Patient-derived tumor xenograft and organoid models established from resected pancreatic, duodenal and biliary cancers

Patient-derived xenograft (PDX) and their xenograft-derived organoid (XDO) models that recapitulate the genotypic and phenotypic landscape of patient cancers could help to advance research and lead to improved clinical management. PDX models were established from 276 pancreato-duodenal and biliary cancer resections. Initial, passage 0 (P0) engraftment rates were 59% (118/199) for pancreatic, 86% (25/29) for duodenal, and 35% (17/48) for biliary ductal tumors. Pancreatic ductal adenocarcinoma (PDAC), had a P0 engraftment rate of 62% (105/169). KRAS mutant and wild-type PDAC models were molecularly profiled, and XDO models were generated to perform initial drug response evaluations. Subsets of PDAC PDX models showed global copy number variants and gene expression profiles that were retained with serial passaging, and they showed a spectrum of somatic mutations represented in patient tumors. PDAC XDO models were established, with a success rate of 71% (10/14). Pathway activation of KRAS-MAPK in PDXs was independent of KRAS mutational status. Four wild-type KRAS models were characterized by one with EGFR (L747-P753 del), two with BRAF alterations (N486_P490del or V600E), and one with triple negative KRAS/EGFR/BRAF. Model OCIP256, characterized by BRAF (N486-P490 del), had activated phospho-ERK. A combination treatment of a pan-RAF inhibitor (LY3009120) and a MEK inhibitor (trametinib) effectively suppressed phospho-ERK and inhibited growth of OCIP256 XDO and PDX models. PDAC/duodenal adenocarcinoma have high success rates forming PDX/organoid and retaining their phenotypic and genotypic features. These models may be effective tools to evaluate novel drug combination therapies.

www.nature.com/scientificreports/ therapies [3][4][5] . There are no validated biomarkers to assess individual tumor susceptibility to such targeted or systemic therapies in these cancers besides targeting DNA repair factors with poly-ADP ribose polymerase inhibitors in Pancreatic Ductal Adenocarcinoma (PDAC) harboring germline BRCA mutations [6][7][8][9] . PDAC is the most common and widely characterized periampullary cancer. The majority (> 90%) of these tumors harbor the oncogene Kirsten RAt Sarcoma virus mutations (KRAS mt ), complex chromosomal rearrangements involving chromothripsis and copy number changes 10,11 . KRAS oncogene is a potent activator of Mitogen Activated Protein Kinase (MAPK) pathway driving PDAC initiation and progression. Despite intensive efforts, current therapeutic advances in this disease is extremely limited 9 . A minority of PDACs has KRAS wild-type (KRAS wt ), but may also exhibit MAPK deregulation via somatic alterations in genes activating Receptor Tyrosine Kinase (RTK) signaling in RTK/RAS/MAPK cascade. Additional pathway mediators that are altered may include BRAF in-frame deletions (~ 1%), NRG1 gene fusions (< 1% to 11%) and ERBB2/HER2 amplifications (2-24%) [12][13][14][15][16] . Targeted therapies to inhibit RTKs, ERBB1/EGFR (e.g. erlotinib) and/or ERBB2 (e.g. afatinib or pertuzumab) in patients with KRAS wt and NRG1 fusion PDAC, have resulted in promising initial responses 14 . New therapies in bile duct and periampullary adenocarcinomas are also not forthcoming 17 . Therefore, new clinically relevant models for these cancers are needed to better understand their biology and for discovery of new therapeutics against these cancers evading treatment.
While few hundreds PDAC PDX models have established worldwide, few studies with limited number of models have reported detailed characterization of their establishment, phenotypic characterization and molecular profiling [18][19][20][21][22][23][24] . Similarly, for PDX models of bile duct cancers 25,26 . Patient tumors and PDX may also be used to establish patient-derived organoid (PDO) models, and these latter models are useful for initial higher throughput functional genomic studies, with subsequent validation in the matched PDX models 27 . Establishing a renewable PDX/PDO resource with relevant clinical annotation, and genotyping is essential for preclinical studies in new drug and biomarker discoveries 11,20 . In this manuscript, we described our initiative to establish 127 PDX from resected pancreato-duodenal and biliary cancers and 10 matching PDAC xenograft-derived organoid (XDO) models. A subset of PDAC was further molecularly profiled for genomic fidelity to corresponding patients by genome-wide copy number and gene expression. In addition, whole exome profiling was performed on 31 PDAC PDXs to identify potential drug actionable alterations. Using a subset of KRAS wt/mt PDAC models, we showed that these PDAC tumors exhibited heterogeneous KRAS-MAPK pathway activation which can be targeted using a rational polytherapy approach in patient-derived organoid and their xenograft models.

PDX model establishment from pancreatic-duodenal-biliary system. Between September 2008
and June 2013, a total of 276 Whipple resected samples were collected for PDX establishment from pancreas (n = 199), duodenum/ampulla of Vater (n = 29), and extrahepatic bile duct (n = 48), as summarized by histological subtypes (Table 1), and inventory tables (Supplementary Tables S1 and S2). The PDX tumors retained the morphological features of their matched patient tumors, as shown in these representative cases, Fig. 1, A-E. Both orthotopic and subcutaneous sites showed fidelity of the poorly differentiated PDAC features to their matched patient tumors (Fig. 1A). Compared to the patient tumors, the PDX tumors at either the orthotopic or subcutaneous site showed a tendency toward enrichment of tumor cells compared to stromal components. We did not observe metastases at either the subcutaneous or orthotopic tumor growths, although in the latter situation, larger tumors may invade into adjacent organs, such as intestines, spleen and liver. Table 1. Number of implanted specimens and engrafted xenograft tumor models. Patient-derived tumor models (XG) after initial passage (P0) in mice and serial passages for at least three passages to evaluate stable xenograft propagation (SXG). PDAC pancreatic ductal adenocarcinoma; IPMN intraductal papillary mucinous neoplasm. *A subset of PDAC were ascites specimens; 6/21 had initial engraftment and also formed SXG models. **15 IPMNs included eight with invasive adenocarcinoma; XG_P0 IMPN cases were associated with invasive adenocarcinoma (4/5) and one with focal invasion; Stable XG IPMN cases were associated with invasive adenocarcinoma (n = 2), and focal invasion (n = 1).  Additional to the subcutaneous site, PDAC patient tumor fragments were implanted at the orthotopic site (Supplementary Table S3), with an engraftment rate of 22% (16/71), as compared to 100% (71/71) when implanted subcutaneously. Among the 169 PDAC tumors, the ability to engraft at the subcutaneous site was not associated with specific clinical pathological features (Supplementary Table S4). However, engraftment was significantly correlated with poorer patient overall survival (Fig. 1F, p-value = 0.0095), but not recurrence rate ( Supplementary Fig. S1B).
Growth characteristics of PDAC xenografts. A subset of PDAC models was further characterized by xenograft tumor latency (n = 34) and growth rates (n = 23). The latency time between tumor implant and palpable detection was observed for more than one mouse per model per passage, with a median of 18 days after initial implantation (P0), and serial passages resulted in shorter but stable median latency times ranging 11-14 days ( Supplementary Fig. S1C). At passage 4 all had palpable tumors within 46 days (Supplementary Table S5A). Most (90%) of the models had P0 palpable tumors within 90 days (Supplementary Table S5B). Latency at P0 or P1 did not correlate with patient overall survival ( Supplementary Fig. S1D,E). However, a tendency for a correlation of shorter latency times in P1 with early recurrences was observed in this small subset of PDAC models (Supplementary Fig. S1F,G).
Tumor growth rates in P1 varied among models of the different pancreatic tumors (Fig. 1G), and for duodenal and bile cancers (Supplementary Fig. S2A-D). Tumor models derived from the duodenum grew significantly faster than the pancreas (p-value < 0.0001, Supplementary Fig. S2E). Since there were only two bile duct PDX models, their growth rates were not analyzable. Tumor latency duration was not associated with tumor cellularity ( Supplementary Fig. S2F).The growth rates in the PDAC subset was not significantly associated with clinical outcome ( Supplementary Fig S3A,B).

Molecular profiling of PDAC models. Copy number and gene expression profiling was performed on 4
randomly selected PDX models to determine the genomic fidelity among serial passages (P0-P4). The majority of the gene copy number changes observed in early passages were retained overtime as profiles clustered within a model ( Supplementary Fig. S4A). Gene expression profiles showed high concordance across passages within a model (R > 0.97, Supplementary Fig. S4B).
KRAS mutations by direct sequencing were detected in 95% (73/77) of the PDAC models. However, four (5%) PDAC PDX and their matched patient tumors were KRAS wt . Whole exome sequencing was performed on 31 PDX models, including the 4 KRAS wt , of which 23 PDXs had corresponding patient tumor profiles. The PDX cohort had a median of 63 somatic mutations compared to 49 mutations in the patient cohort. The median proportion of patient tumors' somatic mutations that were found in their matched. PDX models was 76% (range 18-95%); 21/23 cases had > 55% representation except for 2 cases with 18% and 30% overlaps (Supplementary Table S6). Vice versa, median proportion of PDX somatic mutations that were found in their matched patient was 53% (range 1-94%).
The somatic profile on 31 PDX models were characterized by alterations typically found in the PDAC landscape ( Fig. 2), and included the common PDAC-associated genomic aberrations such as TP53, SMAD4, CDKN2A, BRCA1/2 13 , and 18 of the 32 genes that were significantly mutated in a 456 patient-set reported by Bailey et al. 10 . Somatic with or without germline or germline alone mutations in BRCA1/2 that lead to homologous recombinant deficiency were found in 19% (6/31) models profiled (Supplementary Table S7).

Establishment of organoid lines from PDX models.
Xenograft-derived organoid (XDO) lines were successfully established from 10/14 (71%) of the PDAC PDX models. These models could be maintained longterm (> 5) passages. Histological fidelity between the parent xenografts and derived organoids was preserved as shown with representative models (Fig. 3).
Model XDO 256 BRAF N486_P490del was further assessed for sensitivities to trametinib and a pan-BRAF inhibitor LY3009120, alone or in combination. Western blots showed partial inhibition of pERK by trametinib or LY3009120 alone, and complete inhibition when they were combined (Fig. 5A). This greater inhibition signaling inhibition was also associated with greater growth inhibition in XDO tumor cells when treated with drug combination compared to single agent treatment (Fig. 5B).
The corresponding PDX models also showed that the KRAS-MAPK pathway activation occured independent of KRAS genotype as shown by the phospho-ERK and phospho-AKT levels ( Supplementary Fig. S6). PDX models, which included one KRAS Q61R (OCIP285) and one triple negative, wild-type KRAS/EGFR/BRAF (OCIP180) were both significantly growth suppressed when treated with trametinib, with the KRAS mt model showing greater response than the KRAS wt triple negative model ( Supplementary Fig. S7 www.nature.com/scientificreports/ phospho-ERK profiles, targeted drug combination treatment with trametinib and LY3009120 showed greater inhibition in PDX OCIP256 (BRAF N486_P490del ) as compared to single agent treatments, mirroring the finding in XDO (Fig. 5B,C).

Discussion
We attempted engraftment of 276 resected pancreato-duodenal and biliary cancers patient tumor specimens in immunodeficient mice to generate PDX models. The initial engraftment rate at P0 was highest for duodenum at 86% (25/29), followed by 59% (118/199) for pancreatic tumors, and 35% (17/48) for bile duct adenocarcinoma. Among the 160 initial engrafters, 127 PDXs formed stable models (99 pancreatic, 20 duodenal and 8 bile duct), which were sustainable for at least three serial passages. In four PDAC PDX models, we showed stability of copy number and gene expression profiles up to 4 passages. Whole exome sequencing and copy number profiling of 31 models revealed similar landscape of genomic changes as observed in patient tumors 10,13,28,29 . In addition, organoid lines were successfully established from 10/14 (71%) of PDAC PDXs, a subset of which were subjected to drug testing. Our establishment of PDX models from pancreatic-biliary-duodenal cancers represent the largest effort published by individual institutions. A review of published literatures showed a combined total of 114 pancreatic, and 54 bile duct PDX models (Supplementary Table S8), few of them have had full details on engraftment stability, genome-wide molecular profiling completed, and made publicly available. In addition, the PDXFinder data repository (pdxfinder.org) revealed additional models that may be available from various consortium and commercial sources. The 20 ampullary duodenal carcinoma PDXs we are reporting represent the largest collection known. The establishment rate for duodenal carcinomas is high at 86% initial and 69% stable engraftment. In addition to 54 bile duct (cholangio) carcinoma PDXs that have putatively been established, we generated an additional 8 stable PDXs, with initial (P0) success rate of 35%, which dropped to 17% for stable models. From the study reporting largest number of bile duct PDX establishment, the engraftment rate was independent of corresponding patients' tumor stage and overall survival, but was significantly associated with a shorter time to recurrence. Due to our small collection of bile duct cancers, engraftment and clinic-pathology correlations were not feasible.
We generated a total of 99 pancreatic PDX models, including 97 carcinomas of duct epithelial origin, 1 acinar carcinoma and 1 neuroendocrine carcinoma. The duct epithelial origin carcinomas included 92 ductal, 3 mucinous, 1 squamous and 1 undifferentiated. Furthermore, 3 of the ductal carcinomas were established from IPMN tumor fragments. The majority (80%) of the IPMN specimens failed to form stable PDX. This is consistent with IPMN being mainly in-situ neoplastic proliferation. The three IPMN-associated stable PDXs had corresponding patient tumor with invasive adenocarcinoma or severe dysplasia with focal invasion.
Multiple groups have reported establishment of PDX models from PDAC with engraftment successes ranging from 43% (57 PDXs) 30 to 85% (35 PDXs) 31 , (Supplementary Table S8). Our success rates are comparable for initial (62%) and stable (52%) engraftment. A large factor that influenced engraftment success appears to be tumor www.nature.com/scientificreports/ cellularity and quantity of the implanted fragment. Studies reporting higher success rates used larger implanted fragment (5 mm 3 ) 31 compared to smaller fragment sizes (< 3 mm 3 ) in our study, and others 30 . In our study, lowest engraftment rates were observed for orthotopic implant sites (16/71, 22%), and implant sources from ascites specimens (6/21, 28%), and was likely due to smaller size of tumor tissue implanted in the orthotopic method, and fewer tumor cells in ascites specimens, respectively. Loukopoulos et al. and Perez-Torras et al. 32,33 have suggested that one advantage of the orthotopic compared to subcutaneous implantation has been its metastatic potential to distant organs, otherwise they have been shown to have similar global gene expression 22 . However, we observed that tumors implanted in the pancreas could invade into adjacent organs, such as intestines, spleen, liver etc., when they grow larger, mimicking locally invasive pancreatic cancer in patients (results not shown). The use of mouse strains less immunocompromised than the NOD SCID may impact to lower engraftment rates. For example, the Nude strain yielded a 43% engraftment success (absent thymus and T cells) 30 compared to our set (62% at P0). However, the Nude strain was also reported to yield a high engraftment rate of 71% from ascites (Supplementary Table S8). These observations suggest multiple conditions favoring engraftment, a more immunocompromised mouse strain, aggressive tumor characteristics and high tumor cellularity/larger implant fragments. Specific to our method, Matrigel® may have supported the initial tumor engraftment. Tumor growth characteristics of the PDAC models did not significantly associate with specific clinicopathological features of the patient tumors. However, successful tumor engraftment was highly correlated with poorer overall survival of the patients. This has been a consistent observation across different tumor types including non-small cell lung, head and neck and bile duct cancers [34][35][36] .
Our evaluation of global copy number and gene expression through early serial passages (P0-P4) indicated that the global similarities were retained overtime in a representative four PDAC PDX models. This is consistent with other studies using few representative models to show genomics fidelity between PDX early and late passages, as well as to their matched patient tumors 20,22 , and between primary, metastasis and PDX 24 . www.nature.com/scientificreports/ Whole exome sequencing on paired PDAC patient tumors and PDXs showed high fidelity in their somatic mutation profile. Globally, some discrepancies in somatic mutations were observed between matched patient-PDX tumors. Several factors may contribute to these discrepancies, in particular specific to our dataset, existing data was derived from either whole genome or exome sequencing methods which may have variant detection biases (e.g. target regions, depth of coverage). Additional contributors, also reported by other studies, include tumor enrichment observed in PDX compared to their matched patients 22 , and the incomplete removal of mouse variants orthologous to human genes 24 . Consistent with these reported PDX-patient tumor evaluations, our results also showed that PDAC-associated mutations were represented in PDXs. Excluding BRCA1/2, these somatic mutations are not currently drug targetable.
To increase the utility of PDX models, we successfully generated 10 xenograft-derived organoids (XDO) with a 71% rate of success. Active MAPK pathway signaling as profiled by phospho-ERK and phospho-AKT were observed in both wild-type and mutant KRAS models. XDO drug screens followed by PDX validation, showed that these models had some sensitivity to trametinib regardless of the KRAS status suggesting that MAPK signaling is a core proliferative signaling pathway in PDAC. One of the KRAS wt models, OCIP256 also has an oncogenic in-frame BRAF deletion (N486-P490del). This alteration was previously targeted by a RAF dimer inhibitor LY3009120 in BxPC3 PDAC cell line 37 . Interestingly, XDO/PDX OCIP256 tumor growth was more suppressed with trametinib treatment compared to LY300920 alone (Fig. 5). Our data is corroborated with the recent clinical evidence that oncogenic in-frame BRAF deletions confer sensitivity to MAPK signaling 12 . While one patient carrying BRAF in-frame deletion had a partial response to a MEK1/2 inhibitor, the other patient showed no response highlighting a need to identify effective combination therapy strategies with MAPK inhibition. We confirmed the synergistic effect of the LY3009120/trametinib combination in OCIP256 PDO and XDO models (Fig. 5). To our knowledge, the OCIP256 model is the first patient-derived xenograft and organoid model derived from a PDAC patient with BRAF in-frame deletion and could be a useful model to investigate effective combination strategies with MAPK inhibition. www.nature.com/scientificreports/ The transition between these preclinical platforms would be cost effective for studies that could perform high-throughput drug treatment strategies, followed by validation of the fewer conditions using more complex in vivo PDX environment testing. These pancreas PDX models which are clinically annotated and capture the genetic diversity of the patients will be used as a preclinical tool to better understand the biology of pancreatic cancers and its underlying impact on effectiveness of therapeutic approaches. These PDX and XDO models are an invaluable renewable tumor resource for research.

PDX establishment. The University Health Network (UHN) Human Research Ethics and Animal Care
Committees approved this study protocol (REB# 08-0767 T). Surgical and ascites specimens were collected at the Toronto General Hospital (TGH-UHN) between September 2008 and June 2013 with informed consent from participants. Human research followed the guidelines of Canada Tri-Council Policy Statement, in accordance with Declaration of Helsinki (www. pre. ethics. gc. ca.). Tumor specimens were collected in serum-free Dulbecco's Modified Eagle's Medium (DMEM), and used within 24 h. Fragments (4-8 mm), or cell pellets were mixed with 10% Matrigel™ (Product Number 354234, Corning) at 4 °C, and implanted in the subcutaneous pocket at flank, or orthotopic site in Non-Obese Diabetic Severe Combined Immune-Deficient (NOD SCID) mice 38 . Animal care followed the guidelines of UHN Research Institutes' policies and the guidelines of the Canadian Council on Animal Care, and consistent with ARRIVE guidelines for study design 39 .
Tumor growth was measured twice weekly with a caliper for the length (L, largest diameter) and its perpendicular width (W), including skin fold. The volume was calculated using the formula V = W 2 × L/2. Tumor latency/ lag phase was recorded and defined as the duration between tumor implant to first palpable tumor detection. The initial implant of patient tumor fragment into mouse host was defined as passage 0 (P0), followed by serial   ). Genomic DNA (500 ng) samples were amplified, fragmented, precipitated and hybridized to Illumina Human Omni 2.5 M SNP chip array as specified by manufacture. Intensity files were converted for quantification and normalization in GenomeStudio using HumanOmni2.5-8v1_A.bpm manifest and cluster files from Hg19. Copy gains and losses were detected using ASCAT algorithm 48 . ASCAT utilizes the Piecewise Constant Fitting (PCF) method to define segments and assigns a copy number for each allele. Hence, the total copy number of a segment was derived from the sum of the allele specific copy numbers. Copy gains and losses were defined as having 0.5 copy above (for gain) and 0.5 copy below (for loss) relative to ASCAT estimated sample ploidy. Hierarchical clustering was performed on segmented copy number Log R values using Pearson dissimilarity (1-Pearson correlation coefficient).
Gene expression profiles were generated from RNA (200 ng) of flash frozen xenograft tumors using the Illumina Whole Genome Gene Expression DASL Assay, using manufacturer's protocol for HT-12 V4 BeadChip (Illumina Inc.), and the fluorescence data files were converted for quantification using GenomeStudio® software without normalization or background subtraction. Quality assessment, pre-processing and normalization of the raw expression data were implemented using R/Bioconductor package lumi 49 . Raw data was log2-transformed, back-ground subtracted, and normalized using a variance-stabilizing transformation algorithm. Correlation and fold-change analysis were done on probes detected (p < 0.01) in at least two passages within a series. Pearson dissimilarity (1-Pearson correlation coefficient) was used to perform hierarchical clustering on the expression values of samples with multiple passages.
Generation of XDO models. Xenograft-derived organoids (XDOs) were established from cryopreserved PDX tumor fragments (< 5 mm 3 ). Thawed fragments were dissociated using 100 g/ml Liberase (TH research grade, Catalog #5401135001, Roche Molecular Sysems, Inc.) for 90 min at 37 °C, then further with Gibco Try-pLE Express (Catalog #12,605,036, ThermoFisher Scientific Inc.) for 10-20 min at 37 °C, based on the method described by Fujii et al. 50 . Dissociated cells were collected and embedded in growth factor-reduced Matrigel™ (Catalog #354,230, Corning Inc.), and overlaid with hPDAC growth medium 51 . Standard PMLB protocols were followed to propagate models for a minimum of 5 passages to be considered stable long-term cultures, and quality assurance was performed to evaluate concordance of organoid with matched PDX and patient histology and DNA fingerprinting.

Drug testing on pancreatic tumor models in vitro and in vivo.
An in vitro drug screen was conducted on organoids. Dissociated cells were seeded on top of a thin layer of Matrigel™ in culture medium in a 384-well plate (3,000 cells per well) 52 . All pharmaceutical agents were added the next day, in a 6-point concentration series from 10 uM to 0.001 uM to wells in triplicate. Cell viability was assessed by ATP quantification using the CellTiter-Glo 3D luminescence-based assay (Catalog #G9681, Promega Corp.). Viability values were normalized to vehicle control wells and dose concentrations were log10-transformed. Drug dose response scores were calculated using area under the curve (AUC) values for each drug, and were normalized by dividing the AUC value by the maximum area for the concentration range measured for each drug. CompuSyn software 44 was used to calculate combination indices for combination drug studies.
PDX drug screen was performed using banked cryopreserved PDX fragments that were thawed and implanted in donor mice, tumors were harvested at full size, and fragments were generated and implanted in male mouse replicates for drug treatments. When tumor replicates averaged 100 mm 3 , mice were randomized (n = 5/group) Statistical analysis. Patient outcome endpoints were characterized by survival time, and time to relapse with respect to diagnosis date. Survival was defined as time from diagnosis to death or last follow-up. Any death was considered an event. The probability of survival was estimated using Kaplan-Meier method. The Cox proportional hazards model was used to estimate the hazards ratio for the adjusted effect. The p-values cited in the multivariate analysis were based on the Wald test within the Cox proportional hazards model. The difference in survival distribution between different groups was tested using log-rank test. The relapse was considered as an event and death without relapse as competing risks. The probability of relapse was calculated using the cumulative incidence approach. The difference between the distributions of time to relapse between groups was calculated using the Gray test.
PDX growth rates were estimated using log-linear mixed effects models. The models include fixed effects of week, treatment, PDX and all interactions as well as a random intercept and day effect. This type of models allows the test the difference in the treatment effect across the models used. Tumor volumes and latencies of PDXs were log transformed to stabilize the variance.

Data availability
Our PMLB PDX models are listed in the open global catalogue of PDX models at the PDXFinder repository (pdxfinder.org). Somatic mutations data generated in this work may be accessed using the CBIOPORTAL.CA repository (https:// cbiop ortal. ca/ cbiop ortal/ study/ summa ry? id= pancr eas_ pdx_ 2021).