Establishment of patient derived xenografts as functional testing of lung cancer aggressiveness

Despite many years of research efforts, lung cancer still remains the leading cause of cancer deaths worldwide. Objective of this study was to set up a platform of non-small cell lung cancer patient derived xenografts (PDXs) faithfully representing primary tumour characteristics and offering a unique tool for studying effectiveness of therapies at a preclinical level. We established 38 PDXs with a successful take rate of 39.2%. All models closely mirrored parental tumour characteristics although a selective pressure for solid patterns, vimentin expression and EMT was observed in several models. An increased grafting rate for tumours derived from patients with worse outcome (p = 0.006), higher stage (p = 0.038) and higher CD133+/CXCR4+/EpCAM− stem cell content (p = 0.019) was observed whereas a trend towards an association with SUVmax higher than 8 (p = 0.084) was detected. Kaplan Meier analyses showed a significantly worse (p = 0.0008) overall survival at 5 years in patients with grafted vs not grafted PDXs also after adjusting for tumour stage. Moreover, for 63.2% models, grafting was reached before clinical recurrence occurred. Our findings strengthen the relevance of PDXs as useful preclinical models closely reflecting parental patients tumours and highlight PDXs establishment as a functional testing of lung cancer aggressiveness and personalized therapies.

Here, we report the establishment of a large panel of LcPDXs derived from NSCLC patients, which closely recapitulate and maintain primary tumour features over 10 passages in mouse. Interestingly, we found significant correlations between PDX take rate and patient's prognosis and survival suggesting PDX grafting as a surrogate functional testing to anticipate lung cancer aggressiveness. Targeted mutation sequencing showed maintenance of genetic profile in LcPDX compared to respective parental tumours. Finally, the analysis of growth characteristics allowed us to identify among LcPDXs a group of faster growing models that become stable in mice before disease recurrence in patients. These group of LcPDX may be utilized as "avatars" for testing personalized treatment strategies.

Results
Establishment of a lung cancer PDX perpetual bank. Tumour samples from 97 lung cancer patients (65 adenocarcinomas -ADC, 16 squamous cell carcinomas -SCC and 16 other lung tumours or metastases -OL) have been implanted in both flanks of SCID or nude mice. Overall, 49/97 (50.5%) samples successfully gave rise to a P1 generation in mice. However, grafted PDXs showed variable and generally slower growth rate during the first three passages in mice (P1-P3) and 11 of them were lost in this time frame. In particular, latency period and time to transplantation were generally higher and more heterogeneous in P1-P3 period, whereas beyond P3, PDX growth was more homogeneous (Supplementary Figure 1). At P3, 38 PDXs (39.2%) were considered grafted and suitable for further analysis. In detail, 23/66 ADC (34.8%); 7/16 SCC (43.8%) and 8/15 OL (53.3%), including 2 large cell, 1 sarcomatoid, 3 small cell carcinomas, 2 lung metastases from sarcoma and oral cavity carcinoma successfully grafted (Tables 1 and 2). Furthermore, we set up a freeze/thawing procedure on tumour samples derived from established PDXs that was thoroughly exploited to establish a large and continuously growing collection of frozen samples (Supplementary Table 1).
LcPDX models recapitulate primary tumour features. Histopathology. Pathology examination and molecular analyses of the established PDX models and their parental tumours were carried out to confirm the accuracy of these in vivo models. Immunohistochemistry analysis of 27 PDXs confirmed the parental tumour diagnosis corresponding to 20 ADC, 4 SCLC and 3 SCC. Interestingly, the same profile was maintained for more than ten passages in mice (Fig. 1A, Table 3). In the 20 ADC, comparative analysis of histologic patterns, stroma, necrosis percentage and immunohistochemistry markers (TTF-1, p40, vimentin, Ki-67 antigen and synaptophysin) showed a general maintenance of the relevant profiles over several passages in mice. In particular, heterogeneity of histologic patterns in parental tumour was maintained in the corresponding LcPDXs, although there was in late passages (P > 10) a slight tendency for solid pattern to prevail and for acinar and cribriform patterns to be under-represented. Indeed, solid pattern was appreciable in 6 primary tumours, in 7 PDXs at P ≤ 10 and in 8 PDXs at P > 10; acinar pattern in 4, 3 and 2 and cribriform pattern in 6, 5 and 4, respectively (Table 3 and Supplementary Figure 3). Stromal component was generally lower in LcPDXs compared to parental tumours, but a higher level of stromal cells percentage was present in both early (P ≤ 10) and late stage (P > 10) PDXs than in cell line-derived XG (Fig. 1C, Table 3 and Supplementary Figure 3).
Analysis of marker expression highlighted that, apart from a general similarity (Fig. 1B), there was a slight decrease of TTF-1 and an increase in vimentin in PDX derived from ADC. Indeed, TTF-1 was expressed in 6 primary tumours, 4 LcPDX at P ≤ 10 and 4 LcPDX at P > 10;and vimentin in 4 primary tumours, 6 LcPDX at P ≤ 10 and 8 LcPDX at P > 10. Moreover, for LT63, LT220 and LT268 complete loss of TTF1 and marked acquisition of vimentin and Ki67 labelling index was observed in mice ( Fig. 1D, Table 3 and Supplementary Figure 3). In addition, an analysis of specific tumour cellular subpopulations showed that CD133 + cancer initiating cells (CICs) were maintained in LcPDXs with a content similar to that observed in primary tumours (Fig. 1E).
Response to treatment. Eighteen PDX models were characterized for their responsiveness to cisplatin. Five of them (responders) reached at least a partial response (PR, median of four tumors) upon treatment, whereas in the other 13 a progression of the disease (PD) was always appreciable (non responders). Interestingly, overall survival (OS) of patients from whom responder PDXs were derived was higher than that of patients which gave rise to non-responders PDXs. (Fig. 1F    PDX grafting is associated with patients survival. Clinical characteristics of patients were analyzed to investigate PDXs take determinants. Association of sex, age, smoking habits, lung functionality parameters (FEV1, FEV1/FVC and COPD) and clinical outcome with tumour grafting capability were examined. As shown in Table 1 and Table 2, a worse outcome was significantly associated with an increased likelihood of grafting (p = 0.006). An increased grafting probability was observed also for higher stage tumours (p = 0.041 for Stage II, III, IV vs. Stage I and p = 0.038 for Stage III,IV vs. Stage I,II) and tumours with higher CD133 + /CXCR4 + / EpCAM − subpopulation (p = 0.019) whereas a trend towards increased grafting for tumours with SUV max higher than 8 (p = 0.084) was observed ( Table 1). The correlation between clinical outcome and PDX grafting was further analyzed using Kaplan Meier and Cox analyses. At the time of the present analysis, 42 patients died and the OS rate at 5 years was 51% (95% CI, 39-62%). Interestingly, OS at 5 years was significantly worse (p = 0.0008) in patients with grafted PDXs (36%; 95% CI, 20-52%) than those with not grafted PDXs (61%; 95% CI, 43-74%) ( Fig. 3A), also considering tumour stage (73%; 95% CI, 56-85% in stages I/II vs 20%; 95% CI, 10-36% in stages III/IV, p < 0.0001) (Fig. 3B). In multivariable Cox analysis, grafting PDXs and tumour stage significantly influenced OS (  Table 2). These data highlight that the capability of tumour samples to establish LcPDXs is an inherent characteristic of primary tumours linked to the biological aggressiveness potential, thus acting as an adverse prognostic marker.
Grafted PDX may act as "Avatars" for personalized treatment. We further analyzed LcPDX growth characteristics, categorizing our platform based on average time to transplantation (TT-time from implant to explant) in the first 10 passages (Fig. 3C). Average TT for all LcPDX was 44.95 days; average TT for SCC-derived LcPDXs felt almost all around this value, SCLC-derived LcPDXs displayed a slower TT, whereas ADC-derived LcPDXs were split into two distinct groups (faster growing, fgADC and slower growing, sgADC, Fig. 3D). Although no tumour or patient characteristics clearly discriminated fgADC from sgADC, sgADCs were enriched in Stage I tumours (37.5% vs. 4.4% in sgADC and fgADC respectively) and in tumours bearing CTNNB1 mutations (33.3% vs. 14.3% in sgADC and fgADC respectively) and were mainly derived from heavy smokers (Supplementary Tables 2 and 3).
In order to clarify if a co-clinical approach could have been feasible for LcPDXs, we analyzed their growth characteristics according to patient survival, and found no correlation between the time required to reach P3 (Time to Reach Grafting -TRG) and the patient DFS (Fig. 3E), indicating that faster LcPDXs were not necessarily derived from patients with lower OS or DFS. Indeed, for 24/38 (63.2%) LcPDXs (9 Stage I, 6 Stage II and 9 Stage III/IV), TRG was less than DFS of the relevant patients (Fig. 3E). Thus, these LcPDX may have been exploited for meaningful drug testing, before tumour recurrence occurred in patients. Interestingly, this "Avatar" approach could have been put to use to give insights for the treatment of 25.5% (13/51) patients with progression of disease: 16.7% (1/6) Stage I, 33.3% (4/12) Stage II and 24.2% (8/33) Stage III/IV (Table 4).

Discussion
We reported here the establishment (with 39.2% grafting rate) of a large platform of lung cancer PDXs. Although the study population represents less than 10% of all patients who underwent mediastinal biopsy or pulmonary resection with curative intent for primary lung cancer during the same period, there was no selection by major prognostic factors, and median survival of the 97 patients providing PDXs samples was very similar to the one of the other 953 patients (47 vs 57 months). These models recapitulated all the features of original tumours. In particular, tumour architecture was maintained in our PDXs, with stromal content and main histological patterns being similar to parental tumours at least in early passages (P ≤ 10). Interestingly, cellular patterns linked to non aggressive lung cancer subtypes (i.e lepidic) 25 were lost in our models, whereas a slight tendency toward a selection for more solid, vimentin-expressing tumours was observed in late mouse  passages. This suggests at the one side that grafted models could be preferentially derived from more aggressive tumours and at the other side that a tumour evolution towards a higher aggressiveness was appreciable also during PDX passages in mouse. A tumour evolution was strongly suggested by LT63, LT220 and LT268 behaviour in mouse, indeed in these models an epithelial to mesenchimal transition (EMT) was appreciable, with loss of TTF-1, acquisition of vimentin and Ki67 expression. Interestingly, LT220 and LT268 were the only two analyzed tumours for which no mutations were detected, suggesting either that they could carry other rare mutations or that they are not cell-autonomous and their tumour growth and progression are mainly dependent on tumour adaptation to microenvironment (ME). These data suggest a progressive adaptation of PDXs to murine ME. Since PDXs fully recapitulated primary tumour characteristics, it can be argued that these models could also recapitulate the natural history of tumours. Indeed, in our models, the cross-talk between tumour (human) and microenvironmental (murine) cells seemed to produce a selective pressure that advantaged solid patterns, vimentin expression and eventually complete EMT. Thus, although PDXs did not retain original tumour ME and therefore are not suitable for investigating non-cell autonomous (stromal and immune) drivers of tumour evolution 26 , our observations suggest that murine ME may be partly able to vicariate human tumour ME also in driving malignant progression. Thus, LcPDXs may be an operational model to investigate stromal mechanisms underlying cancer evolution towards dissemination (EMT) and aggressiveness and can therefore be also suitable for preclinical studies involving ME (stroma and innate immunity)-directed treatments.
The analysis of clinico-pathological characteristics indicated a good correlation between PDX establishment and survival (OS and DFS). A multivariable Cox analysis confirmed an higher HR for grafted PDXs also when adjusting for Stage, SUV max , Sex and Age. These data confirmed that aggressive tumours have an advantage in terms of grafting capability, as suggested by pathology analysis and as already reported for early stage lung Pattern ( Table 3. Patients and tumours histopathological analysis. (1) Number of positive models; *Analysis of 10 PDXs that reached P > 10 in mouse; **Resulting from a pattern loss in one PDX and a pattern acquisition in another PDX.
cancer PDXs 19 . A trend for metabolically more active tumours (SUV max > 8) to graft in mice was also observed. A low primary tumour SUV max was reported as a predictor of long-term survival and as a tool to identify NSCLC patients without lymph node involvement 27,28 . Lung tumour mortality is mainly due to metastasis development, and it has been reported that PDXs derived from lung cancer brain metastasis showed a higher take rate compared to PDXs derived from primary tumours 23 . However, we found no correlations between PDX establishment and TNM staging of primary tumour. On the contrary, we found a tendency towards a higher grafting capability for tumours according to the presence of CD133 + /CXCR4 + /EpCAM − cells. We previously reported the heterogeneity of CD133 + cancer initiating cells 29 and in particular the capability of the CD133 + /CXCR4 + /EpCAM − subpopulation to sustain tumour dissemination 30 . All these data suggest that aggressiveness of grafted tumours may mirror an inherent biological trait of tumours and that grafting capability may depend on the presence of cells that are able to seed in a non orthotopic soil. Interestingly, LcPDXs showed enrichment in mutation rate of TP53 in all tumour types and of KRAS in ADC where the frequency of mutations was almost double than that expected   The assumption of KRAS mutations as prognostic markers in lung ADC is controversial [32][33][34][35][36] , but the high rate of KRAS mutations in grafted PDXs seem to support this hypothesis. Grafted PDXs were categorized based on growth rate and ADC-derived PDX were divided in two groups (fgADC and sgADC). No patient or tumour characteristics could discriminate these two groups, probably because of the low number of models (14 fast and 8 slow) under evaluation. Additional work will clarify if these two groups were derived from tumours or patients with distinct biological traits, in particular to confirm a prevalence of stage I tumours from heavy smokers or CTNNB1-mutated tumours to give rise to sgADC or a slightly higher tendency of KRAS-mutated tumours to give rise to fgADC. As a role of aberrant WNT signalling in response to cigarette smoke has already been reported 37,38 , sgADCs could represent tumours of heavy smokers with impaired WNT/beta catenin pathway. Accordingly, sgADCs might represent a tumour subgroup with decreased tumour-microenvironment cross-talk capability, either because of the lack of smoke-induced lung inflammation in mouse or due to a particular mutational status.
One of the main advantages of establishing LcPDXs is that parental tumour traits can be conserved over time, making biological studies available even after patient death (Supplementary Figure 4). This represents a unique opportunity to investigate specific therapies in tumours with different pathological and molecular features as already demonstrated, utilizing our NSCLC PDX platform, for bevacizumab-treated LKB1 mutated tumours 39 . However, to exploit these models as being "avatars" for personalized patient therapy, the time required to stabilize them in mouse has necessarily to be less than DFS and, due to clinical aggressiveness of lung cancer, the use of PDXs for developing personalized therapy strategies is generally unfeasible. This issue coupled with the generally low grafting rate, allows mainly to exploit the broadest spectrum of different mutations of LcPDXs to investigate the efficacy of targeted therapies rather than to implement co-clinical approaches. Nonetheless, we here first show that 63.2% of grafted models was stabilized in mice before occurring of tumour recurrences, thereby offering a chance to personalized treatment in approximately one fourth of all lung cancer patients involved in this study and, interestingly, in one third of Stage II (and one fourth of Stage III) patients with progression of disease.
In conclusion, we reported the establishment of a wide panel of LcPDX, which accurately mirrored primary tumours in terms of subtype, growth pattern, genotype, metabolic activity and cancer stem cell composition. Moreover, PDX take rate correlated with patient OS and DFS, tumour stage, SUV, and presence of CD133 + / CXCR4 + /EpCAM − cells, consistent with more aggressive tumours prone to disseminate. Our LcPDX platform, beside recapitulating a broad spectrum of lung cancer-related mutations useful to test targeted therapies, could also be suitable for developing personalized co-clinical studies and investigating tumour-microenvironment cross-talk. All these observations strengthen the relevance of LcPDXs as an operational and robust pre and co-clinical model closely mirroring parental tumours, which could play a strong clinical role in biological and pharmacologic studies.

Material and Methods
Patients Selection. Tumour samples were collected from May 2006 to July 2013, with a sterile procedure in the operating room, after mediastinal biopsy (9 patients) or anatomical pulmonary resection, such as pneumonectomy, lobectomy or segmentectomy (88 patients). These patients represent 7% (9/135) of all mediastinal biopsies, and 10% (88/915) of all anatomical resections for primary lung cancer, performed at the "Fondazione IRCCS Istituto Nazionale dei Tumori" during the same period. Samples of primary NSCLC were obtained from patients undergoing surgical resection, who gave their informed consent after approval from the Internal Review and the Ethics Boards of the Fondazione IRCCS Istituto Nazionale Tumori and all methods were performed in accordance with istitutional guidelines and regulation.
PDXs establishment. PDXs were established as described in ref. 5 . PDXs models were propagated for three rounds in mice (P1-P3) before to be considered stabilized and then frozen in a solution of 90% FBS and 10% DMSO and stored in liquid nitrogen. The experimental protocol was approved by the C.E.S.A. (Ethical Committee for Animal Experimentation, of the National Cancer Institure Foundation), and animal experimentation was performed following guidelines drawn-up by C.E.S.A. according to ref. 40 .   3 ; stable disease (SD, score:4): MGI < 50% and increase in initial volume (iiV) <25%; progression of disease 2 (PD2, score:2): MGI < 50% and iiv >25% and tumor growth delay (TGD) > 1.5; PD1 (score: 0): MGI < 50% and iiv > 25% and TGD < 1.5. TGD was calculated as (time to event)/ (average time to event of controls), assuming an event as the time a tumor needs to triple its starting volume.
NGS analysis. DNA   Statistical analysis. Continuous variables were presented as mean values ± standard deviation (SD) and median with inter-quartile range (IQR), and categorical variables as numbers and percentages. Comparisons among groups for continuous variables were performed using a two-sided Student's t-test for normally distributed variables and a two-sided Wilcoxon's rank-sum test for variables not conforming to a normal distribution, and for categorical variables using contingency table analysis with the Chi-square test. The primary end-points of the study were overall survival (OS) and disease free survival (DFS). For each end-point, the time to event occurrence was computed from the date of surgery to the date when the event was recorded, or was censored at the date of last follow-up assessment in event-free patients. Hazard ratios (HR) of OS and DFS and the corresponding 95% confidence intervals (CIs) according to age, sex, tumour stage, SUV, and grafting of tumour were estimated using Cox proportional hazard models. A multivariable analyses was performed including terms for all these factors in the same Cox model. Survival curves were estimated using the Kaplan-Meier method and were compared by the log-rank test. Graphical evaluation by Schoenfeld residual plots indicated that the model assumptions concerning proportional hazards were appropriate. All tests were two-sided and a p-value of less than 0.05 was taken as statistically significant. Statistical analyses were performed using SAS 9.2 (SAS Institute, Cary, NC) and the figures were obtained using STATA 11.0 (StataCorp LP, College Station, TX) statistical software.