Patient-derived xenografts of triple-negative breast cancer reproduce molecular features of patient tumors and respond to mTOR inhibition

Both the Stanford University Research Compliance Office’s Human Subjects Research and IRB Panel and Stanford’s Administrative Panel on Laboratory Animal Care (APLAC) approved this study. After obtaining informed written patient consent, breast cancer tissues were obtained fresh from operating rooms at Stanford Hospital and Clinics. In six cases of TNBC (SUTI097, SUTI103, SUTI110, SUTI151, SUTI319, SUTI368), fresh tumor tissue was sterilely obtained from primary breast cancer tissue that was undergoing surgical excision, and in one case (SUTI151M), the tumor tissue was taken fresh from a soft tissue TNBC metastasis to the quadriceps muscle in the thigh that was undergoing biopsy (SUTI151M is from the same patient who had months earlier donated a piece of her primary breast tumor SUTI151). Portions were frozen or placed in formalin and embedded in paraffin for later analyses. Fresh tumor tissue was kept on ice in RPMI 1640 medium supplemented with penicillin/streptomycin and 10% heat inactivated FBS (Invitrogen-Life Technologies, Carlsbad, CA, USA) for transport, minced into one to two millimeter fragments, then sterilely and orthotopically transplanted into the number two mammary fat pads of 5 to 10 female NOD SCID mice (NOD.CB17-Prkdc ^scid /J, Jackson Laboratory West, Sacramento, CA, USA). Briefly, the mice were anesthetized by inhalation of 1 to 3% isoflurane, their hair was clipped, and their skin sterilized with povidone-iodine and alcohol. A small skin incision was made in the lateral flank and minced tumor chunks were mixed with LDEV-free Matrigel (BD Biosciences, San Jose, CA, USA) and implanted into the mammary fat pad by trochar insertion. The incision site was closed with Vetbond tissue adhesive (3 M, St. Paul, MN, USA). Mice were maintained in pathogen-free animal housing. The established xenografts were subsequently passaged from mouse to mouse to expand xenograft numbers; xenograft tumors were also stored frozen in FBS containing 10% dimethyl sulfoxide (DMSO, EMD Chemicals Inc., Billerica, MA, USA) solution for future engraftment. Xenograft tumor tissue was frozen on dry ice for RNA isolation and microarray analysis and for subsequent protein analyses. Tumor fragments were also fixed in phosphate buffered saline with 10% formalin (Sigma-Aldrich, St. Louis, MO, USA) for histological studies. All animal care was performed in accordance with Stanford University and IACUC guidelines.

Formalin-fixed, paraffin-embedded tissue sections of patient or xenograft tumors were cut into 4 μm sections, deparaffinized in xylene, rinsed in ethanol and rehydrated. Staining was performed using the Ventana XT platform and internal antigen retrieval CC1 standard. The antibodies used were rabbit monoclonal antibodies for ERα (clone SP1, 1:25 dilution, Thermo Scientific, Fremont, CA, USA) and PR (clone 1E2, ready to use, Roche-Ventana Medical Systems, Inc., Tucson, AZ, USA). The universal secondary protocol and the DAB MAP kit (Ventana Medical Systems, Inc., Tucson, AZ, USA) were used to detect and amplify the signal. Both biomarkers were scored using a three-tier system: 0 = negative, 1 = weak, and 2 = strong, respectively defined as <1%, 1% to 50%, and ≥50% of tumor cell nuclei staining positively. HER2 protein expression was performed and interpreted using the Ventana PATHWAY HER2 antibody (rabbit monoclonal, clone 4B5; Ventana, Tucson, AZ, USA). The Food and Drug Administration-approved Ventana PATHWAY is scored from 0 to 3+. Staining in <10% of tumor cells is scored as showing no overexpression (0 or 1+). Strong, complete, circumferential membrane staining in >30% of tumor cells is considered overexpression and is designated as strong positive (3+). Strong circumferential membrane staining in <30% of tumor cells, or circumferential but less than strong staining in any proportion of tumor cells, is designated as equivocal (2+). All immunohistochemical assays were conducted in parallel with known positive and negative controls. The slides were observed using a Nikon Eclipse 80i microscope (Nikon Instruments Inc., Melville, NY, USA). Pictures were taken using a Nikon Digital Camera DXM1200F and images were obtained using Nikon ACT-1 software.

Genomic DNA was extracted from patient or xenograft tumor samples using DNeasy Blood & Tissue Kit (Qiagen, Valencia, CA, USA). Array CGH analyses were performed at SciGene (Sunnyvale, CA, USA) using Human Genome CGH Microarray 4x44K (Agilent Technologies, Santa Clara, CA, USA), and processed on SciGene's robotic aCGH workstations (ArrayPrep® Target Preparation System, Mai Tai® Hybridization System, and Little Dipper® Processor, SciGene, Sunnyvale, CA, USA).

Mutations were detected by sequencing PCR products derived from amplification primers in the introns flanking PIK3CA exons 1, 2, 3, 5, 6, 7, 9, 18 and 20 using Ampli Taq Gold DNA polymerase (Applied Biosystems-Life Technologies, Carlsbad, CA, USA). The primer sets used in these reactions are listed in Table S1 in Additional file 1. Exons that had sequence homology with a known PIK3CA pseudogene were not sequenced; however, the sequenced exons included all common mutation hotspots. The reaction was run using a touchdown PCR protocol where the annealing temperature was started at 63°C and decreased for 0.5°C per cycle for 12 cycles. Then the reaction was continued for another 25 cycles at 94°C, 30 sec; 58°C, 30 sec; and 72°C, 30 sec per cycle. PCR products were checked by 2% agarose gel against a GeneRuler 50 bp DNA Ladder (Frementas, Glen Burnie, MD, USA) and sequenced by BigDye Terminator v3.0 Cycle Sequencing Kits (Applied Biosystems-Life Technologies, Carlsbad, CA, USA). The sequencing results were analyzed with Sequencher 4.8 software (Gene Codes Corporation, Ann Arbor, MI, USA).

Frozen tumor tissues from xenografts were cut into small pieces on dry ice. RNA was extracted using RNeasy Plus Mini Kit (Qiagen) following the manufacturer’s instructions. The quantity and purity of the RNA sample was measured using the Agilent 2100 bioanalyzer (Agilent Technologies). RNA samples were submitted to Stanford Protein and Nucleic Acid core facility for microarray analysis using Affymetrix GeneChip Human Genome U133 Plus 2.0 arrays. All xenograft microarray datasets are posted on GEO under accession number GSE47079 [47].

TNBC subtyping was done following the Pietenpol group’s methods [7, 8]. The microarray data of the patient derived xenograft tumors were robust multi-array average (RMA) normalized and log transformed. For genes containing multiple probes, the probe with the largest interquartile range across the samples was chosen to represent the gene. The processed samples were uploaded to the website [48], where the samples were subjected to an ER-filter scrutiny and then assigned TNBC subtypes.

A rapamycin response signature was generated as described previously [49] using Connectivity Map Build 02 gene expression data of 5 rapamycin treated and 18 control MCF7 cell samples [50]. The Connectivity Map studies highlight the ability to use drug treatment on cell lines to identify a set of genes that reflect response to a drug; therefore, independent of the cell type profiled, a “signature” of drug response can be identified from the data that reflects the response to drug treatment [50]. Multiple studies by other groups and our own have shown that such drug response signatures do also function as drug sensitivity/resistance signatures, with sensitive samples having gene expression patterns more like untreated cells and resistant samples having gene expression patterns more like treated cells [49‐52]. Specifically, tumors with dysregulation of genes that are modulated by treatment of rapamycin will be predicted as “sensitive” or “resistant” based on their correlation to gene dysregulation from rapamycin treatment in the Connectivity Map data. This approach, which uses expression data from cell lines before and after treatment with a drug, has the advantage over using expression data from cell lines classified as ‘resistant’ or ‘sensitive’ because cell line data have confounding factors, such as subtype that can affect prediction models. By using prediction models that include genes specific to a particular drug’s response, we are not limited by these confounding factors. Thus, treated cell lines of one subtype (for example, luminal MCF7 cells) may be used to predict drug response in samples of different subtypes (for example, basal-like or HER2-overexpressing cancer cells).

To generate the signature, Mas5 normalized gene expression data were quantile-normalized and log2 transformed and used as the training set. A Bayesian binary regression algorithm was then used on the training set to generate the signature. It was further optimized and internally validated in leave-one-out cross-validation (LOOCV) analysis, which tests each individual sample’s classification by leaving it out of the model and predicting if it is treated or untreated [49, 51].

For signature external validation, CEL files were downloaded from GEO GSE18571, which contains gene expression data for both in vitro and in vivo rapamycin treatment samples on TNBC cell line MDA-MB-468 [26], and from Connectivity Map batches 2, 35, 44, 56, 63, 70, 626, 757 and 767, and analyzed as described by Cohen et al. [49]. We then confirmed the ability of the rapamycin response signature to predict sensitivity to rapamycin in vitro by comparing the EC50 (see below) for a diverse panel of cell lines to the predicted sensitivity, that is, similarity to untreated cells, based on gene expression.

As previously described [49], 18 breast cancer cell lines were obtained from ATCC (HCC38, HCC1806, HCC1428, HCC1143, BT483, BT549, BT474, MDA-MB-361, MDA-MB-157, MDA-MB-435S (now considered to be of melanoma origin), MDA-MB-231, MDA-MB-453, SKBR3, ZR75, CAMA I, MCF7, Hs578t, T47D) and used for dose-response assays. Cells were seeded in 384-well plates (Nunc, Rochester, NY, USA) in MEBM media (Lonza, Walkersville, MD, USA) containing 5% fetal bovine serum (Gibco, Grand Island, NY, USA), at a density to yield 80% confluency in control-treated wells at 96 h post-treatment (as determined by growth curves). After 24 h, rapamycin was added at 10 doses of 0, 0.1 pM, 0.3 pM, 1 pM, 3 pM, 10 pM, 30 pM, 100 pM, 300 pM and 1 nM. A BIOMEK 3000 (Beckman Coulter, Indianapolis, IN, USA) robot was used to seed the cells and dispense the drug. After 96 h, CellTiter-Blue Reagent (Promega, Madison, WI, USA) was added to test cell viability. After 2 h of incubation at 37°C, the fluorescence was recorded (560(20)_Ex/590(10)_Em) using a Victor3V 1420 Multilabel Counter (Perkin-Elmer, Waltham, MA, USA) plate reader. After subtracting background fluorescence, EC50 was calculated using GraphPad Prism v5 (GraphPad Software, La Jolla, CA, USA) to fit a constrained sigmoidal dose-response curve. Predicted sensitivity for these cell lines was computed using gene expression data from Cohen et al. [49] and applying the Bayesian binary regression model. Detailed methods for running the regression model are given in [53]. Predicted sensitivity was compared to actual EC50 using linear regression.

To examine the relationship between intrinsic subtype and rapamycin sensitivity, 1,401 breast cancer samples from eight microarray studies were then analyzed (Table S2 in Additional file 2, duplicate samples were removed from GEO datasets GSE6532, GSE7390 and GSE3494). Intrinsic breast cancer subtypes were assigned as previously described [49, 54, 55]. Rapamycin sensitivity of the patient breast cancer sample was calculated as described by Cohen et al. [49]. A detailed method, the input files, output files and the logistic regression program used in this study are available in [53].

Rapamycin and CCI-779 (LC Laboratories, Woburn, MA, USA) were stored as 50 mg/ml solutions in 100% ethanol at -80°C. The stored solutions were diluted in PBS containing 4% ethanol, 5% polyethylene glycol 400 and 5% Tween 80 for treatment. When tumor xenografts grew to an average between 50 to 100 mm³ in tumor volume, mice were stratified and randomized by tumor volume into treatment groups of 5 to 10 mice each. The treatment groups included: 1) rapamycin (sirolimus) group, receiving intraperitoneal (IP) administration of 7.5 mg/kg rapamycin every other day for up to six weeks; 2) CCI-779 (temsirolimus) group, receiving IP administration of 7.5 mg/kg of CCI-779 every other day for up to four weeks; 3) control group, receiving IP administration of control vehicle for up to four weeks; and, in some sets of experiments, 4) a doxorubicin group, receiving IP administration of 2 mg/kg doxorubicin (Sigma-Aldrich, St. Louis, MO, USA) diluted in PBS once every week for three weeks. Tumors were measured twice a week with a caliper in two dimensions. Tumor volume was calculated by the following formula: tumor volume = (l x w²)/2, where l is the longest diameter of the tumor, w is the shortest diameter of the tumor. Mean tumor volumes were calculated, and growth curves were established as a function of time. The error bars indicated the value of the standard error of the mean. The Student’s t-test was used for statistical analysis. We considered P <0.05 as statistically significant.

For protein extraction, frozen xenograft tumors were gently thawed and washed in ice-cold PBS. They were then homogenized using a glass homogenizer and lysed in radio-immunoprecipitation assay buffer containing protease and complete phosphatase inhibitors (Roche, West Sussex, UK). After quantification using Pierce’s BCA protein assay (Thermo Scientific Pierce, Leicestershire, UK), 5 to 10 μg total proteins were run on 4 to 12% SDS-PAGE gels (NuPAGE Bis-Tris Gels, Invitrogen-Life Technologies, Paisley, UK), then immunoblotted with the antibodies for PTEN, mTOR, p-mTOR (Ser 2448), 4EBP1, p-4EBP1 (Ser 65), S6K1, p-S6K1 (Thr 389), eIF4E and p-eIF4E (Ser 209) at 1:1,000 dilutions. Secondary horseradish peroxidase (HRP)-labeled antibodies were used at 1:5,000 dilutions. Tubulin was used as a loading control. All antibodies were obtained from Cell Signaling Technology (Hitchin, Hertfordshire, UK). Membranes were visualized by the ECL developer system (GE Healthcare Life Sciences, Piscataway, NJ, USA). Protein expression was quantified by analyzing a representative autoradiograph with Image Image J software (public domain software developed at the Research Services Branch, National Institute of Mental Health, Bethesda, MD, USA) [56].

We developed a panel of seven TNBC xenograft tumors from six patients, generated from six primary tumors (SUTI097, SUTI103, SUTI110, SUTI151, SUTI319, SUTI368) and one soft tissue metastasis in the quadriceps muscle of the leg (SUTI151M). In general, these xenografts represented aggressive TNBC. Most patients from whom the xenografts were derived had poor disease-free survival despite treatment with multiple standard therapies, with four of five patients who had follow-up greater than one year showing metastatic recurrence or death from disease (Table S3 in Additional file 3).

We compared the xenograft tumors to the corresponding patient tumors by histology, clinical biomarker expression, genome-wide array CGH, and PIK3CA sequencing. Based on hematoxylin and eosin (H&E) staining (Figure 1A, B), the original TNBC tumors exhibited a variety of histologies that were conserved in the corresponding xenografts. Specifically, there were similarities in cancer cell morphology, mitotic index, stromal abundance and percent necrosis. As expected, all xenografts were confirmed to be triple-negative by ER, PR and HER2 staining using the same clinical laboratory protocols as were performed on the patient samples (Figure 1C-E).

To compare global genomic profiles between patient tumors and their corresponding xenografts, we performed array comparative genomic hybridization (aCGH) on two tumors, SUTI110 and SUTI151. The xenograft tumors faithfully maintained the genomic DNA alterations observed in the corresponding patient tumors (Figure 2 and Figure S1 in Additional file 4). Interestingly, a previously unreported 5q11-12 deletion was observed in both patient tumor SUTI151 and its corresponding xenograft (Figure 2), and was also maintained in the xenograft of the soft tissue metastasis, SUTI151M, that developed months later (data not shown).

PIK3CA sequencing showed that two of the seven (29%) patient primary and metastatic tumors and xenografts contained missense mutations (Figure 3A, B). SUTI097 patient and xenograft samples contained an exon 6 mutation (1173 A > G, I391M); SUTI110 patient and xenograft samples contained an exon 20 mutation (3302 G > C, G1049R). We did not observe any exon 9 mutations, which are common mutation sites in ER positive tumors.

We also detected multiple SNPs in introns flanking the sequenced PIK3CA exons, and these sequence variations were also maintained between all primary tumors and their corresponding xenografts (SUTI097, SUTI103, SUTI319, SUTI368, Figure 3C). In the one soft tissue metastasis, however, the metastatic tumor showed an additional SNP not present in the primary tumor or in the xenografts generated from the patient’s primary or metastatic tumors (SUTI151 and SUTI151M). When sequence variations were analyzed between multiple xenograft tumor passages for primary tumor SUTI319, passage 3 of its xenograft tumor and passage 5 of its xenograft tumor, both before and after rapamycin treatment, there was conservation of three SNPs identified in introns between exons 5 and 6, and between exons 6 and 7 (Figure 3C). In summary, our xenografts recapitulated the histology, biomarker status, genomic profile and PIK3CA sequence of corresponding primary patient tumors.

Human TNBCs have been shown to be heterogeneous, comprised of at least six stable subtypes and a possible seventh unstable subtype [7]. To determine the subtypes of our panel of TNBC xenografts, we performed microarray analyses on the xenograft tumors and classified them according to the method developed by Pietenpol’s group [7, 8]. As shown in Table 1, of our panel of seven TNBC xenograft tumors, five xenografts subclassified into four of the six stable subtypes; two were classified with Pietenpol’s “unstable” group. In particular, SUTI097 belonged to the immunomodulatory (IM) subtype; SUTI103 and SUTI110 were classified as basal-like 1 (BL1); SUTI151 was classified as basal-like 2 (BL2); SUTI151M, the soft tissue metastasis of SUTI151, was identified as mesenchymal (M) subtype; and SUTI319 and SUTI368 clustered with the “unstable” group of TNBCs. We find it anecdotally fitting that the metastasis of a BL2 subtype subclassified as M subtype, which is associated with increased expression of genes involved in cell motility, cellular differentiation, growth pathways and TGF-β signaling [7]. Our xenograft panel thus represents a majority of TNBC subtypes, making it suitable for pre-clinical drug testing.

To explore how commonly rapamycin sensitivity is expected to occur among diverse breast cancers, including TNBC, we developed a rapamycin response signature that would predict sensitivity or resistance based on a cancer’s gene expression. This builds on our previous work with gene expression-based signatures, which were derived using the same approach, and which showed that drug response signatures predict whether a tumor or cell line will be sensitive or resistant to a drug [49, 51, 52]. The rapamycin response signature contained 200 probe sets and represented 175 unique genes (See Table S4 in Additional file 5). Figure 4A shows the heatmap view of the expression pattern of the probe sets from the signature in the training set samples from the Connectivity Map database [50] displaying expression changes in response to rapamycin treatment. When analyzed by leave-one-out cross validation, a method that leaves a single sample out to predict whether it will be classified as treated (probability >0.5 on the y-axis) or untreated (probability <0.5 on the y-axis), 22 of 23 training set samples showed the expected prediction, including all five treated cell line samples (Figure 4B), indicating high consistency and robustness of our signature.

To validate our rapamycin response signature in TNBC, we next tested its accuracy on an independent external dataset containing both in vitro and in vivo rapamycin treatment samples that were generated from the TNBC cell line MDA-MB-468 [26]. For both the in vitro MDA-MB-468 cells and the 22-day xenografts, the treated samples were all correctly predicted to be more like the treated samples in the signature training set, and hence more resistant to further rapamycin treatment, than the untreated MDA-MB-468 cells (P = 0.0017) and xenografts (P = 0.06) (Figure S2A in Additional file 6). These data confirm that the rapamycin response signature can distinguish TNBC cells that have been treated with rapamycin from untreated TNBC cells.

We also tested the rapamycin response signature on other samples in the Connectivity Map that had available treatment data on rapamycin as well as other drugs. As shown in Figure S2B in Additional file 6, only samples treated with rapamycin or PI3K inhibitors showed the expected rapamycin response signature post-treatment pattern whereas the samples treated with random drugs did not (P <0.0001).

Finally, to show that the response signature does predict sensitivity and resistance to rapamycin, we compared predictions of the rapamycin response signature in a panel of 18 breast cancer cell lines to the actual EC50 obtained when these cells were treated with rapamycin. As shown in Figure S3 in Additional file 6, we found a significant correlation between signature prediction and in vitro drug sensitivity as measured by EC50 (r = -0.3; P = 0.02). Therefore, the response signature was confirmed to also be a predictor of sensitivity to rapamycin.

We then used the rapamycin response signature to perform a supervised analysis of eight published gene expression datasets including 1,401 breast cancer samples from patients that were classified into the five intrinsic molecular subtypes [54]. As shown in Figure 4C, D, 94% of BLBCs, 68% of HER2-overexpressing tumors, 18% of luminal A tumors, 58% of luminal B tumors and 37% of normal-like tumors were predicted to be sensitive to rapamycin. The predicted rapamycin sensitivity differed between subtypes, with the more aggressive subtypes - such as basal-like, HER2-overexpressing and luminal B - having much more frequently predicted sensitivities. Among them, BLBC had the highest frequency of rapamycin-sensitive tumors. Signatures for other PI3K pathway inhibitors showed similar patterns of predicted drug response among the different subtypes (data not shown).

We next used our panel of seven TNBC xenografts to evaluate rapamycin sensitivity in vivo, measuring growth inhibition of two mTOR inhibitors, rapamycin (sirolimus) and/or its analog, CCI-779 (temsirolimus) and doxorubicin, a drug widely used to treat TNBC. Compared to untreated tumors, growth rates of doxorubicin-treated tumors in all seven orthotopic xenograft models showed only a minimum to partial response, with growth inhibition ranging from 2 to 52% (Figure 5 and Table S3 in Additional file 3). In sharp contrast, the overall efficacy of rapamycin and CCI-779 was significantly higher than doxorubicin for the treatment of these TNBC xenografts (P = 0.0003, Figure 5). On average, rapamycin inhibited tumor growth by 94% (range 77 to 99%), whereas the average inhibition by doxorubicin was only 36% (Figure 5, and Table S3 in Additional file 3). This supports the hypothesis that TNBCs are highly sensitive to rapamycin. For the four xenografts treated with both mTOR inhibitors, drug efficacy was similar. However, none of the tumors disappeared completely, with most maintaining a small volume of residual tumor, suggesting that additional drugs may be necessary in combination with mTOR inhibitors to totally ablate residual disease.

To identify pathway activation, we performed Western blot analyses on PTEN and other mTOR pathway proteins. As shown in Figure 6, the protein levels of PTEN vary among the xenografts. SUTI097, SUTI110, SUTI319 and SUTI368 have relatively higher PTEN protein levels than xenografts SUTI103, SUTI151, SUTI151M, although all samples exhibited in vivo sensitivity to mTOR inhibitors (Figure 5). The PTEN levels generally remained consistent pre- and post-treatment and its expression did not appear to be associated with any particular TNBC subtype.

Phosphorylated mTOR and its downstream proteins - 4EBP1, S6K1 and eIF4E - were detected in all xenograft samples (Figure 6), demonstrating baseline mTOR pathway activation. Treatment with one or both mTOR inhibitors decreased phosphorylation of mTOR and, to varying extents, its downstream proteins. The exception was metastatic tumor xenograft SUTI151M (Figure 6), although both treatments still inhibited its growth by over 90%, raising the possibility of other potential modes of action of mTOR inhibitors. For all primary tumor xenografts, the overall phosphorylation of mTOR, S6K1, 4EBP1 and eIF4E proteins was decreased by 53%, 33%, 62% and 64%, respectively, in rapamycin-treated tumors compared with tumors in the pretreatment and control groups (Table S5 in Additional file 7), supporting decreased mTOR pathway activity after treatment.