High endogenous CCL2 expression promotes the aggressive phenotype of human inflammatory breast cancer

Inflammatory Breast Cancer (IBC) is a highly aggressive malignancy with distinct clinical and histopathological features whose molecular basis is unresolved. Here we describe a human IBC cell line, A3250, that recapitulates key IBC features in a mouse xenograft model, including skin erythema, diffuse tumor growth, dermal lymphatic invasion, and extensive metastases. A3250 cells express very high levels of the CCL2 chemokine and induce tumors enriched in macrophages. CCL2 knockdown leads to a striking reduction in macrophage densities, tumor proliferation, skin erythema, and metastasis. These results establish IBC-derived CCL2 as a key factor driving macrophage expansion, and indirectly tumor growth, with transcriptomic analysis demonstrating the activation of multiple inflammatory pathways. Finally, primary human IBCs exhibit macrophage infiltration and an enriched macrophage RNA signature. Thus, this human IBC model provides insight into the distinctive biology of IBC, and highlights potential therapeutic approaches to this deadly disease.

I nflammatory Breast Cancer (IBC) is the most aggressive form of breast cancer, with the lowest survival rate compared to all other breast cancers. At diagnosis, these tumors are considered at least stage IIIB, with most patients exhibiting extensive lymph node metastases and nearly 30% distant metastases [1][2][3][4][5] . Clinical symptoms include skin redness, an orange-peel appearance (peau d'orange) and swelling, making it initially challenging to distinguish from non-cancerous inflammatory diseases of the breast, as these tumors also grow diffusely rather than as palpable nodules. Another prominent feature of IBC is the presence of tumor emboli in dermal lymphatic vessels 4 . Diagnosis of IBC is made based on biopsy confirming the presence of cancer cells either in dermal lymphatics or in the breast parenchyma associated with clinical detection of inflammation 5,6 .
These distinct clinical and histopathological features suggest unique molecular drivers of the IBC phenotype. However, despite many efforts over the past decade to find unique molecular characteristics, IBC has not been defined as a separate molecular subtype. Genomic studies have revealed that IBC is a heterogeneous disease comprising all classical breast cancer molecular subtypes 7,8 . Even within triple-negative tumors, the most common IBC subtype, the seven triple-negative molecular phenotypes identified exist in similar proportions in IBC and non-IBC, with no robust differences in the gene expression signature between IBC and non-IBC 4,7,8 .
One of the key barriers in advancing understanding of IBC has been the lack of models which recapitulate all of the unique features of human IBC. To date, there are only a few human IBCderived cell lines [9][10][11][12][13][14] . The paucity of models that reproduce features of human IBC upon xenotransplantation and are amenable to manipulation has been a major impediment to the mechanistic understanding of unique aspects of the IBC phenotype. Here, we describe a triple-negative human IBC cell line that recapitulates all key features of human IBC in a mouse xenotransplant model. We show that IBC tumors mobilize distinct subsets of inflammatory cells locally and systemically, and demonstrate a key role of CCL2 secreted at very high levels in recruiting macrophages into tumors and thereby facilitating tumor growth and metastasis.

Results
Biological characterization of the A3250 human IBC cell line. The A3250 breast carcinoma cell line was established from a primary tumor of a patient with triple negative IBC, as previously described for other A series lines 15 . Short Tandem Repeat (STR) analysis confirmed that this line was unique based on its comparison to all human cell lines in ATCC STR and DSMZ STR databases, with less than 55% similarity to any of the existing STR profiles (Supplementary Table 1). In tissue culture, confluent A3250 cells exhibited epithelial morphology (Fig. 1a). At lower cell densities, we observed two distinct phenotypes: cell colonies consisting of epithelial-like polygonal cells in close contact, and colonies comprised of scattered cells with more elongated appearance (Fig. 1b). A3250 cells exhibited higher colonyforming ability than MDA-MB-231 cells, a commonly studied triple negative breast cancer (TNBC) line and retained the ability to form colonies even at low serum concentration (1% FBS) and at very low seeding densities ( Supplementary Fig. 1a, b). In monolayer culture, A3250 cells exhibited a population-doubling time of 28 h in 10% FBS and proliferation was not significantly reduced upon short-term culture in 1% serum ( Supplementary  Fig. 1c). They lacked detectable estrogen or progesterone receptors (ER or PR), and expressed HER2 (Human Epidermal Growth Factor Receptor-2, ERBB2) mRNA at levels comparable to other TNBC cell lines ( Supplementary Fig. 1d). In comparison, BT474 cells with amplified HER2 expressed more than 70-fold higher levels ( Supplementary Fig. 1d, Supplementary Table 2). A3250 orthotopic tumors also lacked ER, PR or HER2 protein, in contrast to MCF7 or BT474 tumors that stained positive for ER and PR or HER2, respectively (Supplementary Fig. 1e-j). These data confirmed that A3250 was a TNBC cell line. A3250 human xenograft model recapitulates key clinical and histopathological features of human IBC. We engineered A3250 cells to express green fluorescent protein (GFP) and firefly luciferase (Fluc) to allow for analyses at single-cell resolution in tissue sections, and for quantitative analyses of tumor burden by bioluminescence imaging, respectively. Orthotopic injection of 10 6 cells into the mammary fat pad of either SCID or NOD SCID mice led to the formation of primary tumors at 100% incidence, with comparable growth kinetics (Fig. 1c, d). A3250 tumors induced pronounced erythema in the overlying skin (Fig. 1e). Nipple changes observed in IBC patients, such as inversion and protrusion, were also seen early in A3250 tumor development ( Supplementary Fig. 2a-c). Tumors did not form typical palpable nodules characteristic of non-IBCs, but were flat in appearance (Fig. 1e, f). Analysis of GFP-tagged A3250 tumors revealed a diffuse distribution of tumor cells as clusters or single cells, throughout the mammary gland and adjacent skin (Fig. 1g, h, Supplementary Fig. 2d, e). Accordingly, A3250 tumors had abundant stroma and low tumor-stroma ratio (% tumor cell area: 33% ± 10.6; range 18-43%, n = 3). By comparison, both SUM149 IBC and MDA-MB-231 non-IBC tumors formed nodules (Supplementary Fig. 2f-j) and exhibited high tumor-stroma ratio (% tumor cell area: SUM149, 61% ± 3.44, range 55-64%, n = 4; MDA-MB-231, 77% ± 2.0; range 74-79%, n = 4). SUM149 formed many invasive clusters ( Supplementary Fig. 2i), but did not show the dispersed growth pattern observed with A3250. Because A3250 tumors exhibited a diffuse pattern of growth, bioluminescence imaging provided more accurate assessment of tumor burden than volume measurements.
Ingenuity Pathway Analysis (IPA) of mRNA Sequencing (mRNA-Seq) data revealed that pathways implicated in cell growth and protein synthesis, in particular mTOR signaling and its effectors, and translation initiation factors eIF2, eIF4 and p70S6K signaling were among those most highly activated in A3250 cells. Among the top ten were also pathways related to cell metabolism, such as mitochondrial dysfunction, oxidative phosphorylation and sirtuin signaling, as well as integrin signaling and junction remodeling ( Fig. 1n and Supplementary Data 1). The majority of these pathways were conserved in A3250 cells in vitro and tumors in vivo.
A3250 IBC tumors spontaneously metastasize to lymph nodes and distant organs. Rapid and extensive regional and distant metastasis is a salient feature of human IBC. In accordance with the notion that A3250 tumors retained key characteristics of human IBC, A3250 tumor xenografts spontaneously metastasized to multiple organs upon orthotopic injection into the mammary fat pad. This was determined by bioluminescence analysis of organs ex vivo and validated by analysis of GFP signal by stereomicroscopy of whole organs and tissue sections (Fig. 2). A3250 cells metastasized to the sentinel (inguinal) lymph nodes (LNs) at 100% incidence and to more distant LNs (axillary and brachial) at 46 % incidence within six weeks (16/35 LNs positive) (Fig. 2a). Lymph node metastases commonly presented as large, highly proliferative lesions, and in the majority of the cases almost completely replaced the lymph node (Fig. 2b, d-f). c Tumor growth following orthotopic injection of 10 6 A3250 tumor cells in SCID or NOD SCID mice measured by caliper weekly. NOD SCID n = 8 and SCID n = 6 biologically independent mice. Data represent mean ± SEM (Standard Error of the Mean). d Bioluminescence imaging of A3250 tumors in SCID mice at week six (n = 12 mice). Black line denotes mean value. e Representative picture of an A3250 tumor at 6 weeks showing erythema of the skin and flat tumor appearance. f Fluorescence stereo-microscopy of an A3250 tumor (GFP, green) ex vivo. g A3250/GFP 300 μm tumor section imaged by confocal microscopy and reconstructed in 3D. Note diffuse pattern of growth. (GFP, green) (h). Higher magnification of boxed area in (g). H&E staining of paraffin embedded tumor sections (i, j) and overlying skin (k). Note the presence of lymphovascular emboli in the dermis (k, arrows). IHC staining for cytokeratin (l) and for the proliferation marker Ki-67 (m). Images are representative of three independent experiments. n = 4 biologically independent mice. n Top ten A3250 canonical pathways based on mRNA-Seq of A3250 tumors, determined by IPA. Top pathways are ranked based on p values. Scale bars: a, b, 100 μm; f, 2 mm; g-m, 100 μm. Source data are provided as a Source Data file. A3250 tumors also showed a high incidence of metastases to the lung and liver (89%) (Fig. 2a). Lung lesions exhibited a unique growth pattern, with cell clusters distributed diffusely through the lung parenchyma (Fig. 2c), closely resembling the diffuse pattern of primary tumor growth. Lung metastases were large (Fig. 2g), frequently replacing an entire part of the lobe, and were highly proliferative ( Fig. 2h-j). Proliferation was particularly high in small lesions ( Fig. 2h-j). We did not observe solitary disseminated cells or micro-metastases at six weeks postinjection, indicating that upon seeding, A3250 cells proliferated rapidly and colonized the lung. Associated with the high metastatic burden at endpoint, we observed notable weight-loss (Fig. 2k) and muscle-wasting of tumor-bearing animals, consistent with cachexia. Together, these findings indicate that A3250-derived tumors recapitulated the aggressive metastatic phenotype of human IBC.
Transcriptomic profiling reveals activation of multiple pathways associated with inflammation in A3250 cells. To gain insight into potential drivers of its aggressive IBC phenotype, we performed transcriptomic analyses of A3250 cells. We compared the A3250 gene expression profile with that of six human TNBC cell lines; two IBC and four non-IBC. Hierarchical clustering of mRNA-Seq data revealed that the transcriptional profile of A3250 cells was closest to that of SUM149 IBC cells (Fig. 3a). There were 4881 Differentially Expressed Genes (DEGs) in A3250 when compared to MDA-MB-231, 7033 DEGs compared to IBC3, and 4919 DEGs compared to SUM149. There were 1779 DEGs in A3250 when compared to all three cell lines, MDA-MB-231, SUM149 and IBC3 (Fig. 3b).
IPA revealed that in comparison to MDA-MB-231 cells, A3250 displayed remarkable enrichment in pathways associated with inflammation. There were ten canonical pathways activated in A3250 compared to MDA-MB-231, of which eight were associated with inflammation (Fig. 3c, Supplementary Data 2). The wound healing pathway was most significantly upregulated, followed by the tumor microenvironment (-log(p value): 9.99 and 8.12, respectively), whereas TREM1 signaling showed the highest z-score (3.58). The TREM1 pathway triggers and amplifies inflammation through interactions with TLR pathways, and is (e, f). Lung metastases stained by H&E (g) or Ki-67 (h-j). Different sizes of metastases are depicted (h-j). k Weight loss of mice bearing A3250 tumors at six-week endpoint (n = 10 biologically independent mice). Differences between the true means were determined through a one-sample t test (p < 0.0001). Short black line indicates mean % weight loss. Red dotted line indicates the average weight gain of an age-matched, naïve control mouse during the same period. p value was determined by one sample t test. Images are representative of two independent experiments. n = 5 biologically independent mice. ***p < 0.001. Scale bars: b, e, 200 μm; c, d, f-j, 100 μm. Source data are provided as a Source Data file. In comparison to IBC3, A3250 showed activation of pathways related to fibrosis, EMT and cytoskeleton among others (Fig. 3c, Supplementary Data 3). Activation of PI3/AKT signaling and tumor microenvironment, and inactivation of semaphorin signaling pathway were observed in A3250 in comparison to both, MDA-MB-231 and IBC3. A3250 showed only five pathways enriched in comparison to SUM149, of which LXR/RXR canonical pathway, related to lipid metabolism and inflammation was most significantly upregulated (Fig. 3c, Supplementary Data 4).
We next performed Gene Set Enrichment Analysis (GSEA) by using the hallmark collection of molecular signature databases (MSigDB), which capture and refine information from multiple gene sets and show high degrees of association with the corresponding protein profiles 17 . GSEA showed that from 50 hallmark datasets, 34 gene sets were significantly enriched in A3250 compared to MDA-MB-231, 31 gene sets compared to IBC3, and 22 gene sets compared to SUM149. Among the pathways implicated in inflammation, hallmark "TNFA signaling via NFKB" was significantly enriched in A3250 compared to MBA-MB-231, IBC3 and SUM149 (Fig. 3d).
Hallmark "IL6 JAK STAT3 signaling" was significantly enriched in A3250 compared to MDA-MB-231 and IBC3 (Fig. 3d) and only marginally compared to SUM149 (p = 0.098, FDR = 0.12), indicating that JAK/STAT3 pathway is activated in A3250 and in agreement with previous studies implicating the JAK/STAT3 pathway in IBC 5 .
Consistent with our evidence indicating activation of multiple inflammatory pathways in A3250 by IPA and GSEA, analysis of activated upstream regulators showed a striking association with Among the top inactivated upstream regulators was RC3H1 that modulates activity of the NF-κB pathway and restrains inflammation by suppressing TNF expression 18 . TNF was also among the top ten upstream regulators activated in A3250 compared to IBC3 (Fig. 3e, Supplementary Data 6), but not to SUM149 (Fig. 3e, Supplementary Data 7). Taken together, these data demonstrate prominent activation of pathways associated with inflammation and innate immunity in A3250 IBC compared to MDA-MB-231 non-IBC cells, whereas differences in inflammatory pathways between A3250 and IBC3 or SUM149 IBC were less pronounced.
A3250 IBC tumors induce prominent regional and systemic inflammation. Pathologists have divergent views about the extent and significance of inflammatory infiltrate in IBC 19 , and the prevailing view is that inflammation is a consequence of the blockage of dermal lymphatic vessels by cancer cells. Thus, we sought to characterize the kinetics of inflammatory cell accumulation in A3250 tumors in comparison to LVI. Immunostaining demonstrated that A3250 tumors were heavily infiltrated with CD45+ cells four weeks post-injection, in comparison to MDA-MB-231 tumors (Fig. 4a, b). In the dermis, CD45 + cells were already abundantly present at three weeks post-injection (Fig. 4c). Dermal lymphatics were numerous and enlarged at that same time-point, but did not yet contain tumor emboli (Fig. 4d). A high density of CD45 + cells was also observed in the dermis of the skin overlying A3250 tumors, even when the tumor was distant from the skin, separated by a layer of muscle and fat ( Supplementary  Fig. 4a-c). Striking accumulation of CD45 + cells in A3250 tumors was seen as early as a week post-inoculation (Fig. 4e, f), concomitant with lymphatic vessel dilation in the dermis, but prior to detectable tumor invasion of lymphatic vessels. These results demonstrate that inflammatory cells accumulated prior to tumor invasion of dermal lymphatics. Thus, the abundance of inflammatory cells in A3250 xenografts likely resulted from their active recruitment by tumor-derived factors, rather than as a consequence of the blockage of lymphatic vessels by tumor emboli.
We next examined lungs from mice bearing A3250 tumors at four weeks post-injection to gain insights into the Inflammation associated with A3250 tumors by immunostaining analysis. a, b CD45 IHC staining of tumors at four-weeks as indicated. c, d Immunofluorescent staining of the skin adjacent to A3250 tumors for CD45 (red) and lymphatic marker LYVE-1 (green) at three-weeks (DAPI, blue). Note dense inflammatory infiltrate in the dermis and the absence of tumor cells within the lymphatics at this time point. e, f Immunofluorescent staining of A3250 tumor (GFP, green) for CD45 (red) at one-week post-injection. g, h Immunofluorescent staining of A3250 tumor for F4/80 macrophage marker (red) and cytokeratin (green) at four weeks as indicated. i Skin adjacent to A3250 tumor immunostained for F4/80. j, k Immunostainining of SUM149 tumor (GFP, green) for F4/80 (red) at four-weeks. l, m Iba1 IHC staining of tumors at four-weeks as indicated. Images are representative of three independent experiments. n = 3 biologically independent mice. Scale bars: 100 μm. microenvironment of metastases. H&E staining revealed abnormal lung architecture characterized by areas of dense inflammatory infiltrate and loss of alveolar structures ( Fig. 6a-d).
Immunostaining showed that CD45 + inflammatory cells were associated with disseminated A3250 tumor cells, and were more prominent than in the lungs of mice bearing MDA-MB-231 tumors of comparable size ( Fig. 6e-h). Flow cytometry demonstrated a trend towards an increase of Ly6C hi myeloid cells and a~40% increase in Ly6G hi neutrophils compared to the lungs of tumor-naïve mice (Fig. 6i). Fractions of total CD11b + cells, Ly6C hi inflammatory monocytes and Ly6G hi neutrophils were significantly increased in blood in comparison to tumor-free mice (Fig. 6j), and this correlated with the increase of these cell subsets in the lung. In comparison to MDA-MB-231, Ly6G + cells were significantly increased and there was a trend towards increase of Ly6C + cells in blood. The fraction of Ly6G + cells relative to F4/80 + and Ly6C + cells was lower in lungs with A3250 metastases than in the lungs with MDA-MB-231 metastases (Fig. 6k), because the total number of Ly6G + cells was 2.8-fold lower (Avg. number of lung Ly6G + cells/g tissue: A3250, 3.7 × 10 6 vs. MDA-MB-231, 14.2 × 10 6 ). Of note, despite an increase of Ly6G + neutrophils in the blood and lungs, there were very few Ly6G + cells within A3250 primary tumors, suggesting a mechanism for selective exclusion of neutrophils. In contrast, in the MDA-MB-231 tumor model Ly6G + neutrophils were the main cell subset recruited to both, primary tumors ( Fig. 5c, e, f) and to the lungs (Fig. 6k).

CCL2 is highly expressed in A3250 cells and in human IBCs.
To identify factors that might mediate the florid recruitment of myeloid cells by A3250 tumors, we analyzed chemokine expression in a mRNA-Seq data set generated from A3250 tumors and cells in vitro. Analysis of human chemokines 23 Data 8). SUM149 cells expressed several chemokines from the CXCL family that were also expressed by A3250, including CXCL1, 2, 3 5, and 16, but did not express CCL2 (Fig. 7b). CCL2 mRNA was also highly upregulated in A3250 cells not only in comparison to SUM149, but also to IBC3 and to a panel of non-IBC breast cancer cell lines (Fig. 7c, Supplementary Table 4). High levels of CCL2, CXCL5 and CCL20 proteins were confirmed in supernatants of A3250 cells by Olink multiplex immunoassay. SUM149 IBC cells also secreted CXCL5 and CCL20, but not CCL2 protein, whereas MDA-MB-231 expressed either no detectable or very low levels of this chemokine (Fig. 7d, Supplementary Table 5). While these results correlated relatively well with qPCR or mRNA-Seq, any differences could reflect cell specific variations in RNA translational efficiency or protein secretion.
To investigate how the chemokine expression profile of A3250 cells related to that of human breast cancers, we analyzed expression levels of the top three most highly expressed chemokines, CCL2, CXCL5, and CCL20, in the TCGA-BRCA database of human invasive breast cancer samples of which none were annotated as IBC (n = 1222). All three chemokines were expressed at much higher levels in A3250 IBC compared to their median expression levels in TCGA breast cancer samples (Fig. 7e). To interrogate expression of CCL2 in IBC patient samples, we analyzed a Van Laere microarray data set consisting of 41 human IBCs and 55 non-IBCs 24,25 , and an Iwamoto data set consisting of 25 IBCs and 57 non-IBCs 26 . These data sets were chosen because they had most IBC samples and because mRNA-Seq data for human IBCs have ARTICLE not yet been reported. CCL2 median expression levels were ranked significantly higher in IBCs than in non-IBCs in both datasets (Fig. 7f, g). Taken together, these data indicate that human IBCs, like A3250, commonly exhibit high CCL2 expression levels.
Chemokine receptor CCR2 is highly expressed by myeloid cells in the A3250 IBC tumor model. CCL2 exerts its effects through CCR2, which is the main chemokine receptor mediating monocyte and macrophage recruitment to sites of inflammation 27,28 . We investigated the expression of CCR2 in A3250 tumors, lungs and on leukocytes in the blood. In the blood, CCR2 was expressed bỹ 30% of total CD11b + cells, and mainly by Ly6C hi inflammatory monocytes (76% CCR2 + of Ly6C hi on average), suggesting increased mobilization of these cells from the bone marrow. A small fraction of Ly6C lo monocytes (7% CCR2 + of Ly6C lo on average) also expressed CCR2, whereas CD11b + cells that were Ly6C − were also negative for CCR2 ( Fig. 7h-j, Supplementary  Fig. 5c, d). In A3250 tumors, CCR2 was expressed by a majority of CD11b + cells (~90%), and in lungs by~45% of all CD11b + cells. In tumors, CCR2 was expressed by Ly6C hi monocytes and by Ly6C lo F4/80 + macrophages to a comparable extent (Fig. 7h-j and Supplementary Fig. 5e). In the lung, Ly6C hi cells were the main subtype that expressed CCR2, with some expression also by Ly6C lo macrophages ( Fig. 7h-j, Supplementary Fig. 5f). CCR2 was not expressed by CD11b + F4/80 − Ly6C − cells in the blood, tumors or in the lungs (Supplementary Fig. 5c-f). Together, these results demonstrated high levels of CCR2 + inflammatory monocytes in the blood and high CCR2 expression by myeloid cells associated with A3250 tumors and lung metastases. While CCL2 was highly expressed by A3250 cells, other human chemokines that activate CCR2 (CCL7, CCL8, CCL13, and CCL16), were not expressed ( Supplementary Fig. 5g, Supplementary Table 6). These findings implicated CCL2 as the key tumor cell-derived agonist of CCR2 in the A3250 model and strongly suggested an important role for CCL2 in monocyte and macrophage recruitment to both A3250 primary and metastatic sites.
CCL2 knockdown markedly reduces A3250 tumor growth, lung metastasis, and inflammatory phenotype. To directly investigate the function of CCL2 in the malignant phenotype of A3250 tumors, we performed shRNA knockdown (KD) for CCL2. shCCL2 resulted in a~63% reduction in CCL2 mRNA, as determined by qPCR, and in a decrease of secreted CCL2 protein in vitro (Fig. 8a, b). CCL2 protein showed multiple bands on Western; including a 11 kDa native form and 13 kDa and 15 kDa glycosylated forms 29 . In tumors, CCL2 transcripts were reduced by 76% on average, as determined by mRNA-Seq ( Fig. 8c; Supplementary Table 6). Downregulation of CCL2 in A3250 cells did not result in an altered expression of other human CCR2 ligands, which remained undetectable (Supplementary Table 6).
CCL2 KD led to a striking inhibition of tumor growth, as determined by bioluminescent imaging (Fig. 8d). A3250 CCL2 KD tumors also exhibited a striking reduction in clinical signs of skin inflammation characteristic of IBC ( Supplementary  Fig. 6a-c). Erythema in particular was drastically diminished upon CCL2 KD, and this effect was independent of tumor size. By comparison, SUM149 IBC tumors did not show pronounced erythema ( Supplementary Fig. 6a, b). Furthermore, CCL2 KD decreased both the incidence and the overall metastatic burden of lung metastases (Fig. 8e, f). The majority of mice bearing CCL2 KD tumors lacked macro-metastases in the lungs, and mainly presented with disseminated tumor cells, whereas mice with control A3250 tumors mainly presented with macro-metastases. Histopathological analysis of cytokeratin-stained tumor sections revealed that the invasiveness of A3250 tumors was not decreased upon CCL2 KD ( Supplementary Fig. 7a-f), suggesting that CCL2 KD primarily decreased metastatic colonization rather than dissemination.
Primary tumors exhibited an early onset of tumor necrosis upon CCL2 KD. Analysis of endogenous expression of GFP in tumor sections revealed that in CCL2 KD tumors, only the peripheral tumor area was viable (Fig. 8g, h), and H&E staining showed large necrotic areas in comparison to control tumors at the four-week time-point (Fig. 8i-k). This was not due to the inhibition of angiogenesis, since blood vessel densities in tumors were not changed upon CCL2 KD ( Supplementary Fig. 7g-i). CCL2 KD also reduced tumor proliferation by as much as 83%, with an average number of EdU+ cells/mm 2 tumor decreased from 458 to 78 per 6 h ( Fig. 8l-n, Supplementary Table 3). In contrast, CCL2 KD had no effect on proliferation and colony formation of tumor cells in vitro (Fig. 8o, p), arguing that CCL2 did not stimulate A3250 growth in an autocrine manner. Similar in vitro and in vivo data were obtained using A3250 cells transduced with an independent shCCL2 construct (Supplementary Fig. 8). In accordance, A3250 tumor cells lacked detectable levels of the CCR2 receptor by mRNA-Seq (Supplementary Table 7). In contrast, CCR2 was highly expressed by mouse cells in the microenvironment of A3250 tumors ( Supplementary   Fig. 8 Effects of CCL2 knockdown on A3250 tumor growth and metastasis. a CCL2 mRNA expression levels in A3250 control and A3250-shCCL2 cells determined by qPCR. n = 3 biologically independent samples. Data are presented as mean values ± SD. Two independent experiments were performed. b CCL2 Western analysis of A3250 cells in vitro as indicated. c CCL2 mRNA levels in A3250-control and A3250-shCCL2 tumors determined by mRNA-Seq. n = 2 biologically independent samples. Data are presented as mean values ± SD. Two independent experiments were performed. d Bioluminescence imaging (total Flux) of A3250-control (n = 12 tumors from 6 mice) and A3250-shCCL2 tumors (n = 14 tumors from 7 mice) in vivo 6 weeks post-injection. Two tumors were harvested from each mouse, therefore to control for repeated measures, a Linear Mixed Model fit by REML and t test using Satterthwaite's method was used for analysis (p = 0.00115). e Metastatic burden in the lungs of A3250-control (n = 6) and A3250-shCCL2-bearing (n = 7) tumors determined by ex vivo bioluminescence imaging at 6 weeks. Signal intensity is indicated in photons/second (p/s). Incidence of metastases: macro-metastases, ≥10 6 p/s; micro-metastases,  Table 7). These results indicated that CCL2 did not promote A3250 tumor growth through autocrine stimulation and supported the concept that macrophages recruited by CCL2 promoted A3250 tumor cell proliferation and viability in vivo.

CCL2 downregulation in tumors decreases macrophage infiltration and systemic inflammation.
Immunostaining of immune cells in CCL2 KD tumors revealed a striking reduction of CD45+ cells in viable tumor areas, most of which were F4/80+ macrophages (Fig. 9a, b), thus correlating the loss of tumor macrophages with inhibition of tumor growth. To examine the systemic effects of CCL2 downregulation, we examined the lungs of A3250 tumor-bearing mice during the early stages of tumor dissemination by immunostaining (3-4 weeks post-injection). In the lungs of mice bearing A3250 control tumors, hotspots of dense inflammatory infiltrates associated with disseminated tumor cells were observed (Fig. 9c-e). These inflammatory hotspots were not present in the lungs of mice bearing A3250/CCL2 KD tumors, and CCL2 KD led to a striking normalization of the lung architecture (Fig. 9c-e). Flow cytometry revealed a reduction of Ly6C hi cells in the lungs by 75%, to the normal level, and of total CD11b + cells by 58% (Fig. 9f). These results correlated with a decrease of Ly6C hi inflammatory monocytes in the blood to the normal level of naïve age-matched mice (88% reduction) (Fig. 9g). In contrast, the number of Ly6C hi cells in the spleen was unchanged upon CCL2 KD and was not correlated with the levels of Ly6C hi cells in the blood or in the lungs (Fig. 9g). This decrease in inflammatory cell accumulation in the lungs correlated with the decrease in A3250 lung metastases, consistent with a metastasispromoting role of Ly6C+ cells in the lung.
Human IBCs are enriched in macrophages. A recent study identified distinct clusters of human TNBCs based on immune infiltrates that ranged from macrophage-enriched to neutrophilenriched 21 . To examine whether human IBCs are macrophageenriched, we analyzed microarray data sets of 41 human IBCs and 55 non-IBCs 24,25 . We applied a 37-gene signature specific for-TAMs in breast cancer 30 . Hierarchical clustering revealed two groups of breast cancers, enriched or not enriched in TAM genes. Among breast cancers, there were significantly more IBC samples enriched in TAM genes (60% IBCs, 40% non-IBCs), than not enriched in TAM genes (24.5% IBCs, 75.5% non-IBCs) (p value of the chi-square test of 0.007) (Fig. 10a).
We extended this analysis by immunostaining an independent cohort of nine human IBCs for the presence of macrophages (CD68+) and neutrophils (CD66b+) (Fig. 10b and Supplementary Fig. 9). High macrophage content was seen throughout the tumors, both in the stroma and/or within tumor cell clusters, in tight apposition to tumor cells. In IBC cases presenting as tumor cell clusters in dermal lymphatic vessels, we observed a dense accumulation of CD68+ macrophages in the dermis. In stark contrast, neutrophils were rare in all IBC tumors analyzed (Fig. 10b, Supplementary Fig. 8). These results were consistent with the gene expression analysis indicating that human IBCs are macrophage enriched, and in agreement with our findings with A3250 tumors, further underscoring the relevance of the A3250 tumor model to human IBC.

Discussion
IBC remains a major clinical challenge, as this highly aggressive form of breast cancer is the most difficult to diagnose and treat, yet has no specific targeted therapies 31 . Understanding of the cellular and molecular determinants of IBC has lagged far behind that of other breast cancers, largely due to the lack of adequate models for this disease and the relative rarity of this tumor phenotype. Although several human IBC-derived cell lines have been reported [9][10][11]13 , none recapitulates the full spectrum of IBC features in a mouse model. The human triple-negative IBC cell line A3250, described here, closely recapitulates the key features of human IBC in a mouse xenograft model, including skin erythema, flat tumor appearance, diffuse tumor growth pattern, high stroma to tumor ratio, lympho-vascular tumor emboli in the skin, and extensive lymph node and distant metastases. The fact that the SCID mouse strain utilized is deficient in adaptive immunity indicates that T and B lymphocytes are not necessary for the aggressive IBC phenotypic manifestations.
It has been generally assumed that the inflammatory phenotype of IBC, which includes erythema and edema of the breast and overlying skin 32,33 , is a consequence of tumor emboli in dermal lymphatics, which are believed to impede egress of inflammatory cells from the tissue 4 . While dermal lymphatic vessel invasion and clinical symptoms of inflammation coincide in IBC, there have been no studies to date concerning their causal relationship. We demonstrated in the A3250 IBC tumor model that inflammation is driven directly by the tumor, and is not secondary to dermal lymphatic vessel invasion. This conclusion is based on our observations from the longitudinal analysis of tumor development that demonstrated vigorous recruitment of inflammatory cells into the tumor and adjacent skin early in tumor development, prior to dermal vessel invasion and formation of lympho-vascular emboli. This does not preclude, however, that at the later stages of tumor development potential changes in lymphatic vessel functions may further impact tissue inflammation.
IBC is still primarily a clinical diagnosis, and despite the inflamed clinical appearance of the skin, very few studies have investigated inflammatory cells in IBC [34][35][36] , and to what extent they contribute to IBC progression 33 . We observed exceptionally high macrophage content in A3250 tumors and in adjacent skin, and this was confirmed in human IBC tissues. Macrophages were frequently seen in direct contact with IBC tumor cells, whereas in most breast cancers macrophages are typically localized at the tumor periphery or in the stroma. High macrophage densities were also observed in uninvolved skin, in line with a recent report that in patients, breast tissue not involved with IBC was enriched in macrophages 37 . Another study found a significant increase of CD14 + monocytes in blood and in tumors of patients with IBC compared to non-IBC 38 , consistent with our findings. Taken together, our data reveal that the prominent presence of macrophages in close contact with tumor cells is a common characteristic of IBC. Finally, in most human breast cancers and in mouse models, neutrophils are the most abundant subset of myeloid cells 21,39,40 , while in A3250 tumors and in human IBCs neutrophils were sparse. A recent study of immune cell profiles in breast cancer categorized TNBCs based on TAN and TAM frequency 21 . Accordingly, IBC can be classified as a macrophageenriched subtype with low TAN to TAM ratio.
CCL2 is a well-known mediator of the innate response and inflammation that supports breast cancer progression through several mechanisms 27,28,41 . TAMs recruited by CCL2 facilitate early steps of metastases by promoting tumor cell migration and intravasation 42,43 . CCL2 enhances recruitment and retention of metastasis-associated macrophages 27 and has been shown to promote metastatic recurrence after chemotherapy 39,44 . We demonstrated that the high level of endogenous CCL2 expression by A3250 tumor cells was a key factor in driving the inflammatory cell influx in IBC and identified the important function of CCL2 in promoting IBC tumor growth. CCL2 KD, and consequently reduction in TAM densities, led to a striking reduction of tumor cell proliferation and a concomitant increase in tumor necrosis. Thus, A3250 cells were highly dependent on TAMs for primary tumor growth, while exaggerated systemic inflammation driven by exceptionally high levels of CCL2 promoted expansion of metastases. These results make it tempting to speculate that macrophages may directly support tumor cell metabolic needs critical for proliferation and survival. Of note, SUM149, an IBC line, which does not express CCL2 at the RNA and protein level, showed a much less aggressive IBC phenotype, lacking erythema or dermal lymphatic invasion and with significantly less inflammatory infiltrate as compared to A3250. These findings further support the important role of high level endogenous CCL2 production in promoting the florid IBC phenotype. Finally, our evidence that macrophages are a dominant subset in clinical IBCs suggest that immunotherapeutic strategies that deplete or otherwise target macrophages may be promising therapeutic approaches for IBC. Thus, this IBC tumor model provides a discovery tool with which to facilitate further understanding of IBC biology and potentially foster the development of much needed IBCtargeted therapeutics.

Methods
Cells and culture conditions. The A3250 breast carcinoma line was isolated from a primary tumor of a 50-year-old anonymized patient with IBC, provided to the Aaronson lab from a tissue biorepository directed by Dr. Dirk Iglehart, Duke University. We have complied with all relevant ethical regulations. Duke IRB Pro00012025 "Breast Cancer Blood/Tissue Repository and Database" is a continuation of Dr. Iglehart's IRB protocol permitting them to provide anonymized human tumor tissues as byproducts of standard clinical procedures to other investigators for cancer research. The Mount Sinai IRB considers research performed with anonymized human samples as non-human subjects research. The A3250 IBC cell line was established in culture by the Aaronson lab as performed for other A series tumor lines 15,45,46 . Briefly, the tumor was cut into 2 mm pieces in high-glucose DMEM supplemented with 4 mM L-glutamine (Sigma, Cat. # D6429), 10% FBS (Corning, Cat. # 35-010-CV) and 1X antibiotics (100 U/ml penicillin, 100 U/ml streptomycin) (Gibco, Cat. # 15140-122), and pieces were evenly dispersed within several T25 tissue culture flasks (~12 pieces/flask). Cultures were incubated at 37°C and medium was replaced only if pH changes occurred or when significant debris was present. Once confluent, cells were split at a 1:2 ratio.
Retrovirus and lentivirus generation. To generate retroviruses for constitutive gene expression, the expression plasmid of interest was co-transfected with the pCL-ampho packaging plasmid in HEK293T cells. For lentivirus production 46 , HEK293T cells were co-transfected with expression plasmid of interest, pCMV Δ8.91 packaging and pMD VSV-G envelope plasmids using Lipofectamine 2000 according to manufacturer's instructions. Conditioned medium containing viral particles was harvested two, three and four days after transfection and virus stock was pooled and aliquots were stored at −80°C until use.
Generation of stable cell lines. To generate A3250 and MDA-MB-231 cells expressing eGFP (enhanced Green Fluorescent Protein) and firefly luciferase (Fluc) (A3250-eGFP-FLuc and MDA-MB-231-eGFP-Fluc), parental cells were transduced with the pCCBS-EGFP retrovirus, followed by puromycin selection 46 . The cells were then transduced with the NSNI-luc lentivirus 45 and selected with neomycin and puromycin.
Cell line authentication. Short tandem repeat (STR) data were generated on the SeqStudio Genetic Analyzer, at the Dept. of Oncological Sciences Sequencing core. We confirmed that the STR signatures of A3250 cells did not match STRs of known cell lines in the ATCC or DSMZ STR database. SUM149, MDA-MB-157, MDA-MB-231, HCC3153, SKBR3, and MCF7 cell lines were also authenticated using STR analysis. Cells were prepared per the manufacturer's recommendations (FTA Sample Collection Kit, cat. no. 135-XV) and processed as described 48 .
IncuCyte proliferation assay. A3250 cells were seeded to 40% confluence in 6-well plates in complete media with 10% FBS, allowed to adhere for 6 hours, and subsequently changed to media containing 10% or 1% FBS. Cells were imaged every four hours for four days using the Incucyte™ ZOOM (Essen BioScience Inc, GUI version 2018), for a total of 16 10X images per well for each time-point. Media was changed once, 48 hr after the start of imaging. Average percent confluence was calculated for each well at every time point (n = 2 wells per condition).
Colony formation assay. A3250 and MDA-MB-231 tumor cells were seeded in triplicate into 6-well plates at a density of 500 or 250 cells per well (52 or 26 cells/ cm 2 ). Cells were seeded in complete media with either 10% or 1% FBS and incubated for 11 days, with media changes every four days. Cells were then fixed with ice-cold methanol (100%) for 10 min and stained with a 0.2 % crystal violet in methanol for 30 min at room temperature. Images of each well were taken and colony numbers were quantified using ImageJ v. 1.52j.
Mouse experiments and tissue collection. SCID mice were used in all experiments unless indicated otherwise. Animals were housed in climate-controlled barrier conditions with automated light/dark cycles. A3250 (1 × 10 6 cells in 100 µl PBS, Phosphate-Buffered Saline solution), SUM149 (2 × 10 6 cells in 100 µl PBS) or MDA-MB-231 (2 × 10 6 cells in 100 µl PBS) tumor cells were injected orthotopically into the fourth mammary fat pads on both sides of six to eight-week-old SCID or NOD SCID mice (Charles River Laboratories). For most experiments, SCID mice were bred inhouse. Mouse weights and tumor sizes were measured weekly. Mice were sacrificed one to eight weeks post-injection, and primary tumors and organs were collected. Six hours before sacrifice, mice were injected with 100 μg of 5ethynyl-2′-deoxyuridine (EdU). For each mouse, tumors, sentinel lymph nodes (inguinal), distant lymph nodes (axillary and brachial), lungs and liver were collected. Tumors and organs were fresh-frozen in OCT, or fixed 1 hr in formalin and embedded in paraffin. Organs collected from mice with GFP-labeled tumors were fixed for 1 hr in formalin and frozen in OCT, to preserve GFP and allow the evaluation of metastases at a cellular resolution. The percent change in body weight was determined through the difference between the body weights measured during week 1 and week 6 (i.e. the endpoint) of the experiment. Mouse experiments were approved by the Icahn School of Medicine at Mount Sinai's Institutional Animal Care and Use Committee (IACUC) and in compliance with all relevant ethical regulations for animal testing and research. Maximal tumor burden permitted by Mount Sinai's IACUC is 10% of mouse body weight or tumor diameter of 10 mm in any dimension. We confirm that tumors did not exceed maximum approved tumor size or burden. In most experiments tumors were collected when 5-7 mm in diameter.
Bioluminescent imaging of tumors and metastases. Mice were anaesthetized and background luminescence signals were acquired. Mice then received an intraperitoneal injection of D-luciferin dissolved in sterile PBS (150 mg/kg), and images were acquired every two minutes until peak luciferin signal was achieved for each mouse. Subsequent to in vivo imaging, mice were sacrificed and organs were quickly harvested. D-luciferin was added again to excised organs and metastases in organs were imaged ex vivo. Organs from a mouse injected with the parental tumor cell line not labeled with Fluc were imaged ex vivo and used to determine background values for each tissue. Bioluminescent imaging was performed on the IVIS® Spectrum (PerkinElmer) and analyzed using LivingImage software, version 4.5 (PerkinElmer).

Visualization of metastases by GFP in whole organs and in tissues sections.
To evaluate macro-metastases, formalin-fixed whole organs from mice bearing eGFP-Fluc expressing tumors were imaged with the Leica MZ16 FA stereomicroscope and DFC300 FX camera before embedding in OCT. Lymph node and lung metastases were evaluated at single-cell resolution in frozen sections (6 μm and 10 μm, respectively) of formalin-fixed tissues. Each section was mounted on a microscope slide, rehydrated briefly in PBS, and coverslipped with Vectashield HardSet mounting medium with DAPI (Vector Laboratories H-1500). Sections were immediately analyzed for eGFP signal using the Nikon Eclipse E600 microscope and DS-Qi1 monochrome camera. The number of sections analyzed for each lymph node was 4-16, and for each lung was 16 macrometastases (clusters of >10 tumor cells). The bioluminescence value for each tissue was then compared to the metastasis evaluation by GFP, and the following classification emerged: tissues with Fluc values <10 4 contained no metastases detectable by GFP analysis; Fluc values ranging from ≥10 4 to <10 5 represented few isolated tumor cells and some micrometastases; Fluc values from ≥10 5 to <10 6 reflected mainly micrometastases and some isolated tumor cells; and Fluc values ≥10 6 reflected presence of macrometastases, in addition to many micrometastases and isolated tumor cells.
Angiogenesis was quantified by measuring the area of CD34+/LYVE-1-vessels in tumor sections using NIS Elements 3.22 (Nikon). Three 20X images per tumor were quantified (total area of 2.5 × 10 6 μm 2 per tumor, n = 3 tumors per experimental condition) and data expressed as vessel area/total tumor area analyzed (% vessel area).
TANs and TAMs were quantified by measuring Ly6G+ and F4/80+ cell area, respectively, in tumor sections using NIS Elements 3.22 (Nikon). One entire scanned section of each tumor was quantified (total area ranged from 1-3.42 mm 2 per tumor, n = 3 tumors) and data expressed as % positive cell area/ viable tumor area analyzed.
EdU proliferation and GFP quantification in tumors. EdU was visualized in formalin-fixed OCT-embedded frozen tumor sections using the Click-iT™ EdU Imaging Kit with Alexa Fluor 594 (red) (Invitrogen cat. no. A-11037, 1:300). Following EdU staining, sections were incubated in 0.5% TritonX/PBS at room temperature for 20 min, and immunostained with anti-GFP antibody (Abcam ab6556, 1:100) to enhance the GFP signal. Images were captured with the Nikon Eclipse E600 microscope and DS-Qi1 monochrome camera. Tumor cell proliferation was determined by quantifying EdU+ signal in the GFP-positive tumor areas using NIS Elements 3.22 (Nikon). Three 20x images were quantified per tumor (total area of 2.5 × 10 6 μm 2 per tumor, n = 3 tumors per experimental condition). Data were expressed as number of EdU+ cells/mm 2 GFP+ tumor area. To calculate tumor-stroma ratio, GFP images were overlaid onto the corresponding DAPI images to determine tumor boundaries and exclude necrotic areas. GFP+ viable tumor area was quantified as above and data expressed as GFP+ tumor area/ total tumor area analyzed (% GFP + area).
Optical clearing and 3D reconstruction of tumors. A3250-eGFP tumors were fixed in formalin for 1 hr, embedded in OCT and frozen. 300 μm thick sections were optically cleared using the Ce3D method 49 . Images were captured with a Zeiss LSM 880 Airyscan confocal microscope using 20x objective. Images 512 × 512 pixels were acquired at 1X optical zoom, with a voxel size of 3 μm depth. A z-stack image was then exported and processed using Imaris software (Oxford Instruments) to generate a 3D image.
Immunohistochemistry. Immunohistochemistry was performed on paraffinembedded tissue sections using the Leica Bond RX automated immunostainer (Leica Biosystems), according to the Leica staining protocol. Briefly, all slides were deparaffinized using a heated Bond™ Dewax Solution (Leica cat. no. AR9222) and washed with 1X Bond™ Wash Solution (Leica cat. no. AR9590). Epitope retrievals were carried out for 20 minutes using the citrate-based Bond™ Epitope Retrieval 1 solution (Leica cat. no. AR9961). Slides were then blocked with either the Dako Dual Endogenous Enzyme Block (Agilent S2003) for ten minutes prior to staining with the Bond™ Polymer Refine Red Detection System (Leica cat. no. DS9390), or with a 3-4% v/v hydrogen peroxide block for five minutes that is incorporated into the Bond™ Polymer Refine Detection kit (Leica cat. no. DS9800) and used for DAB staining. All antibodies were diluted using Bond™ Primary Antibody Diluent (Leica Biosystems AR9352).
Anonymized human IBC tumor samples, formalin-fixed and paraffinembedded, were obtained from the Mount Sinai pathology archive, under the auspices of the Department of Pathology and Tisch Cancer Institute Biorespository Core, and were independently evaluated by another pathologist. According to ISMMS IRB criteria, human tissue samples collected as byproducts of standard care and anonymized are considered non-human subjects research. Out of nine samples, six were tumors, assessed as highly invasive ductal carcinomas, whereas three samples presented exclusively as tumor emboli in dermal lymphatics. Human IBC tumor samples were immunostained with anti-CD66b (BD Biosciences 555723; 1:600) and anti-CD68 antibodies (Dako M0814; 1:100) on the same sections, using Leica Bond RX immunostainer and according to the protocol for Multiplexed Immunohistochemical Consecutive Staining in Single Slide (MICSSS) 50,51 . Whole-slide scans were acquired with the NanoZoomer S60 scanner C13210-01 (Hamamatsu). Alternatively, images were captured with the Nikon Eclipse E600 microscope and DS-Ri1 color camera.
Flow cytometry. Xenograft mouse tumors~5 mm in diameter, spleens and lungs were minced into 1 mm 3 pieces. Age and sex-matched tumor naïve mice were used as controls. Tissues were digested in collagenase D (Sigma 11088858001) using the MACSmix tube rotator (Miltenyi) at 37°C, and pressed through a 70 µm filter. Lymph nodes were dissociated by mechanical disruption. Blood was collected by cardiac puncture. Erythrocytes in blood, spleen and lung were lysed using RBC lysis buffer (eBioscience 00-4300-54), as per the manufacturer's protocol. Cells were counted using the Countess II Automated Cell Counter (Invitrogen). Samples were stained with antibodies on ice for 30 minutes in FBS Stain Buffer (BD Pharmigen 554656), and run on the BD LSRFortessa X-20 cell analyzer. The antibodies used were all from Biolegend: anti-mouse CD45-BV785 (Biolegend 103149), CD11b-APC (Biolegend 101212), Ly6G-PE (Biolegend 127608), Ly6C-BV421 (128032), F4/80-FITC (123108), and CCR2-BV605 (Biolegend 150615) (all at 1 μg/ml). Dead cells were excluded from the analysis based on DAPI signal (50 ng/ml). Analysis was performed using BD FACSDiva™ software (v.8.0.1) (BD Biosciences). Contour plot and histogram images were generated using FCS Express version 6.0 (De Novo Software). Absolute cell numbers were calculated using AccuCheck Counting Beads (Invitrogen PCB100). Cells were identified on FSC-A/SSC-A plot and doublets were excluded using FSC-A/FSC-H plot. Viable cells were identified using FSC-A/DAPI plot and CD45+ cells were identified on FSC-A/CD45 plot. CD11B+ myeloid cells were identified on CD45/CD11B plot. Boundaries between positive and negative staining cell populations were defined using Fluorescence Minus One controls.
Western Blotting. For CCL2 detection, A3250-control or A3250-shCCL2 tumor cells (~7 × 10 6 cells) were cultured in 5 ml serum-free media for 24 hr. Conditioned media were concentrated 23-fold using Amicon Ultra Centrifugal Filter Units (Ultracel-3K, Millipore) and 20 μl was loaded on a 12% SDS-PAGE gel. Western blot analysis 52 was performed 52 using anti-human CCL2 (R&D Systems MAB679) at 1 μg/ml. Secondary antibody, Alexa Fluor 680 goat anti-mouse IgG (Invitrogen, A-21059) was used at 1:10,000 dilution and signal was detected using the LI-COR Odyssey fluorescence imaging system. Western blot source data is provided as a Source Data file.
Olink multiplex immunoassay. Supernatants of A3250, SUM149 and MDA-MB-231 cultures were collected at 24 h and 48 h after seeding 1 × 10 6 cells per 10 cm plate in 8 ml of standard media for each cell line. The conditioned media was harvested and analyzed for chemokines within the Olink inflammation panel (Olink Bioscience, Sweden) by the Human Immune Monitoring Center, Mount Sinai. The assay uses proximity extension assay (PEA) technology wherein oligonucleotide labeled antibody pairs allow pair wise binding to target proteins. The antibody bound target antigens which correspond to oligonucleotides from an amplicon are then quantified by high throughput qPCR and provided as relative quantification of protein concentration denoted as Olink proteomics arbitrary unit NPX (normalized protein expression) on a log2 scale.
RNA Isolation and qPCR. Total RNA was isolated from subconfluent cultures after 24 hr of plating. RNA was isolated using the RNeasy kit (Qiagen), RNA concentration was estimated using a NanoDrop 1000 spectrophotometer (Thermo Scientific) and cDNA was generated from 1 μg RNA using SuperScript II reverse transcriptase (Invitrogen). qPCR was performed with ViiA Real Time PCR System (Life Technologies, Carlsbad) using Fast SYBR Green Master mix (Applied Biosystems, Ther-moFisher). Transcript expression was normalized to the GAPDH house-keeping gene and mRNA[RU, Relative Units] was calculated using the 2 −ΔΔCT method for qPCR analysis. For comparison of human chemokine expression in A3250 and MDA-MB-231 cell lines, sequences of the PCR primers were as follows: CCL2: fp 5ʹ-ATCAC mRNA-Seq. mRNA-Seq #1 (GSE158974) was performed on A3250-control (n = 3) and A3250-shCCL2 (n = 3) tumors in vivo and on A3250 cells in vitro (n = 1). Tumors were harvested when~4-5 mm in diameter. Samples were preserved in RNAlater™ Solution (Invitrogen AM7020) and processed at the Department of Oncological Sciences Sequencing core for mRNA-Seq (75 bp single-end reads on an Illumina NextSeq 500).
mRNA-Seq #2 (GSE 180788) was performed on A3250, MDA-MB-231 and SUM149 cell lines cultured in vitro in standard culture conditions (n = 3). Briefly, cell pellets from subconfluent cultures of A3250, SUM149, and MDA-MB-231 were collected 24 hr after seeding, RNA was isolated using RNAeasy kit (Qiagen), processed for mRNA-Seq and sequenced at GeneWiz (150 bp paired-end reads on an Illumina HiSeq 2500). The mRNA-Seq datasets generated during the current study are available in the Gene Expression Omnibus database: GSE158974 and GSE180788.
mRNA-Seq data analyses. mRNA-Seq fastq files were processed on Galaxy servers 53 . Reads from human and mouse genes were separated by aligning them to the human (hg38) and mouse (mm10) genomes using RNA STAR (v. 2.7.9a) 54 . Gene expression was estimated using featureCounts (Subread v. 2.0.3) 55 , which was normalized for gene length to give expression in units of Fragments Per Kilobase of exon per Million mapped reads (FPKM).
Differential gene expression analysis was performed using the edgeR 56 package with a biological coefficient of variation value of 0.1. Only genes with a minimum of 10 reads, absolute log2 fold-change > = 2 in A3250 compared to SUM149 and MDA-MB-231 cell lines and FDR < = 0.05 were considered to be differentially expressed. Venn diagrams were constructed based on the overlapping differentially expressed transcripts.
For the heatmaps and trees, mRNA-Seq data was downloaded from Gene Expression Omnibus (GSE27003) 57 for MDA-MB-468, BT20, HCC3153 and BT474 cells, and GSM3145605 for IBC3. Complex Heatmap library in R was used to generate the trees and heatmaps based on transcript expression levels determined by Kallisto 58 .
IPA pathway analysis. We performed canonical pathway analysis and upstream regulator analyses on the differentially expressed genes (DEGs) with an absolute log2 fold change > 2 using Ingenuity Pathway Analysis (IPA, Qiagen, Redwood City, www.qiagen.com/ingenuity). We used IPA canonical pathway analysis to determine pathways significantly enriched among DEGs (p value of Fisher's exact test < 0.05; −log p value > 1.3). We considered canonical pathways with an activation z-score ≥ 2 to be activated and pathways with an activation z-score of ≤ −2 to be inhibited 59 . Ratio is calculated as the number of DEGs for the given pathway divided by the total number of genes that map to the same pathway.
We identified significant upstream regulators (p value of Fisher's exact test < 0.05) and determined whether they were significantly activated (activation z-score > 2) or inhibited (activation z-score < −2) using IPA upstream regulator analysis. We excluded the following groups from the analysis results: biologic drug, chemical -endogenous mammalian, chemical -endogenous non-mammalian, chemical-kinase inhibitor, chemicalother, chemical drug, chemical reagent, chemical toxicant, mature microRNA and microRNA.
Gene set enrichment analysis. Gene set enrichment analysis was performed using the GSEA software (version 4.1.0) 60,61 . We utilized the GSEA Preranked feature to compare enrichment between the samples. For each comparison, log fold change for each gene was computed using DeSeq2 62 , retrieved from https://doi.org/ 10.1186/s13059-014-0550-8. The genes were then organized into upregulated (log FC > 1) lists and downregulated (log FC < 1) lists. Each list was sorted by −log(adjusted p value) with the most significantly changed genes at the top of the list, and both lists were input individually into GSEA Preranked and matched against the Hallmark gene set from MSigDB 17 . Statistical significance (nominal p values) of enrichment scores are estimated using an empirical phenotype-based permutation test procedure within the GSEA software 60 . Gene sets were considered significantly enriched with FDR < 25%, and p < 0.05.
Human IBC transcriptome analyses and statistics. To obtain all publicly available expression data for inflammatory breast cancer (IBC), we conducted a comprehensive search of the Gene Expression Omnibus (GEO) 63 , Short Read Archive 64 and EBI Expression Atlas databases 65 , as well as The Cancer Genome Atlas (TCGA; https://cancergenome.nih.gov/) and International Cancer Genome Consortium (ICGC; https://dcc.icgc.org/) data portals. No mRNA-Seq datasets were found. We selected E-MTAB-1006 and GSE22597 microarray datasets (from EBI Array Express and GEO databases, respectively) because they had a large number of IBC samples, because tumor samples were obtained prior to treatment, and because gene expression profile was generated on breast cancer samples in all different hormone receptor categories [24][25][26]66 .
We downloaded the processed expression data for IBC and non-IBC tumor samples and obtained the normalized expression values for all genes, where the raw data was previously normalized using gcrma package in BioConductor (R version 2.13.0) 67 . We assessed the expression of genes represented by multiple probes using the probe with the maximum signal intensity. We then calculated the percentile rank of signal intensity for each of the analysed genes in each of the samples. Percentile ranks of CCL2 were calculated as described above and the significance of the difference in distributions of percentile ranks in IBC vs non-IBC samples was assessed using Wilcoxon rank sum test. To assess macrophage enrichment, we calculated the percentile rank of gcrma normalized signal intensity in IBC and non-IBC tumor samples for 37 TAM-specific genes 30 . For each gene, the percentile ranks were z-score transformed. The samples were hierarchically clustered using Ward's clustering criterion and Euclidean distance. Samples were divided into two clusters and the difference of frequencies of IBC samples in the two clusters compared using chi-square test.
Statistical analyses. Statistical differences in proliferation (EdU) and F4/80+ cell densities in tissue sections were determined through an unpaired parametric twotailed t test. Differences in necrosis were determined using a one-sample two-tailed t test. The statistical difference in weight loss between tumor-naïve and tumorbearing mice was determined through a one-sample t test. For flow cytometry analysis, significant differences between cell subset fractions were determined using a goodness of fit linear regression model. Kruskal-Wallis non-parametric rank sum test was carried out to determine significance between absolute values of cell subsets. In the case of tumors and lymph nodes, where multiple tissues originated from the same mouse, a linear mixed model fit by REML and t test using Satterthwaite's method were used to control for repeated measures and determine significant differences between groups. Differences between absolute values of cell subsets were determined with a Wilcoxon rank sum test (Mann-Whitney nonparametric test). Tumor bioluminescence data was first log transformed to create symmetry, followed by a linear mixed model fit by REML and t test using Satterthwaite's method to test differences between the groups. Bioluminescence in the organs was first tested for normality, and significant differences were determined by a Mann-Whitney two-tailed t test.