Erratum to: Humanization of the mouse mammary gland by replacement of the luminal layer with genetically engineered preneoplastic human cells

Due to a typesetting error, the labelling was changed and the figures in this article [1] were presented in the order 2, 4, 10, 6, 1, 3, 5, 7, 8, 9, 11, 12, 13, 14 and the supplementary figure links were inverted. The revised version has the figures in the correct order.

Results: Growth of normal mammary epithelial cells in conditions favoring ERBB3/4 signaling prevented squamous metaplasia in vitro. Normal human cells were quickly lost after intraductal injection, but cells infected with lentiviruses expressing CCND1, MYC, TERT, BMI1 and a short-hairpin RNA targeting TP53 were able to engraft and progressively replace the luminal layer in the mouse mammary ducts, resulting in the formation of an extensive network of humanized ducts. Despite expressing multiple oncogenes, the human cells formed a morphologically normal luminal layer. Expression of a single additional oncogene, PIK3CA-H1047R, converted the cells into invasive cancer cells. The resulting tumors were ERα+, Ki67+ luminal B adenocarcinomas that were resistant to treatment with fulvestrant. Conclusions: Injection of preneoplastic human mammary epithelial cells into the mammary ducts of immunodeficient mice leads to replacement of the murine luminal layer with morphologically normal human cells. Genetic manipulation of the injected cells makes it possible to study defined steps in the transformation of human mammary epithelial cells in a more physiological environment than has hitherto been possible.

Introduction
Pioneering studies by the Weinberg group have shown that normal human cells can be fully converted into tumor cells by targeting a small number of processes known to be abnormal in cancer, including telomere maintenance, the restriction point, cell cycle checkpoints and apoptosis [1]. Human mammary epithelial cells transformed in this way have been widely used, but they frequently form tumors with histological features rarely seen in human breast cancer, in particular squamous cell carcinoma [2]. This is probably caused by excessive activation of the epidermal growth factor receptor (EGFR) [3]. Because squamous metaplasia increases with time in culture, current approaches attempt to shorten the transformation process as much as possible, often to as little as 24 hours [4,5]. Squamous cell changes are also commonly seen in transgenic mouse models of breast cancer, where part of the explanation may lie in mistargeting of oncogenes to myoepithelial or bipotent progenitor/stem cells rather than luminal stem cells [6].
About 70% of human breast cancers express the estrogen receptor α (ERα, ESR1), but genetically defined tumors based on transformation of normal human cells are usually ERα-negative (ER−). Because the ESR1 gene is rarely amplified in breast cancer, ERα expression is normally ascribed to a lineage choice that traps the cells in an ERα+ state. The most likely cell of origin for this event is a luminal progenitor or stem cell located in the terminal duct lobular unit (TDLU) [7].
Traditional breast cancer models based on injection of tumor cells directly into the mammary fat pad [8] or under the renal capsule [9] do not take into account the unique features of the microenvironment in which breast cancers develop. Behbod and colleagues recently described an intraductal injection technique that disseminates tumor cells throughout the mouse mammary ductal tree, including the TDLUs [10]. This approach places potential tumor cells at or near the normal point of origin of breast cancer and faithfully reproduces the histology of human ductal carcinoma in situ (DCIS) [11].
In this article, we describe an approach based on the Behbod intraductal injection technique using genetically engineered cells cultured in conditions favoring EGFR homolog 3 (ERBB3) signaling. Cells transduced with vectors expressing CCND1, MYC, TERT, BMI1 and a short-hairpin RNA targeting p53 (TP53) were able to form a morphologically normal luminal layer. Introduction of a single additional activated oncogene, PIK3CA-H1047R, converted them into ERα+ invasive ductal adenocarcinomas.

Flow cytometry analysis
Cells were grown in WIT medium (mammary culture medium developed by Tan Ince [2]), detached using 0.5% trypsin-ethylenediaminetetraacetic acid (Invitrogen, Carlsbad, CA, USA), pelleted, washed and counted. A total of 10 5 cells were incubated for 20 minutes with the following antibodies: allophycocyanin (APC)-conjugated anti-epithelial cell adhesion molecule (EPCAM), phycoerythrin (PE)-conjugated anti-CD49f and PE-conjugated anti-CD10 (BD Biosciences, San Jose, CA, USA). Nonspecific PE-IgG2a and APC-IgG1 (BD Biosciences) were used as control antibodies. The gates marking the quadrants in the figures were set to exclude cells labeled by the control antibodies. Cells were washed three times, then analyzed on a FACSCalibur flow cytometer (BD Biosciences) using CellQuest Pro software.

Bioinformatics
To identify the normal cell types expressing EGFR family members in human reduction mammoplasty tissue, the most variable probes for each EGFR family member were selected from the Lim et al. Gene Expression Omnibus [GEO:GSE16997] series matrix file [14]. The error bars in Figure 2A are standard errors of the mean. The P-values shown are derived from t-tests performed to compare the mature luminal with the indicated populations. The tumor analysis in Figure 2B and 2C was performed on [GEO:GSE6861] Affymetrix GeneChip Human X3P Array (Affymetrix, Santa Clara, CA, USA) gene expression data derived from large operable or locally advanced breast cancer [15,16]. The .CEL files were normalized with robust multiarray average (RMA) in R [17], filtered to remove probe sets spanning a region <56 base pairs, and the most variable probe sets for EGFR family members were selected based on the standard deviation across all the tumors. The classification shown in Figure 2B and 2C into luminal, molecular apocrine and basal-like tumors is based on expression of FOXA1 (positive in luminal and molecular apocrine tumors) and ESR1 (positive in luminal tumors). Both genes show strongly bimodal expression; the cutoff is the nadir between the two peaks of the bimodal distributions. This classification splits breast tumors into the three groups described by Farmer et al. [18]. FOXA1 was used instead of AR because the X3P probes work slightly better (both genes are excellent markers for luminal and molecular apocrine tumors). The plot in Figure 2B shows the raw expression values generated by RMA. These values were converted to a distribution with the density command in R using the default parameters, then the maximum of each distribution was scaled to 1 to make it easier to compare the distributions in Figure 2C.

Cell biology
Reduction mammoplasty samples were donated by healthy premenopausal women between the ages of 21 and 40 with no previous history of breast cancer. The study was approved by the Tayside Tissue Bank Ethics Committee (project TR000015), and patients gave us their written informed consent to participate. Breast primary epithelial cells (BPECs) were prepared by using standard techniques [19] and cultured in a 37°C, 5% CO 2 , 5% O 2 incubator in Primaria dishes (BD Biosciences) in WIT medium containing 10 ng/ml basic fibroblast growth factor and either 10 ng/ml epidermal growth factor (EGF) for pWIT or 5 ng/ml NRG1 and 10 ng/ml AREG for svWIT [2]. Low-passage cultures of BPECs were sequentially infected with lentiviruses and selected for 4 to 10 days in media supplemented with 2 μg/ml puromycin, 50 μg/ml hygromycin B or 500 μg/ml G418, as appropriate. Viral particles were produced by calcium phosphate transfection of 293 T cells [20]. Infections were performed at a multiplicity of 10 infectious units per cell based on the MCF7 titer. The final 4G-shp53-PI3K cells were uniformly positive for CFP and GFP, which are carried by the BMI1 and shp53 vectors, respectively, but showed heterogeneous staining for tdTomato because the tdTomato vector was added last, not selected with an antibiotic and conferred no growth advantage on the cells. For proliferation assays, cells were treated with 2 nM 17β-estradiol or 1 μM fulvestrant and counted at each time point with a Coulter counter (Beckman Coulter, Brea, CA, USA). To test for significant differences between the growth curves in Figure 4B and Figure 10C, the slopes were compared after linear regression on logged data with the lm (linear model) function in R, with a correction in Figure 4B for performing three tests.

Xenografts
The study was performed in accordance with European Community Standards of Care and Use of Laboratory Animals under level 2 containment at the University of Bordeaux. Approval was granted for the animal experiments by the Comite d'Ethique pour l'Experimentation Animal (CEEA50) Ethics Committee, Bordeaux (project number 5012034-A). For xenografts, 100,000 BPECs and 100,000 p53 −/− mouse embryo fibroblasts irradiated with 30 Gy were mixed with 10% growth factor-reduced Matrigel (BD Biosciences) and injected into 6-week-old female NSG mice (nonobese diabetic, severe combined immunodeficiency, interleukin 2 receptor gamma chainnull, strain number 005557; The Jackson Laboratory, Bar Harbor, ME, USA). The injection volumes were 100 μl for subcutaneous injections and 10 μl for intraductal injections [10]. Tumor size after subcutaneous injection was measured with a PhotonIMAGER (Biospace, Paris, France) and scaled to the size at the first time point in each tumor group. To identify significant differences, t tests were performed in R on logged data with a correction for performing five tests in Figure 6A and ten tests in Additional file 2: Figure S2E and S2F. To measure tumor size after intraductal injection, the glands were removed at the end of the experiment and scanned as a single group with the PhotonIMAGER.

Results
A new genetically defined human adenocarcinoma model WIT medium was developed to prevent squamous metaplasia of genetically defined mammary tumors [2]. When we grew reduction mammoplasty cells in WIT medium, transformed them with TERT, BMI1, ESR1 and MYC, and injected them subcutaneously into immunodeficient mice, they formed mixed tumors that rapidly became entirely squamous with the appearance of stratification and keratin pearls. Figure 1B to 1E show the morphology of the tumors. EGFR family members show differences in expression in normal mammary epithelial cell subsets: EGFR expression is significantly lower in mature luminal cells than in other subsets; ERBB3 expression is significantly lower in stem cells than in mature luminal cells; and luminal progenitors have an intermediate phenotype ( Figure 2A) [14,21]. EGFR family members show strikingly bimodal patterns of expression in human breast tumors: EGFR is higher in basal-like and molecular apocrine tumors; ERBB2 is higher in molecular apocrine tumors; ERBB3 is higher in luminal and molecular apocrine tumors; and ERBB4 is higher in luminal tumors ( Figure 2B, and 2C) [15,16,18]. On the basis of these observations, histopathological data showing EGFR expression by myoepithelial cells, keratinocytes and squamous tumors [22,23] and the fact that EGF promotes expansion of the myoepithelial layer in shortterm organoid cultures [3], we reasoned that it might be possible to reduce squamous metaplasia in engineered tumor models by decreasing the activation of EGFR and increasing that of ERBB3 and ERBB4. To test this hypothesis, we grew normal reduction mammoplasty cells in WIT medium containing either EGF (pWIT medium) or amphiregulin (AREG) plus neuregulin (NRG1) (svWIT medium). Parallel cultures in the two media were infected first with TERT and BMI1 viruses and then superinfected with CCND1 and MYC viruses to produce 4G cells. Expression of the transgenes was confirmed by Western blotting (Figure 3A and Additional file 1: Figure S1A). Cells grown in pWIT formed loose colonies of spindle-shaped cells mixed with larger cells, whereas cells in svWIT formed tighter colonies of round cells that were more uniform in appearance ( Figure 3B). Flow cytometry showed that the svWIT culture was homogeneous, with a single population high in EPCAM expression, low in CD10 expression and intermediate in CD49f expression ( Figure 3C). The cells in pWIT contained two populations, one CD49f−/CD10+/EPCAM− and the other CD49f+/CD10−/EPCAM+. The CD10+ population in pWIT could correspond to myoepithelial cells. The EPCAM+ cells in both media are presumably luminal cells. The EPCAM level is higher in svWIT than pWIT, suggesting that the cells are more differentiated in svWIT. Western blotting showed that the switch from EGF to amphiregulin and neuregulin in svWIT decreased the level of EGFR receptor tail phosphorylation, despite an increase in the amount of the protein ( Figure 3A), consistent with weaker activation of EGFR and decreased receptor turnover. Western blotting of the cells in pWIT showed that they express keratin 14 and ΔN-p63α strongly, confirming the suggestion arising from the flow cytometry data that at least one of the populations in pWIT is myoepithelial or squamous. The cells in svWIT expressed luminal markers (AGR2, ERBB3 and keratin 18) ( Figure 3A), as well as ERα and ELF5 ( Figure 3A), but not FOXA1 or GATA3 (Additional file 1: Figure S1B). ER, FOXA1 and GATA3 are markers for mature hormone-sensing cells, whereas ELF5 is a marker for the secretory lineage [24]. Coexpression of ER and ELF5, as seen in the cells grown in svWIT medium, has been proposed to mark luminal progenitors that have not yet committed to one lineage [24]. We conclude that a shift in the balance of EGFR family signaling in favor of ERBB3 selects for cells that share some features with luminal progenitors  and prevents overgrowth by myoepithelial and squamous cells.

Tumor formation by cells grown in svWIT medium
It was not possible to compare the phenotype of tumors formed by 4G cells grown in the two media, because the cells are not tumorigenic. To transform them fully, they were infected with lentiviruses expressing PIK3CA-H1047R and a short-hairpin RNA targeting TP53 (shp53). Western blotting showed that the shp53 vector reduced the amount of p53 ( Figure 4A). Cells infected with the PIK3CA vector showed increased expression of PI3K and increased phosphorylation of AKT on T308, indicating that PDK1 was activated by PI3K ( Figure 4A). Growth curves showed that the shp53 and PIK3CA vectors significantly increased the growth rate ( Figure 4B). As expected for genetically defined models of this type, comparative genomic hybridization (CGH) profiles were completely flat until p53 was silenced, but even cells containing all six vectors had only minor changes (loss of 3p21, gain of 5p14, loss of X) ( Figure 5). Interestingly, the breakpoint on 3p21 truncates PBRM1 (BAF180), a known breast cancer tumor suppressor gene [25]. To determine whether the cells were tumorigenic in vivo, they were injected subcutaneously into immunodeficient mice. Only the cells transduced with both the shp53 and the PIK3CA vectors could form progressively growing tumors ( Figure 6A; the images of mice show how the data were collected, and the graph shows fluorescence normalized to the value one week after injection). Histological examination showed that the tumors were grade 3 adenocarcinomas ( Figure 6B). The tumor cells stained strongly with anti-GFP antibody and with humanspecific antibodies against keratins 7 and 19, confirming that they were carcinoma cells [26] derived from the injected human epithelial cells. Cells in contact with the mouse stroma expressed p63 and keratin 14, suggesting that they might be trying to form a myoepithelial boundary layer ( Figure 6B; the stroma is marked by an asterisk). Cells in the main tumor mass expressed keratins 5/6 and 8/18. The overall pattern of keratin staining indicates that the tumors were adenocarcinomas rather than squamous carcinomas, but the presence of scattered p63-positive cells and the induction of myoepithelial genes at boundaries with mouse stroma suggest that the cells may retain some common progenitor features. Importantly, the tumor cells were positive for ERα and Ki67 but negative for PGR ( Figure 6B). The experiment was repeated with two other mammoplasties (XS03 and XS05), and expression of TERT, BMI1, MYC and CCND1 and activation of PI3K were confirmed by Western blotting (Additional file 2: Figures S2A and S2B). Flow cytometry showed the presence of a single EPCAM+ luminal population (Additional file 2: Figures S2C and S2D).
Subcutaneous xenografts only formed tumors following introduction of the PIK3CA-H1047R vector (Additional file 2: Figures S2E and S2F). We conclude that cells transformed in medium containing amphiregulin and neuregulin form luminal B mammary adenocarcinomas with squamous metaplasia of cells in contact with the mouse stroma.

Orthotopic tumor formation
Behbod and colleagues recently described a new approach to modeling DCIS by injecting cells through the nipple directly into the mouse mammary ductal tree [10]. To test whether exposure to this more physiological environment would correct the transformed phenotype of the 4G-shp53-PIK3CA cells, we mixed them with irradiated mouse fibroblasts and laminin-rich extracellular matrix (Matrigel) and injected the mixture into the ducts. Fibroblasts and Matrigel were used to promote the survival of the epithelial cells in the perioperative period, but we have not formally tested whether they increase the rate of engraftment. In contrast to the Kuperwasser approach [27], the fibroblasts were murine in origin and injected into the ducts, not into the stroma. Six-week-old mice were injected because the majority of the primary ductal tree is present, providing a large surface area for the human cells to colonize, but the ducts are still growing and forming side branches at this age, so the cells can potentially be incorporated into ducts as they are laid down.
Three weeks after injection, mice were killed and their mammary glands were harvested for examination by stereomicroscopy ( Figure 7A). Clusters of fluorescent cells were scattered throughout the injected glands. At higher magnification, the main ducts and side branches are clearly visible ( Figure 7B). The injected cells expressed red, blue and green fluorescent proteins, but the strength of the red signal was variable in different clones because the number of copies of the tdTomato provirus varied between cells. Clusters of yellow cells alternate with clusters of green cells in the right panel in Figure 7B, indicating that different ratios of red and green vector were present. We can infer that hundreds of cells successfully seeded independent tumor foci throughout the gland. The hematoxylin and eosin (H&E) staining showed that the mammary ducts were filled with cells resembling the pattern seen in DCIS ( Figure 7C). Immunohistochemistry (IHC) for GFP confirmed that the tumor cells were derived from the injected human cells. IHC for p63 and SMA showed that the myoepithelial layer was present at this stage (3 weeks after injection) ( Figure 7C), indicating that the tumor was a DCIS, but there were fewer myoepithelial cells than in normal murine ducts ( Figure 8). As expected, the myoepithelial cells did not stain for GFP, indicating that they were murine in origin ( Figure 8: red arrow, normal duct; green arrow, DCIS). Intraductal injection of cells from the other mammoplasties also produced ERα+ DCIS (Additional file 2: Figure S2G). The tumor cells were all positive for keratins 7, 8/18 and 19; most were weakly positive for keratins 5/6; a few were positive for keratin 14 ( Figure 7C). This mixed luminal/basal pattern has been reported to occur in one-fourth of primary operable breast tumors [28]. Interestingly, the cells were positive for ERα and Ki67 but negative for PGR ( Figure 7C; the PGR stain is negative relative to control cells in the same section).
Six weeks after injection, stereomicroscopy showed dilated ducts packed with fluorescent cells. There was again variable red staining, indicating that multiple  clones were present, but large regions showed homogeneous staining, suggesting that individual clones had expanded to fill those regions ( Figure 9A). Eight weeks after injection, the tumor had spread to fill large contiguous parts of the ductal tree ( Figure 9B). H&E staining showed that the tumor had broken out of the ducts and started to invade the mammary fat ( Figure 9C). Keratin 19 and GFP staining confirmed that the tumor cells were derived from the injected human cells. The tumor cells still expressed ERα, but the intensity of keratin 8/18 staining became weaker and that of keratin 14 stronger as the cells migrated from the ducts into the stroma. This was accompanied by histological features of early squamous differentiation (pink cytoplasmic staining, marked by arrows in Figure 9C, without stratification or keratin pearls). We conclude that loss of environmental cues when the tumor breaks free from the ducts allows the cells to evolve toward a more adenosquamous phenotype, although the squamous differentiation is much less pronounced than in our previous model [12].
Because the tumors are potentially a model for highgrade, ERα-positive breast cancer, they were tested for estrogen dependence. Western blotting showed that in vitro treatment with fulvestrant reduced the level of ERα ( Figure 10A), but the growth rate was unaffected ( Figure 10B and 10C). The cells were xenografted intraductally into mice and allowed to form tumors for 3 weeks, after which the mice received weekly injections of fulvestrant for 5 weeks. IHC showed that the level of ERα was lower in the tumors that received fulvestrant than in the vehicle-treated controls ( Figure 10D), confirming that the drug had reached its target in vivo. To quantify the response, tumors were allowed to form in the mice for 3 weeks, treatment was given for 3 weeks, the injected mammary glands were excised and total tdTomato fluorescence was measured ( Figure 10E). There was no significant difference between the treated and control groups ( Figure 10F). The lack of response in vitro and in vivo mimics the resistance of human luminal B tumors to endocrine therapy in patients with breast cancer.

Humanization of the mouse mammary gland
The Weinberg laboratory has shown that "humanization" of the mouse mammary gland by human fibroblasts promotes engraftment of human mammary epithelial cells [27]. The epithelial cells form duct-like structures, but they do not ramify to fill the gland with a normal ductal tree, possibly because it is difficult to distribute human fibroblasts evenly throughout the fat pad. The logic behind this approach is that subtle differences in signaling proteins create a species barrier that prevents human myoepithelial cells from communicating properly with mouse fibroblasts. We reasoned that human luminal cells might be able to communicate effectively with mouse myoepithelial cells; therefore, we tried to use the mouse myoepithelial layer as a scaffold upon which to build a human luminal epithelium. We initially attempted to humanize the ducts of NSG mice by intraductal injection of reduction mammoplasty cells immortalized by TERT and BMI1 ("2G" cells), but the cells were rapidly lost. To give the human cells a proliferative advantage over the resident murine luminal cells, we tested 4G-shp53 cells, which, as we described above, are nontumorigenic ( Figure 6A). Three weeks after intraductal injection, there was much less fluorescence in the injected glands than after injection of their PIK3CA-transformed counterparts ( Figure 11A). Histological examination showed that small clumps of human cells had penetrated the mouse luminal layer and had attached to the underlying murine myoepithelial layer ( Figure 11B). By 6 weeks, large segments of the mammary ductal tree were fluorescent, indicating that viable human cells had engrafted and increased in number. The presence of alternating yellow and green regions showed that the grafts were derived from multiple different clones ( Figure 12A). IHC for GFP confirmed that extensive regions of the ductal tree contained human cells ( Figure 12B). Histologically, the human cells were slightly larger than the murine luminal cells, but there was no evidence of the premalignant changes of the sort usually observed in flat epithelial atypia or atypical ductal hyperplasia [29]. Human cells replaced murine cells in the luminal layer, but the myoepithelial layer was murine, as shown by the absence of cells positive for both p63 and GFP in doublelabeled sections (Figure 13). The myoepithelial layer was intact, as judged by staining for SMA, but the number of myoepithelial cells was lower than in contiguous regions with murine luminal cells. All of the human cells expressed ERα, and many expressed Ki67, but they did not express PGR, FOXA1 or GATA3 ( Figure 12C and 12D), consistent with their in vitro profile. To test whether the cells could differentiate to form secretory cells, they were injected into the mammary ducts of two mice and allowed to replace the luminal layer of the ducts. The mice were then mated and sacrificed on day 1 of lactation, 14 weeks after injection of the human cells. The human cells were restricted to the ducts and did not form alveoli, although some human cells in the ducts did undergo secretory differentiation ( Figure 14: open arrowheads show murine alveoli; filled arrowheads show a nonsecretory humanized duct; and arrows show casein-expressing cells in a humanized duct). Taken together, the data suggest that the cells immortalized by the TERT, BMI1, CCND1, Fulvestrant has no effect on proliferation of the cells measured by crystal violet staining (B) or growth curves (C). Fulvestrant treatment was started at the arrow (day 3). n = 3; DMSO, Dimethyl sulfoxide; NS, not significant. Scale bars = 1 cm. (D) 4G-shp53-PI3K cells were injected into the ducts, allowed to form tumors for 3 weeks and then treated with fulvestrant or vehicle for 5 weeks. Representative immunohistochemical images show a marked reduction in ERα levels in treated tumors. Each image is derived from a separate animal (n = 6 for control, n = 7 for fulvestrant). The slides were all processed together. Scale bars = 50 μm. (E) and (F) 4G-shp53-PI3K cells were injected into the ducts, allowed to form tumors for 3 weeks and then treated with fulvestrant or vehicle for 3 weeks. PhotonIMAGER scans (E) of excised mammary glands after treatment with fulvestrant or vehicle (control) show how the response was quantified. The scanner counts fluorescence emitted by tdTomato. Graph in F shows quantitative data from the glands shown in E (n = 8 for control, n = 9 for fulvestrant).
MYC and shp53 transgenes may be precursors of hormone-sensing or secretory cells trapped at the luminal progenitor stage.

Discussion
We have shown that genetically defined human mammary epithelial cells expressing multiple oncogenes can form a morphologically normal luminal layer in the mouse mammary gland and that expression of a single additional activated oncogene, PIK3CA-H1047R, is sufficient to convert them into invasive ERα+ adenocarcinoma cells. Two factors were probably critical to the success of the approach: the use of oncogenes commonly mutated or amplified in ERα+ breast cancer and the use of culture medium favoring ERBB3 over EGFR signaling to prevent squamous metaplasia in vitro.
Previous attempts to make breast cancer models by transformation of normal human mammary epithelial cells typically led to the formation of overtly squamous tumors that are unrepresentative of breast cancers seen in human patients. Several arguments implicate EGFR signaling in squamous metaplasia. First, EGFR is prominently expressed in the skin [22]. Second, the myoepithelial layer proliferates at the expense of the luminal layer in short-term organoid culture in media containing commonly used levels of EGF [3]. Third, deletion of Erbb3 in mice leads to loss of the luminal layer and expansion of the basal/myoepithelial layer [30]. Fourth, EGFR is preferentially expressed by stem cells and basallike tumors, whereas ERBB3 is expressed by mature luminal cells and luminal tumors; luminal progenitors express both receptors, presumably because of their intermediate status (Figure 2 and [21,30]). To attenuate EGFR signaling, we replaced EGF in the culture medium with amphiregulin, which is 100-fold less active on EGFR [3], and added neuregulin, which activates ERBB3 and ERBB4 [31]. The WIT medium we used was developed specifically to promote the growth of luminal cells [2], but in the conditions we used, cell lines established in EGF-containing WIT medium underwent squamous metaplasia. Furthermore, our cell lines established in amphiregulin/neuregulin medium grow poorly if transferred back into EGF medium, suggesting that the amphiregulin/neuregulin combination not only prevents overgrowth of a confused myoepithelial population but also favors the growth of luminal cells. That said, we fully agree with Ince et al. [2] that WIT medium favors the growth of luminal cells. We suspect that the differences between our results and theirs reflect subtle differences in the formulation of the medium, the origin of the cells or the choice of transforming genes. Whatever the reason, our results indicate that replacement of EGF with amphiregulin and neuregulin is a simple modification that turns WIT into an even more robust luminalizing medium.
Transgenic mouse models frequently produce tumors with histological features rarely seen in human breast cancer patients. The striking feature of PI3K-induced mouse models is the extent to which they express ERα, despite their unpromising histology [32]. This points to a role for PI3K in the regulation of ERα expression, but the cells in our model already formed ERα+ xenografts before infection with the PIK3CA vector. The Kuperwasser group has described a model based on CCND1, PI3K, KRAS and TP53 that produces ERα+ tumors [4,5]. CCND1 is a strong candidate for driving ERα+ tumor formation in both our models because it is commonly amplified in ERα+ tumors and activates ERα in the  absence of hormone [33,34]. In addition to these arguments, the rationale for using CCND1, PI3K-H1047R, MYC, TERT, BMI1 and sh-p53 was based on genomic data in luminal breast cancer and on past experience with mammary transformation protocols. According to the Cancer Genome Atlas Network, PIK3CA mutations are seen in 45% of luminal A tumors and 29% of luminal B tumors, and TP53 mutations are seen in 29% of luminal B tumors [35]. MYC was included in the transformation cocktail because Elenbaas et al. reported that the endogenous MYC gene would amplify spontaneously if MYC was not supplied exogenously in their transformation protocol [36], and TERT was included to prevent telomere erosion [1]. BMI1 is preferentially expressed by luminal tumors [12,37] and cooperates well with MYC [38], but it was included mainly to prevent unwanted differentiation or genomic instability at early times after the cells were put into culture [12,39]. It is possible that in the new medium it may not be necessary. BMI1 prevents activation of the p53 and Rb pathways [40], but the fact that silencing of p53 by RNA interference is required for full transformation in our model indicates that BMI1 is having at best a partial effect. We suspect this is because BMI1 is expressed at a low level by our lentiviral vector. After silencing of p53, the cells from the XS11 mammoplasty acquired a deletion on 3p21 starting within PBRM1, the gene encoding the BAF180 subunit of the polybromo-and BRG1-associated ATPdependent chromatin-remodeling complex [41]. Truncating mutations, as seen here, have previously been described in breast cancer, where they reduce p21 induction by transforming growth factor β and p53 [25]. We have not performed CGH on multiple independently derived cell lines, so we cannot say whether this change is absolutely required for transformation with CCND1, PI3K-H1047R, MYC, TERT, BMI1 and sh-p53, but it seems unlikely. Indeed, the more striking feature is how flat the profiles are after transformation with defined oncogenes, something we also saw in our previous ER+ model [12].
Despite robust ERα expression, there was no response to fulvestrant. CCND1 and mutant PIK3CA can confer resistance to endocrine therapy in cell lines, but the relationship between PIK3CA mutation and the response to endocrine therapy is hotly debated [42]. Paradoxically, PIK3CA mutations are associated with a gene signature of low mammalian target of rapamycin complex 1 (mTORC1) signaling and better outcomes in ERα+ breast cancer [43]. It will be interesting to see whether the genetically defined ERα+ PIK3CA-mutant model we have developed can shed fresh light on this problem.
Genomic studies indicate that DCIS has a mutation profile similar to that seen in invasive cancer [44], and molecular clock studies indicate that breast tumors undergo little clonal expansion or genomic rearrangement until the last driver mutation is acquired [45]. Taken together, these results push the timeline for acquisition of oncogenic mutations far before the formation of DCIS. Our data following intraductal injection of 4G-shp53 cells show that cells with all but the last driver mutation can form a morphologically normal luminal Figure 13 The myoepithelial layer is murine in humanized glands. Ducts humanized with 4G-shp53 cells were costained for p63 (red), green fluorescent protein (GFP, green) and DNA (4′,6-diamidino-2-phenylindole (DAPI), blue) (n = 3). All injected human cells express GFP strongly. Green arrow, human luminal cells in a humanized duct; red arrow, murine myoepithelial cells in a murine duct. Scale bars = 50 μm. epithelium. The 4G-shp53 cells have a growth advantage over the murine luminal cells they replace, but, apart from being bigger than murine luminal cells, they are normal in appearance. They do not survive after subcutaneous injection, indicating that mammary ducts provide a more welcoming environment than skin. The microenvironment of the duct seems to promote glandular differentiation: the transformed cells were predominantly keratin 18-positive/keratin 14-negative while confined to the ducts, but keratin 14 expression increased when these cells migrated into the stroma. A possible interpretation of this is that the murine stroma is programmed to release EGF in amounts appropriate for murine progenitors but excessive for human progenitors [3]. Alternatively, the fact that occasional human cells bordering the necrotic core of the largest DCIS lesions expressed p63 suggests that squamous differentiation may be a cell-autonomous effect that normal myoepithelial cells prevent through release of a diffusible factor.
We see humanization of the luminal layer in the mouse ducts after intraductal injection as a form of tissue engineering in which the murine myoepithelial layer serves as a scaffold on which to build a human luminal layer. The failure of human cells to form a normal mammary ductal tree after xenografting into the mouse mammary fat pad is normally explained by the existence of a species barrier between human myoepithelial cells and mouse stromal cells. The Weinberg group has shown that this barrier can be partially overcome by infiltration of the mouse stroma with human fibroblasts [27]. Our results indicate that it may be better to retain the mouse myoepithelial layer and replace only the luminal layer with human cells. To do this, we used an intraductal xenografting approach developed by Behbod and colleagues, who showed that the approach is strikingly effective for growth of primary human DCIS cells in the mouse [10]. The DCIS histology in the mouse faithfully recapitulated the histology of the human sample from which it was derived and included regions where hyperplastic human cells formed single-and Figure 14 4G-shp53 cells can undergo secretory differentiation but do not form normal alveoli. Mice with human cells engrafted in the ducts were sacrificed on day 1 of lactation (n = 2). Hematoxylin and eosin (H&E) stain and immunohistochemistry for green fluorescent protein (GFP) show that human cells are present in the ducts but do not spread into the lactating alveoli. Immunohistochemistry for casein shows that some of the human cells in the ducts have undergone secretory differentiation. The structures in the middle of the image are mouse alveoli (open arrowheads). They do not stain for casein, because the antibody is human-specific. The structure on the right is a humanized duct containing cells that do not express casein (closed arrowheads). The structure on the left is a humanized duct containing cells that do express casein (arrows). Scale bars 200 μm. multilayered epithelia [11]. Taken together, our results suggest that intraductal grafting is a major advance over previous approaches to model human breast cancer biology in the mouse. We were not able to replace the murine luminal layer with unmodified human cells. Instead, we started with fully transformed cells and removed one oncogene, PIK3CA. Although the 4G-shp53 cells were able to form a morphologically normal luminal layer in the ducts, they were clearly not functionally normal, because they could not form normal alveolar structures in lactating mice. It will be interesting to test cells expressing even fewer oncogenes to determine the point at which human cells lose their competitive advantage over mouse cells, in order to see whether they can fully humanize the gland and produce functioning alveoli.
A weakness of mouse transgenic models in general is the strong tendency for tumor histology not to resemble that of human breast cancer. One reason may be a fundamental difference in the cell of origin of murine and human mammary tumors. Recent work carried out in the Blanpain laboratory has identified a long-lived luminal stem cell [6] whose human homolog is a strong candidate for the cell of origin of luminal, molecular apocrine and basal-like human breast cancers. The ERα+, ELF5+ cell in our cultures could be the preneoplastic counterpart of this cell. Recent work done in the Ormandy laboratory indicates that ERα and ELF5 crossinhibit one another in luminal progenitors, with definitive hormone-sensing or secretory lineage differentiation ensuing when one factor triumphs over the other [24]. This is perfectly consistent with the failure of our transformed cells to respond to inhibition of ERα with fulvestrant. Drawing parallels between normal lineages and tumor cells is fraught with difficulty because tumor cells inevitably fail many of the tests cell biologists use to characterize normal cells, such as differentiation of progenitors into mature, terminally differentiated cells. Nevertheless, taken together, our data suggest that injection of genetically engineered human cells into the mouse mammary ducts may provide a strategy to study the differentiation of normal and malignant human mammary epithelial cells in a more physiological context than has hitherto been possible.

Conclusion
We have shown that altering the balance of EGFR family signaling in favor of ERBB3/ERBB4 prevents squamous metaplasia of mammary epithelial cells in vitro, that human cells with multiple defined oncogenic changes can replace the luminal cell layer in the mouse mammary gland with a morphologically normal luminal layer of human cells, and that addition of a single activated oncogene, mutant PIK3CA, is sufficient to convert these cells into invasive ERα+ adenocarcinoma cells.
for funding to H. Bonnefoi and R Iggo. We thank the Tayside Tissue bank for providing reduction mammoplasty tissue; the Bergonie Institute microarray platform for performing CGH; Birgit and David Lane for providing antibodies; and Addgene, PlasmID, their contributing scientists and Francois Moreau-Gaudry for providing plasmids.