC SILAClabelled cells were immediately washed twice with ice-cold PBS, lysed in TEAB Lysis Buffer (0.5 M TEAB, 0.05% SDS, phosphatase inhibitors) on ice and sonicated for 3 cycles (30 s sonication and 30 s incubation) on ice. Samples were centrifuged at 16,100g for 15 min at 4°C. Supernatants were collected and protein concentration was determined using Pierce BCA Assay Kit. 766.66 µg of "Light"-, "Medium"and "Heavy"-labelled PCC lysates (a total of 2.3 mg) were combined in a 1:1:1 ratio. Subsequently, the pooled proteins were reduced with 10 mM TCEP (at RT for 30 min), alkylated with 50 mM iodoacetamide (ThermoFisher) (at RT for 15 min in the dark), and digested with 46 µg of Trypsin (Promega, V5111) at 37°C for 16 h. Next day, proteins were further digested with 46 µg of Trypsin at 37°C for 4 h. Peptides were acidified in 5% trifluoroacetic acid (TFA), and dried down on Speedvac Concentrator (ThermoFisher) at 60°C for 2 h prior to reverse-phase fractionation.
Reverse phase off-line peptide fractionation. Peptides were fractionated as previously described 77 . Peptides were separated on a Zorbax Extend-C18 column (4.6 × 150 mm, 3.5 µm, Agilent Technologies) at 250 µL/min using the following gradient profile (minutes: %B); 5:0.5, 20:30, 24:40, 26:75, 29:75, 30:0.5 and 55:0.5. The buffers used were Buffer A (LC-grade water supplemented with 0.1× v/v NH 4 OH (pH 10.5)) and Buffer B (100% Acetonitrile). The eluent was collected into 96 round-bottom plates and fractions were collected into each well every 15 s. Only fractions in the elution window with traces of high amounts of peptide material were used and all of the fractions were concatenated into 10 final fractions with each containing 200 µg peptides on average. Fractions were dried using Speedvac Concentrator (Thermo Fisher) at 60°C for 2 h.
Data-dependent acquisition (DDA) LC-MS/MS. Phosphopeptides were analysed using an Ultimate 3000 RSLCnano system (Thermo Scientific) coupled to a LTQ OrbiTrap Velos Pro (Thermo Scientific). Peptides were initially trapped on an Acclaim PepMap (C18, 100 µm × 2 cm) and then separated on Easy-Spray PepMap RSLC C18 colum (75 µM × 50 cm) (Thermo Scientific) over a 156 min gradient from 0 to 98% Buffer B (80% ACN in 0.08% FA) against Buffer A (0.1% FA) at a flow rate of 300 nl/mL. Samples were transferred to mass spectrometer via an Easy-Spray source with temperature set at 50°C and a source voltage of 2.0 kV. The mass spectrometer was operated in data-dependent acquisition mode with Multi Stage Activation for phosphorylation neutral loss detection. Survey full scan of MS spectra (335-1800 m/z) were acquired in the Orbitrap with a resolution of 60,000 and FTMS Full AGF Target of 1,000,000. Top 15 method was selected for fragmentation using collision-induced dissociation. The resulting raw files were searched against the SwissProt Mus musculus database on Mascot server (Matrix Science; 2016) in Proteome Discoverer 2.1 (ThermoFisher Scientific). Search parameters included a precursor mass tolerance of 10 ppm, fragment mass tolerance of 0.8 Da, peptide modifications of carbamidomethylation (C) as static modification and oxidation (M, P and K) as well as deamination (N, Q) and phosphorylation (S, T, Y) as dynamic modification. A decoy database search was performed to determine the peptide FDR with the Percolator module. Phosphosite localisation probabilities were calculated with PhosphoRS node. A 1% peptide FDR threshold was applied, and peptides were filtered for medium peptide confidence, minimum peptide length of 6, and finally peptides without protein reference were removed. All peptides that exhibited a confidence of less than "high" and with less than two uniquely identified peptides were excluded from further analysis.
Phosphoproteomics data analysis. Quantified phosphopeptides used for downstream analysis were filtered by applying the following criteria: (1) peptide quantification in all three SILAC channels (light, medium and heavy), (2) presence of only one unique phosphorylation site per peptide, (3) phosphorylation-site localisation probability of >75%, (4) detection and quantification of each unique phospho-peptide in at least two out of five biological replicates. Phosphopeptide abundances were normalised by the median abundance of the SILAC-channel before a SILAC-ratio was computed. Phosphorylation-motifs were obtained for each phospho-peptide by retrieving the flanking ±6 amino acids around the detected phosphorylation locus by aligning the peptide-sequence to the Uni-ProtKB/Swiss-Prot database (accessed: 10/08/2019). Kinase-recognition motifs were identified by using the PhosphoSitePlus (v6.5.9.3) 79 . MotifAll algorithm by applying a significance threshold of 1e−06 and a support threshold of 0.05 with against background set to automatically selected phosphorylation sites. Kinases were then identified using the MEME-suite CentriMO 80,81 local motif-enrichment algorithm in pre-identified motifs by applying default options against the Eucaryotic Linear Motif (ELM 2018) motif database. For phospho-network analysis and kinase-substrate network visualisation, the Cytoscape software package 82 was utilised by incorporating the PhosphoPath app 75 with default settings by selecting the following databases:(1) PhosphoSitePlus for kinase-substrate (KS) interactions set to human KS interactions, (2) BIOGRID for protein-protein interactions and (3) WikiPathways for pathway information. Pathway-enrichment was conducted by selecting the whole proteome as background and only pathways with a significant enrichment of p-value < 0.05 (Fisher exact test with Benjamini-Hochberg multiple testing correction) were incorporated into the analysis.
Phospho-antibody array analysis. For phospho antibody array analysis of tumour cell signalling, PCCs were plated in 6-well dish and cultured in Minimal Media overnight. CM was freshly generated from mono-and co-cultures of PCC, PSC and MØ, as shown above. PCCs were stimulated with respective conditioned medium for 5 min at 37°C 5% CO 2 , and the lysates were immediately collected and analysed using the PathScan Intracellular Signalling Array Kit (CST, 7744) as per manufacturer's recommendation. Fluorescent intensities were analysed on Image Studio (LI-COR).
CIBERSORT analysis. The CIBSERSORT cell estimation method was utilised to estimate immune cell levels in the TME of the PAAD-TCGA cohort as previously described 83,84 using the CIBERSORT R implementation (v1.04). Only absolute CIBERSORT values for immune-cell populations were used. Spearman-based rank correlation of the z-score normalised OSMR expression levels and the CIBERSORT immune abundances was chosen.
Immunohistochemistry, staining quantification and cell morphology analysis. Immunohistochemistry staining was performed using the BOND RX automated platform (Leica Microsystems). Four micrometre sections of formalin-fixed paraffin-embedded (FFPE) sections were cut and mounted on charged slides. Dewaxing and heat induced epitope retrieval (HIER) of slides was automated on the Bond RX using epitope solution 1 (AR9961) for 20 min at 100°C. Using the Bond Refine Kit (DS98007) as per manufacturer's instructions, briefly endogenous peroxidase was blocked (hydrogen peroxide, 10 min) followed by a non-specific protein block (casein, 20 min) prior to primary antibody application (15 min). Following buffer washes, the post primary linker was omitted and the labelled polymer was applied (8 min) followed by visualisation with diaminobenzidine (10 min) and nuclei haematoxylin counterstain (5 min). The following antibodies concentrations were used: Pan Cytokeratin (Abcam, 9377) 2 µg/ml. For αSMA (Sigma Aldrich, A5228) 0.4 µg/ml, the Agilent animal research kit (K3954) was used as per manufacturer's instructions. Both endogenous peroxidase (3% hydrogen peroxide, 10 min) and non-specific protein (casein 20 min) (Vector, SP5020) blocks were applied. Following this biotin-labelled primary antibody was applied (30 min), followed by streptavidin-peroxidase (30 min) and visualisation with diaminobenzidine (5 min). Nuclei were then counterstained with Gills haematoxylin (30 s). Following counterstaining, all sections were dehydrated through graded alcohol, cleared in xylene and coverslipped. Heamatoxylin and eosin (H & E) staining was performed on 4 µm FFPE sections. Briefly, following dewaxing and rehydration, sections were stained with Gills Haematoxylin (ThermoFisher 6765007) for 3 min and following an extended water wash, counterstained with Eosin (ThermoFisher 6766007) for 1 min. Sections were then coverslipped following dehydration in graded alcohol and clearing in xylene. These stained tumour sections were imaged using Leica SCN400 Slide Scanner (Leica). Quantification of positive marker staining of slides was performed using HALO (Indica Labs). Morphology analysis of PanCK pos cells was performed using the Tissue Classifier machine learning algorithm on HALO. A list of antibodies used can be found in Supplementary Table 5.
Statistics analysis and data presentation. General graphical visualisation of data and statistical tests were performed using Prism 7 (GraphPad Software Inc) and various packages on R (4.0.0). Heatmaps and unsupervised hierarchical clustering (Ward's linkage, Euclidean) were generated using R package, Heatmap.2. PCA plots were made using prcomp function on R package, ggplot2.