Genomic characterization of explant tumorgraft models derived from fresh patient tumor tissue

Although direct transfer xenografts of fresh human tumors, or primary “tumorgrafts”, have been reported as early as the 1970s for testing of new pharmaceutical agents [1‐4], only recently has there been a resurgence of interest in this alternative to the more traditional cell line xenograft models. This has been due, in part, to the growing realization that drugs which work in the traditional cell line xenograft models rarely exhibit comparable efficacy in patients with the anatomically/pathologically equivalent tumor [5‐7]. Indeed, it has been highlighted that the failure of drugs during clinical trials is linked to a lack of testing in clinically relevant preclinical models [8]. The advent of “precision medicine” (PMed), utilizing novel molecular techniques to design patient tumor-specific therapeutic regimens, requires preclinical cancer models that closely reflect the originating human disease to allow evaluation and rapid translation of these molecular based approaches to therapy selection in the clinic [9].

While there have been a growing number of studies reporting comparative chemotherapeutic responses between tumorgraft models and patients [10‐13], there is limited data addressing genomic similarities between tumorgraft models and the originating patient tumors [14‐16]. Previous studies have reported excellent concordance in expression profiles between patient tumors and tumorgraft models of similar histotypes [17, 18]; however, the comparisons were not made in a pairwise fashion between each tumorgraft and the originating tumor from the donor patient.

Given the added cost and time required to develop primary patient tumorgrafts relative to the more classical cell line xenografts, a detailed cellular and molecular characterization of the tumorgraft models showing a high degree of equivalence to the originating patient’s tumor in the human host could provide the justification for the routine use of this model, even at the ectopic subcutaneous site, in translational studies and personalized medicine applications. Our goal was to compare the genomic profiles of a panel of diverse tumorgraft models to the patient tumor tissue from which the models were derived. To address inter-disease variability, the panel of tumorgraft models selected for this study contains a spectrum of neoplastic diseases, rather than a limited set of disease types [14‐16, 19]. Our long term objective is to develop this panel of mixed-type tumorgraft models to evaluate new treatment strategies based on the molecular characterization of an individual disease and the subsequent treatment with targeted therapy, independent of the tumor histology and anatomical location.

Forty-nine tumorgraft models, representing 18 distinct cancer pathologies have been developed by direct implantation of patient tumor tissue into immune-compromised nude mice giving the overall take rate of 27% (Table 1). These 49 models were distributed across eight different cancer location categories, based on the NCI Cancer by Body Location/System (Table 2). Tumorgraft models of the gastrointestinal tract and skin account for ~50% of successfully established tumorgraft models. There were an additional three models that formed 1^st generation tumorgrafts but that failed to re-establish tumorgrafts following cryopreservation of 1^st generation tumorgraft fragments, giving a 94% success rate in tumorgraft model re-establishment following cryopreservation. No association was identified between tumorgraft development and the time between surgical resection and implantation (p = 0.86 by simple t-test on representative sample set, data not shown).

The take rate (the ability of donor tumor to successfully be propagated in mice) was closely related to tumor stage at time of resection. Forty-three percent (35/82) of tumors with a pathological staging of Stage III and Stage IV tumors were successfully propagated in mice, accounting for 73% of fully developed models (Table 2). Only 10% (6/60) of the donor tumors from Stage I or II cancers successfully propagated formed tumorgrafts in mice, accounting for only 12% of the fully developed models. Pathological staging information for the remaining 22% of the tumor tissues implanted into mice (40/182) which formed tumorgrafts was not obtained. A contingency table analysis with stage III and IV as one condition and stage I and II the second condition supported the hypothesis with a p-value = 1.2 E-5 and an odds ratio of 6.7 for this data set (data not shown).

Tumorgraft development was also found to be related to cancer type. Of the human tumor samples used to derive the panel of 49 tumorgraft models, a success rate of >40% was observed for skin (65%), musculoskeletal (54%), head and neck cancers (50%), and gastrointestinal cancers (41%) (Table 1). Lung (26%), genitourinary (22%), and endocrine (20%) cancers exhibited lower take rates for tumorgraft formation. Cancers of the breast (5%), predominantly early stage (data not shown), and the lymphatic system (0%) proved the most challenging to form tumorgrafts using the athymic nude mouse strain and methodologies adopted here. No tumorgrafts were established from tumors of gynecologic, leukemic, or neurologic origins; however, these finding are not necessarily representative of the true take rate for these tumor types as ≤2 attempts for each of these cancer types were attempted.

Hematoxylin and eosin stained sections from matching patient tumors and 1^st and 2^nd generation tumorgrafts from pancreatic cancer, breast cancer, lung cancer and cervical cancer models were examined for histological similarities. In all 4 different tumorgraft models, the initial patient tumors were characterized by histological features consistent with the anatomical site of origin (Figure 1 A-D). The 1^st generation tumor grafts resembled the original patient tumor, including in most instances retention of peritumoral stroma which became progressively replaced in the 2^nd generation tumorgrafts by malignant cell overgrowth. In addition, the relative state of differentiation of the respective cancer type was retained following engraftment, including production of mucin and retention of atypical glandular structures for the adenocarcinomas, even after 302 and 177 days for the pancreatic and breast cancers, respectively. In the lung cancer and cervical cancer models, the 1^st and 2^nd generation tumors revealed viable malignant appearing cells. The 2^nd generation tumorgrafts retained essential diagnostic features at 286 and 345 days, respectively, for the lung and cervical cancer models.

Of the 20 samples (10 patient samples, 10 1^st generation tumorgraft samples) submitted for oncogene mutation analysis, only 15 mutations were identified. Four tumorgraft models each had a single point mutation that was also observed in the originating tumor tissue. Models VARI-LTM-044 colorectal cancer and VARI-LTM-034 pancreatic cancer both had a single KRAS mutation; melanoma models VARI-LTM-086 and VARI-LTM-027 had BRAF and NRAS mutations, respectively (Table 3). Model VARI-LTM-041 colorectal cancer contained three mutations (KRAS and two PIK3CA mutations) that were also present in the matching tumor. The colorectal cancer model VARI-LTM-026, did not exhibit mutations in the patient tumor, but showed a single KRAS mutation in the tumorgraft. In the osteosarcoma model VARI-LTM-021, no mutations were identified in the patient tumor with any genetic drift identified in the 1^st - 4^th generation tumorgrafts (data not shown).

Paired analysis of the 24 tumorgraft/tumor pairs exhibited a high degree of similarity, with Pearson correlation coefficients ranging from r = 0.75 to 0.99 with a median coefficient of 0.93 (Table 4). There were only four pairs, three gastrointestinal and one lung, that had Pearson values r < 0.90; the NSCLC model VARI-LTM-023 had the Pearson correlation coefficient of r = 0.75. For this model, the originating patient tumor RNA had a RIN (RNA Integrity Number) of 4.9 implying degraded RNA, which was likely to be the primary factor driving the low correlation and would typically be excluded from the molecular analysis.

Of the 28 genomes amplified by the Affymetrix method, 12 of the 14 tumorgraft/donor tumor pairs co-clustered in an unsupervised analysis (Figure 2A). Of the 20 genomes amplified by the NuGEN method, 7 of 10 tumorgraft/ tumor pairs coclustered (Figure 2B). 4 of the 5 tumorgraft/tumor pairs that did not co-cluster in an unsupervised fashion exhibited Pearson correlation coefficients r ≤ 0.90 (Table 4). The tumorgraft and originating patient tumor from the NSCLC model VARI-LTM-023 co-clustered despite a low Pearson correlation coefficient of r = 0.75, suggesting significant variance in respect to other tumors in the cohort.

The genomic profiles of two representative osteosarcoma models VARI-LTM-020 and VARI-LTM-021 were observed over 4 generations to assess the potential for genomic drift over time in vivo. The Pearson correlation coefficients were very high, ranging from r = 0.98-0.99 between all 4 generations in both models (data not shown). The largest decrease in Pearson correlation coefficient occurred between the human tumor to 1^st generation mouse tumorgraft (VARI-LTM-020, r = 0.91; VARI-LTM-021, r = 0.94). Genomic fidelity was then maintained in subsequent mouse to mouse generation comparisons. This stability in the genomes across mouse tumorgraft generations, along with the absence of additional somatic oncogene mutations suggests the maintenance of genomic stability during serial in vivo passages and time in vivo for at least up to 186 days (VARI-LTM-021).

Paired analysis of all 24 tumorgrafts to their matching tumors identified 25,806 gene probes that were present in at least 50% of the 48 samples. 395 genes were decreased and 17 genes increased two-fold or greater in the tumorgrafts (Additional file 1: Table S1) compared to the patient tumors (Figure 3A, green and red circles respectively).

Of the 395 genes down-regulated in the murine tumorgrafts, there was a significant enrichment in pathways related to immune response (Figure 3B). Enrichment scores were extremely high, with the top-scored classical complement pathway showing a –log p-value of 25. Of the 17 genes up-regulated in the tumorgrafts, there were only 1–2 genes per map, suggesting no enrichment for any canonical pathway (data not shown). An independent GO biological pathway enrichment analysis was conducted using the down regulated probe list and 14 of the most significantly enriched top 100 pathways p < 4.0E-9 were associated with immune response [23]. Classification of the down-regulated genes with regards to enrichment in protein function, identified a borderline significance (p = 0.05) in receptors and ligands in the tumorgrafts compared to the patient tumors. For the up-regulated genes, the only significant enrichment in protein function was for proteases, accounting for only two of the 17 genes.

Comparative analysis of the patient tumors that formed tumorgrafts to patient tumors that did not form tumorgrafts identified 491 genes that were up-regulated and 691 genes down-regulated 2-fold or greater in those tumors that formed tumorgrafts compared to non-tumorgraft forming tumors (Additional file 2: Table S2). The up-regulated genes in the tumors that formed tumorgrafts were enriched for cell cycle, cell signaling pathways and cytoskeleton remodeling (Figure 4A). One key canonical pathway map, enriched with up-regulated genes that were elevated in those tumors that formed tumorgrafts, is for WNT signaling (Figure 5); genes up-regulated in this pathway map included VEGF-A, Frizzled, β-Catenin, c-Myc, and FAK-1. The down-regulated genes in the tumorgraft-forming tumors were enriched for immune response pathways (8/10 pathways, Figure 4B). One key pathway map enriched with down-regulated genes is that of the role of integrins in NK cell cytotoxicity (Figure 6); note that in the context of this study, the upper Target Cell in this pathway corresponds to the relevant tumorgraft-forming cells.

The goal of this study was to create a panel of tumorgraft models, developed directly from patient tumor tissue from a wide range of heterogeneous tumor types of various stages. Prior investigations have focused on specific cancers, e.g. Secondary liver cancer [15] lung [24], pancreatic cancer [16], pediatric osteosarcomas [14], or pediatric rhabdomyosarcoma [25], rather than a spectrum of malignancies. To examine the suitability of the panel of tumorgraft models generated in our lab as molecularly-relevant preclinical models, we directly compared a number of phenotypic and genotypic features of the tumorgrafts to the originating patient tumors from which they were derived.

We have fully developed 49 models, spanning 18 different cancer types. A critical component of our development process is the ability of the model to re-establish following cryopreservation, alleviating the need for continual propagation, reported by some investigators [16]. Cryopreservation significantly reduces the cost and necessary resources, extending the life of the tumor model, and minimizing genetic drift that could occur following long-term continuous in vivo propagation.

Our overall tumor take rate is similar to previous reports [3, 5]. A number of strategies have been tested to improve tumorgraft take rates, including the severe combined immunodeficient mouse strains CB17SC-M-F scid ^−/− [9, 25, 26] and CB17/Icr scid[12], suppression of the immune system of recipient CBA/CaJ mice by thymectomy, whole body irradiation, administration of 1-β-D-arabinofuranosylcytosine [14, 25], and use of orthotopic models [27]. However, these strategies could increase the cost of model development with minimal improvement in tumorgraft development over the use of subcutaneous implantation into naive athymic nu/nu mice reported in this study and by other investigators [2, 16].

It should be noted use of the athymic nu/nu mouse for use in the development of such tumorgraft does have limitations. The use of more immunodeficient mouse strains, e.g. NOD/scid or NOD/scid gamma (NSG), are necessary for successful engraftment of certain tumor types e.g. leukemia [28, 29], may better resemble the human microenvironment, and could provide better consistency in growth rates following implantation of human tumor tissue into mice. However, in the context of developing a panel of tumorgraft models across a wide range of tumor types, the athymic nu/nu mouse has proven a robust model.

The stage of the cancer at the time of acquisition has an impact on a successful tumorgraft development. Only 10% of Stage I and II tumors successfully formed tumorgrafts following implantation into mice; this is in contrast to the 43% take rate of tumor tissue from patients with Stage III or IV cancer at time of tissue acquisition which was successfully propagated in the murine host. Overall, ~75% of the 49 tumors that formed tumorgrafts were from patients that were diagnosed either Stage III or IV.

In this study, breast cancers had a take rate of only 6% following subcutaneous implantation into mice. Dobrolecki et al [27] reported success in developing breast cancer models by orthotopically implanting the breast cancer tissue into the mammary fat pads of mice. While more labor intensive than the subcutaneous model, an orthotopic approach should be considered if there is need for a tumorgraft model of a specific cancer that cannot be established using the subcutaneous model.

An important conclusion from this study is that these subcutaneous (predominantly ectopic with the exception of melanomas) tumorgraft models recapitulate the expression profiles of the original patient’s tumor within its natural site of site of metastatic spread. This provides some evidence that refutes the lack of relevance of the subcutaneous model for drug discovery and development, which may rather reflect the use of genetically drifted in vitro cell lines in the classical xenograft context. The primary value for any tumor model is to recapitulate the in-vivo malignant tissue state, and accurately predict the efficacy of therapeutic agents in a timely and cost-effective manner. In contrast with commonly used animal models in which established human cancer cell lines are injected subcutaneously and fail to accurately predict human clinical trial results [8], this human tumorgraft model system is more likely to be a successful surrogate preclinical model for several reasons.

First, tumor bearing tissue fragments are transplanted containing not only malignant cells, but also supporting stromal tissue in an anatomically correct hierarchy. Second, the transferred tumors include cells post engraftment retaining their essential cell to cell interactions and relative state of differentiation as assessed by their histological appearance. Third, because of the high success rate in establishing not only 1^st, but also subsequent generation tumorgrafts, there is expansion of the number of tumorgrafts available for testing and the ability to verify retention of histological hallmarks in a serial fashion. Fourth, assessment of tumor growth is relatively straightforward given the subcutaneous location of the xenografts. Finally, since portions of tumors can be cryopreserved and successfully engrafted post thawing, testing of therapeutic agents, or combination therapies not originally considered, can be studied with this experimental and scalable in vivo model system.

In the era of molecularly targeted oncology agents, the effective translation of results in the mouse model to the clinic requires the genome of the tumorgraft model to be high comparable to the original tumor [30]. Significant changes in DNA fidelity or RNA expression levels significantly diminish the translational value of a model. Entire genome characterization of RNA expression rather, than a focused RT-PCR analysis of a select cohort of genes [16], allows for the assessment of transcriptome wide alterations and associated biological systems that occur during the growth of the transplanted human tumors.

The high Pearson correlations seen between the genomes of the 24 pairs of tumorgraft and originating patient tumors implies that the donor tumor genome is largely maintained in the tumorgraft model. Similarly, genotypic fidelity has also recently been reported in a panel of 25 human breast cancer tumorgrafts [27] and on a smaller scale in secondary liver cancers using quantitative PCR analysis of 21 genes related to oncogenesis and cell cycle [15]. The lower Pearson correlations observed in the gastrointestinal cancers could be an artifact of the tissue harvest and the high levels of digestive enzymes present in the patient tumors within the gastrointestinal tract. High levels of RNases present in pancreatic cancer are reported to complicate the extraction of high quality RNA [31]. This finding highlights the importance of an optimized, rapid tissue collection protocol, especially for genomic profiling. It should be noted that poor RNA quality, for genomic profiling does not necessarily impact the ability of the tissue to form a viable tumorgraft. The high degree of clustering by matched tumor/tumorgraft pairs (79%) supports the high degree of similarity of the genomes of the patient tumor and resulting mouse tumorgraft even when the opportunity for RNA degradation may occur.

Genomic stability is an important characteristic to consider prior to the long term use of a tumorgraft model. The high correlation in gene expression across four tumorgraft generations observed in this study and similar findings of others [3, 16, 27, 30] and absence of drift in somatic mutations in known oncogenes, suggest that the genomes of tumorgraft models are stable. Histological integrity of patient-derived tumorgrafts has also been demonstrated for up to ten generations in immune-compromised mice [4] and up to 30 passages in mice without significant changes in growth and morphological characteristics [15]. One study reported tumorgraft models cultured in vivo for 10–12 generations before regenerating the models from earlier cryopreserved generations [3].

The role of somatic oncogene mutations in tumorigenesis, pathogenesis and disease progression can have a profound influence on therapy [32]; mutations in oncogenes RAS [33] and B-RAF [34] are commonly found in a variety of cancers. Therefore it is imperative that such mutations are maintained in the tumorgraft models. Although only a few oncogenic mutations were observed, those identified in the patient tumors were conserved in the matching tumorgraft, consistent with other reports [16] and consistent with known mutations in specific tumor types [20, 21].

There are a number of possible explanations for the enrichment of canonical immune pathways within those genes down-regulated in the tumorgrafts compared to patient tumors. First, a loss of tumor cell immune pathways and proteins not necessary for growth and development when transplanted from the immunecompetent patient into the immunocompromised host. Second, there is an intentional tumor tissue response to transplantation to down regulate those genes responsible for its immunogenic signature presented to the vestigial immune defenses of the immunocompromised mouse. A third explanation, and most likely, is that the gene expression changes are due to a loss of circulating human immune cells which infiltrate and support tumor development, when the tumor was transplanted from the human into the mouse as earlier reported by Neale et al [17].

In those patient tumors that formed tumorgrafts, up-regulated genes show enrichment in pathways considered “usual suspects” in highly proliferative, late stage tumors; these canonical pathways include cell cycle, cytoskeletal remodeling, and those known drivers of tumorigenesis WNT and AKT signaling [35‐37]. These same tumors were elevated for vascular endothelial growth factor (VEGF). This increased level of VEGF in the tumor could support the neovascularization of the tumor once implanted into the mouse; Nisolle et al [38] reported that the survival and growth of human endometrium transplanted into nude mice was correlated with a high VEGF content. Transgenic expression of VEGF into human islet cells followed by transplant into the livers increased the rate of tissue revascularization, survival and function [39]. A recent study demonstrated the pre-treatment of human ovarian tissue with VEGF and vitamin E prior to implantation into the back muscle of immunodeficient mice demonstrated enhanced growth, compared no pre-treatment of the ovarian tissue [40].

The downregulation of immunological pathways in the tumors that formed tumors would suggest that loss of immune pathways is a mark of late stage tumors and provides a selective advantage supporting the establishment of the tumors in a foreign host or location. This downregulation of immunological pathways could be a reflection of the loss of expression signature of patient immune infiltrating cells [17]. Of particular note is the downregulation of natural killer (NK) cell ligands, as identified as a pathway enriched for genes down-regulated in tumors that formed tumorgrafts compared to those tumors that failed to form tumorgrafts (Figure 4B). Although athymic nu/nu mice cannot generate mature T lymphocytes, they maintain the ability to mount a response to T-independent antigens [41] and do produce NK cells as a component of their vestigial immune system. The downregulation of the NK ligands may therefore promote human tumor development in immune compromised mice. While there is little evidence demonstrating the interaction of human integrins and murine NK receptors, Kanwar et al reports interspecies cross-talk via the demonstration that the ectopic expression of human intercellular cell adhesion molecule-1 (ICAM-1) in mouse EL-4 tumors inhibited in vivo tumor growth [42]. The mechanisms of NK cell recognition of self and immunosurveillance in both mouse and humans are complex and involve numerous receptors that recognize many cell surface ligands [43]. Experimental evidence clearly supports the view that murine NK cells can inhibit human tumor cell development and metastasis both in vitro and in vivo[44‐50].

The maintenance of the genome in the tumorgraft, from the human to murine host, supports the concept that tumorgraft models can provide an invaluable and robust tool in the evaluation of precision medicine methodologies and novel treatment strategies prior to use in clinical trials. An important point is that a subcutaneous primary patient tumorgraft models can recapitulate the expression profiles of the original “orthotopic” disease in the human and allow for reasonably high-throughput preclinical screens for drug efficacy. These models may therefore be useful in reducing the rate of attrition in drug development that has been in part been attributed to the use of the classical cell line xenografts.