LXRα limits TGFβ-dependent hepatocellular carcinoma associated fibroblast differentiation

Transforming growth factor β (TGFβ) is deposited in the extracellular space of diverse tissues. Resident fibroblasts respond to TGFβ and undergo myofibroblastic differentiation during tissue wound healing and cancer progression. Cancer-associated fibroblasts (CAFs) communicate with tumor cells during cancer progression, under the guidance of TGFβ signaling. We report that agonist-activated liver X receptors (LXR) limit the expression of key components of myofibroblast differentiation, including the α-smooth muscle actin (αSMA) gene in liver cancer cells. CAFs derived from hepatocellular carcinoma (HCC) express high αSMA and low LXRα levels, whereas hepatocarcinoma cells exhibit an inverse expression pattern. All hepatoma cells analyzed responded to the LXRα agonist T0901317 by inducing fatty acid synthase (FASN) expression. On the other hand, T0901317 antagonized TGFβ-induced fibroblastic marker responses, such as fibronectin and calponin, in a subset of hepatoma cells and all CAFs analyzed. Mechanistically, LXRα antagonized TGFβ signaling at the transcriptional level. Smad3 and LXRα were recruited to adjacent DNA motifs of the ACTA2 promoter. Upon cloning the human ACTA2 promoter, we confirmed its transcriptional induction by TGFβ stimulation, and LXRα overexpression repressed the promoter activity. Hepatosphere formation by HCC cells was enhanced upon co-culturing with CAFs. T0901317 suppressed the positive effects exerted on hepatosphere growth by CAFs. Taken together, the data suggest that LXRα agonists limit TGFβ-dependent CAF differentiation, potentially limiting primary HCC growth.


Introduction
Misregulation of transforming growth factor β (TGFβ) signaling occurs in infectious, cardiovascular diseases and cancer 1 . TGFβ halts cell proliferation and accelerates cell death, as it occurs in immune cells or during liver homeostasis 2,3 . Moreover, TGFβ promotes epithelialmesenchymal transition (EMT) and stimulates fibroblasts to secrete cytokines, chemokines and extracellular matrix (ECM) molecules 2,3 . Cellular responses initiate when TGFβ binds to the type II (TβRII) and type I (TβRI) kinase receptor complex 4 . Within this complex, TβRII trans-phosphorylates TβRI, activating TβRI to phosphorylate Smad2 and Smad3, which, in complex with Smad4, regulate gene transcription. The TGFβ receptor also activates mitogen activated protein kinase (MAPK), the tyrosine kinase Src and phosphatidylinositol-3´-kinase pathways. Upon ligand-induced cleavage, the TβRI intracellular domain regulates transcription 4,5 .
In adult tissues, TGFβ regulates EMT and fibroblast maturation 6,7 . Transcriptional programs become activated in epithelial cells by TGFβ, mediated for example by Snail and Slug, leading to repression of cell-cell contact genes, induction of cytokine genes as well as ECM genes, such as fibronectin 8 . EMT generates mesenchymal-like cells that partially resemble fibroblasts and contribute to chronic tissue fibrosis and cancer progression 9 . Cancer-associated fibroblasts (CAFs) secrete TGFβ and other cytokines to assist tumor cell invasion and suppress anti-tumor immune responses 10,11 . Myofibroblast differentiation is a key response to TGFβ signaling, mediated by Smad complexes with β-catenin and parallel MAPK signaling 6,12 . The contractility of CAFs depends on microfilament networks built by α smooth muscle actin (αSMA) and associated myosins, tethered via membrane receptors to specialized collagen fibers, all induced by TGFβ signaling 6,12 .
Among several malignancies, TGFβ regulates the progression of hepatocellular carcinoma (HCC), a cancer with high-mortality rate worldwide 13,14 . TGFβ suppresses HCC by arresting hepatocyte proliferation at the early G1 phase of the cell cycle, and inducing apoptotic responses 13,14 . TGFβ promotes EMT, survival and CAF differentiation, contributing to HCC invasion and metastasis [14][15][16] . Similar to TGFβ, the nuclear liver X receptors (LXRα/NR1H3 and LXRβ/NR1H2), either suppress or promote cancer, by inhibiting cell proliferation or helping tumor cells to escape from immune detection, respectively 17 . The LXRs are expressed in hepatocytes among other tissues; agonistic LXR ligands include the oxysterols, metabolic derivatives of cholesterol, inducing the transcriptional activity of LXRs 18 .
We previously demonstrated that LXRα suppresses TGFβ-induced differentiation 19 , and, in the context of HCC, we described a role for the transcription factor Snail as a target of the crosstalk between TGFβ and LXRα 20 . Due to the established role of TGFβ on CAF biology, we here analyzed signaling crosstalk between TGFβ and LXRα in human HCC fibroblasts

Results
LXRα expression is enriched in epithelial, whereas the TGFβ-induced myofibroblast gene αSMA is enriched in mesenchymal, HCCs We screened HCC cell lines previously classified based on their response to TGFβ signaling, as epithelial-like cells with an early TGFβ target gene signature (Huh7, PLC/PRF5, Hep3b, HepG2) and as mesenchymal-like cells with a late signature (SNU398, SNU423, HLF, SNU449) 21 . Using E-cadherin (epithelial) and vimentin (mesenchymal) as marker genes, we found good concordance of mRNA and protein expression profiles with the classification (Fig. 1a-c), but also detected vimentin mRNA and protein expression in Huh7 and Hep3b cells (Fig. 1a-c). Snail mRNA expression was used as an intermediate mediator of mesenchymal differentiation, and was found to be expressed by most HCC cells (Fig. 1a).
All epithelial HCC cells expressed high, whereas the mesenchymal HCC cells expressed low, LXRα mRNA and protein levels (Fig. 1a, d, e). Fatty acid synthase (FASN) mRNA, a direct readout of LXRα transcriptional activity 17 , had a rather similar to the LXRα mRNA expression profile (Fig. 1a). The response of the HCC cells to the wellestablished LXR agonist, T0901317 17 , was measured by analyzing FASN protein and mRNA expression. The epithelial HCCs expressed endogenous FASN protein (Fig.  1d), reflecting the mRNA profile (Fig. 1a, Suppl. Fig. S1), and T0901317 stimulation induced FASN in all epithelial HCCs examined, which was easier to detect at the mRNA (Suppl. Fig. S1) than at the protein level (Fig. 1d). As previously reported 19,20 , T0901317 stimulation enhanced LXRα levels by 2-5-fold in epithelial HCCs (Fig. 1d densitometry). TGFβ stimulation had no appreciable effect on FASN expression, and combination of TGFβ with T0901317 normalized FASN levels to basal in Huh7, Hep3b and HepG2 cells (Fig. 1d). LXRα stabilization appeared somewhat reduced after co-treating the cells with TGFβ and T0901317, but only in Hep3b cells (Fig. 1d).
In the mesenchymal HCCs SNU398, SNU423, HLF and SNU449, basal LXRα protein expression was at the limit of detection, and T0901317 stimulation did enhance LXRα levels so that they became detectable (Fig. 1e). Accordingly, the mesenchymal HCCs expressed basally endogenous FASN (Fig. 1a, Suppl. Fig. S1) and T0901317 stimulation induced FASN mRNA levels to a comparable degree as in epithelial HCC cells (Suppl. Fig.  S1); this effect appeared weaker when FASN protein levels were measured (Fig. 1e). Combination of TGFβ and T0901317 stimulation resulted in relatively weaker induction of FASN by T0901317 in two, but not the other, mesenchymal HCCs examined (Fig. 1e, Suppl. Fig. S1). The data indicate that many HCC models respond to LXRα agonist and TGFβ stimulation; in certain HCCs, TGFβ partially antagonizes the stimulatory effect of T0901317 on FASN expression and LXRα stabilization.

Activation of LXRα suppresses myofibroblastic genes induced by TGFβ
αSMA represents a hallmark gene of activated fibroblasts 6,12 ; in contrast to the LXRα expression profile, αSMA levels were low and high in epithelial and mesenchymal HCCs, respectively (Fig. 1a). Only HepG2 cells expressed high LXRα, FASN and αSMA mRNA levels (Fig. 1a, blue bar differentiates HepG2 from other HCC cells). In agreement with the mRNA profiles, only HepG2 and SNU398 cells expressed αSMA protein, whose levels only slightly changed upon TGFβ or T0901317 stimulation (Fig. 1f, g).
We also examined fibronectin and calponin expression, as additional readouts of TGFβ response and fibroblast activation (Fig. 1f, g). In epithelial HCC cells, TGFβ induced fibronectin (all cells tested) and calponin (all cells except PLC/PRF5) and T0901317 reduced this response primarily in Hep3b cells (Fig. 1f). In the mesenchymal HCC cells (except SNU398), TGFβ induced fibronectin whereas T0901317 did not exhibit any appreciable effect (Fig. 1g); TGFβ also induced calponin and T0901317 normalized the induction to basal level (Fig. 1g). Our observations suggest that T0901317 can suppress TGFβ-induced myofibroblastic markers in about 50% of the HCC cell models. We then examined recently isolated human HCC primary CAFs (AKH12, AKH14 and AKH38) 22,23 . All three CAF models responded to T0901317, as revealed by FASN induction (Fig. 2). The CAFs expressed detectable LXRα and LXRβ mRNA levels. TGFβ stimulation enhanced LXRβ mRNA expression only in AKH12 cells (Fig. 2). Myofibroblast gene profiling indicated that αSMA (AKH14 being an exception), SM22α and calponin mRNA expression were induced by TGFβ; in AKH38 cells, T0901317 suppressed the TGFβ effects, whereas suppression of αSMA was measurable but not significant in AKH12 cells, where calponin expression was even enhanced after co-stimulation (Fig. 2).
The mRNA expression analysis was corroborated by analysis of the corresponding proteins (Fig. 3a). Induction of αSMA, calponin and fibronectin by TGFβ was clearly suppressed in the presence of T0901317 in AKH12 CAFs, less potently in AKH38 CAFs, but not in AKH14 cells (Fig. 3a). FASN protein analysis in the CAF models responding to T0901317 stimulation confirmed the mRNA data ( Fig. 2), and TGFβ suppressed the positive effect of T0901317 on FASN (Fig. 3a). The very low endogenous LXRα protein levels in primary CAFs were difficult to assess, but in AKH38 cells, stabilization due to T0901317 stimulation made LXRα detectable (Fig. 3a).
Microscopic analysis of actin-based cytoskeletal organization confirmed abundant accumulation of stress fibers containing β-actin and αSMA and formation of actinsupported membrane ruffles in the three CAFs (Fig. 3b). CAFs stimulated with TGFβ, T0901317 or both exhibited similar phenotype (Fig. 3b). The gene and protein expression data support the responsiveness of patient-derived HCC CAFs to the LXRα agonist T0901317, and an antagonistic effect of LXRα signaling against TGFβ-mediated myofibroblast gene expression in certain HCC CAFs.
TGFβ and LXRα regulate the human αSMA/ACTA2 gene promoter We examined a possible mechanism by which LXRα signaling might regulate the myofibroblastic response to TGFβ by focusing on the αSMA/ACTA2 gene. Since previous studies of ACTA2 gene regulation focused on mouse or rat cells, we cloned a fragment of the human ACTA2 gene spanning the transcriptional start site, extending from −1400 to +50 bp relative to the start site (Fig. 4a).
Bioinformatic analysis of the promoter sequence identified putative TGFβ/Smad-binding elements (CAGA motifs), and LXR-responsive elements (Fig. 4a). ChIP experiments in the HCC CAFs could not succeed due to the low number of primary CAFs available, and primary human foreskin AG1523 fibroblasts were chosen as a good approximation based on established antagonism between TGFβ and LXRα in this cell type 19 . ChIP assays in AG1523 fibroblasts revealed specific binding of Smad2/ Smad3 to three adjacent regions of the ACTA2 promoter, containing CAGA motifs (Fig. 4b). As a positive control, the plasminogen activator inhibitor 1 (PAI-1) gene promoter, a well-characterized target of TGFβ signaling, was analyzed under the same conditions (Fig. 4b). ChIP assays for LXRα could not generate positive results due to the insufficient quality of our antibodies.
In order to bypass the above deficiency and examine binding of LXRα on the ACTA2 promoter, we analyzed direct binding of Smad3 and LXRα to a synthetic human ACTA2 gene promoter DNA fragment, which encompasses a cluster of CAGA and LXRα motifs (Fig. 4a). Using protein extracts from human embryonic kidney 293-T (Fig. 4c, d) or human HCC HepG2 cells (Fig. 4e), we detected binding of both Smad3 and LXRα. When Smad3 and LXRα were overexpressed (Fig. 4c), or when only LXRα was overexpressed and Smad3 was endogenous (Fig.  4d), TGFβ or T0901317 stimulation did not affect significantly their binding to the ACTA2 fragment (Fig. 4d, quantification). In HepG2 HCC cells, endogenous Smad3 and LXRα binding to the αSMA/ACTA2 fragment was recorded, and stimulation with TGFβ, T01901317 alone or together did not significantly affect this binding (Fig. 4e). As positive control for robust TGFβ-dependent Smad3 binding to DNA, a concatameric CAGA DNA fragment was used (Fig. 4f), demonstrating binding in agreement with the PAI-1 promoter ChIP data of Fig. 4b.
Transient transfection experiments of the ACTA2 gene promoter-luciferase construct (Fig. 5a) in HepG2 cells revealed its inducible activation by TGFβ stimulation (Fig. 5b). Smad3 and Smad4 co-expression also exhibited a trend of inducing this promoter, but did not score (see figure on previous page) Fig. 2 The LXRα agonist T0901317 suppresses TGFβ-induced myofibroblast gene expression in CAFs. AKH12, AKH14 and AKH38 CAFs were treated in complete medium for 74 h; TGFβ1 (5 ng/ml) was administered with or without LXR agonist (5 μM T0901317). Real-time PCR analysis of the indicated gene expression normalized to the expression of GAPDH. Mean ± SD values are plotted. Experiments were performed in biological triplicates (n b = 3), each of them in technical triplicate (n t = 3). Statistical comparison (two-sided t-test) indicates significant differences, *p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001 significant (Fig. 5c). The ACTA2 promoter exhibited a trend for reduced activity in the presence of two different LXRα agonists, T0901317 and GW3965 (Fig. 5d), whereas upon LXRα overexpression at two different concentrations, the promoter activity was reproducibly and more significantly repressed (Fig. 4e). Combining TGFβ and T0901317 stimulation resulted in the same negative trend relative to the positive effect of TGFβ stimulation, yet the repressive effect of LXRα was not significant (Fig. 5f). We also analyzed the CAGA 12 -luciferase promoter, a synthetic promoter that potently responds to TGFβ/Smad signaling 24 . In the AG1523 fibroblasts, as in HCC, the synthetic promoter was activated by TGFβ, whereas T0901317 co-administration exhibited a trend to diminish the promoter activation, which was not significant (Fig. 5g). Thus, increasing LXRα expression causes a reduction of the ACTA2 promoter activity. All observations support a hypothesis whereby LXRα may negatively regulate the αSMA/ACTA2 gene promoter.
HCC CAFs promote hepatosphere growth, which is reduced upon LXRα activation CAFs contribute to tumor progression in response to TGFβ signaling 10,11 and during HCC development 15,16 . In  Fig. 3 The LXRα agonist T0901317 limits TGFβ-induced myofibroblast differentiation in CAFs. AKH12, AKH14 and AKH38 CAFs were treated in complete medium for 74 h; TGFβ1 (5 ng/ml) was administered with or without LXR agonist (5 μM T0901317). a Expression of the indicated proteins was determined by immunoblotting. Experiments were performed in biological duplicates (n b = 2). An arrow marks the correct protein band and a star a non-specific protein recognized by the LXRα antibody. Molecular size markers are indicated in kDa. b F-actin (red) and α-SMA (green) microfilaments and DAPI (blue) in CAFs, after 48 h of the indicated treatments. Bars represent 10 µm. Experiments were performed in biological triplicates (n b = 3) this context, the role of LXRα remains unexplored. In order to simulate tumor tissue organization, we employed 3D culture conditions that lead to the development of tumor spheroids (hepatospheres) and examined the role of CAFs and LXRα signaling. Hep3b cells assembled typical hepatospheres, while SNU449 formed multicellular aggregations (Fig. 6a). Co-culture with AKH12 CAFs in a cell-to-cell ratio of 1:3 or 1:6 (AKH12 to HCC cells), resulted in larger Hep3b hepatospheres, which formed with higher frequency; AKH12 co-culture with SNU449 showed a more pronounced phenotype as these co-cultures generated architecturally well-organized hepatospheres that formed with higher frequency relative to the SNU449 cells cultured alone (Fig. 6a). We pretreated AKH12 CAFs with T0901317 for 24 h prior to co-culture with HCC cells, leading to reduced hepatosphere size and number (Fig. 6b). As a specificity control, we pretreated CAFs with a chemical inhibitor of the TGFβ receptor type I kinase (LY2157299), which caused a small increase in hepatosphere size or number (Fig. 6c). Thus, activation of LXRα suppressed 3D hepatosphere growth that was positively promoted by coculturing CAFs and mesenchymal HCCs.

Discussion
Based on an unbiased screen for chemical compounds aimed to block TGFβ-induced EMT, we previously explored mechanisms by which oxysterols, the physiological ligands of the LXR transcription factors, antagonize a subset of responses to TGFβ signaling 19 . LXR agonists also block Snail-dependent mesenchymal differentiation, survival mechanisms and Snail-independent pro-apoptotic responses to TGF-β 20 . By focusing on HCC models, we now established that such antagonism regulates CAF responses to TGFβ signaling, at least in part via a gene regulatory mechanism whereby LXRα limits the TGFβ inducibility of target genes, such as αSMA/ACTA2 (Figs. 1-5). A correlate of this antagonism in the context of liver cancer is the ability of LXRα agonists to limit the growth and expansion of HCC 3D spheroids, a phenotype positively regulated by CAFs when co-cultivated with HCC cells (Fig. 6). We propose that oxysterols acting on HCC CAFs have the ability to down-modulate responses to TGFβ, which impacts on HCC survival and proliferation (Fig. 7).
A majority of epithelial HCC cells expressed detectable LXRα levels and responded to the LXR agonist T0901317 (Fig. 1d, Suppl. Fig. S1), whereas mesenchymal-like HCC cells and HCC CAFs, expressed weakly detectable LXRα levels, and responded only weakly to T0901317 stimulation (Figs. 1e, 2, 3a, Suppl. Fig. S1). Our recent study suggests that LXRα signaling counteracts TGFβ-mediated Snail expression 20 ; independent in vivo studies corroborate the above, since in the LXRα −/− mouse model, high Snail expression levels were measured in the prostate epithelium 25 . Furthermore, studies in A549 lung adenocarcinoma cells revealed that oxysterol signaling via LXRα/β promoted Snail expression 26 . Thus, in the context of cancer, the impact of LXRα signaling needs to be evaluated with a specific focus on the diverse cellular components of the tumor microenvironment. In the present study, we propose that LXRα signaling limits the TGFβ-mediated mesenchymal and fibroblast differentiation.
Smad3 and LXRα recognize distinct CAGA and LXRE DNA motifs, respectively 5,17 , which are encompassed in the cloned αSMA/ACTA2 promoter (Fig. 4a). By analyzing the regulation of the αSMA/ACTA2 gene (Figs. 4 and 5) we provided evidence that LXRα can associate with the αSMA/ACTA2 promoter; possibly, such binding limits the accessibility of the promoter to TGFβ/Smad3-mediated co-activators (Fig. 7). Previous work showing a synergy between Smad2, Smad3 and LXR signaling via the co-(see figure on previous page) Fig. 4 Smads and LXRα bind to the α-SMA gene promoter. a Diagram of the hACTA2 gene promoter indicating the transcriptional start site (arrow, +1), putative Smad-(CAGA motif, red boxes) and LXR-binding sites (blue boxes) on the promoter, with numbering relative to the start site corresponding to the first base pair of each CAGA motif and the ends of the promoter fragment. Promoter fragments detected by ChIP (green) and used for DNAp (grey) are shown. b ChIP-qPCR analysis of endogenous Smad2/3 binding to three different fragments of the hACTA2 gene promoter as indicated in (a), or to the human PAI-1 gene, used as positive control. Mean ± SD values are plotted. Experiments were performed in biological duplicates (n b = 2), each of them in technical triplicate (n t = 3). Statistical comparison (two-sided t-test) indicates significant differences, ***p < 0.001. c DNA-precipitation (DNAp) assays using an hACTA2 gene promoter-specific oligonucleotide and human 293-T cell lysates after transient transfection with pCDNA3-Flag-Smad3 and pCMX-LXRα in the absence (−) or presence (+) of stimulation with 5 ng/ml TGFβ1 for 2 h followed by sequential immunoblotting for transfected Smad3 and LXRα. d DNAp using lysate from human 293-T cells transfected only with pCMX-LXRα and stimulated with 5 ng/ml TGFβ1 for 2 h and/or 5 µM T0901317 (T09) for 16 h followed by sequential immunoblotting for endogenous Smad3 and transfected LXRα. Densitometric quantification of the Smad3 and LXRα protein bands is shown. e DNAp using lysate from untransfected HepG2 cells stimulated with 5 ng/ml TGFβ1 for 2 h and/or 5 µM T0901317 (T09) for 18 h followed by sequential immunoblotting for endogenous Smad3 and endogenous LXRα. f DNAp using a multimerized Smad-binding element (CAGA) oligonucleotide as positive control followed by sequential immunoblotting for endogenous Smad3 and LXRα. Total cell lysates (TCL) indicate expression of each protein, relevant proteins are marked with an arrow, molecular size markers are indicated in kDa, and grey boxes with a plus sign mark control DNAp with streptavidin beads in the absence of specific oligonucleotide to indicate non-specific, background precipitation. Experiments were performed in biological duplicates (n b = 2), each of them in technical triplicate (n t = 3) activator protein RAP250 27 does not explain the observations obtained in HCCs. In several HCC cells, regulation of additional myofibroblastic genes, including calponin, SM22α and fibronectin by LXRα was evident (Figs. 1, 2, 3a). It is therefore possible that LXRα may exert a more general mechanism counteracting TGFβ-mediated differentiation responses.
HCC and CAF 3D co-culture assays were employed in order to simulate an in vivo representative biological context (Fig. 6). Epithelial Hep3b cells with an early TGFβ gene signature 21,23 , generated spheroids independent of the presence of CAFs. Mesenchymal SNU449 HCCs with a late TGFβ gene signature, proposed to express stem cell genes 21,23 , formed "disorganized hepatospheroids" (Fig. 6). Unexpectedly, SNU449-CAF co-culture positively influenced the growth and size of the hepatospheroids (Fig. 6), suggesting that CAFs promote the acquisition of a more regular organoid architecture. The positive trophic and architectural effect of AKH12 CAFs over the SNU449 HCC cells in forming hepatospheres, was reduced by administration of T0901317 only to the CAF population (Fig. 6). This suggested that possible mediators produced by CAFs, which promote hepatosphere architecture and growth, are repressed by LXRα signaling. This mediator  Fig. 4a, including the luciferase reporter. b-f HepG2 cells were simultaneously transfected for 48 h with hACTA2-luc reporter and β-gal plasmid alone (b, d, f) or together with empty vector pCDNA3 or pCDNA3-Flag-Smad3 and pCDNA3-Flag-Smad4 (c) or pCMX-LXRα full length (e). Cells were treated in starvation medium with 5 ng/ml TGFβ1 for 16 h or with T0901317 (5 μM), GW3965 (5 µM) or combination of TGFβ and T0901317 for 16 h, prior to luciferase and β-galactosidase activity detection. g Primary human foreskin AG1523 fibroblasts were transfected as above with the CAGA-luc reporter and β-gal plasmid followed by treatment in starvation medium with vehicle (control), 5 ng/ml TGFβ1, T0901317 (5 μM) or combination of the last two for 16 h, prior to luciferase and β-galactosidase activity detection. Mean ± SD values are plotted. Experiments were performed in biological triplicates (n b = 3), each of them in technical quadruplicate (n t = 4). Statistical comparison (two-sided t-test) indicates significant differences, *p < 0.05; **p < 0.01; ***p < 0.001; ns p > 0.05. Immunoblotting for detection of co-transfected LXRα is provided as control (e) generated by CAFs may not be TGFβ itself (Fig. 6c), although TGFβ is important in colon cancer CAFs 11 and in HCC stellate cells 15 . Further analysis about the signaling pathways, CAF-secreted cytokines and/or cell-cell contact proteins are required to address this important aspect. It should also be noted that the primary patientderived CAFs exhibit to some extent patient-specific expression profiles and responses to signaling molecules. Thus, we propose that oxysterols provide anti-TGFβ inputs via LXRα, which limit myofibroblastic differentiation and communication between CAFs and carcinoma cells in liver tumors.
Cells were serum-deprived for 16 h; 5 ng/ml TGFβ1 (PeproTech EC Ltd, London, UK) in the absence or presence of 5 μM T0901317 or GW3965 (provided by Timothy C. Gahman and Andrew K. Shiau, Ludwig Cancer Research, La Jolla, CA, USA) were added at 80% confluency for 48 or 72 h, as specified. CAFs were stimulated with ligands without prior serum deprivation. Dimethylsulfoxide (DMSO), vehicle to T0901317 and GW3965, was added as control.

Three-dimensional hepatosphere culture
Hep3b and SNU449 cells (1.5 × 10 3 per well) were cultured alone or co-cultured with AKH12 CAFs (ratio 3:1 or 6:1 parenchymal cells:CAFs; both conditions resulted in similar hepatospheres), for 96 h in 10% FBS/MEM (Hep3b) or RPMI-1400 (SNU449), respectively. For hepatosphere formation, the 96-well plate Gravity-PLUS TM Hanging Drop System (InSphero AG, Schlieren, Switzerland) was used according to the manufacturer´s instructions. Where indicated, 5 μM T0901317 was added separately on SNU449 and AKH12 cells for 24 h prior to sphere formation, in order to assess compound effects on a single population. LY2157299 (2 μM, Sigma-Aldrich, Stockholm, Sweden) was administered during sphere formation. Hepatospheres were photographed in the InSPHERO plates using a Zeiss Axioplan-2 microscope with 10 × objective lens, at ambient temperature and without immersion oil, via a Hamamatsu C4742-95 CCD digital camera and acquisition software Volocity® (Perki-nElmer Waltham, MA, USA). Hepatospheres were assessed as the ratio of wells containing uniform spheres larger than 50 µm in diameter (cell aggregations of diverse and non-uniform shape were not analyzed), relative to the total number of wells. Sphere size was determined via Image J analysis (NIH, USA). Triplicate (n b = 3) biological experiments were performed in 10 technical replicates (n t = 10) per condition.

Statistical analysis
Experiments were performed in biological replicates (n b ), most of which included at minimum technical triplicates (n t = 3), as indicated in the figure legends. Sample size depended on the assay type: for sphere quantification assays a minimum of 25 spheres was measured per experiment to reach statistical power for discrimination between conditions. For all groups that are statistically compared, the variance within each group was very similar, close to identical. Representative results are reported as averages minus-plus standard deviation, s.d. Statistical analysis was performed with GraphPadPrism (La Jolla, CA, USA) using two-sided t-test, and statistical significance was assigned with *p-value < 0.05; **p < 0.01; ***p < 0.001.