Ability to Generate Patient-Derived Breast Cancer Xenografts Is Enhanced in Chemoresistant Disease and Predicts Poor Patient Outcomes

Background Breast cancer patients who are resistant to neoadjuvant chemotherapy (NeoCT) have a poor prognosis. There is a pressing need to develop in vivo models of chemo resistant tumors to test novel therapeutics. We hypothesized that patient-derived breast cancer xenografts (BCXs) from chemo- naïve and chemotherapy-exposed tumors can provide high fidelity in vivo models for chemoresistant breast cancers. Methods Patient tumors and BCXs were characterized with short tandem repeat DNA fingerprinting, reverse phase protein arrays, molecular inversion probe arrays, and next generation sequencing. Results Forty-eight breast cancers (24 post-chemotherapy, 24 chemo-naïve) were implanted and 13 BCXs were established (27%). BCX engraftment was higher in TNBC compared to hormone-receptor positive cancer (53.8% vs. 15.6%, p = 0.02), in tumors from patients who received NeoCT (41.7% vs. 8.3%, p = 0.02), and in patients who had progressive disease on NeoCT (85.7% vs. 29.4%, p = 0.02). Twelve patients developed metastases after surgery; in five, BCXs developed before distant relapse. Patients whose tumors developed BCXs had a lower recurrence-free survival (p = 0.015) and overall survival (p<0.001). Genomic losses and gains could be detected in the BCX, and three models demonstrated a transformation to induce mouse tumors. However, overall, somatic mutation profiles including potential drivers were maintained upon implantation and serial passaging. One BCX model was cultured in vitro and re-implanted, maintaining its genomic profile. Conclusions BCXs can be established from clinically aggressive breast cancers, especially in TNBC patients with poor response to NeoCT. Future studies will determine the potential of in vivo models for identification of genotype-phenotype correlations and individualization of treatment.


Introduction
Neoadjuvant chemotherapy (NeoCT) is the standard of care for patients with locally advanced and inflammatory breast cancer, and it is increasingly being used in operable breast cancer both to decrease extent of surgery, and as an in vivo indicator of sensitivity to systemic therapy. Ten to 30% of patients will not respond to NeoCT, while 3% of patients have tumor progression while on therapy [1]. Patients with tumor progression during chemotherapy or with significant residual cancer burden after NeoCT are at high risk of relapse [1][2][3]. There is a pressing need to understand mechanisms of resistance and to identify novel therapeutics that are effective in patients who do not respond to standard NeoCT as well as in those with significant residual disease after NeoCT.
Triple negative breast cancers (TNBC) lack the expression of estrogen, progesterone and HER2 receptors and are not sensitive to estrogen and HER2-targeted therapies. Patients with TNBC have significantly higher pathologic complete response (pCR) rates with chemotherapy compared with non-TNBC [4] and if pCR is achieved, patients with TNBC and non-TNBC have similar survival. In contrast, TNBC patients with residual disease (RD) after chemotherapy have worse overall survival compared with non-TNBC. Therefore there is an urgent need to identify novel therapies for chemoresistant TNBC.
Currently, in vivo models available to test novel agents are limited. Cell lines may diverge from their tumors of origin during adaptation to in vitro growth conditions and when they are injected to form xenografts. Responses to treatments in cell lines often are not concordant with those observed in patients. Genetically engineered mice can demonstrate specific aspects of oncogenesis, but do not recapitulate the heterogeneity of clinical tumors and rely on the assumption that the same genetic alterations transform both mouse and human cells [5].
Recently there has been great interest in developing patient-derived xenografts (PDXs) from several tumor types including colon cancer, lung cancer, ovarian cancer, head and neck tumors, leukemia as well as breast cancer [6][7][8][9][10][11][12][13][14][15][16][17]. Hidalgo et al. described an in vivo pancreatic cancer patient-derived serial transplantation model and initially reported an engraftment rate of 80%, and a "take"-rate of passagable from the tumors of 93% [18][19][20]. On expansion of their series, they reported an engraftment rate of 61%, and that engraftment was greater in tumors derived from patients with worse prognosis [21]. Further, a pilot clinical trial suggested a potential role for preclinical testing in PDXs to individualize patient therapy [22].
We hypothesized that tumors from breast cancer patients can be used to establish patientderived breast cancer xenografts (BCXs) with high fidelity to their original tumors, providing in vivo models for TNBC and other chemoresistant tumors. Such a model would require relative stability in the characteristics of the tumor as it is passaged through several generations of immunodeficient mice. To test this hypothesis, we transplanted chemo-naïve breast tumors or residual breast cancers after NeoCT. We determined feasibility of generating passagable BCXs, from residual tumors after NeoCT, and the relative stability of the genomic profile of the patients' tumors with that of serial generations of BCXs.

Ethics statement
The research involving human participants was approved by the MD Anderson Institutional Review Board. All patients signed informed consent prior to participation (IRB protocol # LAB07-0950). All animal studies were approved by the MD Anderson Animal Care and Use Committee (ACUF protocol # 02-08-00833). This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health.

Patient population
Patients with histologically-confirmed invasive breast cancer were recruited for this study. All patients had greater than 1 cm tumor on physical exam or imaging prior to surgery. Fifty-five patients were consented and underwent breast-conserving surgery or mastectomy for local therapy with sentinel node biopsy with or without axillary lymph node dissection (ALND) for regional therapy; one patient had contralateral lymph node recurrence and had an ALND alone; one had bilateral breast cancer. Three patients underwent surgery in the presence of known distant metastases.
From the first fifty-five breast cancer patients enrolled in the study, eight patients were excluded based on lack of adequate tumor on gross pathology assessment, or hematoxylin and eosin (H&E) staining showed no viable tumor tissue in the donor specimen. Therefore, 48 viable tumors were implanted from 47 patients (one with bilateral breast cancer).

In vivo tumor implantation and histopathologic analysis
The tumor specimen was placed in sterile tissue-culture medium on ice and brought immediately to the animal facility. Tumor was implanted into five 6-week old female BALB/c nu/nu mice under isoflurane anesthesia, and all efforts were made to minimize suffering. Two~0.3 cm skin incisions were made with subcutaneous pockets on the left and right mid-back. One tumor piece (~1-3 mm) was inserted into each pocket and the skin was closed. When tumor diameter reached 1.5 cm, mice were euthanized, tumors were excised, cut into~1x1x1 mm fragments, and passaged to successive generations of 5 mice. At each passage, remaining tumor was snap frozen in liquid nitrogen and stored at -80°C. Samples of these tumors were fixed in formalin and embedded in paraffin blocks. All primary tumors and BCX samples were assessed by H&E staining for morphology. In selected samples immunohistochemistry for CD20, CD3, CD45, and cytokeratin was performed in the Department of Veterinary Medicine and Surgery and Department of Pathology.
In vitro growth of conditionally reprogrammed cells and reimplantation A 1.5 cm diameter BCX-010 xenograft was excised from the mouse and finely minced in Fmedia as described in Liu et al. [23]. Tumor fragments were digested for 4 hr at 37°C in Fmedia using collagenase/hyaluronidase (Stemcell Technologies; 1X) with occasional light mixing. The slurry was filtered, pelleted, washed, and plated in F-media supplemented with 10 μM Y-27632 (Enzo). The media was changed after 24 hours. Once the culture reached~70-90% confluence, the cells were split at a~1:4 ratio. After 4 passages, the cells were collected and 1x10 8 cells were subcutaneously injected with or without 50% Matrigel into the flank of athymic nude (MD Anderson) or CIEA NOG (Taconic) mice. STR fingerprinting was used to confirm the human/BCX-010 origin of the cultured cells and the single cell-based xenografts.

DNA isolation and DNA fingerprinting
Genomic DNA was isolated from frozen or FFPE tumor tissue using the QIAamp DNA Mini Kit (Qiagen). Short tandem repeat (STR) DNA fingerprinting was performed with the AmpFℓSTR Identifiler kit (Applied Biosystems) at the Characterized Cell Line Core Facility.

Fluorescence in situ hybridization
FISH was performed at the Molecular Cytogenetics Core Facility. The all mouse centromere probe (Kreatech, Netherlands) or the human centromere 16 probe (Abbott Molecular) was applied and slides were stained with DAPI.

Copy number and somatic mutation determination
Genomic DNA from five of the patients (1, 2, 4, 5, and 6) and corresponding P3 BCXs were submitted to Affymetrix for molecular inversion probe (MIP) assay with copy number and somatic mutation determination. The Affymetrix platform interrogated 300,008 single nucleotide polymorphisms and 412 somatic mutations in 45 genes. The MIP assay was performed as previously described [25].
Briefly library prep was performed using KAPA library prep kit (Kapa Biosystems, Inc), and equimolar amounts of DNA were pooled (8-12 samples per pool) for capture of 201 genes that are clinically relevant in cancer (S1 Table). The captured libraries were sequenced on a HiSeq 2000 (Illumina Inc.). In addition, whole exome sequencing was performed for normal tissue or blood, primary tumors and corresponding P0 and P1 for BCX-010, -017, -022, and -024. For data analysis, reads were aligned to human reference assembly hg19 using BWA software and duplicated reads were removed using samtools. Single nucleotide variants (SNVs) and small indels were called using VarScan2 [29].

Functional proteomics
Reverse phase protein array (RPPA) was performed as previously described [30] to compare the proteomic profile of P0 patient tumors and corresponding BCXs (P1, P2 and P3), in models BCX-001, 002, 004, 005, and 006. A logarithmic value reflecting the relative amount of each protein in each sample was generated. The median polish normalized RPPA data set consisted of 154 proteins.

Statistical analysis
Descriptive statistics were used to describe engraftment rates. BCX success rates in different tumor types and clinical scenarios were compared using Fisher's Exact 2-tailed analysis. Linear mixed effects (LME) model was used to assess the differences in passage time among passage numbers. The LME model included fixed effect of passage number (1 vs. 2 vs. 3) and random effect of model.
For RPPA, unsupervised clustering was applied to all tumor samples across all of the proteins, using Spearman Correlation and complete linkage. Paired t-test was performed to compare changes in relative protein amounts between, P0, P1, and P3. The p values were fitted to a-beta-uniform mixture-BUM model to calculate the corresponding false discovery rate (FDR) [31][32][33]. FDR <0.2 was considered significant [34]. To compare protein expression levels across multiple generations, a g-statistic was used [35]. R packages including GeneCycle, q value and ClassComparison were employed [34].

Ability to generate patient-derived breast cancer xenografts is enhanced in TNBC and in chemoresistant disease
Forty-eight viable tumors were implanted from 47 patients: one patient had a bilateral tumor implanted as separate models. Implants were followed for 6 months after implantation. The resultant BCXs were designated MDA-BCX-001 to -055 (MD Anderson Breast Cancer Xenograft 001 to 055). At the time of diagnosis, the median age of patients was 55 years (range, . At presentation, 32 tumors were hormone receptor-positive (HR+), 13 were TNBC, and 3 were HER2-positive (HER2+). At presentation, average tumor size (± standard deviation) was 4.3±3.1 cm (range, 0.8-15). Twenty-four of the tumors (50%), came from patients who received NeoCT. Of the 23 patients who had NeoCT with intact primary tumors, 7 had progressive disease, 7 had a partial response, and 9 had stable disease. One patient had an axillary cross-metastasis and that responded to therapy.
Tumor from the surgical specimens (designated as "P0") was implanted ( Fig 1A) and mice with tumor growth at the site of implantation were considered to have successful engraftment. Tumors derived from 13 of the 48 specimens (27%) were successfully engrafted.
BCXs that reached 1.5 cm in size were passaged to another set of mice (P2; Fig 1A). Time to tumor size of 1.5 cm diameter (i.e. passage from P1 to P2) was an average of 2.6 months (range 0.8-7 months; median 1.9 months) ( Fig 1B). P1 tumors were passaged to P2 generation and then were transferred into the third passage (P3) in all 13 models. When the time from implantation to first passage (P1) was compared to the time from first passage to second passage (P1-P2 interval) or to the time from second passage to third passage (P2-P3 interval), was significantly longer than both latter passages (p<0.001 for both; Fig 1C). Table 1 demonstrates comparison of patient and tumor characteristics and success of BCX engraftment. BCX engraftment was higher in TNBC compared to HR+ tumors (7 of 13, 53.8% vs. 5 of 32, 15.6%, p = 0.02). One of three HER2+ tumors which had received NeoCT engrafted.
Of note, two TNBC BCX models were derived from tumors not classified as TNBC prior to chemotherapy. For the BCX-011 model, the pre-NeoCT core biopsy was ER+, but the post-NeoCT and the emerging BCX-011 P1 samples were ER-. In addition, the pre-NeoCT core biopsy was HER2+ and ER+ for BCX-017, but the surgical sample after neoadjuvant chemotherapy was ER+, HER2-and the emerging BCX-017 P1 sample was TNBC. Notably this latter patient received paclitaxel with concomitant trastuzumab followed by 5-FU, epirubicin, and cyclophosphamide with concomitant trastuzumab.
At a median follow-up of 25 months (range, 1.8-39.4) after surgery, 12 of 44 patients with unilateral Stage II or III disease at surgery developed distant metastasis; the patient with bilateral tumors was disease-free at last follow up. For 5 patients, passagable tumors were derived prior to their development of distant metastases (Table 2). Among the 44 patients with unilateral localized disease, patients whose tumors developed BCXs had significantly lower recurrence-free survival (p = 0.015; Fig 2A) and distant recurrence-free survival (p = 0.004; Fig 2B) compared to patients whose tumors did not develop BCXs. Three patients with Stage IV disease underwent surgery; BCXs developed from the tumors of two patients. When all 47 patients are considered, patients whose tumors developed BCXs had significantly lower overall survival (p<0.001, Fig 2C).
DNA fingerprinting demonstrated donor origin of patient-derived xenografts STR DNA fingerprinting was used on all models that were serially transplantable to evaluate specific regions (loci) within nuclear DNA, to confirm that the xenografts were patient-derived. The STR profiles of the 13 passagable BCX models were all unique, showing no matches to known DNA fingerprints in American Type Culture Collection, the Deutsche Sammlung von Mikroorganismen und Zellkulturen, Cell Line Integrated Molecular Authentication, MD Anderson Cancer Center databases or to each other. The P0 genomic DNA from each patient resulted in a match with the corresponding BCX specimens for these 13 models (BCX-001, 002, 004, 005, 006, 010, 011, 017, 022, 024, 042, 051, and 055).

Histopathologic and molecular characterization identified murine origin tumors in two mice
A breast pathologist and a veterinary pathologist compared H&E stained slides of surgical specimens and BCXs. For BCX models designated successful BCX models, no histologic differences were found. The tumor derived from Patient 3 took the longest time (8.4 months) to develop a 1.5 cm mass. STR characterization of P0 demonstrated a unique human tumor. However, STR from P1 showed lack of amplification with seven of eight STR primer sets and discordance with P0 in the only one primer set that was successfully amplified. Mouse implanted with tumor from Patient 8 established a tumor; time from P0 to P1 and P1 to P2 took 5.2 and 4.5 months, respectively. However, tumor growth stopped at P2. STR from P1 showed lack of amplification. Therefore, STR results raised concern that both BCX-003 and -008 series xenografts may not be of human origin.
Although STR characterization of P0, P1 and P3 confirmed that BCX-001 series was indeed derived from Patient 1, histologic evaluation of late passages of the BCX-001 revealed morphologically undifferentiated tumors that lost the characteristics of epithelial/breast tumor cells, raising concerns about the fidelity of the model.
To determine whether the BCXs were of human origin, we performed PCR using two sets of primers designed to amplify a target sequence in which human (HBB) and murine (Zfp42) DNA could be distinguished (S2A Fig) [24]. We tested serial passages collected from BCX-001 and BCX-003 models as well as BCX-004 as a control. As expected, all three patient tumors (P0) showed no amplification of murine DNA, but amplification was seen with human primers (S2A Fig). In BCX-004, the mouse sequence was detected in P1 and P3, but at significantly less intensity than the human DNA product; this may be attributable to murine stroma and leukocytes within the xenograft. In contrast, for BCX-003, PCR of DNA from P1 and P3 xenografts demonstrated only amplification with murine DNA primers, suggesting the tumors were solely of murine origin. For BCX-001, there predominantly was amplification of human DNA at early passages including P4, but at P5, there was an abrupt loss of human DNA and amplification of murine DNA exclusively. Interestingly, this corresponded to an increase in growth rate (S2B Fig). The other model, BCX-008 P0 and P1 were tested similarly. Human sequence was detected in P0, and only mouse sequence detected in P1 (data not shown). When we examined H&E stained slides of BCX-001, passages 1 to 4, showed well differentiated epithelial tumors, whereas passages 5 to 8 showed tumors composed of dense neoplastic cells with smaller round nuclei and minimal basophilic cytoplasm suggestive of spontaneous murine lymphoma (S2C Fig). However, immunohistochemistry for CD20, CD3, CD45 and cytokeratin were negative on P5, suggesting this is an undifferentiated tumor of unknown tissue origin.

Functional proteomics demonstrated alterations in PI3K signaling
Unsupervised hierarchical clustering of RPPA was used to evaluate the functional proteomic profile of five of the ten patient's surgical specimens and corresponding serial transplanted in vivo tumors. 135 unique proteins and phosphoproteins were evaluated; 19 were done in duplicate, 2 in triplicate as part of the quality control process. Upon unsupervised clustering, patient tumors clustered together, separate from the BCXs, suggesting that there are significant differences between the primary tumor and the BCXs (Fig 3A). However, each BCX lineage clustered together demonstrated that their proteomic profile remains relatively stable for several in vivo passages. When protein expression was compared between the original breast tumors (P0) versus those from the first BCX passage (P1), levels of 103 of 154 proteins were differentially expressed at a FDR of 0.2 (S2 Table). Eight of these proteins are displayed in Fig 3B: EGFR, merlin, YB1 S102, Cyclin E1, 4E-BP1 T37/46, mTOR S2448, PRAS40 T246 and S6 S240/244. Importantly, a number of these proteins are part of the same PI3K pathway indicative either of increased pathway activity or alternatively of decreased human stromal content. All patient tumor-BCX differences remained significant when the P0 tumors were compared to both P1 and P2 and the first three transplants (P1, P2 and P3).

Copy number analysis determined loss of PTEN in two BCX models
MIP genotyping assays detect genetic alterations including candidate point mutations between specimens and are particularly well designed for copy number analysis [36]. A total of 412 somatic mutations in 45 genes were analyzed in genomic DNA extracted from the surgical specimens from patients 1, 2, 4, 5, and 6 and compared to corresponding BCXs. A heterozygous loss of an allele of PTEN was identified in the P0 tumor derived from patient 2. In the corresponding P3 of BCX-002, the second PTEN allele was lost, resulting in a homozygous PTEN loss (Fig 3C). Consistent with this finding, when PCR was performed with primers to human   (Fig 3D). However, another interpretation could be that human stroma PTEN is present in P0 and that the loss of human stroma allows a homozygous PTEN loss that was present to begin with to be identified.

Genomic sequencing demonstrated relative genomic stability of somatic mutations in BCX models
Whole exome sequencing was performed in the mouse host DNA as well as the matched P0-P1 -normal host DNA in 4 models: BCX-010, -017, -022, and -024 (S4 and S5 Tables). Unsupervised clustering clustered each BCX P1 with the P0 parental tumor (Fig 4A).
As described above, targeted exome sequencing was performed in 6 models: BCX-006, -010, -011, -017, -022, and -024. The mutations detected are listed in S6 Table. Upon unsupervised clustering each BCX clustered with its parental tumor (Fig 4B). Of 28 mutations detected on targeted exome sequencing in P0 tumors, 19 (68%) were preserved in P1 tumors (Fig 4C). Of 21 high allelic frequency mutations (allelic mutation frequency 10% or higher) detected in P1, 18 (86%) were also detected in the parental tumors (Fig 4D). In five of the models, serial passages were sequenced. With serial passaging, mutations were preserved, including potential driver mutations such as PIK3CA mutations, but additional alterations were observed in some models. In the BCX-006 model, pretreatment and on-treatment biopsies obtained during NeoCT with a rapalog-containing combination therapy regimen were also available. These demonstrated that the PIK3CA mutation was detected at a relatively high allelic frequency in the pretreatment and 12 week samples as well as in the P0 sample, and was preserved upon multiple BCX passages (Fig 4E). Although the allelic frequency increased in BCX passages, this was accompanied by an increase in mutant allelic frequency of other mutations as well, suggesting that this may be attributable to increase in tumor cellularity (S5 Table). Thus there did not appear to be selection of the mutation during therapy.

Somatic mutation profile is retained with in vitro culture and implantation
An experimental challenge is experimentally manipulating the PDX cells for functional target validation. Thus we also explored the possibility of developing conditionally reprogrammed cells (CRC) from PDXs and, and the effect of serial transplantation and in vivo re-implantation using the rapidly growing BCX-010 model. We cultured and expanded cells from BCX-010 [23]. After 4 passages, we injected these BCX-010 cultured cells into two strains of immunedeficient mice (nu/nu and NOG) with and without Matrigel. Rapidly growing tumors developed in all cases. STR fingerprinting confirmed identity of the CRC and the CRC-derived xenograft. The cultured cells as well as the CRC-derived xenograft maintained a mutation profile very similar to the originating PDX (Fig 4F).
shows the PTEN gene ideogram followed by the copy number aberration plot and the allele frequency plot for P0 and P3. A heterozygous PTEN loss is detected in P0 (single red line), and a homozygous PTEN loss is detected in P3 (double red line). Each blue dot corresponds to an individual probe on the array. The brown and purple lines mark the thresholds for loss of heterozygosity (LOH) and allelic imbalance regions, respectively. (D) PCR confirmed PTEN deletion in genomic DNA from Patient 2 tumor and BCX-002 P3. PTEN was undetectable in P3 but present in P0. RB1, another tumor suppressor gene, is detected in both samples and included for comparison. (E) PTEN loss demonstrated by next generation sequencing in BCX-024. P0 on the left, P1 on the right.

Discussion
PDXs have considerable promise for drug testing as they contain heterogeneity that is absent in cell line-derived xenografts [37]. They represent high-fidelity models for research and potentially even for individual patient treatment planning. In our study, we have demonstrated a good engraftment rate and rate in initiating passagable BCXs from chemotherapy-resistant breast cancers. Indeed, the take rate in chemotherapy treated BCX was increased compared to treatment naïve tumors consistent with the enrichment of stem cell-like cells in chemotherapy treated tumor [38]. We did observe modest alterations in the molecular profile, including a potential loss of PTEN, and activation of PI3K/mTOR signaling on functional proteomics. We also identified a previously under-reported problem, that of murine transformation in three models. However, the overall genomic profile of BCXs was relatively stable suggesting that they have the potential to function as useful in vivo models.
Engraftment efficiency among PDXs varies in the literature. For example, engraftment rate for pancreatic cancers has been at over 60% [21], whereas for breast cancers at 12-50%, with approximately 5-21% achieving serial transplants [5,11,12,17,39]. The disparity in transplantation efficiency is likely influenced by patient tumor type and host mouse strain [13], but it may also be a reflection of the aggressive biology of donor tumors [14,15]. A comparison of the patient tumors that formed and that did not form xenografts identified upregulation of cell cycle, cell signaling and cytoskeleton pathways, but downregulation of immune response in successfully formed xenografts [40]. In our hands, the tumor take rate was significantly higher from TBNC patients, and patients who had received NeoCT. Of note we actively enrolled chemoresistant patients; progressive disease on NeoCT is seen in only 3% of patients and is associated with a significantly worse prognosis [1]. Strikingly BCXs were successfully developed in six of seven patients who had progressive disease on chemotherapy. Of the patients in whom we established models, two already had distant metastases at the time of surgery, and six developed metastases within an average of 4.4 months. This is consistent with reports that the ability to successfully engraft correlates with tumor aggressiveness [41]. This is further supported in our study with the finding that patients whose tumors developed BCXs had significantly lower recurrence-free survival, distant recurrence-free survival and overall survival.
It has been proposed that PDXs may hold promise to prioritize treatment options in patients with advanced disease [22,42]. Some studies have reported preliminary data correlating antitumor efficacy in patient derived xenograft models with patient outcomes [6]. It is notable that in several patients we were able to develop serially transplantable tumors from patients prior to the patients' development of distant relapse. This timeline would allow us to explore the utility of these models for in vivo testing (AKA as avatar models).
It is important to determine if serially passaged xenografts remain representative of the original patient tumors. In previous series of pancreatic, breast, and colorectal cancer xenografts, no major variations between primary tumors and xenografts had been found [6,9,10,14,19,41,43]. However, Ding et al., identified genetic variation between a primary basal-like breast cancer and the corresponding metastatic tumor, as well as a serially transplanted xenograft from the same patient [44]. In our series both the genomic and proteomic profiles were, for the most part, preserved between the primary tumors and BCXs.
With three BCX implantations we observed development of a mouse tumor. The possibility of inducing a host transformation has been reported in cell line-derived xenografts in the immunocompromised host [45,46]. Further, others have also reported lymphoproliferative tumors of patient origin [16]. Thus, we recommend that once a PDX is established, H&E, and STR should be performed to confirm the fidelity of histopathologic characteristics and to confirm a unique human model that matches the patient tumor is established, respectively. Further, key molecular features should be intermittently re-assessed.
Comparing the primary tumor with the BCXs, we did observe increased activity of the PI3K/Akt/mTOR pathway. This pathway has been implicated in breast cancer development and progression. In our study, we did identify a change in the amount of loss of PTEN in one tumor during adaptation to growth in mice, however as indicated below whether this represents, loss, selection or a change in the amount of human stroma present has not been determined. On RPPA, we also demonstrated several alterations in PI3K signaling consistent with increased PI3K pathway activity in BCXs. Cancer cells evolve and acquire a more resistant phenotype with progression. Recently several studies, including our own, demonstrated a discrepancy in molecular markers between primary tumors and metastases [26,[47][48][49]. Models to delve into mechanisms of molecular evolution are lacking. Our studies demonstrate that there may be some molecular evolution upon in vivo implantation and upon subsequent passaging. It is possible that PI3K pathway activation may give BCXs a growth or survival advantage, leading to accumulation of, or selection for PI3K pathway aberrations. One explanation of our findings is clonal selection [8,11,44]. It is unclear whether the PTEN changes we observed represents enrichment already present in a small subpopulation of cells-clones-in the original patient tumor, or whether it was an acquired alteration under selection pressure for growth in the murine environment. However alternate explanation of our findings may be technical, such as enrichment for tumor DNA in BCX-002 serial passages with the loss of human stroma (with normal PTEN copy number, which would be detected on the MIP arrays) [37,50] and replacement with mouse stroma (with a mouse PTEN DNA which would not be detected by to the MIP arrays) [51]. Similarly increase in tumor cellularity in the BCX may result in an appearance of increased PI3K activity in the overall tumor specimen without a change in the tumor component. The apparent increase in PI3K activity could also represent difference in sample handling and conservation of phosphorylation specific events. There are differences in post-resection processing of primary tumors and BCXs; primary tumors are inked and potentially radiographed for margin analysis and only then specimens retrieved for research. In contrast BCXs are immediately frozen after harvesting; this difference could contribute to apparent higher protein phosphorylation in BCX. This is especially important while assessing PI3K signaling as we have already shown that there are differences in several phospho-residues between matched primary tumor core biopsies and surgical samples obtained after similar processing [30]. This is also consistent with the finding that once BCXs are established, there functional proteomic profiles remained stable, with passages of each BCX model clustering together on unsupervised clustering. Further work is needed to define which of the changes between P0 and P1 represent tumor evolution, and whether BCXs can be used to better delineate the interplay between tumor heterogeneity, drug response, and molecular evolution.

Conclusions
We have developed an in vivo breast cancer model with primary breast tumors from patients with significant residual disease or progression after NST who are at high risk of distant relapse. Their chemoresistant tumors were used to create serially transplantable BCXs. Some genetic and proteomic changes were identified in the process of engraftment and serial passage, especially aberrations in the PI3K/mTOR signaling pathway. However, the majority of the genomic alterations were retained. Future work will determine whether these changes recapitulate molecular evolution that occurs with tumor progression and metastases. This model holds promise as a novel platform to discover molecular aberrations that can be targeted for therapy in chemoresistant disease.
Supporting Information S1 Fig. Determination of human versus mouse origin of BCXs and characterization of lineage. A. Two sets of primers were used to amplify a target sequence [Zfp42, mouse (M) and HBB, human (H)] from (P0) and subsequent BCX (P1, etc.) from Patients 4, 3, and 1. As expected, amplification of the mouse sequence was absent in all patient tumors (P0). For Patient 4, but not for Patient 3, human DNA bands were preserved in the BCX. For BCX-001, human DNA was amplified in early passages but there was loss of human DNA in P5 onwards. B. BCXs derived from Patient 1's tumor were passaged from the surgical specimen (P0) to subsequent generations. Graph shows time to passage of tumor to the next generation when tumors reach 1.5 cm in diameter, and demonstrates a distinct increase in growth speed beginning with P5. C.   Table. Copy number alterations detected in all tested passages of six models by targeted exome sequencing (BCX-006, -010, -011, 017, -022, -024). High amplification cutoff was 4 (highlighted in red) and high deletion cutoff was 1 (highlighted in green). (XLSX) S4 Table. Copy number alterations detected in P0 and P1 of four models by whole exome sequencing (BCX-010, -017, -022, -024). High amplification cutoff was 4 (highlighted in red) and high deletion cutoff was 1 (highlighted in green). (XLSX) S5 Table. Somatic mutations detected in P0 and P1 of four models by whole exome sequencing (BCX-010, -017, -022, -024). (XLSX) S6 Table. Somatic mutations detected in all tested passages of six models by targeted exome sequencing (BCX-006, -010, -011, 017, -022, -024). (XLSX)