Targeting stromal remodeling and cancer stem cell plasticity to overcome chemoresistance in triple negative breast cancer

The cellular and molecular basis of stromal cell recruitment, activation and crosstalk in carcinomas is poorly understood, limiting the development of targeted anti-stromal therapies. In mouse models of triple negative breast cancer (TNBC), Hh ligand produced by neoplastic cells reprogrammed cancer-associated fibroblast (CAF) gene expression, driving tumor growth and metastasis. Hh-activated CAFs upregulated expression of FGF5 and production of fibrillar collagen, leading to FGFR and FAK activation in adjacent neoplastic cells, which then acquired a stem-like, drug-resistant phenotype. Treatment with smoothened inhibitors (SMOi) reversed these phenotypes. Stromal treatment of TNBC patient-derived xenograft (PDX) models with SMOi downregulated the expression of cancer stem cell markers and sensitized tumors to docetaxel, leading to markedly improved survival and reduced metastatic burden. In the phase I clinical trial EDALINE, 3 of 12 patients with metastatic TNBC derived clinical benefit from combination therapy with the SMOi Sonidegib and docetaxel chemotherapy, with one patient experiencing a complete response. Markers of pathway activity correlated with response. These studies identify Hh signaling to CAFs as a novel mediator of cancer stem cell plasticity and an exciting new therapeutic target in TNBC. SIGNIFICANCE Compared to other breast cancer subtypes, TNBCs are associated with significantly worse patient outcomes. Standard of care systemic treatment for patients with non-BRCA1/2 positive TNBC is cytotoxic chemotherapy. However, the failure of 70% of treated TNBCs to attain complete pathological response reflects the relative chemoresistance of these tumors. New therapeutic strategies are needed to improve patient survival and quality of life. Here, we provide new insights into the dynamic interactions between heterotypic cells within a tumor. Specifically, we establish the mechanisms by which CAFs define cancer cell phenotype and demonstrate that the bidirectional CAF-cancer cell crosstalk can be successfully targeted in mice and humans using anti-stromal therapy.

patient survival and quality of life. Here, we provide new insights into the dynamic interactions between heterotypic cells within a tumor. Specifically, we establish the mechanisms by which CAFs define cancer cell phenotype and demonstrate that the bidirectional CAF-cancer cell crosstalk can be successfully targeted in mice and humans using anti-stromal therapy.

INTRODUCTION
Carcinogenesis draws many parallels with developmental biology. During development, dynamic interaction between stromal and epithelial cells drives patterning and function. Cell fate specification occurs through activation of transcriptional cascades in response to extracellular signals from developmental signaling pathways such as Hedgehog (Hh), Wnt, Notch, BMP (bone morphogenetic proteins) and FGF (fibroblast growth factor) (1,2). These pathways direct developmental processes either by direct cell-to-cell contact or through secreted diffusible factors (paracrine signaling). They can act individually or in concert with each other. For example, the interaction between Hh and FGF signaling pathways has been shown to mediate tracheal and lung branching morphogenesis (3,4). In mature, differentiated tissues, these pathways are quiescent but may be reactivated to drive repair and regeneration to maintain tissue homeostasis.
More specifically, the Hh developmental pathway is reactivated in a subset of cancers. Binding of Hh ligand to its receptor Patched (PTCH) enables Smoothened (SMO)-mediated translocation of Gli1 into the cell nucleus to drive the transcription of Hh target genes (5). Mutations in Hh pathway components are oncogenic drivers in "Gorlin's-like" cancers such as medulloblastoma and basal cell carcinoma (BCC), where tumors rely on cell-autonomous Hh signaling (6). Small molecule inhibitors of SMO, Vismodegib and Sonidegib, are well tolerated and clinically approved for the treatment of these lesions (6,7). In contrast, many other solid tumors, including breast cancer, predominantly exhibit ligand-dependent pathway activation (5,6,8). While Hh signaling is quiescent in the adult mammary gland, Hh ligand expression is reactivated in a subset of breast cancers, particularly the poor-prognosis TNBC subtype (8). Breast cancer patients with a paracrine Hh pathway signature, defined by high epithelial SHH ligand expression in combination with high stromal GLI1 expression, have a greater risk of metastasis and breast cancer specific death (8).
Neoplastic cells co-opt components of the tumor microenvironment (TME) to further their progression. The TME is a complex ecosystem comprising a myriad of neoplastic and non-malignant cells embedded in a glycoprotein-rich extracellular matrix (ECM). Prominent cell types include the endothelium, cells of the immune system and cancer-associated fibroblast (CAFs). In addition to its role as a physical scaffold to support tissue architecture, the ECM also functions as a signal transducer between the different TME cell types (9). The stiffness of the ECM and the abundance of fibrillar collagen immediately adjacent to epithelial lesions provide mechanical signals that facilitate tumor development and progression (10)(11)(12). Not surprisingly, the TME has emerged as a major determinant of cancer phenotype. In breast cancer, stromal metagenes, in particular those associated with ECM remodeling, strongly predict prognosis and response to chemotherapy (13,14).
Whilst it is now apparent that Hh signals in a paracrine manner in animal models of TNBC (8) and in isolated cancer stem cells (CSCs) (15), a detailed study of the dynamic crosstalk within the TME is required to make clinical progress in integrating anti-stromal therapies into breast cancer treatment. Progress has been impeded by the field's limited understanding of the mechanisms underlying tumorstromal interactions, a limited repertoire of well-tolerated agents to target the TME, and an absence of predictive biomarkers for response to TME-directed therapies (16).
In this study, we investigated whether, and how, SMO inhibitors (SMOi) could be used for therapeutic reprograming of the TME in human TNBC.

Hh-regulated epithelial-stromal crosstalk mediates a reversible breast cancer stem-like phenotype
To investigate the mechanistic basis for Hh-dependent tumor growth and metastasis in TNBC, we used the murine M6 allograft model of low grade TNBC, in which transgenic Hh expression drives invasion, metastasis and high-grade  Table S1).
Gene Set Enrichment Analysis (GSEA) and Gene Ontology (GO) analysis of the purified epithelial fraction highlighted enrichment for genes specifically and almost exclusively associated with mammary stemness and invasion, consistent with the morphologically undifferentiated phenotype previously observed in Hhoverexpressing tumors (8) (Fig. 1C Fig. 1E). M6-Hh tumors also had elevated expression of the stemness markers Id3, Gpc3, Thy1, Sox10 and Krt6 (19)(20)(21)(22), validating the RNA-Seq data (Fig. 1B,F). Following transplantation of low numbers of sorted primary M6-Hh and M6-Ctrl cells into naive recipients, tumor latency was shorter and penetrance markedly higher in the M6-Hh group (Fig. 1G) 1088; Fig. 1H). Importantly, the proliferation and expression of CSC markers were indistinguishable between M6-Ctrl and M6-Hh cells in monoculture, indicating that Hh expression in M6 cells does not regulate CSC properties in a cell autonomous manner (Supplementary Fig. S1E).
Immunohistochemical detection of the mammary progenitor marker cytokeratin 6 (CK6; product of the Krt6 gene) (24) localized cells with a stem/progenitor signature specifically to the tumor-stromal interface (Fig. 1I). SMO inhibition reduced the expression of Id3, Gpc3, Thy1, Sox10 and Krt6 and significantly reduced the number of cells positive for CK6 and the mitotic marker phospho-Histone H3 at the tumor-stromal interface (Fig. 1F,I). These data demonstrate that paracrine Hh signaling results in the induction of a reversible stemlike phenotype preferentially at the tumor-stromal interface.

Stromal Hh signaling leads to marked ECM-related gene expression changes and associates with poor prognosis in patients with TNBC
RNA-Seq analysis of the stromal fraction revealed 185 genes that were differentially expressed (> 2-fold, P < 0.001), with 146 upregulated and 39 downregulated genes in the stroma of M6-Hh tumors compared to M6-Ctrl and M6-Hh tumor + SMOi (Supplementary Table S1). A number of genes were markedly upregulated by Hh signaling, in particular the growth factor gene Fgf5 at more than 290-fold, St8Sia2 (> 40-fold) and Tspan11 (> 4-fold; Fig. 2A). The large majority of gene expression changes in Hh-activated stroma returned to baseline following treatment with SMOi ( Fig. 2A and Supplementary Table S1), suggesting that the stromal transcriptional changes are SMO-dependent and reversible.
To determine the prognostic value of the Hh-activated stromal gene signature (HSGS), we examined its impact on overall survival using The Cancer Genome Atlas (TCGA) breast invasive carcinoma cohort. The HSGS was not predictive of patient outcome in the unstratified patient cohort (Fig. 2D) but was associated with significantly lower patient overall survival uniquely in the basal breast cancer subtype, where Hh ligand is most frequently overexpressed (8) (Hazard ratio = 9.7 (1.9 -48.2); P < 0.001; Fig. 2D).
Accumulating evidence suggests that CAFs contribute to tumor growth upon Hh ligand activation (15,26). However, the stroma of M6 tumors is composed of multiple cell types, any of which may be responsible for Hh-dependent gene expression changes. We used a single-cell RNA sequencing (scRNA-Seq) approach to determine the cell population/s responding to paracrine Hh signaling. The microfluidic 10X Chromium system was used to comprehensively profile gene expression at cellular resolution in thousands of cells isolated from freshly dissociated M6-Ctrl and M6-Hh tumors ± SMOi. In total, we compared 6,064 FACS-isolated cells from M6-Ctrl tumors, 6,200 cells from M6-Hh tumors, and 2,686 single cells from M6-Hh tumor treated with SMOi.
As shown in Figure 3A, unsupervised clustering analysis of 14,950 cells revealed populations of myeloid, neoplastic, endothelial, CAF and natural killer cells within the breast TME (Fig. 3A). Importantly, the upregulation of canonical Hh target genes Gli1, Ptch1, Ptch2 and Hhip and ECM genes such as Col4a1, Tspan11, St8sia2 and Fgf5 was observed exclusively in the CAF population of M6-Hh compared to M6-Ctrl tumors (Fig. 3A), and not in other stromal cell types. More specifically, the ECM signature detected in the stroma of Hh-expressing tumors via 'bulk' RNA-Seq was driven by CAF gene expression (Fig. 2B,C; Fig. 3B). This scRNA-Seq approach also confirmed the lack of autocrine Hh pathway activation within the neoplastic cells ( Fig. 3A). Treatment with SMOi almost completely reversed the Hh-dependent gene expression changes observed in CAFs without affecting gene expression in other stromal cell types ( Fig. 3A; Supplementary Fig. 2D and Supplementary Table S2), highlighting the on-target activity of SMOi at the single cell level.
Co-culture of primary CAFs with M6-Hh cells was sufficient to recapitulate this dynamic epithelial-stromal crosstalk resulting in the induction of Hh target gene expression in the CAFs (Supplementary Fig. S3A,B) and concomitant upregulation of CSC markers Id3, Gpc3, Itgb3 and Krt6b in M6-Hh cells compared to M6-Ctrl + CAF co-culture systems (Supplementary Fig. S3A,C). Importantly, this stromalepithelial malignant crosstalk was blocked by SMOi (Fig. 3A, Supplementary Fig.   S3B,C). These data allow us to conclude that Hh signaling occurs solely in a paracrine manner in this murine model of TNBC and CAFs are the therapeutic target of SMOi in TNBC.
Hh activated CAFs drive fibrillar collagen deposition and remodeling, resulting in mechanosignaling and a stem-like phenotype in adjacent neoplastic cells Bulk and scRNA-Seq data suggested that stromal Hh-signaling drives collagen remodeling in the local ECM (Fig. 2B,C and Fig. 3), which is known to associate with breast cancer progression (27,28). We employed second harmonic generation (SHG) microscopy, a sensitive label-free method for quantifying fibrillar collagen density and orientation in tissues (29). SHG analysis revealed a ~3-fold increase in fibrillar collagen density at the tumor stromal-interface of Hh-expressing tumors (Fig.   4A), but not in the tumor center (data not shown). The increase in collagen abundance was confirmed by chromogenic staining using Picrosirius red (Fig. 4B). Further detailed analyses of the distribution and orientation of collagen fibers as described by Mayorca-Guiliani et al. (30) and by Gray level co-occurrence matrix (GLCM) analysis (31,32) revealed changes in texture and cross-linking of the ECM with linearization of collagen fibers adjacent to epithelial lesions in M6-Hh tumors, a hallmark of breast tumor growth and invasiveness (10,30) (Fig. 4A). These features of the collagen ECM in Hh-expressing tumors were ameliorated in mice treated with SMOi ( Fig. 4A), demonstrating an ongoing dependency on SMO activation.
Sites of collagen deposition and cross-linking at the stromal-epithelial interface were also marked by increased phosphorylation and activation of focal adhesion kinase (FAK) on cancer cells, a key signaling intermediate downstream of integrin receptors (Fig. 4C). These cells also expressed elevated cytokeratin 6 ( Fig.   4D), correlating fibrillar collagen content to FAK signaling and the acquisition of a stem-like phenotype in the neoplastic cells. Importantly, mechanosignaling and cancer stemness exclusively occurred in close proximity to rich dense collagen regions and were not observed in the core of the M6-Hh tumors (Supplementary Fig. S4A,B).
To directly assess the sufficiency of collagen mechanosignaling to promote stemness, the clonogenic potential of M6-Ctrl and M6-Hh cells was assessed using 3dimensional cultures encapsulated within Alginate-Collagen I Inter-Penetrating Network (IPN) hydrogels. The enrichment for fibrillar collagen in this in vitro model recapitulates the features of stromal collagen matrix deposition observed in Hhexpressing models. Increased content and presence of highly bundled fibrillar collagen significantly increased the clonogenic capacity of M6 cells, a functional surrogate for CSC activity (33), independently of Hh ligand expression (Fig. 4E).
Furthermore, increasing collagen I abundance also increased expression of the stem cell markers Id3, Itgb3 (CD61) and Krt6 (CK6) (Fig. 4F). These data demonstrate that Hh-dependent stromal ECM remodeling is sufficient to foster a CSC phenotype.

Paracrine Hh-FGF5 signaling also contributes to CSC plasticity
To identify additional mechanisms by which stromal signaling promotes the acquisition of a CSC phenotype, we turned our attention to Fgf5, which was strongly upregulated in Hh-activated stroma compared to controls ( Fig. 2A,C and Supplementary Table S1 and S2). qRT-PCR analysis of whole tumors confirmed ~60-fold upregulation of Fgf5 mRNA in M6-Hh tumors, which was reversed upon SMOi treatment (Fig. 5A). Notably, a subset of Hh-activated CAFs exhibited robust expression of Fgf5 at the single-cell resolution, reflecting the spatial localization of these CAFs, in close proximity with M6-Hh cells (Fig. 5B). Immunohistochemical analysis of phospho-FGFR revealed potent receptor activation in M6-Hh tumors, primarily in epithelial cells adjacent to stroma, which was reversed upon SMOi treatment (Fig. 5C). To explore the role of FGF5 in the acquisition of the CSC phenotype, M6-Ctrl cells were treated in vitro with recombinant FGF5 protein and proliferation, stem cell marker expression and sphere forming capacity were evaluated. FGF5 treatment led to a modest increase in proliferation under serum and growth factor deprivation (Supplementary Fig. S5A). In contrast, the stemness markers Id3 and Sox10 were robustly upregulated (Fig. 5D) and primary and secondary sphere forming capacity increased by ~3-fold ( Fig. 5E and To test whether the induction of sphere forming capacity by FGF5 treatment was epigenetically stable or plastic, we tested the impact of addition or removal of FGF5 to primary and secondary sphere cultures. Increased sphere formation in response to FGF5 was observed in secondary cultures regardless of whether those cells were pre-treated with FGF5 during primary cultures ( Fig. 5E and Supplementary Fig. S5B). Furthermore, removal of recombinant FGF5 decreased secondary sphere formation to levels comparable to those of cells never treated with FGF5 ligand (Fig. 5E and Supplementary Fig. S5B). Consistent with this result, treatment of spheres with a small molecule inhibitor of FGFR signaling (NVP-BGJ398) prevented FGF5 mediated tumorsphere formation in primary or secondary cultures (Supplementary Fig. S5C). These results demonstrate that stromal-derived FGF5 is sufficient to promote reversible transition to a CSC phenotype, rather than through the expansion of a sub-population of CSC.
The CSC phenotype is associated with resistance to cytotoxic chemotherapy in TNBC (34). To test whether the FGF-dependent increase in CSC alters the sensitivity of TNBC cells to chemotherapy, the efficacy of docetaxel was evaluated in M6 cell lines in vitro. Monocultures of M6-Hh cells did not display differential sensitivity to docetaxel when compared to M6-Ctrl cells as expected (Supplementary Fig. S5D).
However, stimulation of M6-Ctrl and M6-Hh models with recombinant FGF5 ligand rescued these cells from docetaxel cytotoxicity (Fig. 5F). The FGFR inhibitor NVP-BGJ 398 abrogated drug resistance conferred by FGF5 (Fig. 5F). Similar results were observed in the human MDA-MB-231 cell line model of TNBC ( Supplementary   Fig. S5E). This result suggests that FGF5, released by Hh-activated CAFs, creates a "chemo-resistant niche" at the tumor-stromal interface that can serve as a reservoir for eventual tumor relapse in TNBC. It also suggests that targeting both tumor and stromal compartments with chemotherapeutic regimen and SMOi, respectively, may be an effectively therapeutic strategy.

Stromal SMO inhibition combined with chemotherapy markedly improves preclinical TNBC outcomes
To directly test these findings in more clinically relevant models, we turned our analysis to xenograft models of human TNBC. All three patient-derived xenograft (PDX) models tested were Hh ligand-positive as was the MDA-MB-231 cell line model (Supplementary Fig. S6A). We found convincing evidence of exclusively stromal-restricted Hh signaling, using species-specific RT-PCR and sensitivity to SMOi (Supplementary Fig. S6B,C). In addition, in vitro treatment of MDA-MB-231 cells with SMOi did not alter Hh target gene expression or proliferation ( Supplementary Fig. S6D,E), consistent with a paracrine requirement for Hh signaling in these models. The HCI-002 and MDA-MB-231 models were chosen for further investigation, as they are well-accepted models for TNBC (35,36).
Analysis of the ECM in HCI-002 xenografts by SHG and picrosirius red staining revealed abundant fibrillar collagen exclusively at the tumor-stromal interface (Fig. 6A,B) that was highly linearized and densely packed as depicted by orientation and GLCM analysis (Fig. 6A). Areas of high fibrillar collagen density were also associated with concomitant FAK phosphorylation and expression of the human CSC marker ALDH1 (Fig. 6C). As observed in the transgenic model, collagen density and orientation, FAK phosphorylation and CSC marker expression were reduced following SMO inhibition (Fig 6A-C).
Based on these observations we predicted that SMOi would sensitize tumors to cytotoxic chemotherapy. HCI-002 PDX and MDA-MB-231 xenografts were then treated with SMOi +/-docetaxel ( Fig. 6D-F). Compared with vehicle treatment, either SMOi or docetaxel monotherapy slowed tumor growth. However, the most robust and durable therapeutic effect occurred with combined therapy (Fig. 6D,F). Interestingly, the proportion of mice with metastatic disease at ethical endpoint (based on primary tumor size) was doubled in the docetaxel treated group, an observation previously made with paclitaxel in TNBC mouse models (37) (Fig. 6E). Combination therapy reduced the frequency of mice with metastatic disease to below that seen in the vehicle control group, despite these mice being alive much longer compared to those in the other treatment groups. Similar therapeutic benefit was observed in MDA-MB-  Table S3).
To assess if stromal Hh pathway activation determines clinical response, we evaluated epithelial SHH ligand and stromal GLI1 expression in treatment naïve surgical tissue by immunochemistry. Only 10 patient tumor samples were evaluable.
Three out of 10 tumors had high paracrine Hh Pathway Activation Signature (HPAS), characterized by high epithelial SHH in combination with high stromal GLI1 expression ( Fig. 7B and Supplementary Table S3). Of these, two patients experienced a clinical response whereas all patients with low HPAS expression had progressive metastatic disease on the prescribed treatment regimen (Supplementary Table S3). An additional patient experienced clinical benefit, but the status of Hh pathway activation was unknown as her tumor sample was not available for analysis (Supplementary Table S3).
Downstream analysis of the effect of paracrine Hh signaling revealed moderate to high phospho-FGFR expression, high collagen deposition and fiber linearization in treatment naive tumor specimens of the two responders with biopsy material available (Fig. 7C). This correlated with elevated phospho-FAK mechanosignaling and ALDH1 positive cells at the tumor-stromal interface (Fig. 7C). In contrast, the non-responder with high paracrine HPAS exhibited weak phospho-FGFR expression, low collagen content and minimal/no evidence of mechano-signaling and breast cancer stem cells (Fig. 7C). We therefore conclude that these additional tumor factors may represent adjunct biomarkers of therapeutic response for patient selection for anti-SMOi based combination therapies in Hh-expressing TNBC.

DISCUSSION
In certain settings, CSCs are responsible for metastasis to distant organs (39)(40)(41)(42) and are frequently enriched in residual tumors following chemotherapy (43)(44)(45), reflecting a role in therapeutic resistance. The hierarchical model for CSC maintenance proposes that CSC behave like tissue-resident physiological stem cells, self-renewing and undergoing asymmetric divisions to generate differentiated progeny (46). However, evidence from cell culture models has challenged the hierarchical CSC model, by suggesting that cancer cells can transition into a CSC state under specific culture conditions (46)(47)(48)(49). In support of this notion, we now demonstrate that stromal cues from Hh-activated CAFs, forming a supportive niche enriched for FGF and fibrillar collagen-rich ECM, are capable of inducing and maintain a stem-like phenotype in TNBC cells in vivo. By combining a murine gainof-function model, small molecule inhibitor studies in human xenografts with powerful in vitro systems, we have demonstrated the plastic characteristics of breast CSCs that can be successfully targeted using anti-stromal therapies, reducing metastatic growth and sensitizing to taxane chemotherapy.
Increased stromal collagen content correlates with stemness in the epidermis, both in the cancer and homeostatic contexts (50,51). It also enhances CSC properties of breast cancer cells in vitro (52,53). However, the impact of ECM collagen content and matrix mechanical properties on the biology of CSCs is not well defined. Our work provides new mechanistic insights, demonstrating that increased collagen density and fiber linearity at the tumor-stroma interface are associated with FAK activation and increased CSC number, dependent upon Hh paracrine signaling.
Notably, we report a relationship between collagen abundance and clonogenicity in vitro and in vivo. Suppressing collagen production using SMO inhibitors was associated with decreased Krt6 + and ALDH1 + CSCs, respectively, in both murine and human models of TNBC. Interestingly, recent data links mammographic fibrillar collagen density to breast cancer risk, raising the possibility that breast cancer progenitors in these patients may have expanded in response to a dense collagen matrix (28,54,55).
FGF signaling has been shown to drive malignant processes including stem cell self-renewal, multipotency and therapeutic resistance (22,56,57). In metastatic breast cancer, resistance to anti-cancer treatment is primarily due to FGFR gene amplification (58). Here, we demonstrate a novel ligand-driven mechanism by which FGFR activation mediates both breast cancer stemness and chemoresistance, downstream of activation of the Hh signaling pathway. Importantly, our findings strongly suggest that CAF targeting using small molecule inhibitors of SMO is sufficient to prevent FGF ligand signaling and may overcome resistance to chemotherapies. Interestingly, FGF5 has been reported to be upregulated in prostate CAFs relative to normal fibroblasts (59), where it is also a target of Hh-Gli signaling (60). Thus this axis may be operational, and of therapeutic value, in tumor types beyond TNBC.
How FGFR activation and high FAK mechanosignaling lead to the establishment of a stem-like phenotype remains to be determined, but they are associated in vitro and in vivo with upregulation of transcription factors previously implicated in mammary physiological and cancer stem cells, including ID3 and SOX10 (21,22). The mechanisms underlying Id3 and Sox10 transcription are unknown. However Id3 may be induced through Erk-EGR1 signaling, as observed in activated T cells, downstream of both FAK and FGFR (61). Our data also reveals the cooperative activity of ECM remodeling and FGF signaling in driving malignancy and drug resistance, recapitulating the interaction seen between these pathways during development and wound healing (62).
Importantly, many elements of Hh paracrine signaling to CAFs are active during embryonic development in mammals, though have not previously been linked.
Dhh is highly expressed in a subset of epithelial cells of the mammary end bud, an invasive and proliferative structure responsible for ductal elongation in the developing mouse mammary gland (63). Consistent with our observation in TNBC, stromal but not epithelial Hh signaling is required for appropriate ductal morphogenesis (64). A number of FGF ligands are secreted by mammary stromal cells, and activation of epithelial FGFR1/2 is required for mammary ductal elongation and stem cell activity (65)(66)(67). In addition, mammary stromal fibroblasts secrete and remodel ECM components including collagens (68). Similar to our results in neoplastic cells, increased collagen density and mechanosignaling via FAK is sufficient to inhibit mammary epithelial cell differentiation and increased clonogenic potential (68). Thus the paracrine Hh signaling we observe in TNBC most likely represents the dysregulation and chronic activation of a process that is important for normal mammary ductal morphogenesis.
Using high-throughput single cell RNA-sequencing, we demonstrate that CAFs are the only stromal cell type responding to Hh ligand, and that SMO inhibitors act 'on-target' to reverse CAF gene expression changes induced by Hh signaling.
Surprisingly, long-term (up to 3 months) daily treatment with SMOi did not alter the stromal cell composition of mammary tumors. This result contrasts markedly to that recently observed in pancreatic, colon and bladder cancer models, where chronic SMO inhibition was associated with marked changes in stromal cellular composition and shorter survival for mice receiving long-term SMOi treatment (69)(70)(71)(72). The basis for this difference is not known, but may be explained by the divergent epi/genomics contexts of these cancer types, resulting in the evolution of distinct tumor microenvironments (11). Alternately, differences in the origin or phenotype of CAFs (73) in these endodermally-derived tumors versus ectodermally-derived mammary carcinomas may be relevant.
The benefit from therapeutic targeting of CAFs is two fold. Firstly, others and we have provided evidence for the crucial role of CAFs in supporting CSC selfrenewal and resistance to chemotherapy (74)(75)(76)(77)(78)(79). Therefore, targeting the CAF population and the subsequent abolition of the CAF-neoplastic cell interaction represent a practical strategy to improve cancer outcomes. Secondly, unlike neoplastic cells, CAFs have not been reported to exhibit genomic instability and are therefore less likely to acquire resistance to therapy over time, making them good targets for combination cancer therapies. Combined therapy with SMOi + docetaxel was well tolerated by mice and humans, and effective in treating a proportion of women with metastatic disease who had previously failed on taxane chemotherapy, including one patient who experienced a complete response. These remarkable results provide the first evidence to our knowledge for clinical benefit from a CAF-directed therapy.
Treatment response in patients correlated with high levels of paracrine Hh signaling, FGFR activation and fibrillar collagen deposition, suggesting that the mechanism of action in patients may be consistent with that in mouse models. Hh, FGFR or collagen pathway activation may have value as predictive biomarkers of response to SMOi.
Whilst phase I clinical trials are not designed nor powered to assess therapeutic efficacy, these data suggest an exciting new therapeutic strategy for drug-resistant or metastatic TNBC which should proceed to prospective assessment through Phase II clinical trials.

Cell culture
M6 murine mammary carcinoma cells derived from the C3(1)/SV40 Tag mouse model (gift from J. Green, NIH (80)) were cultured as previously described (8) PDX tumor tissues, acquired from the laboratory of A. Welm (35) were serially passaged as 2mm 3 fragments in the cleared fourth mammary fat pads of pre-pubescent NSG mice according to established protocols (35). When tumors became palpable, they were measured three times weekly in a blinded manner using electronic calipers to monitor growth kinetics. Tumor volume was calculated using the formula (π/6) x length x width 2 . Upon reaching ethical or predefined experimental endpoints, mice were euthanized and primary tumor and any associated metastases were collected.

In vivo drug treatment experiments
SMOi GDC-0449 (S1082, Selleckchem) and NVP-LDE225 (Novartis, Australia) were dissolved in 0.5% methylcellulose, 0.2% Tween ® 80 (Sigma-Aldrich) and 0.5% methylcellulose, 0.5% Tween ® 80, respectively, and then delivered by oral gavage (100 mg/kg/bid, GDC-0449; 80 mg/kg/day, NVP-LDE225). Chemotherapy (Docetaxel, McBeath Australia) was diluted in 5% dextrose then delivered by intraperitoneal injection (15 mg/kg/week). Tumor bearing mice were randomly assigned into respective treatment groups once tumor volume reached 100 mm 3 (n = 7 -8 mice per group). Tumor growth was calculated for each individual tumor by normalizing to the tumor volume at day 0. In short-term studies examining the molecular and histological impact of Hh pathway activation and inhibition, mice were treated between 8 to 14 days then euthanized. At euthanasia, primary tumors were harvested and macroscopic metastatic lesions were scored. For the long-term therapeutic study, mice were treated to endpoint. Animals were excluded from overall survival analysis if they had to be sacrificed for poor body conditioning, unrelated to tumor size endpoint. Animal technicians, who were blinded to the experiment treatment groups, independently monitored the mice.

Next Generation Sequencing
We isolated by FACS the stromal DAPI -/GFP -/EPCAMand epithelial DAPI -/ GFP + /EPCAM + cell fractions from at least 5 M6-Ctrl and M6-Hh tumor models treated with vehicle (0.5% methylcellulose, 0.2% Tween ® 80) or with SMOi. RNA was isolated using the miRNeasy kit (Qiagen). For standard input samples, 1 µg of total RNA was used as input to the TruSeq RNA Sample Preparation Kit v2 (Illumina). The samples were prepared according to the manufacturer's instructions, starting with the poly-A pulldown. The number of PCR cycles was reduced from 15 to 13, to minimize duplications. The samples were sequenced on the HiSeq2000 using v3 SBS reagents (Ramaciotti Centre for Genomics, University Of New South Wales (UNSW)). Low input RNA stromal samples were firstly amplified using the Ovation ® RNA-Seq System V2 kit (4 ng of total RNA input; Nugen Integrated Sciences Pty.

Ltd.) according to the manufacturer instructions. 1 µg of the cDNA was sheared with
Covaris to fragment sizes of ~200bp. The material was used as input to the TruSeq RNA Sample Preparation v2 kit, starting at the end-repair step. The number of PCR cycles was reduced from 15 to 10. All the samples were sequenced on the HiSeq2000 using v3 SBS reagents (Ramaciotti Centre for Genomics, University Of New South Wales).

Bioinformatics and computational analysis of RNA-sequencing data sets
Analysis of the RNA-Sequencing data was conducted on the high-performance computing cluster at the Garvan Institute following a standard four step approach, cleaning, aligning, counting and differential expression with an additional normalization step. FASTQ files were quality checked using FastQC version 0. 11 (81). This aims to remove unwanted variation and produce more reliable pair-wise comparisons when calculating differential expression. In this instance, RUVr with a K of 3 was found to be the most effective method based on the suggested diagnostics, e.g. plots of P-value distributions and PCA. Differential expression analysis was performed within the RUV analysis using edgeR (82).
Single cell RNA-Sequencing using the Chromium Platform and bioinformatical analysis M6 tumors were processed into single cell suspensions as described previously. Differential gene expression analysis in Seurat was performed using the 'bimod' likelihood-ratio test.

Data and Code Availability
All RNA Sequencing files that support the findings of this study have been deposited in GEO with the accession code PRJNA369574. The RNA-Seq pipeline and the analysis scripts can be found on the respective websites: https://github.com/elswob/rna-seq-pipe and https://github.com/elswob/Hh.

Gene Set Enrichment Analysis (GSEA)
Gene-sets used in GSEA were extracted from version 3.  Table S1). The gene list was further assessed for survival analysis using the TCGA breast invasive carcinoma cohort. The processed TCGA data was downloaded from cBioPortal (86) based on the TCGA study (87).
The gene signature score was defined by a weighted average method for each sample in the TCGA cohort. Survival curves were estimated using the Kaplan-Meier method, with overall survival used as the outcome metric.

RNA isolation, reverse transcription, quantitative RT-PCR and Fluidigm array experiments and analysis
Individual stromal CAFs and epithelial malignant M6 cancer cells were FACS-  Supplementary Table   S5. Relative mRNA expression levels were normalized to β-actin, GAPDH or HPRT and quantification was performed using the comparative C T method described by Livak and Schmittgen (88).

Immunohistochemistry, immunofluorescence and histological analysis
Tissues were fixed in 10% neutral buffered formalin at 4°C overnight then processed for paraffin embedding. For histological analysis, 4 µm tissue sections were stained with haematoxylin and eosin using standard methods. Immunohistochemical, immunofluorescence and picrosirius red staining were performed on paraffinembedded tissue sections using standard protocol. Full details of each antibody used and their relative staining protocols for immunochemistry are described in

Supplementary table S6.
Histological analysis of the proliferative marker phospho-histone H3, ALDH-1 and the progenitor cell marker CK6 were carried out by digitizing entire images using the Aperio CS2 digital pathology slide scanner (Leica Biosystem) at 20x magnification.
Cells that stained positively for phospho-Histone H3, ALDH-1 or CK6 within a distance of 200 µm from the CAFs at the tumor-stromal interface were then counted and averaged over at least 5 fields using the Aperio Imagescope software (Leica Biosystem). The limit of 200 µm reflects the well-established diffusional distance for Hh ligand in mammalian models (89). Picrosirius red stain was analyzed as previously described (90). Two specialist breast pathologists, who were blinded to the experiment treatment groups, independently scored the remaining IHC stains. Areas software. Images were processed using ImageJ (National Institutes of Health) as previously described (51,91). Two cell biologists from various institutes, who were blinded to the treatment groups, independently scored the IF stains for CK6, ALDH1 and phospho-FAK.

Second-harmonic generation (SHG) microscopy, Gray-level co-occurrence matrix and orientation analysis
Formalin-fixed, paraffin embedded sections stained with hematoxylin and eosin and mounted in DPX (Sigma) were imaged using a 20x 1.0 NA objective on an upright fixed-stage two-photon laser scanning microscope system (Zeiss). The excitation source was a Ti:Sapphire femto-second laser cavity (Newport Mai Tai), coupled into a LSM 710 scan module. An excitation wavelength of 890 nm was used to collect SHG signal (435 ± 20 nm) from collagen. Maximum collagen coverage values derived from SHG signal (by depth (line graph) and at peak value (histogram inset)) was used as a measure of collagen abundance and density. Signal was acquired from three separate areas measuring 320 x 320 µm 2 across each sample. Bright-field transmission images were co-acquired with SHG data.
ImageJ (NIH, Bethesda MD, USA) was used to calculate percentage area covered by SHG signal per image, after conversion to a binary image based upon a single manually determined threshold value applied across all images as previously described (50,91). Results were expressed as medians, ranges and quartiles across all data sets.
Gray-level co-occurrence matrix (GLCM) analysis was carried out as previously described (31). Briefly, collagen fiber organization was assessed using GLCM analysis to characterize the texture of a sample and determine the correlation of the SHG signal within the matrix. The correlation plots represent the similarity in signal strength between pixels. A slower decay shows a more organized and correlated network of collagen fibers than in samples with a faster decay. GLCM analysis was performed in ImageJ.
Orientation Analysis was carried out as previously described (30). Briefly, fiber orientation analysis was performed on SHG images using an in-house ImageJ (NIH) macro where structure tensors were derived from the local orientation and isotropic properties of pixels that make up collagen fibrils. Within each input image, these tensors were evaluated for each pixel by computing the continuous spatial derivatives in the x and y dimensions using a cubic B-spline interpolation. From this, the local predominant orientation was obtained. The peak alignment (measured in degrees) of fibers was then determined, and the frequency of fiber alignment calculated.
Colony forming ability was assessed at day 5, 8 and 12.

Tumorsphere assays
Low passage M6 cells grown to 70 -80% confluency as adherent monolayer were trypsinized, quenched in normal culture media, washed three times with large volumes of calcium-magnesium free PBS then passed through a 40 µm cell strainer to obtain a single cell suspension. Cell number was determined using the Countess™ Automated Cell Counter (Invitrogen) then seeded in sphere-promoting culture at density of 2.5 × 10 3 cells/mL in ultra low-adherent 6-well plates (Corning ® LifeSciences). Cells were grown at 37°C in a 5% CO 2 incubator. Primary sphere formation efficiency was determined after 5 days. Spheres larger than 40 µm were counted manually using a light microscope and automatically using the IncuCyte ZOOM ® Live Cell System (Essen BioScience). Primary spheres were then collected by gentle centrifugation and washed with calcium-magnesium free PBS prior to dissociation into single cell suspension. Cell number was determined as above then seeded in triplicate at density of 5 x 10 2 cells/mL in ultra low-adherent 6-well plates. We defined a Hedgehog Pathway Activation Signature (HPAS) predictive for clinical response to sonidegib (LDE-225) in combination with docetaxel as cases with both high SHH expression in the tumor epithelium (SHH Hscore > 150) and intense GLI1 expression in the tumor stroma.

Statistical analysis
All statistical analyses of all data were performed using GraphPad Prism 6.0c