Tumor specimens were obtained from BC patients with their informed consent. Tumor fragments were removed during surgery, as previously described 8. Briefly, fresh tumor fragments were grafted subcutaneously into the interscapular fat pad of female Swiss nude mice under anesthesia. Mice were kept in pathogen-free animal housing (Institut Curie) and received estrogen (8 μg/mL) diluted in drinking water. Xenografts appeared at the graft site two to eight months after grafting. They were subsequently transplanted from mouse to mouse. The experimental protocol and animal housing were in accordance with institutional guidelines as proposed by the French Ethics Committee (Agreement B75-05-18, France).

The morphology of patient tumor tissues was compared with that of the corresponding xenografts by examining paraffin-embedded sections according to standard protocols 8. Tumors were removed from mice and immediately fixed in 10% formaldehyde solution for immunohistologic examination.

Immunostaining was performed according to previously published protocols 9. Briefly, 4 μm tissue sections were prepared from a representative sample of the tumor. After rehydration and antigen retrieval in citrate buffer (10 mM, pH 6.1), tissue sections were stained for estrogen receptor (ER), progesterone receptor (PR) and ERBB2/neu (HER2). Staining was revealed with the Vectastain Elite ABC peroxidase mouse Immunoglobulin G kit (Vector, Burlingame, CA, USA) using diaminobenzidine (Dako A/S, Glostrup, Denmark) as chromogen. Positive nuclear staining for ER and PR was recorded according to standardized guidelines. ER (clone 6F11; 1/200; Novocastra, Rungis, France), PR (PR; clone 1A6; 1/200; Novocastra) and ERBB2 (clone CB11; 1/1,000; Novocastra) expression was evaluated. For ERBB2, only membranous staining was evaluated, as previously defined 10.

Genomic DNA was extracted as previously published 11. Co-hybridization was performed between extracted DNA (primary BC or corresponding xenograft) and normal DNA. Genome-wide resources of 3,261 or 5,244 fluorescence in situ hybridization-mapped sequenced bacterial artificial chromosome (BAC) and P1-derived artificial chromosome (PAC) clones were represented as immobilized DNA targets on glass slides, allowing a mean resolution of 0.5 Mb throughout the genome. Each clone was spotted in quadruplicate on slides prepared by Integragen™ (Evry, France). DNA samples, each 1.5 μg, were digested with DpnII enzyme (Ozyme, Saint-Quentin-en-Yvelines, France) and labeled with random priming using a Bioprime DNA labeling kit (Life Technologies, Villebon sur Yvette, France) with the appropriate cyanine dye (Cy3 or Cy5; Perkin-Elmer, Courtaboeuf, France). The control and test DNAs were co-precipitated with Cot-1 DNA (Life Technologies, Villebon sur Yvette, France), denatured and re-suspended in hybridization buffer. After 24 hours of hybridization, slides were washed with sodium dodecyl sulfate buffer and saline citrate, dried and scanned with a GenePix 4000B scanner (Axon Instruments Inc., Union City, CA, USA). Image analysis was performed with GenePix 5.1 software (Axon) 12. Any BAC with less than two replicates flagged for not meeting qualitative spot criteria was excluded. Normalization was performed with the MANOR algorithm 12. Spots showing a low signal-to-noise ratio or poor replicate consistency were discarded. Status assignment (loss, normal, gain and amplification of chromosome copy number) was performed by using the GLAD algorithm 13.

Hierarchical clustering was performed on profiles based on probe status. The group average was used as the similarity measure and the Pearson algorithm was used as the agglomerative method. Separation into groups was proposed on the basis of the structure of the dendrogram. Data visualization and computation of clustering were performed according to the VAMP (Vizualization and Analysis of Molecular Profiles) analysis procedure 12. Pearson's correlation coefficient was calculated on altered clones based on their clone status (clone status, that is, GNL: gain, normal, loss). For analysis of GNL differences, chromosomal segments were defined as genomic regions with a constant copy number for all samples.

Prior to RNA extraction, a tissue section from the tumor fragments was stained with hematoxylin and eosin to evaluate tumor cellularity. All tumors analyzed comprised more than 40% of tumor cells on the tissue section. Total RNA was isolated from 15-65 mg of frozen tissue using TRIzol reagent (Invitrogen, Cergy-Pontoise, France) according to the manufacturer's instructions. The RNA concentration was measured by absorbance at 260 nm. The quality of each RNA sample was determined on the Agilent 2100 bioanalyzer (Agilent Technologies, Palo Alto, CA, USA. RNA was processed on chips only when the following criteria were met: RIN (a measurement of RNA quality) ≥ 7, (28S/18S) ≥ 1.4, (260 nm/230 nm) ≥ 1.8, and (260 nm/280 nm) ≥ 1.8.

Hu et al. defined and validated the centroids of 306 genes to discriminate between five previously identified BC molecular subtypes (Luminal A and B, Basal, HER2 positive, Normal) 16. The UniGene ID (Build204) gene list was matched to the HG-U133 Plus 2.0 Affymetrix^© platform annotation. Each sample was assigned to the nearest subtype/centroid as determined by the highest Spearman rank order correlation between the gene expression values of the molecular subtype probe sets and the five subtype centroids. A sample with a maximum correlation score less than 0.2 was considered to be unclassified.

The gene expression profiles of the pairs (xenograft and corresponding primary tumor; xenograft at different tumor passages) were assumed to be similar (null hypothesis). The scatter plots of the whole gene expression data set for pairs of xenografts and corresponding primary tumors and pairs of xenografts from different passages confirmed this hypothesis. A linear regression model was fitted to analyze the variation in gene expression of each pair (xenograft and corresponding primary tumor or xenograft after several generations). Linear regression models were built to define the 5th and 95th percentiles of the residual distribution. Normality testing was limited to observation of the density plot and a quantile-quantile plot of the residual values. A residual above the 95th percentile or below the 5th percentile was considered to be an outlier.

This analysis was performed to determine whether specific gene sets (that is, functional groups) were overrepresented in the various gene lists. The DAVID (Database for Annotation, Visualization and Integrated Discovery) Gene Ontology web site was used to test the significance of enrichment in specific gene ontology annotations 17.

A preliminary histologic analysis of xenografts compared their morphology and pathological classification (based on HER2, ER and PR receptor expression) with those of the primary tumors. Results based on immunohistochemistry data are summarized in Table 1 and have been previously published in part 818. Figure 1 illustrates the histology of patient- and xenograft-derived tumors for each of the following BC subtypes: luminal human breast cancer xenograft (HBCx-3), triple-negative (HBCx-8, HBCx-12A and HBCx-10), HER2+ (HBCx-5), and lobular (HBCx-19). The primary HBC-3 BC was an infiltrating ductal adenocarcinoma organized in cell cords with a small in situ component. The xenografted tumors showed a similar architecture with an abundant stromal component. The triple-negative HBCx-8 and HBCx-10 tumors were diagnosed as infiltrating ductal carcinoma with a trabecular architecture that was reproduced in the xenograft. The HBCx-12A was a poorly differentiated ductal carcinoma with a large number of mitotic figures and large cells with abundant cytoplasm. Images at a magnification of 2.5 × of the patient's luminal HBCx-3 and HER2+ HBCx-5 tumors showed a high content of stromal component (Figure 1A, D). Organization of stroma around the tumor cell nests in HER2+ tumors was reproduced in the xenograft.

To evaluate the similarity of genomic profiles between the tumors and their corresponding xenografts, a BAC aCGH analysis was performed on 18 pairs, of which 2 were HER2+, 4 were ER positive and 12 were triple-negative (Table 1).

Copy number alterations (CNA) of each pair of patient tumors and xenografts were compared by calculating Pearson's correlation coefficients R. Fourteen pairs had a correlation coefficient greater than 0.50, indicating similarity (Table 1). The correlation coefficient was less than 0.5 in four pairs: one pair was HER2+ and three were triple-negative. No correlation with the percentage of high quality clones analyzed was identified (data not shown).

Two examples of paired aCGH profiles are shown in Figure 2. In Figure 2A, both the primary ER positive tumor and the xenograft displayed a slightly altered profile. Most of the alterations were conserved in the xenografts, such as the 1q gain, Xp22 amplification, 2q loss, and alterations in 14p, 15q, and 17. A gain in the 11p region was observed in the xenograft but not in the original tumor. Profiles of the triple-negative tumors are shown in Figure 2B. Most of the chromosomes were affected by genomic alterations. Large genomic gains were observed in almost all chromosomes. Three large genomic losses, in chromosomes 3, 13 and X, were conserved in the xenograft. Slight amplifications of genomic alterations were observed in the xenograft profile, for example, in chromosomes 1, 6, 7 and 17. Non-altered regions were fairly similar in both profiles.

Unsupervised hierarchical clustering analysis of CNA was performed on the whole sample set (18 pairs of primary tumors and corresponding xenografts) in order to determine whether tumors and xenografts were clonally related (Figure 3A). Sixteen of the 18 patient-xenograft pairs clustered together, indicating greater differences in CNA between primary tumors than between a primary tumor and its corresponding xenograft. Luminal and HER2+ BC subtypes clustered into two separate subgroups that were distinct from the triple-negative subgroup. In two cases, xenografts derived from the primary tumor and axillary lymph node from the same patient (HBCx-12A/B and HBCx-13A/B) clustered closely with the corresponding primary tumors (Figure 3B). This result highlights the fact that no major differences in CNA were observed between xenografts derived from breast tumors or the corresponding axillary node metastases. In addition, the two HBCx-12A/12B and HBCx-13A/13B pairs clustered into two small subclusters, indicating a higher degree of similarity between xenografts derived from primary tumors and metastases than between patient tumors and xenografts. An example of the similarity between primary tumor-derived xenografts and metastasis-derived xenografts is shown in Figure 4A. The Pearson correlation coefficients between the patient tumor and the corresponding xenograft were 0.86 and 0.82 (for a primary tumor-derived xenograft and a metastasis-derived xenograft, respectively). A very strong correlation was also observed between the two xenografts, with a correlation coefficient of 0.85. Importantly, the HER2 amplicon, localized on chromosome 17q21-22, was preserved in both the primary tumor-derived xenograft and the metastasis-derived xenograft (Figure 4B). A second amplicon, frequently associated with HER2 amplicons, localized on chromosome 8 (8p12-p11), was also found in the three genetic profiles.

To study the stability of xenografts in serial transplantation, the genetic profile of four xenografts was analyzed from different passages and compared to the profile of the original patient tumor. The genomic profiles of xenografts remained very stable throughout sequential in vivo passages. An example of aCGH profiles at different passages is shown in Figure 4C; the HBCx-21 profiles at p1 and p6 (30 months later), demonstrate a strong homology between the two tumors.

CNA frequency plots were analyzed for differences between patient tumors and xenografts (Figure 5A). The patient tumor plot showed highly rearranged profiles, reflecting the intrinsic genetic instability of BCs. This general CNA frequency pattern was reproduced in the xenograft plot with additional alterations observed for several chromosomes. To analyze the chromosomal regions that differed between xenografts and patient tumors, a "GNL difference" was calculated for each chromosomal segment: a segment was defined as a region with a constant copy number for all samples. The 1,351 segments analyzed are listed in Additional file 1, Table s1 and the frequency distributions of neutral, positive and negative GNL differences are shown in Figure 5B. A positive GNL difference indicates an increased copy number in the xenograft versus the patient tumor, while a negative GNL difference indicates a decreased copy number. Differences were calculated for 20 pairs corresponding to 18 patients and 20 xenografts (xenografts were generated from the primary tumor and, in addition, from a metastasis in two patients) and regions present in at least 10 comparisons were analyzed. The chromosomal distribution of these differences (corresponding to 992 segments) is illustrated in Figure 5C. The analysis was restricted to chromosomal segments showing negative or positive GNL differences in at least 30% of pairs (and present in at least 10 comparisons) in order to identify recurrent changes. It showed that 178 regions were lost and 202 regions were gained (Additional file 1, Table s2 and Table s3, respectively). The chromosomal distribution of these recurrent changes (Figure 5D) shows that gains mainly involved chr 1 and chr 17 (35% and 17%, respectively), while CNA losses occurred in chr 1, chr 5, chr 12 and chr 18 (23%, 19%, 16%, respectively). The chromosomal regions involved in recurrent gains and losses are shown in Table 2. Some regions (chromosomes 1p, 15q, 16p, 18p) were involved in both gains and losses, while other regions were mainly associated with either gains or losses.

The molecular classification of patient tumors and xenografts was determined using 306 genes that allow for discrimination between the five BC molecular subtypes (Luminal A and B, Basal, HER2, Normal); these genes were defined and validated by Hu et al. 16. Matching the UniGene ID (Build204) gene list to the HG-U133 Plus 2.0 Affymetrix^© platform annotation identified a total of 296 (729 probe sets) of the 306 genes. In a group of 28 tumors (including 10 patient tumors and 18 xenograft tumors at various later generations in vivo), 12 samples were classified as Basal, 6 as Luminal A, 3 as Luminal B and 2 as Normal (Table 3). Spearman correlation coefficients were less than 0.20 for seven of these tumors and the corresponding samples were consequently considered "unclassified". In 6 of the 10 comparisons, the molecular signature was concordant and in two cases the patient's tumor was classified as Normal and the xenograft was classified as Basal.

Linear regression analysis was performed to study the global gene expression pattern in xenografts versus primary tumors. An example of the correlation between gene expression patterns is shown in Figure 6. Probe sets that were not differentially expressed fitted a linear model (black dots), while red dots represent extreme residual values (overexpressed or under-expressed). The gene expression scatter plot between the patient tumor and the xenograft indicates that the great majority of probes were not differentially expressed (Figure 6A). Linear regression analysis performed of xenografts from different tumor passage (passage 3 versus passage 6) showed very few probe sets with differential expression (Figure 6B), indicating that gene expression did not undergo any major changes during the course of mouse-to-mouse grafts. Analysis of all paired samples (primary tumor vs corresponding xenograft) showed a significant decrease in the number of probe sets differentially expressed exclusively in one pair (5,530 out of 29,683; 18.6%) or common to 13 pairs (225 out of 29,684; 0.7%). A similar, significant trend was observed for the proportion of overexpressed versus under-expressed genes (Chi-square test, P = 1e-16) with a decreasing number of overexpressed genes when comparing the list of probe sets exclusive to one pair (3,114 out of 5,530; 56.3%) to the list of probe sets common to 13 pairs (6 out of 225; 2.6%) (Figure 6C). Gene Ontology analysis was performed on four probe set lists, that is, common to 1 to 3 pairs, 4 to 6 pairs, 7 to 9 pairs and 10 to 13 pairs (Additional File 2, Tables S1, S2, S3 and S4). A significant enrichment in annotations corresponding to the immune system, response to external stimuli, response to wounding, cell adhesion, inflammatory response, blood vessel formation, skeletal development and cell motility was observed in the group of probe sets common to 10 to 13 pairs (Table 4 and Additional file 3, Table S1). The group common to one to three pairs was mainly enriched in annotations corresponding to protein transport, messenger RNA processing, RNA splicing, cell adhesion, regulation of cell proliferation, regulation of cell death and apoptosis, but no annotation related to the immune system was identified. Analysis of these ontology annotations in the four groups showed an interesting pattern. Significant annotations in the first group progressively became non-significant in the second and third groups and, inversely, significant annotations in the fourth group progressively became non-significant.

For instance, gene expression analysis performed on xenografts after multiple passages identified only 11 probe sets (0.03%) differentially expressed in more than 50% of the paired samples (data not shown).

Altogether, these data show that xenografts display a relatively similar gene expression profile compared with their corresponding human tumors.