Characterization of a Large Panel of Patient-Derived Tumor Xenografts Representing the Clinical Heterogeneity of Human Colorectal Cancer

See commentary by Kopetz et al., p. 5160

This article is featured in Highlights of This Issue, p. 5155

Despite great advances in our understanding of the molecular basis of colorectal cancer (CRC) and the increasing number of targeted therapies, the treatment of advanced CRC frequently remains disappointing, particularly when facing metastases. This is linked to the natural history of the tumors and/or patients, considering (i) the interpatient variability in drug exposure, (ii) the diversity of CRC with respect to molecular profile and sensitivity to a specific agent, and (iii) the intratumoral heterogeneity, and the highly variable tumor cell doubling time. All these sources of variability make difficult the validation of markers predictive for the response to chemotherapy.

The interest of patient-derived xenograft (PDX) models in preclinical testing is more and more emphasized. These xenografts derived directly from patient samples, without in vitro manipulation, provide a more accurate depiction of human tumor biologic characteristics than cell line-derived xenografts with hundreds in vitro passages and serial passing across several generations of mice (1). Moreover, as these models might better reflect a clinical response (2–8), several groups have established disease-specific panels of xenografts directly from patient tumors (2, 3, 7–10). The ability to passage patient fresh tumor tissues into large numbers of immunodeficient mice provides possibilities for better preclinical testing of new therapies.

The above considerations emphasize the need for new in vivo models of CRC representing their genotypic and phenotypic diversities. Here, through a consortium effort from hospitals, academic groups, and biotech and pharmaceutical companies, we have developed a large collection of CRC models directly derived from tumor samples collected during patient surgery. We have extensively characterized these xenograft models at the molecular, histologic, and pharmacologic levels to ensure that they represent the diversity of the clinical situations. All model characteristics and clinical patient history are being compiled in a web-based database for efficient features search and interconnection. To assess whether the histopathology of human CRC is maintained in those models, we investigated the level of differentiation, the presence of mucus and showed the efficient replacement of the human stroma by host cells. The high diversity of the molecular profile and pharmacologic response of CRC highlight the difficulties to classify colon tumors upon molecular profile. Nevertheless, such collection allowed us to set up a phase II trial to investigate the sensitivity profile of the panel and see whether it correlates with clinical data. Our analysis confirms the link between cetuximab sensitivity and KRAS wild-type status observed in the clinical setting. This collection could be used to assess new therapeutic approaches, to better understand the basis for sensitivity of tumors from individual patients, and potentially help the stratification of patients according to molecular markers.

CRC samples were collected, after patient's informed consent, in 3 medical centers: Curie Institute (Paris, France), Gustave Roussy Institute (Villejuif, France), and Lariboisière Hospital (Paris, France). Immediately after surgery (1 hour after resection in average, except for rectal tumors for which it can go up to 4 hours), 2 fragments were snap frozen in liquid nitrogen, then stored at −80°C, for molecular characterization, 2 fragments were fixed in alcohol-formalin-acetic acid solution and paraffin-embedded for histopathologic analysis, and 2 fragments were transferred in culture medium including DMEM with 10 mmol/L HEPES, 4.5 g/L glucose, 1 mmol/L pyruvate sodium, 200 U/mL penicillin, 200 μg/mL streptomycin, 200 μg/mL gentamicin, 5 μg/mL ciprofloxacin, 20 μg/mL metronidazole, 25 μg/mL vancomycin, and 2.5 μg/mL fungizone or DMEM with Nanomycopulitine (Abcys) for engraftment done in the 24 hours after resection. After engraftment, residual fragments were frozen in DMEM including 10% dimethyl sulfoxide and 10% fetal bovine serum. Similar process of sample conservation was conducted on tumor fragments collected from mice and rats. A thawing test from P3 frozen samples was done in a period of 3 months after freezing.

Animals were maintained in the animal facilities of each institution following standard animal regulation and strict health controls allowing transfer between members of the consortium. Swiss-nude and CB17-SCID female mice, as well as NIH-nude rats were bred at Charles River France. Mouse weights were more than 18 g and rat weights were more than 160 g at the start of experiments. Their care and housing were in accordance with institutional guidelines, as well as national and European laws and regulations as put forth by the French Forest and Agriculture Ministry and the standards required by the UKCCCR guidelines (11).

After 2 to 24 hours following the patient surgery, the tumor samples were engrafted on 2 Swiss nude mice. Small fragments (∼50 mm³) were subcutaneously engrafted into the scapular area or on the flank of anesthetized mice (xylazine/ketamine or isoflurane protocol). Tumor growth was measured at least once a week and serial fragment grafts of each given tumor were conducted on 3 to 5 Swiss nude or CB17-SCID (after 3 passages) mice when the tumors reach a volume of 800 to 1500 mm³. Engraftment on nude rat was conducted for each tumor after 3 successfully passages on nude mice. To increase tumor take rate, nude rats were whole body irradiated at 7Gy with a γ-source, 24 to 48 hours before tumor grafting. For orthotopic implantation, analgesia was induced 30 minutes before implantation by intraperitoneal (i.p.) injection of Flunixin (5 mg/kg body weight; Finadyne, Schering-Plough). Then, mice were anesthetized by i.p. injection of xylazine (5 mg/kg body weight; Rompun, Bayer) and Tiletamine (30 mg/kg body weight; Zoletil, Virbac). For implantation, tumor fragment was attached to the serosa of the cecum to be entirely surrounded by the serosa.

Frozen fragments were cut in a cryostat at −20°C then beginning and end sections were stained with Haematoxylin-Eosin-Saffron (HES) for histologic control and evaluation of tumor cell percentage by pathologists. Only samples with an average percentage above 40% were used for DNA extraction. Genomic DNA was extracted according to QIAamp DNA mini-Kit protocols (Qiagen). Total RNA was extracted from tumor samples using a TRIzol method (Invitrogen), purified with an RNeasy mini kit (Qiagen).

Serial 5-μm-thick sections were cut from each paraffin-embedded tumor. Examination of tumor morphology was conducted on slices processed for HES staining. Tumor differentiation, necrosis, stroma, and presence of mucus were scored for each xenograft and correspondent patient's tumor biopsy. Eight to ten different samples were reviewed for each xenograft model at 2 different passages (between P6 and P12). Aperio Scan Scope XT scanner was used to acquire whole sections images at ×200 magnification. Alu probe in situ hybridization was processed. On 5-μm-thick sections of paraffin-embedded tumors in a Discovery XT biomarker platform (Roche Diagnostics) according to the manufacturer's specifications and using its reagents, including the Alu Probe SN59560. Briefly, sections are deparaffinated, rehydrated, pretreated by protease digestion (20 min at 37°C), denaturated 8 min at 85°C, incubated with the Alu probe 1 hour at 50°C, washed, postfixed, stained by an immunoenzymatic method using alkaline phosphatase and NBT/BCIP (nitro-blue-tetrazolium/5-bromo-4-chloro-3-indolyl-phosphate) as a substrate, and counterstained with Nuclear Fast Red.

Microsatellite sequence instability (MSI) testing was conducted according to the National Cancer Institute guidelines. A 5-microsatellite consensus panel was used as (12, 13).

On the basis of the literature, we selected exons 2 to 11 of TP53; exons 9 and 16 of APC; exon 3 of AKT1, exons 11 and 15 of BRAF, exons 4 to 11 of FBXW7, exon 3 of CTNNB1, exons 2 and 3 of KRAS, exons 10 and 21 of PIK3CA, exons 18 to 21 of EGFR, exons 2, 26–27 and 30 of KDR, exon 4 of FCGR2A and FCGR3A, and exon 3 of ERCC1 for direct sequencing analysis, conducted after PCR amplification (Supplementary Table S1). Purified PCR products were sequenced using BigDye Terminator Cycle Sequencing Kit (Applied Biosystems). Sequencing reactions were analyzed on 48-capillary 3730 DNA Analyzer in both sense and antisense directions. All mutations detected were controlled with independent amplification at least once. Sequences reading and alignment were conducted with SeqScape software (Applied Biosystems).

Genomic DNA was analyzed using the Human Genome CGH Microarray-244A (Agilent Technologies). In all experiments, sex-matched DNA from a pooled human female or male (Promega) was used as a reference. Oligonucleotide aCGH processing was conducted as detailed in the manufacturer's protocol (version 4.0; http://www.agilent.com). Data were extracted from scanned images using the Feature Extraction software (v10.2.1.2 to 10.7.3.1, Agilent), along with protocols CGH_v4_10_Apr08 to CGH_105_Jan09. Acquired signals were normalized according to their dye and local GC% content using in-house scripts under the R statistical environment (http://cran.r-project.org). Resulting log2(ratio) were segmented using the circular binary segmentation (14) algorithm implementation from the DNA copy package for R. Aberration status calling was automatically conducted for each profile according to its internal noise [variation of log2(ratio) values across consecutive probes on the genome]. All genomic coordinates were established on the UCSC human genome build hg19 (15). To compensate the possible variation of dynamics in the profiles of paired samples, a simple linear regression fit was conducted under R, to increase the lower dynamics profile to the higher dynamics one. Hierarchical clustering was conducted under R on segmented data, using Euclidean distances and Ward's construction method. The gene expression analysis was conducted using a GeneChip Expression 3′-Amplification Reagents Kit and U133A GeneChip arrays (Affymetrix). All data were imported into Resolver software (Rosetta Biosoftware) for database management and quality controls. Moreover, the raw data (.CEL files) were imported in the BrB Arrays Tools (http://linus.nci.nih.gov/BRB-ArrayTools.html) and normalized with a Robust Multi-array Average (RMA) procedure (http://www.bioconductor.org) to conduct the class prediction analyses and functional analysis. The microarray data related to this article have been submitted to the Array Express data repository at the European Bioinformatics Institute (http://www.ebi.ac.uk/arrayexpress/) under the accession number E-MTAB-991 for Gene Expression and E-MTAB-992 for aCGH.

5-fluorouracil (5-FU) (ICN) and oxaliplatin (Sanofi) were diluted in a 5% (w/v) solution of glucose in water. Cetuximab (Imclone) was diluted in a phosphate buffer solution. Irinotecan (Pfizer) was diluted in water.

Protocol design, chemotherapy techniques, and methods of data analysis were essentially equivalent to those described previously (16). On the first day of treatment, the animals bearing 100 to 200 mm³ tumors were unselectively distributed to the various treatment and control groups (n = 8–10 per group).

Oxaliplatin and 5-FU were IV administered at a dosage corresponding to 70% of the highest nontoxic dose (HNTD) in mice (50 and 5 mg/kg/injection, respectively) with a Q4Dx2 schedule of administration (i.e., 2 injections with a 4-day interval). Irinotecan (CPT-11) was IV injected at a dose of 22 mg/kg/injection, Q2Dx3 (70% of HNTD). In nude rats, 5-FU, oxaliplatin, irinotecan were injected IV at the HNTD (5-FU at 30–40 mg/kg/injection, Q7Dx3, oxaliplatin at 4 mg/kg/injection, Q4Dx3, and irinotecan at 40 mg/kg/injection, Q7Dx3). Cetuximab was administered at the dose of 40 mg/kg/injection (IP, Q4Dx4) in mice and 10 mg/kg/injection (IV Q7Dx3) in rats. Tumors were measured with a caliper 2 to 5 times weekly. Tumor volumes were estimated from 2-dimensional tumor measurements: Tumor volume (mm³) = [length (mm) × width (mm)²]/2. For ethical reasons the animal bearing an excessive tumor volume (>2,000 mm³ for mice and 10,000 mm³ for rats) or tumors impinging too much their behavior were euthanized.

Because tumors were measurable at the start of therapy, the initial tumor burden was taken into account in the calculation of the tumor growth inhibition (ΔT/ΔC value): ΔT/ΔC (%) = [(median T_DayY − median T_DayX)/(median C_DayY − median C_DayX)] × 100 (where DayY is the day of evaluation, and DayX is the day of initiation of therapy for treated [T] and control [C] tumor volumes).

Tumor regressions can be either partial (more than 50% reduction in tumor volume) or complete (tumor regression below the limit of palpation).

The tumor doubling time (in days; Td) was estimated from the plot of the log linear growth of the control group tumors in exponential growth (100–1,000 mm³ range).

Recursive partitioning method was conducted using SAS JMP v9 software as described in ref. 17. The Fisher's test, χ² test, and all logRank analyses were conducted using Everstat V5 (Sanofi based on SAS 8; SAS Institute Inc.).

A total of 85 CRC samples from primary tumors, peritoneal carcinoses, or distant metastases were collected from unsupervised patients and subcutaneously engrafted into nude mice. Patient clinical characteristics are presented in Table 1. The PDX collection nicely reflects the diversity of colorectal clinical cases (18) in terms of gender, age, carcinoembryonic antigen (CEA) dosage, primary tumor location, lymph node, and distant metastasis status as determined at primary stage. It has to be noted that the lowest stage of the disease, namely, stage 0 (e.g., in situ), is not present within the collection.

Of the 85 engraftments, 54 led to the establishment of PDX models, i.e., a tumor take rate of 63.5%. We noted a higher tumor take rate for metastases than primary tumors, albeit it was not statistically significant. No impact of the cold ischemic time (less than 24 hours) was noticed and the median time to reach 300 mm³ after the first engraftment was 59.4 days (see Supplementary Table S2). A biobank was established for all models and the thawing success rate reaches 92.3%, close to our tumor take rate on fresh samples, 97.2% at P3 to P6. Among this collection, 5 coupled primary tumor/associated synchronous hepatic metastasis, 1 coupled peritoneal carcinoses/splenic metastasis, and 1 triplet primary tumor/peritoneal carcinoses and mesenteric lymph node metastasis derived from the same patient have been established. Overall, the collection contains 35 primary tumor, 5 peritoneal carcinoses, and 14 metastatic (12 hepatic, 1 splenic, and 1 mesenteric lymph node) models.

Recursive partitioning was applied to identify parameters associated with high or low probability of in vivo tumor take rate (Table 1). None of the prior treatments impact significantly the tumor take rate. Two or three combined parameters were associated with higher probability of in vivo tumor take rate: (i) pN+, CEA ≥ 6 ng/mL (probability = 1 while combined with primary tumor location = proximal or distal colon, n = 21; probability = 0.87 while combined with primary location = rectum, n = 8) and (ii) pN+, CEA < 6 ng/mL, pM0 (probability = 0.83, n = 6). Conversely, the signature pN0+ primary tumor location = rectum is associated with a low probability of in vivo tumor take rate (probability = 0.12, n = 8), which could be impacted to a longer hot ischemic time.

Histopathologic characterization of our panel showed a concordance between xenografts (2 different passages between P6 and P12) and the corresponding patient's tumor in term of tumor differentiation and mucus secretion (Fig. 1A and Fig. S1). Using standard clinical parameters, xenografts were classified in well to moderately differentiated (92%) and moderately differentiated to undifferentiated (8%). Ten percent of the xenografts were classified as mucinous adenocarcinomas, a rare morphologic subtype of colorectal adenocarcinoma in which more than 50% of the tumor volume is composed of mucin. These results matched the classification of their corresponding patient tumors. We determined as well the percentage of stroma and showed that the tumor cells are clearly enriched in xenografts compared with patient samples (Supplementary Fig. S1). This enrichment has been observed after the first passage (data not shown), but once established, a constant ratio of stroma in the tumor among the different passages was maintained. Finally, in situ hybridization with Alu probe specific to human cells showed the complete loss of human stroma in early passages (Fig. 1B). Overall, these observations confirm that our collection recapitulates the histologic profile and the diversity of CRC.

To investigate molecular abnormalities displayed by our PDX models and their potential variation occurring during xenograft take and passages, we determined MSI status and conducted mutational analysis, aCGH, and gene expression profiling. When the tumor cellularity of patient samples was insufficient, the molecular characterization was conducted on the corresponding first passage xenograft (P0).

Sequence of exons described in Material and Methods was analyzed in all established xenografts localized at hotspots and described in the COSMIC Database (ref. 19; Fig. 2). TP53 was found to be mutated in 38 samples (70%). We detected 25 different mutations, most of them reported as deleterious in IARC database (www-p53.iarc.fr). Full concordance was found when TP53 status was determined by functional analysis of separated alleles in yeast (FASAY, described in ref. 20). About APC, we identified 30 different mutations in 32 mutated samples (59%) including 29 mutations resulting in a truncated protein. For KRAS, we found 9 different mutations in 26 mutated samples (48%) corresponding to frequent hotspot mutations in codons 12, 13, and 61. Two infrequent mutations in codons 19 and 20 (L19F and T20S) were found in a single sample, CR-IC-0018P. In the case of PIK3CA, classical mutations E545K and H1047R, in exons 10 and 21 respectively, were detected in 7 samples (13%). Six samples harbored FBXW7 mutations corresponding to 3 missense mutations and 2 mutations resulting in a truncated protein. A mutation of BRAF was observed in 4 models (7%) corresponding to 3 hotspot V600E mutations and 1 infrequent mutation (D594N). Other observed gene mutations include infrequent deletions or missense mutations in the β-catenin encoding gene (CTNNB1) and EGFR (D761N), respectively, and unknown mutations in ERCC1 or FCGR2A (data not shown).

Overall, frequency of mutations in our collection is similar to the frequency reported in the literature for colon cancer (Table 2). Identical mutation profile was found in xenografts derived from different tumor sites (primary tumor and carcinosis, synchronic, or metachronic metastasis) of the same patient

The 244K-oligonucleotide arrays were used to analyze copy number alterations. The analyzed samples displayed more than 50% of tumor cells and consisted of 43 early passages (P for patient sample, P₀ or P₁) and 39 late passages (P₆–P₉) xenografts including 38 early/late paired samples. As depicted in Fig. 3A, all the samples displayed abnormal aCGH profiles with highly recurrent gains and losses. These recurrent chromosomal abnormalities include a whole loss of chromosomes 4, 14, 15, 21, and 22; a whole gain (more than 3 copies) of chromosomes 7, 13, and 20; and a loss of the long arm and gain of short arm in chromosomes 8 and 17 (Supplementary Fig. S2). About the MSI, 13/53 samples (24.5%) were MSI-Low (MSI-L) or MSI-High (MSI-H) and no correlation was found between MSI status and number of alteration on aCGH or percentage of genome with abnormalities (data not shown).

We examined the status of chromosomal regions corresponding to genes classically associated to colon tumorigenesis (Fig. 2). Although single copy modifications were frequent at these regions, no high amplification, or total loss were observed. EGFR, MET, and BRAF were each gained in more than about a half of samples (56%, 51%, and 51%, respectively), and rarely lost (2% or less of samples). KRAS and PIK3CA were gained in 28% and 19% of samples, and lost in identical proportion (29% and 19%, respectively). TP53, FBXW7, AKT1, APC, CTNNB1, and PTEN were lost in 70%, 53%, 51%, 47%, 40%, and 29% of samples, respectively and very rarely gained.

To examine the genetic stability of the xenografts through passages, we conducted unsupervised clustering based on aCGH results obtained with the 38 paired samples. Only 4 paired samples did not cluster together (Supplementary Fig. S3). Among them, 2 samples displayed a very low number of abnormalities that could result in misclassification. The 2 other samples displayed very high number of abnormalities and could be considered as highly unstable. We also directly compared profiles of early and late passages after dynamic adjustment (Fig. 3A). Most of the models displayed very similar profile with some dynamic modification for abnormalities. Nevertheless, some new abnormalities could be identified in late passages and could be attributed to either a clonal enrichment or to de novo acquisition. Overall, our data show that CRC xenografts display, in general, a high genomic stability through at least the 10 first passages.

To further characterize our models, we conducted gene expression analysis with Affymetrix U133A-microarrays on 115 samples corresponding to 40 patient tumors, 39 early passage xenografts (P₀ or P₁, 37 models) and 36 late passages xenografts (P₇–P₉, 29 models). Unsupervised clustering based on Pearson correlation with gene filtered on P value over 10⁻³ clearly separates patient tumors from xenografts but no major differences between early and late passage were observed (Fig. 3B). Xenografts from the same model clustered together as for aCGH. To identify the set of genes that discriminate patient tumors and xenografts, we conducted class comparison between tumors and early passage xenografts. Analysis revealed an 800 probe set (614 genes) more than 22,000 on array, differentially expressed with a P value under 10⁻³ and with fold change more than 2. Gene ontology analysis of gene differentially expressed was conducted on 203 annotated gene ontology functions. Among the genes downregulated in xenografts with respect to patient's tumors, we observed, as expected, enrichment in genes encoding for extracellular matrix components, collagens, and immune system regulators (Supplementary Table S4). In parallel with the lack of human stroma observed by immunohistology, these results strongly suggest that modification of gene expression is to be largely attributed to the loss of the human stromal components occurring during engrafting and tumor homing of host cells such as immunity and stromal cells. Interestingly, the differences in gene expression through the passages were relatively low, consistent with a remarkable stable gene expression profile of these models.

To identify the intrinsic chemosensitivity of each tumor model to single agent treatments, CB17-SCID mice bearing established PDX were treated with either cytotoxic chemotherapy (5-FU, oxaliplatin, or irinotecan), or with cetuximab (Fig. 2). With doubling times ranging from 2.8 to 17.0 days (median: 6.0 days), the tumor growth rate of the xenografts was slower than most of cell line-derived xenografts. The cachexia induced by the PDX was variable and was not related to tumor growth kinetics, with a median body weight loss of about 8% (range = 0–19%).

As a single agent, 5-FU inhibited the tumor growth in 35/52 tumor models. Nevertheless, taking into account clinical endpoints, such as tumor stabilization (score ++) or tumor regressions (score +++), only 7 models were found sensitive to 5-FU, with 10% to 40% partial regression being observed in 2 models. The antitumor activity of oxaliplatin was modest in this panel of colon tumors, with ΔT/ΔC values ranged from 12% to 38% in only 5 tumor models. Conversely, irinotecan showed a significant activity in 90% of the models (44/49) regardless the origin of the tumor. Tumor stabilization was observed in 20% (10/49) whereas tumor regression was achieved in 39% (19/49) of the models. Finally, 25/52 tumor models (48%) responded to cetuximab: tumor stabilization was achieved in 4 tumor models, and tumor regressions in 14 tumor models, long-term tumor free survivors being observed in 5/14 xenografts.

The chemosensitivity of these tumor models was also evaluated in nude rats. In the 44 established models, the tumor growth rate in nude rats was similar to that of severe combined immunodeficient (SCID) mice. Although we studied the chemosensitivity of only 4 nude rat models, we have observed similar ranges of tumor responses in mice and rats with the 3 cytotoxic agents (data not shown). For cetuximab sensitivity, only 1 model (CR-LBR-0022P) was evaluated in nude rat and was moderately sensitive (ΔT/ΔC = 19%) whereas CR-LRB-0022P model was not sensitive to cetuximab in SCID mice (ΔT/ΔC = 78%). Taking together, these results show the heterogeneity in tumor response to chemotherapy and the lack of simple clustering in regards to clinical, histologic, or molecular profiles.

As illustrated in Fig. 4A, cetuximab was shown very active against wild-type KRAS tumors (e.g., CR-IC-0002P), but also on mutated KRAS ones (e.g., CR-IC-0013M). We investigated the relationship between molecular profile and cetuximab sensitivity (Fig. 4B). The survival curves, corresponding to the time to reach 750 mm³, are presented for cetuximab-treated versus nontreated mice in wild-type KRAS xenografts (Fig. 4B-1), for cetuximab-treated versus nontreated mice in mutant KRAS xenografts (Fig. 4B-2). A significant difference was observed in the 2 cases using a logRank test for factor group for parameter J750 for each model; follow by a Chi2 global test (P < 0.0001; P < 0.0001). For cetuximab-treated mice bearing wild-type KRAS xenografts versus mutant KRAS xenografts (Fig. 4B-3), a significant difference was observed using a logRank test (P < 0.0001).

Nevertheless, about 42% (12/28) of the PDX displaying wild-type KRAS tumors were not responsive to cetuximab. As previously suggested by Sartore-Bianchi and colleagues (21), alterations in other genes of epidermal growth factor receptor (EGFR) pathway could also explain the resistance to anti-EGFR therapy. As shown in Table 3, the absence of response to cetuximab was significantly correlated with the mutational status taking altogether the mutations in BRAF, PIK3CA, and KRAS genes (P = 0.02, Fisher's test). We can also outline that among the tumors for which no mutation in KRAS, BRAF, and PIK3CA have been found, there are 5 tumors harboring a loss of at least 1 copy of PTEN (Fig. 2) suggesting that other genes are responsible for cetuximab resistance observed.

Of 41 models fully analyzed following engraftment into the cecum of SCID mice, 13 (32%) gave rise to metastases (at least 1 metastatic foci) mainly in mesenteric lymph nodes, liver, and lung (see Fig. 4C and Supplementary Table S3). Metastases appeared at various time, mainly between 2 and 3 months following engraftment. From the panel, no clear correlation appears between the nature of the samples (primary tumor vs. metastasis) and their ability to metastasize. It is noteworthy that the secondary sites of metastatic dissemination of theses colorectal tumors are the same as those commonly seen in the clinic (i.e., lymph nodes and liver mainly). Hence, some orthotopic PDX could be used as spontaneously metastasis models, which is of great importance for the evaluation of antimetastatic compounds. Further characterizations are ongoing to evaluate the metastatic penetrance and potential discrimination with nonmetastatic PDX models.

Appropriate preclinical experimental models are necessary to evaluate the efficacy of new anticancer drugs, therapeutic combinations, and to identify biomarkers for sensitivity. The human colorectal cancer cell line–derived xenograft models have been shown to fail in adequately predicting clinical response both in disease and compound-oriented settings (22). To implement predictive CRC animal models, we have generated, through a consortium, a large panel of tumor xenografts established by directly grafting patient tumor fragments into immunodeficient mice. Our panel includes 54 different models and constitutes the most well-annotated panel of colorectal PDX. This panel is representative of the heterogeneity of human colorectal cancer in terms of clinical parameters, histopathology, molecular pattern, and sensitivity to approved drugs. Importantly, some of the models retain the ability to develop spontaneous metastases that is highly needed for preclinical evaluation of new compounds.

Colorectal tumors display a good tumor take rate in immunodeficient mice (more than 60%) as compared with breast cancers (10–37%; refs. 7, 10) or prostate cancers (less than 5%; ref. 23). As described for other xenograft series (9, 10), tumor stage seems to play a major role in tumor take rate. Therefore, node infiltration, advanced stage, and elevated CEA in serum positively correlate with tumor take rate in mice. In addition, no major bias with respect to colorectal cancer subgroups (as defined by localization, histologic, or molecular parameters) was introduced by grafting. This is not the case in breast cancer in which triple negative tumors are selected in the grafting process (10). Importantly, the proportion of xenografts exhibiting: (i) some of the mostly described somatic mutations (APC, KRAS, BRAF, TP53, FBXW7, or PIK3CA) and (ii) low or high microsatellite instability is roughly similar to the proportion found in human CRC.

As already described for other PDX (4, 6–10), the histologic pattern of our panel is well preserved when compared with patient tumors and is consistent with the clinical diversity of human colorectal pathology. The conservation of the architecture of original tumors, including the lymphatic and blood vasculature (unpublished data), is a key advantage of these models with respect to the cell line–derived xenografts. For instance, only 8% of PDX were moderately differentiated to poorly differentiated and 10% mucinous adenocarcinomas, which is consistent with the low prevalence of these histologic subtypes. It has been shown that the stroma component of PDX becomes mainly murine after a few passages (8). We confirmed this observation, nonetheless the tumor cell enrichment, striking similarity of tumor xenograft architecture, and original tumors suggest that cancer cells could educate the microenvironment by reprogramming murine stroma cells to their benefit.

High-density aCGH and gene expression profile microarrays data allowed us to address whether significant differences exist between patient tumors and their respective xenografts. Unsupervised clustering based on aCGH data showed that most of the xenograft samples clusterized with their patient counterpart. Although the genomic alterations are similar, we could detect in the xenografts alterations absent in primary tumors. This being already described in other xenograft series (24, 25), one can speculate that minor cell populations of the primary tumors might be amplified in the xenografts by the grafting process, or alternatively that the mutational process continued to evolve in the xenograft. It has been recently reported that a triple negative breast cancer xenograft retained all primary tumor mutations and displayed a mutation enrichment pattern that paralleled the patient metastasis (26).

Gene expression profile displayed by primary tumors and xenografts were very similar. The set of genes differentially expressed reflect most probably the absence of human stroma in the xenografts. It would be very interesting to examine the gene expression profile of the murine stroma by hybridizing the xenograft samples in mouse arrays. In breast cancer xenografts, a partial recapitulation of stroma-related gene expression by murine stroma has been found (25). The stability of xenograft models with passages is crucial for establishing a preclinical platform. When early and relatively late passages of xenografts were compared, no major differences could be observed in terms of copy number alterations or gene expression profile. This clearly suggests the stability of colorectal xenograft models, as depicted for breast and pancreatic tumor xenografts. Therefore, these models could be used for drug candidate evaluation.

Pharmacologic studies conducted with drugs currently used in CRC allowed us to validate the value of our xenograft panel for evaluating novel drugs. Single agents were used and revealed a high degree of pharmacologic heterogeneity of the panel. Overall, the best responses were found with irinotecan with tumor regressions in almost 40% of the models. As a single agent in patients with newly diagnosed CRC, irinotecan has been reported to generate response rates in the range of 19% to 32% (27). Treatment with cetuximab resulted in tumor regressions in 23% of the models. In light of the clinical data, better responses in the wild type KRAS tumors (48%) were expected. It is interesting to note that a similar proportion (42.4%) of KRAS wild-type cetuximab nonresponsive tumors has been found by Bertotti and colleagues in their collection of metastatic CRC (7). The comparison of the survival curves of cetuximab-treated mice (Fig. 4B) shows a better survival for wild-type KRAS versus mutated KRAS tumors was observed. Interestingly, a positive correlation was observed between the lack of response to cetuximab and the mutational status KRAS, BRAF, and PIK3CA genes.

In our goal to develop a platform approach, these tumors were also successfully engrafted on nude rat, known to exhibit metabolism and drug pharmacokinetic/pharmacodynamic profiles closer to human ones than mice (28). Although nude rats are less immunodeficient than nude mice, take rate, tumor growth, and pharmacologic profiles were comparable. Therefore, these PDX could be used in the rat for drug efficacy studies and pharmacokinetic/pharmacodynamic studies.

Pancreatic PDX have been reported to constitute a very valuable pharmacologic tool for drug development, in particular through a remarkable correlation between drug activity in xenografts and clinical outcome, both in terms of resistance and sensitivity (6). Moreover, such PDX models have been used to identify CRC patient subpopulation that could benefit from HER2-targeted therapy (7). All together, these results support the use of well-characterized PDX models as a powerful investigational platform for efficient therapeutic decision-making steps in the clinic through biomarkers identification. This panel is currently being used in a prospective way for target expression and evaluation of personalized therapies by the members of the consortium.

For commercial purposes, an exclusive license for the patient-derived tumor xenografts described in this article has been granted by the CReMEC consortium to Oncodesign. No potential conflicts of interest were disclosed by the other authors.

The authors gratefully acknowledge the patients who accepted to contribute to this research program and their family. We also thank the biologic resource centers and departments of histopathology of Curie Institute, Gustave Roussy Institute, and Lariboisière Hospital for their most important contribution in the patient tumor collection. The authors are thankful to Jean-Jacques Fontaine, Sophie Chateau-Joubert, and Jean-Luc Servely for their support in ALU hybridization experiment. The authors are grateful to Cancéropôle Île-de-France and Medicen Paris-Region biocluster for providing support in the consortium coordination. The authors really appreciated the help of Emmanuel Canet, Antoine Bril, Philippe Genne, Gilbert Lenoir, Christine Perret, Sylvie Robine, Christophe Thurieau, and Gilles Vassal for discussions, advice, and continuous support. The authors are thankful to Phil Kasprzyk and Francis Bichat for comments on the manuscript.

All the authors are members of the CReMEC (Center of Resource for Experimental Models of Cancer) consortium, which has benefited from a financial support of the French Ministry of Industry. This grant was part of the “Fonds unique interministériel” program (FUI) and initially selected by Medicen-Paris Region biocluster. In addition, the work conducted at the Gustave Roussy Institute was supported by “Département du Val de Marne.”

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.