Embryonic barcoding of equipotent mammary progenitors functionally identifies breast cancer drivers

Summary

Identification of clinically relevant drivers of breast cancers in intact mammary epithelium is critical for understanding tumorigenesis, yet has proven challenging. Here we show that intra-amniotic lentiviral injection can efficiently transduce progenitor cells of the adult mammary gland and use that as a platform to functionally screen over 500 genetic lesions for functional roles in tumor formation. Targeted progenitors establish long-term clones of both luminal and myoepithelial lineages in adult animals, and via lineage tracing with stable barcodes we found that each mouse mammary gland is generated from a defined number of ~120 early progenitor cells that expand uniformly with equal growth potential. We then designed an in vivo screen to test genetic interactions in breast cancer and identified candidates that drove not only tumor formation but also molecular subtypes. Thus, this methodology enables rapid and high-throughput cancer driver discovery in mammary epithelium.

Graphical Abstract

Ying and Beronja employ lentiviral targeting and quantitative lineage tracing of murine mammary epithelium to establish that early progenitors are bi-potent and with equal long-term growth potential. Using these insights they design a large-scale genetic screen that can identify functional drivers of tumor initiation and subtype shift in breast cancer.

Introduction

Sequencing studies have identified more than 15,000 genes altered in breast cancer but not in unaffected germline (Cancer Genome Atlas, 2012; Cerami et al., 2012; Lefebvre et al., 2016; Pereira et al., 2016). Due to their commonality and tumor specificity, they are obvious candidates for functional studies using genetic mouse models (Bissell et al., 2011; Hollern and Andrechek, 2014). Yet, the majority of mutations and copy number alterations (CNAs) in breast cancer are found at low frequency, with 88% of lesions occuring in fewer than 5% of patients (Cerami et al., 2012). These so-called “long-tail” lesions, named after the shape of the frequency histogram of their occurrence, number in the thousands and are yet to be tested as drivers or passengers of tumorigenesis. This challenge to functional annotation of cancer genome data prompted us to investigate if we could develop a method for rapid testing of patient-derived lesions in an immunocompetent physiologically-relevant tissue model.

Mouse mammary epithelium is a powerful system for cancer gene discovery in a complex native environment that is similar to our own. As breast tumors can initiate from both luminal and myoepithelial progenitors (Keller et al., 2012; Polyak, 2007; Visvader, 2009), an unbiased experimental platform should target both lineages with high efficiency. However, the complex architecture of adult mammary tissue has hindered previous attempts at rapid gene targeting methods in vivo. For example, infection of adult mammary epithelium via intra-ductal injection of lentivirus is strongly biased towards luminal lineage, whereas ex-vivo infection of isolated cells is biased towards myoepithelial lineage. Neither method is suitable for large-scale gene function studies (Bu et al., 2009; Hines et al., 2015; Welm et al., 2008). Smaller adenovirus can infect both lineages, but is not suitable for long-term genetic screens due to the transient nature of infection (Tao et al., 2014). An additional challenge to gene targeting in the adult tissue is that such approach is inherently linked with potential immunological response to viral coat protein, viral backbone, or gene targeting construct itself (Nayak and Herzog, 2010).

A method that targets mammary progenitors with gene modifying agents early in morphogenesis prior to establishment of immune tolerance, may overcome the obstacles of the adult tissue. Mammary glands originate from ectodermal progenitors that form placodes between E10.5 and E11.5, develop into a primitive ductal tree by E18.5, and form a branched epithelium of the adult tissue during postnatal development (Robinson, 2007). These early progenitors are thought to give rise to fetal mammary stem cells with superior regenerative and growth potential as compared to more restricted bi- and uni-potent postnatal progenitors (Makarem et al., 2013; Rios et al., 2014). In addition, fetal injection of foreign antigens has been shown to induce sustained immunological tolerance in adult mice (Mold and McCune, 2012). Therefore, early ectodermal progenitors may be the optimal cell population for lentiviral targeting, as they are expected to establish long-term clones in adult mammary glands that would not be a subject to immune attack. Presently, targetability of ectodermal mammary gland progenitors, their number and long-term growth potential remain to be determined. Such basic characterization is critical to establish the method and define the parameters and limitations of scaling up single-gene studies in the mammary epithelium to a high-throughput approach. In addition, we expect that characterization of the tissue’s growth dynamics will enrich our basic understanding of the contribution of early ectodermal progenitors to the morphogenesis and homeostasis of the adult mammary gland.

In the current study we use whole mount 3-dimensional imaging of fetal and adult mammary gland to demonstrate that intra-amniotic injection of lentivirus into E9.5 embryos results in efficient, stable and unbiased transduction of the mammary epithelium. Moreover, we use sequencing-based long-term lineage analysis of the mammary epithelium transduced with a library of stable barcodes to show that each gland is derived from and maintained by a defined number of bi-potent ectodermal progenitor cells, which exhibit a relatively uniform long-term growth potential. Using our targeting approach and detailed understanding of the mammary tissue hierarchy, we design a genetic screen to identify drivers of tumorigenesis. We conduct our study in Pik3ca^H1047R mice, which harbor the most common mutation in breast cancer (Cancer Genome Atlas, 2012), and test 520 gain- and loss-of-function genetic lesions found in patients. Our screen identifies a spectrum of candidate cancer drivers, which enhance tumor formation when compared to oncogenic Pik3ca alone. Interestingly, in addition to faster tumor initiation, a number of candidates shift the molecular signature of oncogenic Pik3ca tumors from Luminal-like to more proliferative Her2-like subtype. Among them is Tsc22d1, a long-tail breast cancer associated gene contained within a 13q14 deletion in ~2% of patients, which our analyses show is a bona fide tumor suppressor of Pik3ca^H1047R-driven mammary tumorigenesis. Together, our study demonstrates that targeting of the long-term ectodermal progenitors of the mammary gland allows for rapid screening of hundreds of clinically relevant mutations to functionally identify drivers of breast cancer.

Results

Intra-amniotic injection of lentivirus results in stable, efficient, and unbiased transduction of adult mammary epithelium

Ultrasound-guided injection of lentivirus into amniotic cavity of mid-gestation embryos results in efficient targeting of the surface ectoderm (Beronja et al., 2013; Beronja et al., 2010), and leads to establishment of long-term clones in epidermis. Interestingly, transduction of appendages such as the hair follicle is less efficient, suggesting that progenitors of ectoderm-derived mini-organs may be relatively inaccessible due to their internalization during placode formation or early barrier formation (Beronja et al., 2010).

To test if we can target ectodermal cells that give rise to the mammary placode and adult mammary gland, we injected a mix of H2B-RFP and H2B-GFP-expressing lentivirus into amniotic cavity of wild type embryonic day (E) 9.25 to 9.75 mice (Figure 1A,B). We collected injected animals at E18.5 and processed their mammary glands for whole mount 3-dimensional imaging (Rios et al., 2014) (Figure S1A). We readily identified transduced cells in fetal glands based on expression of nuclear GFP and RFP (Figure S1A). In addition, we observed double-positive GFP⁺RFP⁺ cells, which represent cells transduced by both lentiviruses (Figure S1A).

Figure 1. Intra-amniotic injection of lentivirus efficiently targets mammary epithelium.

A and B, Ultrasound image (A) and schematic (B) of intra-amniotic injection of lentivirus and infection of ectodermal mammary progenitors. C and D, Maximum intensity projection of whole mount 3D image of E18.5 (C) and P60 (D) mammary gland of R26^mTmG mouse infected at low and high multiplicity of infection (MOI) with lentivirus expressing Cre (LV-Cre). Stars: luminal cells; arrows: myoepithelial cells. Scale bars, 100 μm, apply to all panels. E and F, Schematic (E) and statistics (F) of relative distribution of cells infected at low MOI, with their locations on each branch projected to a standardized rectangle. G, Spatial Kolmogorov-Smirnov test of randomness of distribution of LV-H2B-RFP infected cells in mammary glands.

To test if intra-amniotic transduction of the mammary epithelium was restricted to a subset of progenitors, as reported for transduction of mammary epithelial cells in culture (Hines et al., 2015) and via intra-ductal injection (Bu et al., 2009; Welm et al., 2008), we compared the observed rates of infection with the random sampling model described by the Poisson distribution (Beronja et al., 2013). Using whole mount imaging, we assessed the number of infected and uninfected cells and plotted the proportion of cells that received two or more lentiviruses against the proportion of all infected cells (Figures S1A and S1B). We observed that the transduction dynamics were not different from the curve predicted by Poisson distribution across a range of transductions (Figure S1B; P=0.9601), suggesting that infection of mammary gland progenitors is unbiased.

Next, we evaluated the efficacy of mammary gland targeting. We injected lentivirus expressing Cre recombinase (LV-Cre) into amniotic cavity of R26^mTmG Cre-reporter mice (Muzumdar et al., 2007), which switch membrane-associated RFP (mRFP) with membrane-associated GFP (mGFP) upon Cre expression (Figures 1C and 1D). By varying the viral titer, we observed transduction ranging from a few cells to near complete mammary gland at E18.5 (Figures 1C, S1A and S1B). This broad spectrum of tissue transduction suggested that fetal mammary gland is derived from several to many early progenitors. Importantly, similar transduction range was also observed in adults (P60; Figure 1D), suggesting that adult mammary epithelium is maintained by diverse lineages derived from early progenitors. Using whole mount imaging of adult glands targeted with low titer lentivirus, we observed expression of mGFP⁺ in both cuboidal and stellate-shaped mammary cells (Figure 1D). Co-staining with luminal epithelium enriched E-cadherin and myoepithelial marker SMA (Figure S1C), confirmed that our method transduced both luminal and myoepithelial lineages in the adult gland.

Finally, to understand the spatial distribution of transduced clones in the adult mammary gland, we projected the relative location of mGFP⁺ cells along every branch to a standardized rectangle and performed spatial KS test (Baddeley et al.) (Figures 1E–G). The results indicated that transduced cells follow a random distribution pattern although slightly concentrated on the distal side. This phenomenon is consistent with significant cell dispersal reported to accompany branching morphogenesis during puberty (Hannezo et al., 2017; Scheele et al., 2017). In addition, we expect that early lineage mixing, suggested by spread distribution of rare GFP⁺RFP⁺ cells in fetal glands transduced at low multiplicity of infection (MOI) (Figure S1A), further contributes to random spread of transduced cells throughout the adult gland.

Together, our data suggest that intra-amniotic injection of lentivirus can efficiently infect ectodermal mammary progenitors which give rise to cells of both luminal and myoepithelial lineages and contribute to long-term growth of the adult mammary gland.

Adult mammary gland originates from a defined number of ectodermal progenitor cells with uniform long-term expansion potential

We next set out to determine how many early progenitors contributed to the adult tissue, and whether this contribution was equitable. The random distribution pattern of transduced cells (Figures 1F and 1G) prevented the use of imaging to quantify adult mammary epithelial cells derived from a single progenitor. Instead, we devised a lineage-tracing approach based on identification and quantification of lentivirus-encoded stable barcodes by sequencing. We generated a pool of ~2000 barcodes, by inserting a random 10bp sequence into LV-H2B-RFP construct (Figures 2A and S2A–S2E). To ensure that each progenitor is labeled by a single barcode, we titrated the lentivirus pool to result in ~10% transduction rate of the mammary buds at E12.5 (Figures S2F and S2G). We hypothesized that if early progenitors had an unequal long-term potential to contribute to the adult mammary epithelium we would see few barcodes within the adult gland (Figure 2A). Conversely, if ectodermal progenitors exhibited similar long-term potential, we expected to see higher barcode complexity in the adult gland (Figure 2A). We injected the barcode pool into K14-Cre; R26^yfp/yfp animals, where mammary epithelium is uniformly labeled by YFP and infected cells are marked by YFP/RFP co-expression (Figure 2B). We isolated epithelial cells from each P90 mammary gland and sorted out the YFP⁺RFP⁺ infected population (Figure 2B). We pre-amplified the barcode sequences from gDNA of YFP⁺RFP⁺ cells and subjected this to sequencing (Figures 2C and S2H). The absolute counts of each barcode were then normalized to T0 count, to correct for any sequencing bias (Figures 2C and S2H).

Figure 2. Adult mammary glands originate from a defined number of ectodermal progenitors with uniform long-term growth potential in homeostasis, physiological expansion and regression.

A, Schematic of a strategy to quantify growth potential of ectodermal mammary progenitors using barcoded lentivirus pool. B, Representative image and FACS plot of YFP⁺RFP⁺ (LV-H2B-RFP pool infected) cells in the mammary gland. Scale bar, 50 μm. C, Strategy to profile barcode composition of transduced YFP⁺RFP⁺ mammary cells using sequencing of pre-amplified barcodes followed by normalization to T=0 barcode pool. D, LV-H2B-RFP infection rate, transduced and sorted cell number, and barcode profile from a representative set of mammary glands #1–3 and combined glands #4+5 (See Figure S2I for additional samples). E-G, Statistics of the infection rate (E); number of individual barcode clones (F); extrapolated number of mammary gland progenitors (G) from 6 sets of mammary glands #1–3 and combined glands #4+5 of 3 animals. X-axis indicates gland grouping, same color dots come from the same animal. H, Extrapolated cell number within each clone from 30 mammary glands from three LV-H2B-RFP infected animals. I, Schematic of tissue sampling of LV-H2B-RFP infected mammary glands during lactation and involution. J-L, Statistics of the total number of sorted YFP⁺ cells (J); extrapolated number of mammary gland progenitors (K); and extrapolated cell number within each barcoded clone (L), from 30 mammary glands of LV-H2B-RFP infected animals (n=3 mice of each condition), during lactation (lac) and involution (inv). X-axis indicates gland grouping, dots of the each color come from the same animal. P-values generated using a two-tailed t-test.

For anterior-most mammary glands 1–3 we established the overall transduction rate, the number of transduced cells and the barcode content of each gland from left and right side of three animals (Figures 2D and S2I). Posterior glands 4 and 5 could not be separated and were processed together (Figures 2D and S2I). This led us to several key observations. First, adult glands maintained similar transduction level to that observed in E12.5 buds (Figures 2E, S2F and S2G). This suggests that there is little selection for a dominant early clone, and is consistent with a multi-clonal origin of the adult gland (Shackleton et al., 2006; Stingl et al., 2006). Second, with mammary transduction in the range between 5–15% (Figure 2E), adult mammary glands 1–3 contained between 6–19 long-term barcode clones, while combined glands 4+5 contained between 27–38 distinct barcode lineages (Figures 2F). We also observed that increase in tissue transduction rate was associated with a linear increase in barcode content (Figure S2J). This simple relationship suggested that we could extrapolate the absolute number of progenitors that contribute to each gland by normalizing the number of distinct barcode lineages to infection rate (Figure S2K). Indeed, although different glands had slightly different infection rates (Figure 2E), after normalization we found that there was a conserved number of ~121 (95%CI: 116–126) ectodermal progenitors that contribute to adult glands 1–3 (Figure 2G). Similarly, combined glands 4+5 were calculated to contain approximately twice as many ectodermal progenitors (~247, 95%CI: 241–253; Figure 2G).

We also noted that normalized abundance of each barcode was similar within each mammary gland, and also across the whole set (Figure S2L). Together with the observation of stable transduction rate between mammary buds and adult mammary gland (Figures 2E and S2G), this suggested that the expansion pattern of ectodermal progenitors may be uniform. We next attempted to establish the absolute number of cells within each long-term clone by multiplying the total number of YFP⁺RFP⁺ cells with normalized fraction of each individual barcode (Figures 2H and S2M). Our calculation suggested that, within each gland, a majority of clones were made up of a similar number of epithelial cells (several hundred to several thousand). Interestingly, larger glands 3, 4 and 5 had clones that were larger but in proportion to the overall gland size, as compared to smaller clones in glands 1 and 2 (Figure 2H).

Next, we set to contrast mammary progenitor dynamics in adult tissue homeostasis with hormone-driven tissue expansion and regression as seen during parity (Cowin and Wysolmerski, 2010; Knight and Peaker, 1982). We injected barcoded LV-H2B-RFP virus into animals and profiled barcode content in mammary glands 10 days into lactation (P80, lac) and 5 days into involution (P95, inv; Figure 2I). We collected individual glands at each time point, and processed them as before (Figure 2C). Total number of YFP+ mammary epithelial cells showed a significant increase at P80 (lac) when compared to P90 homeostatic size (hom), and reduction by P95 (inv) when compared to P80 (lac) (Figure 2J). Interestingly, the infection rate across glands remained the same (Figure S2N) suggesting that the burden of expansion and regression was equally born by mammary progenitor clones. We directly addressed this through analysis of barcode complexity, which established that the number of long-term clones (Figure S2O), and estimated number of early progenitors (lac ~122, 95%CI: 119–125; inv ~125, 95%CI: 122–128; Figure 2K) are maintained during parity. Further calculations of cell numbers in each clone indicate that lactation increased while involution decreased overall clone size compared to age-matched non-pregnant females, but in a uniform way rather than through selection of specific clones (Figure 2L). Together, this suggests that equipotency of early mammary progenitors is maintained through pregnancy.

Each mammary progenitor contributes to and maintains both myoepithelial and luminal lineages in homeostasis and during oncogenic growth

Whole mount imaging of mammary epithelia transduced at low MOI suggested that, as a population, ectodermal progenitors have relatively equal potential to develop into myoepithelial and luminal cells (Figures 1D, 3A and 3B). To test if bi-potency is a hallmark of all early progenitors or represents a transient property of a more restricted population, we used our barcode-strategy to profile myoepithelial/luminal composition on a per-clone basis. We isolated CD29^high CD24^low myoepithelial and CD29^low CD24^high luminal cells by FACS from barcoded LV-H2B-RFP infected mammary glands (Figure S3A) and profiled their barcode content. Our analysis showed that each barcode clone comprised of both myoepithelial and luminal cells (Figures 3C and S3B). Moreover, there was a considerable agreement in the relative contribution to either lineage, with clones containing ~5 (95%CI: 3.57–6.07) luminal for every myoepithelial cell. Together, this established that our method had no transduction bias towards either of the two lineages (Figure 3D).

Figure 3. Early mammary progenitors contribute to and maintain both myoepithelial and luminal lineages in homeostasis and during oncogenic growth.

A and B, Maximum intensity projection (A) and statistics of log2 fold-change of proportion of infected myoepithelial cells relative to infected luminal cells (B) suggest no infection bias between the two lineages. P-value generated using one tailed t-test (n=6 mammary glands). Scale bars, 100 μm. C, Number of myoepithelial and luminal cells within each barcoded clone identified in a set of LV-H2B-RFP infected wild type (K14-Cre; R26^yfp) mammary glands (See Figure S3B for additional samples). D, Statistics of log2 fold-change of proportion of myoepithelial cells relative to luminal cells of each clone from 30 mammary glands from three LV-H2B-RFP infected wild type animals. E and F, Extrapolated number of mammary gland progenitors (E) and cell number within each clone (F) from 6 sets of LV-H2B-RFP infected mammary glands #1–3 and combined glands #4+5 of 3 wild type (K14-Cre; R26^yfp) and 3 Pik3ca^H1047R (K14-Cre; Pik3ca^H1047R/+; R26^yfp) animals. X-axis indicates gland grouping, dots of each color come from the same animal. P-values generated using a two-tailed t-test. G, Number of myoepithelial and luminal cells within each barcoded clone identified in a set of LV-H2B-RFP infected Pik3ca^H1047R mammary glands (See Figure S3E for additional samples). H and I, Representative flow cytometry plot (H) and statistics (I) show significant increase of myoepithelial population in Pik3ca^H1047R glands. P-values generated using a two-tailed t-test. J, Statistics of log2 fold-change of proportion of myoepithelial cells relative to luminal cells of each barcoded clone from 30 mammary glands of 3 LV-H2B-RFP infected Pik3ca^H1047R animals. P-value generated using a one tailed t-test.

To investigate how the long-term potential of early mammary progenitors may be affected by a common oncogene we focused on PIK3CA, which encodes the catalytic subunit of PI3K and: (i) is the most frequently mutated gene in breast cancer (Cancer Genome Atlas, 2012); (ii) its constitutive activation is a sufficient driver of mammary hyperplasia and tumor formation in mice (Adams et al., 2011; Van Keymeulen et al., 2015; Yuan et al., 2013); and, (iii) has an available knock-in mouse model that expresses the oncogenic form of Pik3ca (Pik3ca^H1047R) at physiological levels and under endogenous regulatory control (Hare et al., 2014; Kinross et al., 2012; Tikoo et al., 2012). We generated Pik3ca^H1047R expressing mammary epithelium infected by barcoded LV-H2B-RFP at low MOI, and isolated transduced myoepithelial and luminal cells from each gland. Our analyses indicated that the infection rate (Figure S3C), the number of long-term clones (Figure S3D) and the extrapolated absolute number of progenitors per gland were not affected by expression of Pik3ca^H1047R (Figure 3E). In contrast, the clone size was significantly increased in oncogene-expressing mammary epithelium (Figure 3F), which confirms that Pik3ca^H1047R drives mammary gland hyperplasia. Further analysis of lineage bias in Pik3ca^H1047R glands showed that all clones contained both myoepithelial and luminal cells (Figures 3G and S3E). Although the proportion of myoepithelial population was significantly expanded in Pik3ca^H1047R glands (Figures 3H and 3I), as previously reported (Van Keymeulen et al., 2015), there was no significant clonal bias towards either of the two lineages (Figure 3J). Therefore, the lineage bi-potency is a property of every mammary progenitor under both homeostasis and oncogene-driven hyperplasia.

Together, our analyses of the mammary epithelium transduced with a barcoded LV-H2B-RFP virus demonstrate that we can achieve stable transduction of ~120 long-term and equipotent mammary progenitors in each gland (Figures 1–3). Epithelial clones derived from these progenitors have relative equal growth potential in homeostasis and physiological expansion during lactation, and are maintained through involution. Our results are consistent with the reported observation that mammary branching follows a stochastic, self-organized model of action (Hannezo et al., 2017; Scheele et al., 2017). They also suggest that, in the context of a pooled-format lentivirus transduction at MOI≤1, we can conduct a genetic screen in vivo that could test 120 gene targeting constructs in each injected animal (based on 10% targeting of ~120 progenitors of each of the 10 glands per mouse).

Genetic screen uncovers physiological drivers of Pik3ca^H1047R-dependent breast cancer

Next, we set to explore if we can use our lentivirus targeting approach to probe genetic interactions that drive tumorigenesis in mammary epithelium expressing activated form of Pik3ca, which is altered in ~35% of breast cancer patients (Figure 4A). We mated Pik3ca^{H1047R/H1047R} mice with ectoderm-specific K14-Cre transgenic to generate animals expressing a single-copy of oncogenic Pik3ca^H1047R in skin and mammary epithelia (Figure 4B). K14-Cre; Pik3ca^H1047R/+ (Pik3ca^H1047R) mice showed no epidermal phenotype yet developed mammary tumors with complete penetrance starting at ~16 months (Figure 4C). Interestingly, this was comparable to a 12–15 month tumor latency of mice where Pik3ca^H1047R expression was induced postnatally using K5- or K8-CreER transgenes (Van Keymeulen et al., 2015). Moreover, comparison of expression profiles of K14-Cre and K5/K8-CreER induced Pik3ca^H1047R mammary tumors using Gene Set Enrichment Analysis (GSEA) showed that they were highly similar, suggesting that the window of time when Pik3ca^H1047R is activated has a minimal effect on the dynamics of tumor initiation (Figures 4C, S4A and S4B).

Figure 4. In vivo genetic screen identifies drivers of Pik3ca^H1047R–dependent mammary tumorigenesis.

A, Frequency and rate of co-occurrence of mutations in PIK3CA, PTEN and CDKN2A in breast cancer patients of TCGA dataset. B, Schematic of a test for enhanced mammary tumor initiation following intra-amniotic injection of lentivirus expressing shRNAs against Pten or Cdkn2a. C, Kaplan–Meier survival curves of animals transduced with lentivirus expressing Pten or Cdkn2a shRNAs in Pik3ca^H1047R (K14-Cre; Pik3ca^H1047R/+; R26^yfp) and wild type (K14-Cre; R26^yfp) animals. Statistics based on n=8 WT, n=7 WT shPten, n=6 WT shCdkn2a, n=17 Pik3ca^H1047R, n=7 Pik3ca^H1047R shPten and n=6 Pik3ca^H1047R shCdkn2a animals. P-values generated using a Log-rank test. D, Schematic of a test for functional identification of drivers of mammary tumorigenesis using intra-amniotic injection of lentivirus pool modeling patient-derived breast cancer lesions. E, Kaplan–Meier survival curves of wild type (K14-Cre; R26^yfp) and Pik3ca^H1047R (K14-Cre; Pik3ca^H1047R/+; R26^yfp) animals injected with PIK3CA-associated lesion pool. Statistics based on n=17 Pik3ca^H1047R, n=31 Pik3ca^H1047R + LV pool, n=21 WT, and n=32 WT + LV pool animals. P-value generated using a Log-rank test. F, Tumor promoting gene-targeting constructs are identified based on pre-amplification, sequencing and normalization of barcodes from tumor samples. G, Dominant gene targeting constructs from representative mammary tumors generated by the current screen. H, Putative drivers of enhanced Pik3ca^H1047R mammary tumor initiation identified by the current screen.

We next hypothesized that knockdown of known tumor suppressors Pten or Cdkn2a, which are mutated in a sub-set of patients with PIK3CA alterations (Curtis et al., 2012; Pereira et al., 2016) may further enhance mammary tumor initiation (Figures 4A and 4B; also see Figures S4C and S4D). We first confirmed that shRNAs can efficiently knockdown Pten or Cdkn2a in cultured cells, using qPCR, as well as transduced mammary epithelia, using immunofluorescence staining (Figures S4E and S4F). Next, we transduced wild type (WT) and Pik3ca^H1047R mammary glands with shRNAs targeting either Pten or Cdkn2a (Figure 4B). We observed that neither shRNA was sufficient to induce mammary tumors in WT mice, consistent with previous reports (Serrano et al., 1996; Shen-Li et al., 2010). Importantly, both were potent drivers of mammary tumorigenesis in Pik3ca^H1047R mice, accelerating tumor initiation to as early as 11.25 and 8.5 months, respectively (P<0.0001; Figure 4C). Taken together, our results suggest that Pik3ca^H1047R mice have a sensitized mammary epithelium where tumor formation can be accelerated by concomitant targeting of known cancer-driver genes. We further hypothesized that Pik3ca^H1047R mammary epithelium coupled with our lentiviral targeting method may be a powerful system to functionally identify cancer drivers among the myriad genes mutated in breast cancer patients (Figure 4D).

To test this hypothesis, we constructed a barcoded lentivirus library that targeted 520 patient-derived genetic lesions found to co-occur with activating mutations in PIK3CA (Figures 4D, S4G and S4H). Our library modeled putative tumor suppressors using shRNA-mediated knockdown, and putative oncogenes using ORF overexpression (Figure S4G). We combined individual lentivirus constructs into a single pool, which we titrated to achieve mammary transduction rate of ~10%. We injected 32 wild type and 31 tumor-sensitized Pik3ca^H1047R female mice to achieve a ~8X screen coverage. We monitored these mice for up to 24 months and observed that Pik3ca^H1047R mice transduced with lesion pool showed dramatic enhancement in mammary tumor formation (P<0.0001; Figure 4E). We harvested each tumor that formed with reduced latency, pre-amplified and sequenced resident lentiviral barcodes from its gDNA, and normalized barcode counts to T=0 reads (Figure 4F).

We classified tumors as having a single dominant insert if one barcode comprised more than two-thirds of all counts (Figure 4G). We also identified tumors with multiple dominant inserts (Figures S4I and S4J), which we analyzed further. Although the initial transduction rate of ~10% was chosen to promote progenitor targeting with a single shRNA or ORF, it was possible that some mammary progenitors received multiple constructs that collectively drove tumorigenesis. Alternatively, tumor tissue with multiple inserts could be a result of initiation events driven by a single construct but co-occurring in adjacent regions of the same gland, resulting in tumor cell mixing over time. To differentiate between these two scenarios, we sampled distinct regions from each multiple insert tumor via punch biopsy of 30 μm-thick tissue sections (Figures S4I and S4J). Barcode profiling showed that individual biopsies contained a single dominant insert, indicating that these tumors are likely of a single-driver origin (Figures S4I and S4J). We identified lesions as putative drivers of tumorigenesis if they were detected as dominant in a tumor, and if they were observed as such more than once (Figure 4H). Among lesions that fit these criteria, 7 known cancer drivers were detected along with 17 genes not previously associated with breast cancer. Of note, several of the putative breast cancer drivers have been implicated as minor regulators of known oncogenic and tumor suppressive processes. Our most commonly recurring candidate tumor suppressor USP47 (Figure 4H), was shown to suppress Sonic Hedgehog-dependent proliferation in cerebellum and medulloblastoma (Bufalieri et al., 2019), and maintain genomic stability under culture conditions (Parsons et al., 2011). The most commonly recurring putative oncogene HMCLL1 (Figure 4H), was shown to promote viability of leukemia cell lines and colony forming efficiency of murine CML progenitors (Park et al., 2019). These lines of evidence suggest that our genetic screen may be capable of uncovering functional drivers of mammary tumor initiation.

Tumor subtype shift accompanies enhanced Pik3ca^H1047R tumor initiation

PIK3CA mutations, although enriched in the Luminal A/B tumors, are seen in all breast cancer subtypes (Lefebvre et al., 2016; Pereira et al., 2016), and activation of Pik3ca^H1047R was shown to increase cell plasticity in the mouse mammary epithelium (Van Keymeulen et al., 2015). Moreover, recent findings suggest that cellular state and breast cancer subtype can be dynamically reprogrammed by specific genetic/epigenetic lesions (Dravis et al., 2018; Jordan et al., 2016). This prompted us to investigate whether mutations that co-occur with PIK3CA could influence the molecular subtype of Pik3ca^H1047R tumors.

Starting with Pik3ca^H1047R mammary tumors with concomitant depletion of Pten or Cdkn2a, we noted that while knockdown of Pten, which further enhances PI3K signaling, resulted in morphologically similar tumors to Pik3ca^H1047R alone, depletion of Cdkn2a, which functions independently of PI3K, resulted in tumors with different morphology (Figure S5A). To test if this was accompanied by changes in expression of molecular markers of mammary tumors, we performed immunofluorescence staining of ER, PR and HER2. We observed that Pik3ca^H1047R tumors exhibit HER2⁻ and ER/PR⁺ pattern of marker expression (Figures 5A, Figure S5B) (Tikoo et al., 2012; Van Keymeulen et al., 2015), which was unchanged with further depletion of Pten (Figures 5A). In contrast, Pik3ca^H1047R shCdkn2a tumors showed significant de novo expression of HER2 along with ER and PR (Figure 5A). Next, we used gene expression profiling and Principle Component Analysis (PCA) with the PAM50 gene signature and 12 known markers of breast cancer subtypes to compare Pik3ca^H1047R, Pik3ca^H1047R shPten, and Pik3ca^H1047R shCdkn2a tumors with human breast cancer expression profiles from The Cancer Genome Atlas (Cerami et al., 2012; Parker et al., 2009; Perou et al., 2000) (Figures 5B and S5C). Our analyses confirmed previous observations that Pik3ca^H1047R tumors cluster with luminal breast cancer subtypes (Van Keymeulen et al., 2015), and further showed that depletion of Pten had no effect on tumor subtype (Figures 5B and S5C). Interestingly, Pik3ca^H1047R shCdkn2a tumors clustered away from the luminal- and together with HER2-breast cancer subtype (Figures 5B and S5C). We next performed low-pass RNA-seq of all enhanced tumors generated by the screen and compared it to gene expression in human breast cancer patients (Figures 5C and S5D). Our analysis shows that majority of the enhanced tumors shifted away from the luminal subtype and clustered with normal or HER2 but not basal subtype (Figures 5C and S5D). This suggested that our genetic screen can identify putative drivers of not only enhanced tumorigenesis, but also of mammary tumor subtype.

Figure. 5. Drivers of enhanced tumor initiation can promote subtype shift of Pik3ca^H1047R mammary tumors.

A, ER, PR and HER2 staining in Pik3ca^H1047R (K14-Cre; Pik3ca^H1047R/+; R26^yfp), Pik3ca^H1047R shPten (LV-shPten; K14-Cre; Pik3ca^H1047R/+; R26^yfp) and Pik3ca^H1047R shCdkn2a (LV-shCdkn2a; K14-Cre; Pik3ca^H1047R/+; R26^yfp) tumors. Scale bar, 50 μm, applies to all panels. B, Principle Component Analysis (PCA) of PAM50 gene signature contrasts expression profiles of Pik3ca^H1047R, Pik3ca^H1047R shPten and Pik3ca^H1047R shCdkn2a tumors with human breast cancer samples. C, PCA of PAM50 genes shows that expression profiles of many tumors generated by the screen segregate away from Luminal A/B cluster. D, Mutational frequency of known and newly identified candidate drivers of enhanced tumor initiation. E, Kaplan–Meier survival curves of Pik3ca^H1047R (LV-Cre; Pik3ca^H1047R/+; R26^yfp); Pik3ca^H1047R shTsc22d1#1 (LV-shTsc22d1#1-Cre; Pik3ca^H1047R/+; R26^yfp); Pik3ca^H1047R shTsc22d1#2 (LV-shTsc22d1#2-Cre; Pik3ca^H1047R/+; R26^yfp) tumors. Statistics are based on n=8 Pik3ca^H1047R, n=10 Pik3ca^H1047R shTsc22d1#1, n=9 Pik3ca^H1047R shTsc22d1#2 animals. P-values generated using a Log-rank test. F, PCA of PAM50 gene signature shows that depletion of Tsc22d1 results in induction of HER2-like breast cancer expression profile.

Tsc22d1, a long-tail CNA gene, is a bona fide driver of initiation and subtype switch of Pik3ca^H1047R tumor

In order to probe a putative driver of enhanced tumorigenesis and subtype shift identified by our genetic screen, we focused on Tsc22d1 because: (i) Tsc22d1 is one of the long-tail lesions in breast cancer that is associated with chromosomal loss of 13q14 in ~2% of patients (Cerami et al., 2012) (Figure 5D), thus not likely to emerge among the top candidates based on statistical methods; (ii) we observed Tsc22d1 as a screen hit 6 independent times (Figures 4G and 4H); (iii) tumors with Tsc22d1 depletion had among the shortest tumor latency (7.5 months); and (iv) previous findings are consistent with TSC22D1 being a tumor suppressor in tissues; down regulation of TSC22D1 was observed in salivary gland tumors (Nakashiro et al., 1998), gliomas (Shostak et al., 2003), prostate (Rentsch et al., 2006) and colon cancers (Qin et al., 2018). Moreover, evidence for TSC22D1 as a mediator of TGF-β-driven cell death (Ohta et al., 1997), BRAF-induced senescence (Homig-Holzel et al., 2011), and stabilization of p53 (Yoon et al., 2012) suggest it as a potentially broadly-relevant candidate tumor suppressor.

To test if Tsc22d1 is a tumor suppressor in mammary epithelium, we constructed a lentivirus to express shRNAs against Tsc22d1 together with Cre recombinase, allowing us to simultaneously deplete our target and activate expression of Pik3ca^H1047R. We identified two shRNAs that efficiently depleted Tsc22d1 in cultured cells as well as transduced mammary glands (Figures S5E and S5F). We monitored transduced Pik3ca^H1047R mice for 15 months and found that both shRNAs enhanced tumor formation relative to Pik3ca^H1047R alone (5.75 and 7 months vs. 15 months; P<0.001; Figure 5E). We next compared the expression profile of Pik3ca^H1047R; shTsc22d1 tumors with the human breast cancer gene signature, and observed that they clustered with HER2-like human tumors (Figures 5F and S5G). Together, these results show that Tsc22d1 is a bona fide suppressor of Pik3ca^H1047R-driven tumor initiation, whose depletion is capable of shifting the molecular subtype of tumors from Luminal A/B to HER2-like.

To probe how depletion of Tsc22d1 promotes tumorigenesis we compared Pik3ca^H1047R and Pik3ca^H1047R; shTsc22d1 mammary epithelia prior to tumor emergence. We observed that early on (P90), knockdown of Tsc22d1 promoted hyperplastic growth of Pik3ca^H1047R mammary ducts (Figure 6A). To understand what drove this enhanced growth, we probed for proliferation and apoptosis via EdU incorporation assay and activated-caspase3/TUNEL staining, respectively in P21 and P90 mammary glands (Figures 6B, 6C and S6A–J). While knockdown of Tsc22d1 did not affect apoptosis (Figures S6D, S6F, S6H and S6J). it significantly enhanced cell proliferation (Figures 6B and 6C), suggesting a cellular mechanism of faster tumor initiation.

Figure 6. Depletion of Tsc22d1 results in increased proliferation and activation of HER2-driven transcriptional program in Pik3ca^H1047R mammary tumors.

A, Whole mount imaging of mammary ducts at P90. Wild type (LV-Cre; R26^yfp); Pik3ca^H1047R (LV-Cre; Pik3ca^H1047R/+; R26^yfp); Pik3ca^H1047R shTsc22d1#1 (LV-shTsc22d1#1-Cre; Pik3ca^H1047R/+; R26^yfp); Pik3ca^H1047R shTsc22d1#2 (LV-shTsc22d1#2-Cre; Pik3ca^H1047R/+; R26^yfp). Scale bar, 50 μm, applies to all panels. B and C, EdU incorporation rate in Pik3ca^H1047R, Pik3ca^H1047R shTsc22d1#1 and Pik3ca^H1047R shTsc22d1#2 mammary epithelia of P21 (B) and P90 animals (C). Statistics are based on n=4 animals of each condition. Two-tailed t-test, P-value as indicated. D, Her2 staining in mammary epithelia and tumors. Insets show images with merged channels. Scale bar, 50 μm, applies to all panels. E, qPCR measuring relative Erbb2 copy numbers in mammary epithelium and tumors. Two-tailed t-test, P-value as indicated. F and G, Gene Set Enrichment Analyses (GSEA) indicate that Erbb2 transgenic mammary tumor signature (Erbb2 Tg Tumor Hi) was enriched in Pik3ca^H1047R shTsc22d1 tumors (F) and in ERBB2 wild type patients with TSC22D1 deletion (G). Enrichment score (ES), normalized enrichment score (NES), false discovery rate (FDR). H and I, HER2 status (outer circles) scored by IHC (H) and PAM50 (I) in patients sequentially stratified by their genetic status of ERBB2 (center circle) and TSC22D1 (middle circles). Statistics based on TCGA breast cancer data set. P-values generated using a Chi-square test.

To test if a shift to a HER2 subtype is a result of expression changes caused by Tsc22d1 depletion or by selection of Erbb2 amplification during tumorigenesis, we first probed for HER2 expression in Pik3ca^H1047R alone and Pik3ca^H1047R; shTsc22d1 mammary epithelia and tumors. We found that broad HER2 overexpression was evident in the Tsc22d1-depleted mammary epithelium before tumor formation (Figure 6D). Moreover, using qPCR analysis of Erbb2 copy number, we found no evidence of genomic Erbb2 amplification in Tsc22d1-depleted mammary tumors (Figure 6E). Together, this suggests that transition to a HER2 tumor subtype following Tsc22d1 knockdown is independent of diagnostic Erbb2 copy number changes.

To further explore the functional impact of Tsc22d1 mediated HER2-like phenotype switch we used GSEA to compare transcriptional profile of Pik3ca^H1047R and Pik3ca^H1047R; shTsc22d1 tumors with the available profile of Erbb2 transgenic mammary tumor (Zhu et al., 2011). Our analysis shows that Pik3ca^H1047R; shTsc22d1 tumors are significantly enriched for the Erbb2-driven tumor signature when compared to Pik3ca^H1047R alone (Figure 6F). This indicates that tumor subtype switch following Tsc22d1 knockdown is not only associated with expression of PAM50 markers, but also with functional transcriptome reprogramming of HER2-driven tumors. To investigate if loss of TSC22D1 in human breast cancer is also associated with HER2-like phenotype, we analyzed the TCGA breast cancer dataset. We identified patients whose tumors had a normal copy number of ERBB2, as scored by GISTIC (Mermel et al., 2011). We further stratified this group into TSC22D1 wild type or TSC22D1 deleted sets and compared their transcriptional profiles to Erbb2-driven mammary tumors (Figure 6G). Our results show that patients with deletion of TSC22D1 are significantly enriched for transcriptional signature of HER2 driven tumors (Figure 6G). We also compared how the genetic status of TSC22D1 and ERBB2 correlated with tumor classification, scored by IHC and PAM50 analyses (Figures 6H and 6I). We observed that patient tumors with deletion of TSC22D1 were significantly more likely to be classified as HER2 positive by IHC (Figure 6H). Similarly, TSC22D1 deletion was significantly association with breast cancers scored as HER2-like by PAM50 analysis (Figure 6I). Importantly, this was observed in patients irrespective of the ERBB2 amplification status (Figures 6H and 6I), suggesting that TSC22D1 loss results in a general potentiation of Her2 expression and ERBB2 signature. Therefore, TSC22D1 deletion may act as a non-mutational driver of HER2-like breast cancer subtype in human patients.

Loss of Tsc22d1 promotes MDM2-dependent p53 depletion and results in HER2-dependent but not addicted mammary tumors

To investigate the therapeutic implications of our findings and uncover the mechanism of how Tsc22d1 depletion mediates HER2-like subtype switch we generated primary cell cultures from Pik3ca^H1047R and Pik3ca^H1047R; shTsc22d1 mammary tumors. Pik3ca^H1047R cells maintained ER⁺/PR⁺/HER2⁻ expression whereas Pik3ca^H1047R; shTsc22d1 cells showed increased HER2 levels (Figures 7A and S7A), recapitulating the pattern observed in luminal-like and Her2-like tumors they originated from. We also employed CRISPR/Cas9 and induced Tsc22d1 deletions in Pik3ca^H1047R tumor cells in culture (Figures S7B and S7C). As with shRNA-mediated Tsc22d1 depletion, these cells showed ER⁺/PR⁺/HER2⁺ expression (Figure S7D). We next tested the susceptibility of isolated mammary tumor cells to HER2 inhibitor Lapatinib, an established therapy reagent used in clinical practice (Burris, 2004; Geyer et al., 2006). Our results showed that Lapatinib reduced proliferation of HER2⁺ Pik3ca^H1047R; shTsc22d1 and Pik3ca^H1047R Cas9 sgTsc22d1 cells as compared to vehicle treated controls (Figures 7B and S7E). The Lapatinib effect was entirely dependent on Tsc22d1 suppression, as proliferation of Pik3ca^H1047R cells showed no sensitivity to the drug (Figures 7B and S7E). Interestingly, the Lapatinib-mediated growth suppression stood in contrast to the loss of viability observed upon treatment of HER2-amplified cell line SK-BR3 (Figure 7B) (Holliday and Speirs, 2011). This suggests that while Tsc22d1 mutation-mediated acquisition of HER2 like phenotype is responsible for elevated proliferation it does not result in HER2 addiction observed in HER2-amplified breast cancer.

Figure 7. Loss of Tsc22d1 promotes MDM2-dependent p53 depletion and results in HER2-dependent but not addicted mammary tumors.

A, Immunofluorescence staining of HER2 in primary cells from Pik3ca^H1047R (LV-Cre; Pik3ca^H1047R/+; R26^yfp), Pik3ca^H1047R shTsc22d1#1 (LV- shTsc22d1#1-Cre; Pik3ca^H1047R/+; R26^yfp) and Pik3ca^H1047R shTsc22d1#2 (LV- shTsc22d1#2-Cre; Pik3ca^H1047R/+; R26^yfp) tumors. Scale bar, 50 μm, applies to all panels. B, Cell culture proliferation assay of Pik3ca^H1047R, Pik3ca^H1047R shTsc22d1#1 and Pik3ca^H1047R shTsc22d1#2 tumor cells treated with lapatinib or vehicle control. SK-BR3 cells were used as HER2 addicted positive control. P-values generated using a two-tailed t-test. C, Western blots show that nutlin 3a treatment or overexpression of Tsc22d1 ORF can restore decreased p53 level induced by Tsc22d1 silencing. D, Immunofluorescence staining shows that overexpression of Tsc22d1 ORF suppresses HER2 expression in Pik3ca^H1047R shTsc22d1 cells but not in Pik3ca^H1047R p53^−/− (LV-Cre; Pik3ca^H1047R/+; p53^fl/fl; R26^yfp) cells. Scale bar, 50 μm, applies to all panels. E, Cell culture proliferation assay of Pik3ca^H1047R p53^−/−, Pik3ca^H1047R shTsc22d1 + Tsc22d1 (LV- shTsc22d1#1-Cre; LV-Tsc22d1; Pik3ca^H1047R/+; R26^yfp) and Pik3ca^H1047R p53^−/− + Tsc22d1 (LV-Cre; LV-Tsc22d1; Pik3ca^H1047R/+; p53^fl/fl; R26^yfp) tumor cells treated with lapatinib or vehicle control. P-values generated using a two-tailed t-test.

Lastly, we set to probe the mechanism of Tsc22d1-mediated suppression of HER2-like phenotypes, including HER2 expression, increased proliferation, and sensitivity to Lapatinib. In cervical cancer cell lines, TSC22D1 can directly interact with p53 and block its MDM2 binding site to prevent MDM2-dependent P53 degradation (Yoon et al., 2012), and Pik3ca^H1047R p53^−/− mammary tumors also exhibit a HER2-like phenotype (Van Keymeulen et al., 2015). Together this suggested that loss of Tsc22d1 could promote HER2 expression signature by suppressing P53 (Figure S7F). Consistent with this hypothesis, depletion of Tsc22d1 in Pik3ca^H1047R cells resulted in p53 protein reduction that could be suppressed by treatment with MDM2 inhibitor nutlin 3a or overexpression of Tsc22d1 (Figure 7C). Next, we established a primary culture of cells isolated from Pik3ca^H1047R p53^−/− mammary tumors (Figure S7G). We observed that Pik3ca^H1047R p53^−/− cells phenocopy the HER2-like phenotypes seen in Pik3ca^H1047R shTsc22d1 culture (Figures 7D, 7E and S7H). Importantly, while overexpression of Tsc22d1 in Pik3ca^H1047R shTsc22d1 cells, which we showed restored p53, was sufficient to suppress HER2-like phenotypes, Pik3ca^H1047R p53^−/− cells overexpressing Tsc22d1 maintained HER2 marker expression, increased proliferation and Lapatinib sensitivity (Figures 7D, 7E and S7H). This test of genetic epistasis establishes that p53 is downstream of Tsc22d1 and is required for suppression of HER2-like features of Pik3ca^H1047R mammary tumors. Together our study supports a model where infrequent CNAs that result in deletion of Tsc22d1 promote mammary tumor initiation and loss of p53-mediated suppression of HER2-like phenotype.

Discussion

To date, sequencing of cancer genomes has uncovered tens of thousands of genomic lesions, and their correlations with common “founder” mutations in human tumors. However, the vast majority of these lesions and their effects on cancer initiation and progression remain completely unknown. Mouse models of human breast cancers provide a platform to explore these effects, within an intact tissue and in the physiological microenvironment that is similar to our own. However, conventional mouse genetic models take a significant time and effort to establish, and as a result, fewer than 30 breast cancer lesions have been functionally validated (Bissell et al., 2011; Hollern and Andrechek, 2014).

In vivo lentiviral screens have provided a non-invasive alternative to conventional genetics, and allowed for rapid testing of oncogenic lesions in epidermal squamous cell carcinoma (Beronja et al., 2013; Schramek et al., 2014) and non-small cell lung cancer (Chen et al., 2015; Rogers et al., 2017). In the mammary epithelium however, efforts to develop large-scale screening strategies have been hampered by low efficiency and lineage-specific transduction bias (Hines et al., 2015; Welm et al., 2008). In addition, introduction of gene targeting constructs in the adult has the potential to activate the animal’s immune response against foreign antigens (Nayak and Herzog, 2010). In the current study we demonstrate that intra-amniotic injection of lentivirus can efficiently target ectodermal progenitors of mammary gland before lineage specification and morphogenesis. As a result, when targeted progenitors develop into adult mammary gland, both luminal and myoepithelial cells are efficiently transduced, allowing us to test loss- and gain-of-function lesions in both mammary lineages.

Previous studies of growth dynamics in the mammary epithelium focused on the branching process during puberty and lactation cycles in adult animals (Hannezo et al., 2017; Scheele et al., 2017; Visvader and Stingl, 2014). In both scenarios, long-term growth potential was shown to be restricted to a small fraction of stem/progenitor cells but not the majority of differentiated mammary epithelial cells. However, whether a similar tissue hierarchy existed among ectodermal progenitors remained elusive. Here, using whole mount 3-dimensional imaging and lineage tracing we found that transduced mammary cells do not form densely packed clones, but are sparsely distributed across mammary gland even before branching process. Further long-term lineage tracing using barcoded lentivirus indicated that every mammary gland originates from ~120 ectodermal progenitors that exhibit relatively uniform long-term potential. This observation lays the foundation for our screen strategy, as it establishes that at low MOI which ensures single virus integration per cell, we can test up to ~120 lesions in the 10 mammary glands of a single mouse.

Molecular subtyping of breast cancer not only precisely assigns breast cancer patients into distinct prognostic groups but also provides treatment guidance (Parker et al., 2009; Perou et al., 2000; Sorlie et al., 2001; van ‘t Veer et al., 2002). Studies of breast cancer genomes have identified significant enrichment of certain lesions in specific cancer subtypes (Cancer Genome Atlas, 2012; Pereira et al., 2016), although the causal relationship between single or combinations of lesions and breast cancer subtypes is still far from understood. Activating lesions in PIK3CA, the most commonly mutated gene in breast cancer, are concentrated in Luminal subtypes (Cancer Genome Atlas, 2012; Pereira et al., 2016). This link is further supported by studies of mouse models of breast cancer, where mammary tumors driven by membrane-associated form of wild type Pik3ca (Renner et al., 2008; Sheen et al., 2016) or physiological levels of constitutively active Pik3ca^H1047R generally expresses luminal like markers (Adams et al., 2011; Stratikopoulos et al., 2019; Tikoo et al., 2012; Van Keymeulen et al., 2015). Moreover, PIK3CA mutations are also found in other subtypes, suggesting that additional factors may provide the necessary context for phenotypic evolution of the tumor. Indeed, in a model driven by tetracycline-induced overexpression of Pik3ca^H1047R a subset of adeno-squamous carcinomas was observed (Liu et al., 2011), implying that Pik3ca dosage alone may contribute to development of mammary tumors subtypes. Using our in vivo screening strategy, we found that additional gene lesions can promote Pik3ca^H1047R tumor initiation and subtype shift from Luminal to HER2-like. These observations demonstrate that our method can be used not only to rapidly identify tumor drivers but can also reveal critical contributors to breast cancer evolution. The biological basis of breast cancer subtype shift, which has been recently highlighted as a mechanism of treatment resistance in human Luminal breast cancers (Jordan et al., 2016), should be an important focus of future studies.

STAR METHODS

LEAD CONTACT AND MATERIALS AVAILABILITY

Requests for further information should be directed to and will be fulfilled by Slobodan Beronja (beronja@fredhutch.org). All unique/stable reagents generated in this study are available from the Lead Contact with a completed Materials Transfer Agreement.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Mice

All animals, including Pik3ca^{H1047R/H1047R} (donated by Wayne A. Phillips (Hare et al., 2014; Kinross et al., 2012; Tikoo et al., 2012)), K14-Cre, R26^yfp/yfp and R26^mTmG/mTmG Cre-reporter mice (Jackson Laboratories), were crossed to the C57BL/6 background. All animals were immune competent and were not previously used in any experimental procedures or treatments. Developmental stage of female mice used for mammary gland processing was indicated in the Figures. Mice were housed and cared for in an AAALAC-accredited facility at Fred Hutchinson Cancer Research Center, and all animal experiments were conducted in accordance with Fred Hutchinson Cancer Research Center IACUC-approved protocol, project license number 50814.

Primary cell culture and treatments

Tumors were minced and digested in 0.25% collagenase plus 2mg/ml dispase at 37 C° for 1 hour and 0.25% Trypsin-EDTA at 37 C° for 0.5 hour to release epithelial cells. Lentivirus infected epithelial cells, labeled by R26^yfp Cre reporter, were isolated out via FACS and cultured in DMEM medium with 10% FBS. Cells were cultured for less than 5 passages before lentiviral infection and selection. Lapatinib (Apexbio) and nutlin 3a (Tocris Bioscience) were dissolved in DMSO and further diluted in culture medium to a final concentration of 5μM with 0.1% DMSO; 0.1% DMSO was used as vehicle control. Cells treated with Lapatinib were counted at 0, 24 and 48 hours after seeding. Cells treated with nutlin 3a were harvested for western blot 1hr after treatment.

METHOD DETAILS

Lentivirus production and transduction

Large volume (~100 ml) of viral supernatant was produced in 2 T225 flasks of 293TN cells transfected using calcium phosphate with lentiviral backbone and helper plasmids pMD2.G and psPAX2 (Addgene plasmid 12259 and 12260). Supernatant was collected 48 hours post transfection, and concentrated using Centricon Plus-70 centrifugal filter unit (100 kDa, Millipore). This first concentrate was subsequently pelleted by ultracentrifugation at 45,000 rpm for 1 hr using a MLS 50 rotor (Beckman Coulter). Viral pellet was resuspended in 60 μl of Viral Resuspension Buffer (VRB; 20 mM Tris pH 8.0, 250 mM NaCl, 10 mM MgCl2, 5% sorbitol) to generate lentiviral stock suitable for intra-amniotic injections. Lentivirus stocks were serially diluted in VRB and titrated in transduced mouse skin and mammary epithelia using LV-GFP infected epidermis as a qPCR standard, to result in mammary tissue transduction rate of ~10%. Lentiviral transduction of keratinocytes and mammary cell cultures was done on primary cells plated in 12-well dishes at 70,000 cells/well and incubated with lentivirus and polybrene (40 μg/ml) overnight. After 2 days, infected cells were positively selected with puromycin (1 μg/ml) for 4–7 days, and processed for analyses. In utero-guided lentiviral transduction in vivo was performed as previously reported (Beronja and Fuchs, 2013; Beronja et al., 2013; Beronja et al., 2010) with specific modifications. Briefly, to ensure that lentiviral targeting was done on E9.25–9.75 day embryos, pregnant mice were imaged and the embryos precisely staged one day before the surgery using Vevo 1100 animal ultrasound imager with a MS550D transducer (Visualsonics). During the surgery, titrated lentivirus (1 μl) was injected into amniotic cavity on the ventral side of the embryo, using a glass capillary needle fitted on Celltram Vario micro injector (Eppendorf).

Lentiviral constructs

H2B-RFP and patient-derived lesion ORF overexpression was achieved using pLX EF1 Barcode vector described previously (Ying et al., 2018). The pLX EF1 Barcode vector was modified from pLX302 (Addgene plasmid 25896), in which we replaced: i) PGK-puroR cassette between PpuM1/Xba1 sites with annealed oligos containing randomized 10bp barcode (tcagcnnnnnnnnnnc, ctaggnnnnnnnnnngc), ii) CMV promoter between Xho1/Nde1 sites with an EF1α promoter from pEF-BOS (Addgene plasmid 21924). ORFs encoding genetic lesions and H2B-RFP were subcloned into entry vectors and recombined into pLX EF1 Barcode vectors using LR clonase (Invitrogen). Barcode identity of each pLX EF1 vector encoding individual patient-derived genetic lesion was determined using capillary sequencing before and after recombination. Barcodes of LV H2B-RFP pool were determined using Illumina sequencing of pre-amplified barcode regions. RNAi-mediated gene depletion was achieved using pLKO1 shRNA vectors from the mouse TRC1.0 shRNA library (Sigma-Aldrich). To construct the lentiviral pools and ensure equal lentivirus representation, plasmids were mixed together in equimolar ratios. Gene ontology analysis of genes targeted by the lentiviral pool was performed using PANTHER program (Mi et al., 2010). Cre-shRNA expression was achieved using pLKO-Cre vectors described previously (Beronja et al., 2013; Beronja et al., 2010). sgRNA targeting Tsc22d1 was designed using CRISPRko (Sanson et al., 2018) and sub-cloned into lentiCRISPR v2 backbone (Addgene plasmid 52961) (Sanjana et al., 2014). Indel analysis was performed using Synthego ICE tools (Hsiau et al., 2019).

Barcode lineage tracing and in vivo genetic screens

For barcode lineage tracing analysis, barcoded H2B-RFP lentivirus pool infected K14-Cre; R26^yfp/yfp (WT) mammary fat pads were minced and digested in 0.25% collagenase plus 2mg/ml dispase at 37 C° for 1 hour and 0.25% Trypsin-EDTA at 37 C° for 0.5 hour to release epithelial cells. YFP⁺ RFP⁺ cell population (Figure 2B) was isolated using BD FACSAria II machine (BD Biosciences). To isolate barcoded H2B-RFP lentivirus pool infected myoepithelial and luminal cells of K14-Cre; R26^yfp/yfp (WT) and K14-Cre; R26^yfp/yfp; Pik3ca^H1047R (Pik3ca^H1047R) glands, mammary epithelial cells were stained with CD24 (30-F1, Biolegend) and CD29 (HMβ1–1, Biolegend) and gated out of YFP⁺ RFP⁺ cell population (Figure S3A). For Pik3ca^H1047R screen, mammary tumors were harvested when they reached 1.5 cm diameter. Further clonal sequencing of multi-barcode tumors was done on 2 mm diameter punch biopsy cut from 30 μm thick tissue sections. gDNA from all samples was extracted using DNeasy Blood & Tissue Kit (Qiagen). Barcode pre-amplification, sequencing, and data normalization were performed as previously described (Beronja et al., 2013; Schramek et al., 2014) and illustrated in Figures 2C, 4F and S3C. Briefly, a 30-cycle pre-amplification (Phusion, NEB) of the barcode region generated PCR products that were made into sequencing libraries using NEBNext Ultra II DNA Library Prep Kit (NEB). Sequencing of barcode libraries was performed on Illumina HiSeq 2000 using paired-end 50bp read lengths at the Fred Hutchinson Cancer Research Center Genomic Core facility. Short reads were trimmed down to barcode region by FASTX-Toolkit (Hannon, 2010) and mapped to predetermined barcode list using BWA (Li and Durbin, 2009). Liner normalization of barcode counts from samples were performed against T=0 counts to correct for barcode composition and amplification bias. All barcode lineage tracing analysis were performed in 6 sets of 5 mammary glands from 3 animals.

Whole mount 3-dimensional imaging and analysis

Embryonic and adult mammary tissues were stained with Alexa Fluor 647 Phalloidin (A22287, Life Technologies; 4U/ml) and cleared using previously described method (Rios et al., 2014). Mammary tissue was mounted in 80% glycerol and imaged for fluorescent proteins and Phalloidin using Zeiss LSM700 confocal microscope with 1.5μm increment on Z axis. Maximum intensity projection of Z-stack and tile scan images were further processed using Imaris (Bitplane) and ImageJ software (Schindelin et al., 2012). Spatial KS test of infected cells were performed using spatstat package (Baddeley et al.) in R (R-Core-Team, 2018). Extrapolation of the number of cells that received more than two viruses were done using method described before (Beronja et al., 2013; Beronja et al., 2010). All whole mount imaging was performed in mammary glands from at least 3 animals; number of replicates is indicated in Figure legends.

Immunofluorescence assays and western blot

The following primary antibodies were used: Chicken anti-GFP (ab13970, 1:1000 (IF); Abcam); Rat anti-Ecadherin (ECCD-2, 1:500 (IF); Thermofisher), and anti-K8 (TROMA-I, 1:200 (IF); DSHB); Rabbit anti-SMA (ab5694, 1:200 (IF); Abcam), anti-ER (E115, 1:200 (IF), 1:500 (WB); Abcam), anti-PR (D8Q2J, 1:600 (IF); Cell Signaling), anti-HER2 (29D8, 1:400 (IF); Cell Signaling), anti-Tsc22d1 (A303–582A, 1:200 (IF), 1:400 (WB); Bethyl), and anti-p53 (CM5, 1:200; Leica); Guinea pig anti-K14 (BP5009, 1:1000 (IF); Origene); Mouse anti-Pten (CPTC-PTEN-1, 1:200 (IF); DSHB), and anti-Cdkn2a (2D9A12, 1:100 (IF); Abcam). Tissue slides were processed for immunostaining as previously described (Beronja et al., 2013; Beronja et al., 2010), and mounted in ProLong Gold antifade reagent with or without DAPI (Life Technologies). Confocal images were taken on a Zeiss LSM700 system. Images were processed using Zen (Zeiss) and ImageJ software (Schindelin et al., 2012). Western blot was performed using iBlot system (Invitrogen) and imaged using Odyssey Fc (LI-COR).

EdU incorporation and apoptosis assays

5mg/kg of EdU was administered via intraperitoneal injection into animals 4 hours before processing. EdU staining was performed on OCT embedded slides using Click-iT™ EdU Imaging Kit (Invitrogen). Activated-caspase3 was detected with rabbit anti-activated caspase3 antibody (AF835, 1:1000; R&D systems). TUNEL assay was performed using TUNEL BrdU-Red Kit (Abcam). For data analysis we took more than 10 fields of images from mammary glands harvested from each animal and combined the counts from all images of a single animal as a replicate for statistical analysis (n=4 animals of each condition).

RNA-seq and data analyses

Tumor tissues were snap-frozen in LN2, mechanically pulverized, and lysed in Trizol (Invitrogen). We extracted RNA using phenol/chloroform protocol, and purified it using RNeasy Mini Kit (Qiagen), per manufacturer instructions. RNA quality was assessed using Agilent 2100 Bioanalyzer, with all samples passing the quality threshold of RNA integrity numbers (RIN)>8. Library was prepared using NEBNext Ultra RNA Library Prep Kit for Illumina (NEB), and cDNA libraries were sequenced on Illumina HiSeq 2000 using paired-end 50bp read lengths at the Fred Hutchinson Cancer Research Center Genomic Core facility. Reads were mapped to mm10 build of the mouse genome using TopHat2 (Kim et al., 2013), and transcript assembly and differential expression were determined using Cufflinks (Trapnell et al., 2010). PCA analysis and visualization of PAM50 gene set were done using Minitab software, FPKM values were rank transformed to normalize data from human and mouse samples. Gene Set Enrichment Analysis (Subramanian et al., 2005) was performed using GSEA2. Reference gene expression signatures were obtained from previously published expression profile of Erbb2 transgenic mouse model of mammary tumors (Zhu et al., 2011) and Pik3ca^H1047R K5/K8-CreER mammary tumors (Van Keymeulen et al., 2015).

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistics and Reproducibility

All quantitative data were collected from experiments performed in at least a triplicate, and expressed as mean ± s.d. One and two-tailed student t-test, Chi-square test, KS-test and liner regression were performed using Prism 6 (GraphPad software). A P-value of <0.05 was considered statistically significant. No randomization or blinding methods were used in the experiments and no data points were excluded from the analyses. The investigators were not blinded to allocation during experiments and outcome assessment.

DATA AND CODE AVAILABILITY

Data Deposition and Access

RNA–seq data that support the findings of this study have been deposited in the Gene Expression Omnibus (GEO) under accession code GSE122658. Previously published microarray data that were re-analyzed here are available under accession code GSE23938 and GSE69290.

KEY RESOURCES TABLE.

Reagent or Recourse	Source	Identifier
Antibodies
Chicken anti-GFP	Abcam	ab13970; RRID:AB_300798
Rat anti-Ecadherin	Thermofisher	Clone ECCD-2; 13-1900; RRID:AB_2533005
Rat anti-K8	DSHB	Clone TROMA-I; RRID:AB_531826
Rabbit anti-SMA	Abcam	ab5694; RRID:AB_2223021
Rabbit anti-ER	Abcam	Clone E115; ab32063; RRID:AB_732249
Rabbit anti-PR	Cell Signaling	Clone D8Q2J; 8757; RRID:AB_2797144
Rabbit anti-HER2	Cell Signaling	Clone 29D8; 2165; RRID:AB_10692490
Rabbit anti-Tsc22d1	Bethyl	A303-582A; RRID:AB_11125734
Rabbit anti-p53	Leica	P53-CM5P; RRID:AB_2744683
Guinea pig anti-K14	Acris	BP5009; RRID:AB_979615
Mouse anti-Pten	DSHB	Clone CPTC-PTEN-1; RRID:AB_2617320
Mouse anti-Cdkn2a	Abcam	Clone 2D9A12; ab54210; RRID:AB_881819
APC Rat anti-CD24	Biolegend	Clone 30-F1; 138505; RRID:AB_2565650
PE/Cy7 Hamster anti-CD29 antibody	Biolegend	Clone HMβ1-1; 102222; RRID:AB_528790
Rabbit anti-activated caspase3	R&D systems	AF835; RRID:AB_2243952
Chemicals and Enzymes
Alexa Fluor 647 Phalloidin	Life Technologies	A22287; RRID:AB_2620155
Collagenase, Type1	Gibco	17100017
Dispase II	Gibco	17105041
5-Ethynyl-2’deoxyuridine (EdU)	Sigma-Aldrich	900584
Lapatinib	ApexBio	A8218
Nutlin 3a	Tocris Bioscience	6075
Cell lines
293TN	System Biosciences	LV900A-1
SK-BR-3	ATCC	HTB-30
Plasmids
pMD2.G	Addgene	12259
psPAX2	Addgene	12260
pLX302	Addgene	25896
pEF-BOS	Addgene	21924
lentiCRISPR v2	Addgene	52961
Critical Commercial Assays
DNeasy Blood & Tissue	Qiagen	69504
Kit
RNeasy Mini Kit	Qiagen	74104
Phusion HF DNA Polymerase	NEB	M0530L
Click-iT™ EdU Imaging Kit	Thermofisher	C10338
TUNEL BrdU-Red Kit	Abcam	ab66110
NEBNext Ultra II DNA Library Prep Kit	NEB	E7645L
NEBNext Ultra RNA Library Prep Kit	NEB	E7530L
Experimental Models: Organisms/Strains
Pik3caH1047R/H1047R	donated by Wayne A. Phillips	NA
K14-Cre	Jackson Laboratories	004782; RRID:IMSR_JAX:004782
R26yfp/yfp	Jackson Laboratories	006148; RRID:IMSR_JAX:006148
R26mTmG/mTmG	Jackson Laboratories	007676; RRID:IMSR_JAX:007676
Deposited Data
RNA-seq	This study	GEO: GSE122658
Software and Algorithms
R	(R-Core-Team, 2018)	R-Core-Team
PANTHER	(Mi et al., 2010)	http://pantherdb.org/
CRISPRko	(Sanson et al., 2018)	https://portals.broadinstitute.org/gpp/public/analysis-tools/sgrna-design
Synthego ICE tools	(Hsiau et al., 2019)	https://www.synthego.com/products/bioinformatics/crispr-analysis
Imaris	Bitplane	https://imaris.oxinst.com/
ImageJ	(Schindelin et al., 2012)	https://imagej.nih.gov/ij/
Spatstat package	(Baddeley et al.)	https://spatstat.org/
BWA	(Li and Durbin, 2009)	http://bio-bwa.sourceforge.net/
FASTX-Toolkit	Hannon, G.J. Lab	http://hannonlab.cshl.edu/fastx_toolkit/
TopHat2	(Kim et al., 2013)	https://ccb.jhu.edu/software/tophat/index.shtml
Cufflinks	(Trapnell et al., 2010)	http://cole-trapnell-lab.github.io/cufflinks/install/
Minitab	Minitab	https://www.minitab.com
GSEA2	(Subramanian et al., 2005)	http://software.broadinstitute.org/gsea/index.jsp
Prism 6	GraphPad software	https://www.graphpad.com/

Highlights.

• Intra-amniotic lentiviral injection transduces early mammary progenitors.

• ~120 bipotent progenitors with equal long-term growth maintain adult mammary gland.

• Large-scale targeting of mammary progenitors can identify drivers of breast cancer.

• Tsc22d1, lost in 2% of breast cancers, promotes tumor initiation and subtype shift.