A patient tumor transplant model of squamous cell cancer identifies PI3K inhibitors as candidate therapeutics in defined molecular bins

Abstract

Targeted therapy development in head and neck squamous cell carcinoma (HNSCC) is challenging given the rarity of activating mutations. Additionally, HNSCC incidence is increasing related to human papillomavirus (HPV). We sought to develop an in vivo model derived from patients reflecting the evolving HNSCC epidemiologic landscape, and use it to identify new therapies. Primary and relapsed tumors from HNSCC patients, both HPV+ and HPV−, were implanted on mice, giving rise to 25 strains. Resulting xenografts were characterized by detecting key mutations, measuring protein expression by IHC and gene expression/pathway analysis by mRNA‐sequencing. Drug efficacy studies were run with representative xenografts using the approved drug cetuximab as well as the new PI3K inhibitor PX‐866. Tumors maintained their original morphology, genetic profiles and drug susceptibilities through serial passaging. The genetic makeup of these tumors was consistent with known frequencies of TP53, PI3KCA, NOTCH1 and NOTCH2 mutations. Because the EGFR inhibitor cetuximab is a standard HNSCC therapy, we tested its efficacy and observed a wide spectrum of efficacy. Cetuximab‐resistant strains had higher PI3K/Akt pathway gene expression and protein activation than cetuximab‐sensitive strains. The PI3K inhibitor PX‐866 had anti‐tumor efficacy in HNSCC models with PIK3CA alterations. Finally, PI3K inhibition was effective in two cases with NOTCH1 inactivating mutations. In summary, we have developed an HNSCC model covering its clinical spectrum whose major genetic alterations and susceptibility to anticancer agents represent contemporary HNSCC. This model enables to prospectively test therapeutic‐oriented hypotheses leading to personalized medicine.

Highlights

• This HNSCC xenograft model includes all clinical subtypes by location and HPV status.

• Mutation profile is representative of seminal sequencing reports (Including NOTCH).

• Morphology, gene expression and drug sensitivity are stable over generations.

• Cases with PI3KCA alterations were sensitive to the novel PI3K inhibitor PX‐866.

• PI3K inhibition was effective in two cases harboring NOTCH1 mutations.

1. Introduction

Head and neck squamous cell carcinoma (HNSCC) causes 12,000 US deaths per year (Jemal et al., 2010). The epidermal growth factor receptor (EGFR) inhibitor cetuximab is the only targeted drug for HNSCC (Bonner et al., 2006; Vermorken et al., 2008). Human papillomavirus (HPV)‐related HNSCC is increasing in young adults (Chaturvedi et al., 2011; Gillison et al., 2000; Kreimer et al., 2005), has a different biology than HPV− HNSCC (Agrawal et al., 2011; Stransky et al., 2011), and thus representative models of this HNSCC subtype are an unmet need.

Targeted therapeutics development in HSNCC is challenging given the rarity of activating oncogene mutations and mutations in less “druggable” tumor suppressor genes, including TP53 and NOTCH1 (Agrawal et al., 2011; Stransky et al., 2011). A model integrating a multi‐gene and/or pathway‐driven background would enhance our understanding of HNSCC; studying the relevance of genetic aberrations in a multi‐mutated system is more consistent with our current view of cancer as a profoundly complex entity (Gerlinger et al., 2012; Jones et al., 2008).

Conventional drug development uses cancer cell line‐based in vitro screens followed by in vivo testing (Boyd, 1997; Johnson et al., 2001), poorly predicting clinical efficacy, perhaps because cell lines become homogeneous (De Wever and Mareel, 2003; Engelholm et al., 1985; Hausser and Brenner, 2005). Direct patient models using immune‐deficient mice preserve key features that culture cells irreversibly lose (Daniel et al., 2009), are appropriate tools for drug development (Garrido‐Laguna et al., 2011; Sivanand et al., 2012), and enable novel disease models in response to changing epidemiology. Finally, they allow confidence in their origin and the clinical annotation enables robust hypothesis testing (DeRose et al., 2011; Sivanand et al., 2012; Tentler et al., 2012).

The implantation of HNSCC samples in vivo is feasible with engraftment rates between 29% and 44% (Chen et al., 1996; Hennessey et al., 2011; Kimple et al., 2013; Prince et al., 2007; Wennerberg et al., 1983; Zatterstrom et al., 1992). No differences in tumor biology or clinical findings, including survival, were observed in patients with engraftment (Chen et al., 1996), regardless of mouse background (Prince et al., 2007). However, a patient‐derived model of skin squamous cell carcinoma (SSCC) is unavailable, and how it compares molecularly with HNSCC remains undetermined. We sought to construct an HNSCC and SSCC in vivo model reflective of the evolving epidemiologic landscape, characterize it molecularly, and identify new therapeutics.

2. Materials and methods

2.1. Patient samples

Excess, non‐diagnostic fresh tumor tissue from HNSCC patients consented at the University of Colorado Hospital in accordance with the protocol approved by the Colorado Multiple Institutional Review Board (COMIRB #08‐0552) were collected. A clinical pathologist grossed tumors after resection and non‐diagnostic, non‐necrotic portions were utilized. In situ hybridization of HPV low and high‐risk types was conducted using the Ventana INFORM HPV II and HPV III automated assays (Ventana Medical Systems) on 4 μm‐thick paraffin embedded tissue sections.

2.2. Xenograft generation

Animal care and procedures were approved by the Institutional Animal Care and Use Committee Office of the University of Colorado Anschutz Medical Campus. Tumors were placed in collecting medium consisting of RPMI supplemented with 10% fetal bovine serum (FBS), 200units/mL penicillin, and 200 ug/mL streptomycin, and cut into 3 × 3 × 3 mm pieces. Up to five mice (10 tumors) were implanted per patient case in this initial engraftment phase (F1 generation). The right and left hind flanks were sterilized and small incisions on the right and left hind flank create subcutaneous pockets. Prepared 3 × 3 × 3 mm tumor pieces were dipped in Matrigel (BD Biosciences) and inserted into the pocket. When tumors reach 1500 mm³ they were passed to a second colony of animals (5–10 animals per case; F2 generation) just as the fresh human tumors were. Once tumors in the second colony reach ∼1500 mm³, part of the tumors are stored for further analysis and the tumor is passed to a third colony of animals (F3 generation).

To generate orthotopic tumors we used a modified floor‐of‐mouth (FOM) or base‐of‐tongue (BOT) implantation protocol. We obtained tumor from a successful F1 case to F2 passage and cut it in 3 × 3 × 3 mm pieces (one per mouse). We then minced the tumor with a scalpel to create a cell suspension, filtered it with a 40 μM cell strainer (BD Falcon), re‐suspended it in 50 μL of injection vehicle (50% Matrigel, 50% RPMI + FBS), and injected it in the floor of the mouth or tongue base with a 22 g needle.

2.3. In vivo therapy

We tested cetuximab (Imclone; acquired commercially) to get the efficacy background data. We implanted 20 tumors in 10 mice heterotopically and when tumors reached 200 mm³ they were distributed in 2 groups (n = 10 tumors per group): control, or cetuximab 40 mg/kg 2/week IP, for 4 weeks. Similar experiments were run using control and PX‐866 (Oncothyreon Inc.) 2 mg/kg daily by oral gavage. Tumor size was evaluated 2/week using the formula: volume = [length × width²]/2. Sensitivity was defined as T/C of 20% or less. On the final day of treatment (6 h post‐treatment) tumors were harvested and immediately frozen in liquid nitrogen or processed for paraffin embedding.

2.4. Histologic analyses and immunohistochemistry (IHC)

Tumor specimens were routinely cut, stained with hematoxylin‐eosin (H/E), and examined microscopically. Histologic grades were defined as well differentiated (presence of features of epithelial stratification including keratin pearls), moderately differentiated (some evidence of epithelial differentiation but lack of keratinization and pleomorphism of replicating cells), or poorly differentiated (lack of any significant features of epithelial differentiation, most cells proliferating).

IHC analyses were performed on tissue arrays constructed using a manual Tissue Puncher (Beecher Instruments). For IHC slides were de‐paraffinized and re‐hydrated in graded concentrations of alcohol by standard techniques before antigen retrieval in citrate buffer pH 6.0 (Dako Corp.) at 105 °F for 20min. All staining were done in a Dako Autostainer and slides were incubated in 3% H₂O₂ for 10min, followed by primary antibody (60min at RT). The primary antibodies for IHC were: EGFR, pAKT (Ser473), pS6, PTEN (Cell Signaling Technology). Staining was developed using the following conditions: EnVision + Dual Link System HRP (Dako) for 30 min and substrate‐chromogen (DAB+) Solution (Dako) for 7 min. Slides were then counterstain with Automated Hematoxylin (Dako) for 5 min. Both the intensity (0–3+) and the percentage (0%–100%) of cells positive were interpreted by a pathologist that was blinded to the case and/or assigned treatment. For statistical analyses, an index of intensity × percentage were calculated and compared with Student's t‐test for comparison of the means of continuous variables.

2.5. FISH analysis

Each sample was incubated at 56 °C for 4 h, soaked in CitriSolv 3× for 5 min each, dehydrated and allowed to air dry. The slides were incubated in pretreatment solution at 80 °C for ∼12min, in protease solution IV at 37 °C for ∼20min, washed in milli‐Q water at RT, dehydrated and air‐dried. The 3‐color probe mixture was applied to the selected hybridization areas. DNA co‐denaturation was performed at 76 °C for 5min in a thermocycler and hybridization was allowed to occur at 37 °C for 40–48 h. Post‐hybridization washes were performed through incubations in 2xSSC/0.3% NP‐40 at 74 °C for 2 min and 2xSSC for 2min each, followed by dehydration. Finally, 14 μl of DAPI/anti‐fade (0.3 ug/ml Vectashield mounting medium) was applied and covered with a coverslip. Analysis was performed on epifluorescence microscopes using single interference filters sets for green (FITC), red (Texas red), blue (DAPI), gold, dual (red/green), and triple (blue, red, green) band pass filters. For each interference filter, monochromatic images were acquired and merged using CytoVision (Leica Microsystems Inc).

2.6. Gene mutation analyses

DNA was isolated using the DNeasy Tissue Kit (Qiagen). Extracted gDNA was PCR amplified using the GeneAmp High Fidelity PCR System (Applied Biosystems) containing 6% DMSO. Primers were synthesized by Integrated DNA Technologies. Reactions were carried out in 96‐well ABI Veriti thermocycler (Applied Biosystems) using a touchdown PCR protocol (Sjoblom et al., 2006). PCR amplification and Sanger sequencing were performed using primer sets renamed by the authors to better reflect the exon sequenced based on current GenBank identification (Agrawal et al., 2011). Previously undocumented mutations were confirmed using new primer sets referenced. PCR product was directly sequenced using the BigDye Terminator Cycle Sequencing Ready Reaction kit version 1.1 (Applied Biosystems). In the case of difficult templates, this Ready Mix it was combined with an aliquot of the dGTP BigDye Terminator Cycle Sequencing Ready Reaction kit (Applied Biosystems). The standard PCR parameters were; denaturation (5 min at 94 °C), 30 cycles (96 °C for 10 s), annealing (50 °C for 5 s), and extension/termination (60 °C for 4 min), followed by incubation at 10 °C. The products were sequenced on a fluorescent capillary automated sequencer (Applied Biosystems/Hitachi 3730 Genetic Analyzer).

Analyses of DNA sequences were done with Sequencing Analysis version 5.2 and Sequence Scanner version 1.0 (both from Applied Biosystems). Alignments of DNA sequences were done with Sequencher 4.8 (Gene Codes Corporation) and/or SeqScape (from Applied Biosystems). PCR product sequence comparison to GenBank reference sequence and mutation identification were accomplished using Mutation Surveyor version 4.0.4 (SoftGenetics).

2.7. Western blotting

50 mg tissue portions were thawed in a 4X volume of RIPA Buffer (Cell Signaling Technology) and homogenized using single‐use plastic pestles and centrifuged at 16,000 rpm at 4 °C. Protein concentration measurements were taken using the ELx800 Absorbance Microplate Reader and Gen5 software (BioTek) according to the manufacturer's instructions. 30 ng of protein was loaded per well into NuPage Novex 4–12% Bis‐Tris Midi Gel (Invitrogen), transferred using the iBlot Gel Transfer Stack System (Invitrogen) then processed according to the manufacturer's instructions. Primary antibodies were purchased from Cell Signaling Technology: phospho‐EGF Receptor (Tyr1068) (1H12), EGF Receptor (15F8), phospho‐Akt (Ser473) (D9E), Akt (pan) (40D4), phospho‐p44/42 MAPK (Erk1/2) (Thr202/Tyr204), p44/42 MAPK (Erk1/2), phospho‐PI3 Kinase p85 (Tyr458)/p55 (Tyr199), phospho‐S6 Ribosomal Protein (Ser240/244), S6 Ribosomal Protein (5G10), and Pan‐Actin, and used in dilutions recommended by the manufacturer. Secondary anti‐rabbit IgG was purchased from Immuno Research, and used a 1:50,000 dilution. The signal was visualized using Immuobilon Western chemiluminescent HRP substrate (Millipore).

2.8. RNA extraction

30 mg tumor pieces were placed in 300 μL of QIAzol, and homogenized using the MP Biomedicals Fast Prep 24. Samples were centrifuged at 12,000 g for 10min at 4 °C. The liquid homogenate was incubated at room temperature (RT) for 5min. 60 μL of chloroform was added and vortexed for 15 s then incubated at RT for 3min and centrifuged at 12,000 g for 15min at 4 °C. The top phase was transferred to 150 μL of isopropanol and vortexed incubated at RT for 10 min then was centrifuged at 12,000 g for 10 min at 4 °C. 300 μL of 75% ethanol was added to the pellet and centrifuged at 7,600 g for 5min and the RNA pellet was left to air dry. The pellet was dissolved in 100 μL of RNase‐free water and 350 μL buffer RLT and 250 μL 100% ethanol and was added to an RNeasy Mini spin column and run per manufacturer's directions. The RNA concentration and quality was measured using the Nanodrop.

2.9. mRNA sequencing (RNAseq)

Libraries were constructed using 1 μg total RNA following Illumina TruSeq RNA Sample Preparation v2 Guide and the cDNA library was validated on the Agilent 2100 Bioanalyzer using DNA‐1000 chip. Cluster generation was done on the Illumina cBot using a Single Read Flow Cell with a Single Read cBot reagent plate (TruSeq SR Cluster Kit). Sequencing of the clustered flow cell was performed on the Illumina HiSeq 2000 using TruSeq SBS v3 reagents. The sequencer was programmed with a single read at 100 cycles. Sequencing images were generated through the sequencing platform (Illumina HiSeq, 2000). The raw data were analyzed in four steps: image analysis, base calling, sequence alignment, and variant analysis and counting. An additional step was required to convert the base call files (.bcl) into *_qseq.txt files. For multiplexed lanes/samples, a de‐multiplexing step is performed before the alignment step.

2.10. Bioinformatics strategy for RNAseq analysis

On average, we obtained about 54 million (range = 40–90 million) single‐end 100 bp sequencing reads per sample. Reads were mapped against the human genome using Tophat (version 1.3.2) (Trapnell et al., 2009). Here, we used the NCBI reference annotation (build 37.2) as a guide, and allowing 3 mismatches for the initial alignment and 2 mismatches per segment with 25bp segments. On average, 83% (65–95%) of the reads aligned to the human genome. The remaining reads (unmapped to human genome) were mapped against the mouse genome (NCBI reference annotation build 37.2). Next, we employed Cufflinks (version 1.3.0) (Trapnell et al., 2010) to assemble the transcripts using the RefSeq annotation as the guide, but allowing for novel isoform discovery in each sample. Isoforms were ignored if the number of supporting reads was less than 30 and if the isoform fraction was less than 10% for the gene. The data were fragment bias corrected, multi‐read corrected, and normalized by the total number of reads. The transcript assemblies for each sample were merged using cuffmerge. We next computed the transcripts' FPKM values by rerunning Cufflinks using the merged assembly as the guide. The final output of this analysis step is a PxN matrix, where P is the number of samples and N is the number of transcripts, respectively. Gene expression was estimated by summing the FPKM values of multiple transcripts that represent the same gene. Subsequent data analyses of RNAseq were performed on this matrix. All other analyses were performed in R/Bioconductor (R version 2.14.1 (2011‐12‐22)). Singular Value Decomposition was performed using the base R function. Hierarchical clustering was performed finding the euclidean distance matrix and using Ward's minimum variance method for clustering.

Significant genes were identified using the R package LIMMA (Smyth, 2004). The FPKM values from cufflinks were log transformed and fitted with a model factoring origin, HPV, TP53 and NOTCH mutation status, PI3KCA event, and generation. Significant genes were identified using the lmFit, eBayes, and decideTests functions (FDR < 0.1 and fold change >1.5). The statistics were moderated using empirical Bayes shrinkage (via the eBayes function), global multiple testing strategy, and Benjamini & Hochberg adjustment. For Gene Set Enrichment Analysis (GSEA), we used the human pathways obtained from KEGG as genesets. Enriched pathways were identified by running GSEA using 1000 permutations as a stand‐along java app (version 2.07).

In the networks analysis of differentially expressed genes identified by LIMMA analysis for each binary comparison, neighboring genes of these genes were queried using STRING database (Szklarczyk et al., 2011). Enrichment analysis was performed on the gene networks using Cytoscape plug‐in module BiNGO (Maere et al., 2005) with Gene Ontology Biological Processes. P‐values were computed by hypergeometric test and corrected by Benjamini‐Hochberg method from BiNGO.

2.11. Statistical analysis

Descriptive statistics were used to describe the dataset. Categorical comparisons were conducted with Chi‐square or Fisher's test as appropriate. Survival analyses were done with Kaplan Meier's log‐rank test. SPSS ver.19 was used for all analyses. Statistical values only shown if P value or log rank are <0.1.

3. Results

3.1. Patient summary and engraftment

Of 46 implanted tumors (8 SSCC and 38 HNSCC), 23 were newly diagnosed and 23 had locoregional relapse (no therapy >3 months; Table 1). Regarding HNSCC, 27 (71%) were HPV− and 11 (29%) were HPV+; smoking and alcohol use were reported in 81% and 71%, respectively. These figures are consistent with the US HNSCC population (Joseph and D'Souza, 2012).Twenty‐five cases (54%) yielded a xenograft that grew ≥2 passages (Table 2). The resulting tumors from implantation (recovery rate) averaged 30% in F1 and subsequent passages (≥F2) had a stable recovery rate ≥75% (Figure 1A).

Table 1.

Patients implanted and engraftment. Engraftment was similar between oral cavity vs rest (70% vs 42%, P = 0.09), primary vs relapse (52% vs 57%), or HPV− vs HPV+ (57% vs 45%). Over the 36 months following implantation 26 (56%) relapsed, and 18 (39%) died (13 from progression, 5 from other causes). HPV+ patients had a (non‐significant) longer progression‐free survival (PFS) compared to HPV− (698 vs 429 days), as did newly diagnosed compared to relapsed subjects (653 vs 354 days, P = 0.03).

	Number (%)	Engrafted (%)
Cases
Consented	50	25 (50)
Implanted	46 (100)	25 (54)
Site
Lip	1 (2)	1 (100)
Floor of mouth	9 (20)	6 (67)
Mobile tongue	7 (15)	5 (71)
Base of tongue	4 (9)	2 (50)
Tonsil	6 (13)	2 (33)
Pharynx	3 (6)	2 (67)
Larynx	8 (17)	3 (38)
Skin	8 (18)	4 (50)
Cases
Primary	23 (50)	12 (52)
Relapse	23 (50)	13 (57)
HPV
Negative	35 (76)	20 (57)
Positive	11 (24)	5 (45)

Table 2.

Characteristics of successfully engrafted patients, describing the stage at diagnosis and type of sample whether the implanted sample was from the initial diagnosis or from a subsequent relapse. The average time to reach 1500 mm3 in F1 was 174 days (range 26–483 days; ≤120 days, 6 cases; 121–240 days, 15 cases; ≥241 days, 4 cases). BOT: base of tongue; Cet: treated with cetuximab. FOTM: floor of the mouth; HPV: human papillomavirus; HNSCC: head and neck squamous cell carcinoma; PX‐866: treated with PX‐866; R: resistant; S: sensitive: SSCC: skin squamous cell cancer.

Case	Gender	Origin	Site	Initial stage	Type	HPV	Smoking
CUHN002	Male	HNSCC	Tongue	T2N2M0	Primary	−	Yes
CUHN004	Male	HNSCC	FOM	T3N1M0	Relapse	−	Yes
CUHN007	Male	HNSCC	Pharynx	T4N1M0	Primary	+	Yes
CUHN009	Male	SSCC	Scalp	T4N0M0	Relapse	−
CUHN011	Male	HNSCC	Tongue	T2N0M0	Primary	−	Yes
CUHN013	Female	HNSCC	FOM	T3N2cM0	Primary	−	Yes
CUHN014	Male	HNSCC	BOT	T1N2bM0	Primary	+	Yes
CUHN015	Male	HNSCC	Larynx	T4N1M0	Relapse	−	Yes
CUHN016	Male	SSCC	Scalp	T3N0M0	Relapse	−
CUHN022	Male	HNSCC	Tonsil	T2N2bM0	Primary	+	No
CUHN024	Male	HNSCC	Lip	T2N2M0	Primary	−	Yes
CUHN025	Male	HNSCC	Larynx	T3N0M0	Relapse	−	Yes
CUHN026	Female	HNSCC	Tongue	T3N0M0	Primary	−	Yes
CUHN027	Male	HNSCC	BOT	T3N2bM0	Relapse	−	Yes
CUHN029	Male	HNSCC	FOM	T4N0M0	Relapse	+	Yes
CUHN032	Female	HNSCC	Tongue	T2N2cM0	Primary	−	Yes
CUHN033	Male	HNSCC	Larynx	T3N0M0	Relapse	−	Yes
CUHN036	Male	HNSCC	Pharynx	T4N2bM0	Primary	−	Yes
CUHN039	Male	SSCC	Cheek	T3N0M0	Relapse	−
CUHN040	Female	HNSCC	FOM	T3N2M0	Relapse	−	Yes
CUHN041	Male	SSCC	Ear	T4N0M0	Relapse	−
CUHN042	Male	HNSCC	FOM	T3N0M0	Primary	−	No
CUHN043	Female	HNSCC	Tonsil	T4N0M0	Primary	+	No
CUHN044	Male	HNSCC	FOM	T2N0M0	Relapse	−	Yes
CUHN049	Male	HNSCC	BOT	T3N2M0	Relapse	−	Yes

Figure 1.

Assessment of engrafted tumors, A. Schema representing the model. For implantation we used our standard technique (Jimeno et al., 2009; Rubio‐Viqueira et al., 2006), which maintains a narrow control on size and number of pieces of the first (F1) implantation. The number of pieces implanted (6, 8, and 10 tumor pieces in 9%, 22%, and 69% of cases) in F1 did not influence the success of engraftment. B. Orthotopic tumors with locoregional nodes (in the base of tongue [BOT] model the neck is dissected for clarity). Nodes from mice with orthotopic tumors grew >5 mm, compared with <1 mm in normal nodes. Immunohistochemistry (x40; bar 50 μM) with human EGFR and cytokeratin 5, mouse CD45, and dual FISH labeling of mouse cot (green) and humand cot (red) on neck nodes from mice that never had a human tumor (above), and a floor‐of‐mouth (FOM) implantation (below). C. Paired microphotographs (x20; bar 100 μM) of F0 and F2 hematoxilyn‐eosin and EGFR staining of three cases with well (above), moderate (middle) and poor (lower) degree of squamous differentiation. The tumors propagated on mice maintained these features and the EGFR staining across the spectrum. D. All cases showed EGFR positivity by immunohistochemistry (IHC), whose intensity, percentage and H‐score ranged from 1 to 3 (average±SD, 2.5 ± 0.6), 50%–100% (89% ± 18%), and 100 to 300 (226 ± 71). The plot of the correlation between the F0 and F2 H‐score for EGFR staining shows a high degree of correlation (R2 = 0.91, P < 0.0001). Although there were divergences in the H‐score in 9 of 25 paired samples, all were <20% and all cases were classified in the same category. E. Gene expression variations between F0 vs F2 vs F4 generations were compared for the CUHN014 and CUHN022 cases using the expression value of genes that significantly mapped to the human genome in at least one case and/or generation. The greatest variation for both cases was between the F0 patient sample and the F2, with stabilization further. Expression nearly completely stabilized once established on mice with R‐squared values for F2 vs F4 of 0.938 and 0.959 for CUHN014 and CUHN022.

Engraftment was similar in SSCC vs HNSCC (50% vs 55%), early vs advanced stages (50% vs 55%), and primary vs relapse (52% vs 57%). As opposed to breast, pancreatic or renal cancer (DeRose et al., 2011; Garrido‐Laguna et al., 2011; Sivanand et al., 2012), engraftment had no prognostic value for relapse or death (PFS 453 vs 502 days, and OS 682 vs 672 days for engrafted vs failure). Within the 38 HNSCC, engraftment was similar regardless of location, HPV status or stage. We applied two orthotopic techniques, floor‐of‐mouth (FOM) and base‐of‐tongue (BOT), to determine if this model developed locoregional disease in four cases (patients CUHN013,CUHN049 = locoregional relapse and CUHN002/CUHN004 = no relapse). Only CUHN013 and CUHN049 bearing mice nodes became clinically evident and had HNSCC metastasis (Figure 1B).

3.2. Histology, molecular markers and correlation with patients

Next, we determined the histologic and molecular markers relevant to HNSCC. Four cases were well differentiated, 16 moderately differentiated and 5 were poorly differentiated and all cases were positive for EGFR by IHC (H‐score 226 ± 71). Because the EGFR pathway is the focus of current HNSCC therapeutics, we studied downstream activation at the Akt/MAPK/S6K axis. EGFR was similar in HPV+ and HPV− cases (221 ± 60 vs 220 ± 52), but HPV+ cases had higher phospho‐Akt (P = 0.01) and phospho‐S6K (P = 0.02), and non‐significantly lower Pten expression, confirming previous findings (Charette and McCance, 2007; Contreras‐Paredes et al., 2009). Phospho‐Akt and phospho‐S6K expression were strongly correlated (R ² = 0.68, P < 0.001).

To interrogate model stability, patient and xenograft tumors were compared by histology and EGFR expression. Remarkable similarity in cellular structure, architecture, and differentiation was maintained across all 25 cases (Figure 1C). The EGFR expression correlation was robust between patient and xenograft tumors (R ² = 0.91, P < 0.0001; Figure 1D). Gene expression changes between F0 vs F2 vs F4 were compared in CUHN014 and CUHN022. The first 5 cases were assayed by Affymetrix and RNAseq, finding the latter superior in detecting genes with lower expression (Supplementary Figure 1). The greatest variation was between F0 and F2 with R‐squared values of 0.793 and 0.785 for CUHN014 and CUHN022 respectively (Figure 1E), in line with previous reports (Garrido‐Laguna et al., 2011), and altered pathways were not cancer‐associated but immune‐related. Variation between the F2 and F4 samples was small with R‐squared values of 0.94 and 0.96 respectively, suggesting expression stability in established tumors.

3.3. Key gene alteration analyses

We determined the mutation and amplification status of key genetic alterations from the seminal reports (Agrawal et al., 2011; Stransky et al., 2011), first as a benchmark to our collection, but also to test if we can advance the field by using our model to prospectively test genomic‐based hypotheses. To contextualize anti‐EGFR therapy we did hotspot sequencing of EGFR, PTEN, KRAS and NRAS, finding one KRAS mutation (G12D; CUHN016).

Direct exon sequencing of PI3KCA, TP53, NOTCH1 and NOTCH2 was completed the 25 cases (Supplementary Figure 2) and identified one PI3KCA mutation (CUHN026) catalogued as activating (Lee et al., 2005) (Figure 2A). TP53 somatic mutations were found in 48% of cases. Nonsense mutations were identified in 50% of the cases, 33% of which were homozygous. All missense mutations (67% of total) resided within the DNA binding domain. HPV operates by inactivating Tp53 and there is an inverse correlation between HPV status and TP53 mutation (Agrawal et al., 2011; Stransky et al., 2011; Westra et al., 2008). CUHN029, a smoker with HPV+ oral cancer harbored a homozygous nonsense mutation, consistent with the presumptive dual etiologic origin of this cancer.

Figure 2.

Mutation profiling, A. Positions of previously documented and undocumented (*), nonsense (red) and missense (black) mutations in PI3KCA, NOTCH1, and TP53. Fifty percent of the TP53 nonsense mutations occurred upstream of the DNA binding domain. Two mutually exclusive mutations occurred at a higher frequency. A G > T transversion resulting in G245V, and a C > T transition resulting in R282W were observed in 17% of TP53 mutants. Regarding NOTCH1 mutations, previous studies report putative loss of function mutations disrupting the ligand‐binding EGF‐like repeats in the N‐terminal extracellular domain of the receptor at frequencies between 11 and 15%; our observed rate is higher, but this needs to be balanced with our smaller sample size. Interestingly, in CUHN041 the two alleles have different mutations within the same codon producing a homozygous P1770F amino acid change. Mutations in NOTCH1 HD are activating in T acute lymphocityc leukemia (T‐ALL) (Malecki et al., 2006) and lung cancer (Westhoff et al., 2009). However, a R1608S mutation in the HD domain did not increase Notch1 activity in T‐ALL (Malecki et al., 2006), suggesting that the R1608 mutation in CUHN049 is possibly inactivating. B. Prevalence of NOTCH1 germline mutations in mutant xenografts. C. Samples hybridized with EGFR (red) and PI3KCA (green) FISH probes showed a wide spectrum of amplification. Insert: representative cell; –a: amplified; –na: not amplified. D. Schemata depicting tumors with key genetic alterations in TP53, PI3KCA, and NOTCH1.

Nine NOTCH1 and seven NOTCH2 mutations were identified (32% and 24% of cases, respectively; Figure 2A). All mutations occurred upstream of the C‐terminal active domain of Notch. Seven NOTCH1‐mut were located in the EGF‐like repeats suggesting loss of function (Agrawal et al., 2011; Stransky et al., 2011) and were scored as probably damaging using PolyPhen‐2 (Adzhubei et al., 2010), except for the R1279H substitution in CUHN040 that was previously annotated (Agrawal et al., 2011). The final two mutations reside in the heterodimerization domain (HD) and in the RBP‐J_K/CBF1‐associated module (RAM) domain. Three NOTCH2 mutations were nonsense while the rest were missense. The T197I mutation (CUHN040, CUHN041) was not predicted to be damaging (HumVar = 0.4) but is a polar > nonpolar transition within the EGF‐like repeats domain.

We next sequenced matching normal DNA samples and germline alterations were identified in 56% and 33% of tumor NOTCH1‐mut (Figure 2B) and NOTCH2‐mut cases respectively. Prior investigators using NextGen sequencing (Agrawal et al., 2011; Stransky et al., 2011; Wang et al., 2011) ignored mismatches not present in >0.5% of tags in normal tissue but by using direct full‐exonic Sanger sequencing we defined mutation as any event (tumor or germline) that scored as damaging (Adzhubei et al., 2010) or was previously annotated. The CUHN049 tumor demonstrated loss of heterozygosity compared to germline, suggesting a Notch1 tumor‐suppressing role in HNSCC. Interestingly, CUHN049 had suffered aortic valve dysfunction secondary to calcium deposit, commonly associated with germline inactivating NOTCH1‐mut (Acharya et al., 2011; Garg et al., 2005). To explore whether germline mutations were HNSCC‐specific, germline DNA of 30 stage I melanoma patients showed one inactivating NOTCH1 mutation. The germline NOTCH1‐mut rate comparing HNSCC vs non‐HNSCC reached significance (P = 0.02). Patients with germline NOTCH1‐mut were all smokers and younger than those with germline NOTCH‐wt (48 vs 62 years; P = 0.06). Together it can be hypothesized that NOTCH germline mutations may predispose subjects to HNSCC, occurring earlier than the general population with a common triggering event such as smoking.

3.4. Gene amplification events and integrative analysis of gene‐altering events

Given the paucity of EGFR and PI3KCA mutations and the high prevalence of gene copy gain abnormalities reported in HNSCC (Fenic et al., 2007; Licitra et al., 2011), FISH for EGFR and PI3KCA were obtained in 25 cases documenting high level copy gain or amplification in 4 (16%) and 6 (24%) cases, respectively (Figure 2C). Amplification of both EGFR and PI3KCA occurred in CUHN007 and CUHN0049. Overall, PI3KCA genetic events occurred in HNSCC only (Figure 2D).

3.5. Gene expression integrative characterization

Gene expression analysis was obtained to categorize the differences between clinical and molecular subtypes, first looking at individual over‐expressed genes, and then at pathways. Analyses of the 25 cases for the five binary comparisons (SSCC vs HNSCC, HPV, TP53, NOTCH and PI3KCA status) were run using Linear Models for Microarray Data (LIMMA) (Smyth, 2004). Variation between SSCC and HNSCC was less than between HNSCC samples alone and supported pooling both sample sets (Supplementary Table 2).

We developed a strategy whereby the binary relationship of differentially expressed genes and their neighboring gene interactions/relationships were analyzed (Supplementary Figure 3). DNA replication and cell cycle related biological processes were enriched in the HPV and TP53 status comparisons (Supplementary Table 3). When we focused on two “bins” defined by NOTCH and PI3KCA status, we found that HES1 is significantly higher in NOTCH1‐wt. Erbb3 and Rasal1 expression was also increased in NOTCH1‐wt and may promote tumor cell proliferation (Baselga and Swain, 2009) and progression (Ohta et al., 2009). The fibronectin receptor gene (Itgb1) was up‐regulated in NOTCH1‐mut, and is associated with cell motility in oral HNSCC (Hunt et al., 2011), and could constitute a novel therapeutic target. PI3KCA gene duplication or activation increased expression of PI3K signaling mediators including PRKCI that is crucial for proliferative and pro‐survival PI3K signaling (Desai et al., 2011, 2012).

3.6. Cetuximab treatment and stability

Treatment susceptibility can be considered an integrative response of a complex system, and can be used to query a model's validity and stability; also, it is important that a model neither over‐ or under‐estimates drug efficacy. Eleven cases were treated with cetuximab with 7 cases showing growth reduction ≥80% (4 had tumor shrinkage; Figure 3A). The response rate applying stricter clinical criteria (shrinkage ≥30%) was 9%, similar to cetuximab's clinical efficacy (Vermorken et al., 2007) but this correlation remains to be prospectively validated. Sensitivity did not change for three cases treated with cetuximab in multiple generations demonstrating stable susceptibility (Figure 3B). Markers predictive of cetuximab efficacy in HNSCC have been elusive (Langer, 2012; Licitra et al., 2011) and no EGFR pathway‐related proteins were predictive of cetuximab efficacy in the xenografts (Figure 3C). Global gene expression analysis between resistant and susceptible cases identified the Ras/Raf/MAPK pathway upregulation in cetuximab‐susceptible cases while PI3K/Akt/mTOR, Tp53 and cell cycle pathways were upregulated in resistant tumors (Supplementary Table 4).

Figure 3.

Treatment efficacy of cetuximab, A. Eleven cases received weekly cetuximab infusions or vehicle, and showed a wide spectrum of efficacy. Seven cases had a growth reduction over 50% compared to untreated, and four had actual tumor shrinkage below baseline tumor size. These four were heterogeneous in their clinical profile (two primaries and two relapses; two FOTM, one tongue, one BOT) but all four were smokers and HPV−. The three cases with lowest susceptibility correspond to an SSCC with a KRAS mutation (CUHN016) and the only two HPV+ HNSCC cases (CUHN014 and CUHN022). The tumor line plot of the highest and lowest susceptible cases CUHN002 and CUHN016 are shown. B. Repeated treatment at different generations of susceptible, intermediate and resistant cases indicated stable response to cetuximab. C. The Western blot analyses evidenced that the level of baseline expression of EGFR, phospho‐EGFR, Akt, phospho‐Akt, MAPK, or phospho‐MAPK was not predictive of efficacy. Pharmacodynamic changes after therapy were also unrelated with susceptibility to cetuximab. As previously reported, HPV+ cases (CUHN014 and CUHN022) had higher baseline phospho‐Akt. Bars are normalized to untreated control in each case, with calculated standard error (SE).

3.7. Prospective hypothesis testing: PI3K targeting in HNSCC

Using the above characterization we attempted to validate generated hypotheses with a therapy development focus. Given the lack of PIK3CA alterations in SCC, we focused in HNSCC. Tumors with a PI3KCA event had PI3K pathway activation with higher phospho‐Akt (P = 0.004), and non‐significantly higher phospho‐S6K (P = 0.09) or Pten expression by IHC. Given the overexpression of the PI3K pathway in cetuximab‐resistant cases and rate of PI3KCA abnormalities we first investigated the efficacy of PI3K‐directed therapy in the presence of PI3KCA genetic abnormalities. Secondly, because Notch1 pathway activation confers resistance to PI3K inhibitors and Notch1 inhibition reverses this effect (Muellner et al., 2011), we tested whether NOTCH1 mutations (CUHN027 and CUHN040) would induce a PI3K/Akt‐dependent state be susceptible to pharmacologic inhibition.

PX‐866, a wortmannin derivative, irreversibly inhibits PI3K by interacting with lysine‐802 in the ATP catalytic site, has a favorable pharmacological profile compared with wortmannin (Wipf et al., 2004), is well tolerated in humans (Hong et al., 2012), and is early in clinical development in HNSCC. PX‐866 was given in vivo to 9 cases, and in CUHN015 and CUHN022 moderate growth arrest was seen. Tumor growth inhibition (≥80%) occurred in 4 cases (Figure 4A). One case, CUHN014 has PI3KCA amplification, CUHN026 harbors a PI3KCA mutation, and CUHN040 and CUHN027 have NOTCH1 mutations. PX‐866 resistant cases had higher Pten (P = 0.05), lower phospho‐S6K (P < 0.01), and similar phospho‐Akt expression measured by IHC (Figure 4B). Treatment concomitantly decreased phospho‐Akt and phospho‐S6K only in CUHN026 and CUHN027; these had the highest basal phospho‐PI3K levels (Figure 4C).

Figure 4.

Treatment efficacy with PI3K inhibition, A. PX‐866 is a pan‐isoform inhibitor of PI3K with IC50s of 39 ± 21 nM, 88 ± 27 nM, 124 ± 26 nM, and 183 ± 25 against PI3Kα, PI3Kβ, PI3Kδ, and PI3Kγ, respectively. Mice were dosed for 28 days, and a wide range of efficacy was documented. The tumor line plot of the highest and lowest susceptible cases CUHN027 and CUHN002 are shown. B. Phospho‐S6K was highest in CUHN026 and CUHN027, and phospho‐Akt in CUHN026, as expected given it is the only case with an activating PI3KCA mutation. Greater than 80% decreases in post‐therapy phospho‐Akt and phospho‐S6K were documented in the susceptible CUHN027 and CUHN026, indicating on‐target inhibition and underscoring their value as surrogates of driver pathway abrogation. In CUHN014 and CUHN022 phospho‐Akt decreases between 50 and 80% were evidenced. Micro‐photographs are x40; bar 50 μM. C. Western blot analyses confirmed the IHC findings; additionally phospho‐PI3K showed the highest decrease in CUHN026.

Geneset enrichment RNA‐seq analysis identified opposing scenarios predicting PX‐866 or cetuximab susceptibility. We queried the STRING database for high‐confidence interactions (combined score >0.9) among the set of core genes identified from the top 50 pathways of GSEA with annotation as signal transduction, cellular processes and cancer‐related by KEGG. The global network represents core genes identified from the Gene Set Enrichment Analysis of Cetuximab and PX‐866 resistant vs. susceptible cases. We highlighted the first neighbors interacting with EGFR, PIK3CA, NOTCH1 and TP53 and illustrated those as cancer gene specific network modules Here, Ras/Raf/MAPK and Notch pathways including Notch target genes MYC and CCND1 were enriched in PX‐866‐resistant cases. Conversely, Tp53 pathways were more expressed in susceptible cases compared to resistant tumors. Figure 5B shows a summary of the complementary gene expression enrichment according to anti‐EGFR and anti‐PI3K susceptibility (Supplementary Table 4).

Figure 5.

Core genes for cetuximab and PX‐866 susceptibility. Pathway enrichment analysis of cetuximab (above) and PX‐866 (below) from RNA‐seq according to susceptibility (sensitive vs. resistant). Red and green colors represent enrichment in sensitive and resistant group, respectively. Enrichment maps were abstracted from the pathways with P < 0.05 (GSEA analysis) by comparing RNA‐seq of sensitive vs. resistant models, and were manually adjusted to highlight the most relevant pathways for visualization purposes.

4. Discussion

We present an HNSCC and SSCC patient‐derived model including all relevant clinical and molecular subtypes, is stable over several passages, and provides insight into the genetic similarities between SSCC and HNSCC. Importantly, HPV+ HNSCC seems less susceptible to EGFR inhibition than HPV− HNSCC, the potential identification of PI3K inhibitors as a promising class of anticancer drugs with activity in HNSCC with PI3KCA genetic alterations and NOTCH1 inactivating mutations.

Similar engraftment rate among clinical subgroups has yielded a model that covers the HNSCC spectrum. Supporting this, key gene mutation rates were similar to seminal HNSCC genetic landscape reports (Agrawal et al., 2011; Stransky et al., 2011). SSCC and HNSCC are increasingly considered to be biologically alike and the response rate to cetuximab clinically and in our model is similar, and could be seen as integrative of its genetic makeup (Maubec et al., 2011; Vermorken et al., 2007). Because SSCC is not formally recognized and there are limited drug development efforts to study SCCs as a common thread, a model able to identify commonalities and areas where differences are relevant will help move the field forward. One cetuximab resistant case was a KRAS mutant, a known factor in colon cancer, however this is less frequent genetic event in SCC and a clinical correlation has not been established.

PI3KCA mutation or PTEN loss cause pathway activation and dependence and predict response to PI3K inhibitors including PX‐866 (Ihle et al., 2009; Tanaka et al., 2011; Yuan et al., 2011). Furthermore, a correlation between resistance to PI3K inhibition and high PTEN is supported by our dataset. PI3K pathway activation in the presence of HPV is a novel finding in HNSCC, but is supported by current knowledge in HPV biology. The HPV protein E7 enhances keratinocyte migration in a PI3K/Akt‐dependent manner (Charette and McCance, 2007), whereas E6 both modulates PI3K signaling (Contreras‐Paredes et al., 2009), and activates PI3K/Akt conferring cisplatin resistance in HPV+ lung cancer (Wu et al., 2010).

Recent studies have reported inactivating mutations of NOTCH in HNSCC and SSCC pointing to a possible tumor suppressor role (Agrawal et al., 2011; Stransky et al., 2011; Wang et al., 2011) and we find nine NOTCH1 and seven NOTCH2 inactivating mutations supporting this conclusion. Interestingly, 56% of NOTCH1‐mut and 33% of NOTCH2‐mut cases harbored germline mutations while one corresponding patient suffered from a secondary germline NOTCH1‐mut related illness, suggesting a population with a predisposition for HNSCC. Notch inhibition was associated with SCCs in elderly patients taking semagacestat, a gamma‐secretase inhibitor (Extance, 2010), further supporting Notch's oncogenic role.

NOTCH1‐mut CUHN027 and ‐040 demonstrating sensitivity to PX‐866 is a hypothesis‐generating observation. Supporting our findings, Notch1 signaling confers resistance to PI3K inhibitors in other tumor types, and Notch1 inhibition with a gamma‐secretase inhibitor reversed this effect (Muellner et al., 2011). Gene expression analysis supports this by associating PX‐866 resistance to Notch1 pathway enrichment and this approach analyzing the gene sets/pathways analysis followed the concept of incorporating biological knowledge to boost signal detection over noise (Ideker et al., 2011). Here, we integrate gene‐neighboring information from target differentially expressed genes for computing the enrichment maps (Shojaie and Michailidis, 2010) over differentially expressed gene testing alone. Further association between the Notch1 and PI3K pathways has been documented in T‐ALL where NOTCH1 activating mutations lead to HES1 over‐expression, resulting in the suppression of PTEN (Palomero et al., 2007). In addition, PTEN loss is associated with resistance to Notch1 inhibition (Palomero et al., 2007). The PI3K inhibitor wortmannin up‐regulates Notch1 pathway signaling (Kolev et al., 2008). Our observation linking NOTCH1 mutational status to susceptibility PI3K inhibition has immediate translational applicability, given that PX‐866 is being developed in HNSCC‐specific clinical trials (with cetuximab [NCT01252628] or docetaxel [NCT01204099]).

In summary, we have developed a model that covers the clinical spectrum of HNSCC, and whose major genetic alterations and susceptibility to anticancer agents represent a contemporary clinical series of HSNCC. The number of cases does not allow for strong statistical conclusions, but we have made interesting observations including that PI3K inhibitors may be active in cases driven by PI3KCA genetic abnormalities, and that NOTCH1 inactivating mutations may induce a state of enhanced susceptibility to PI3K inhibitors. While it cannot compete with broader established tumor resources for gene discovery, this model's value may enable prospective and functional testing of hypotheses leading to better understanding of HNSCC and SSCC, and personalized medicine advances.

Author contribution

S.B.K. and R.T.A. set up and maintained the tumorgraft lines and database, processed tissue for histology and molecular analyses, conducted the molecular analyses, and were responsible for most of the drug trials. D.P.A. performed the bioinformatics and statistical analyses under the supervision of A‐C.T. B.W.V, J.J.M, J.P.P., P.N.L., J.R.E‐S., S.L.T. and D.B.S assisted with animal experiments, processed, managed, and extracted nucleic acids and protein from tissues and performed gene and protein expression, and mutational analyses. S.L.K conducted DNA copy number analysis under the supervision of M.V‐G. D.W.B., M.A.P., S.M.M., and R.M.H. obtained consent and followed patients. S.P., and D.F.H. provided experimental drug input. M.A.G., S.S., and F.R.H. prepared, stained, and analyzed, together with A.J., all histological sections, including immunohistochemical stains. T.H.L., J.A.G., and J.I.S. consented patients and performed surgeries. J.J.A., W.A.M., W.A.R., D.R. and X‐J.W. provided input and materials. A.J. conceived the study, obtained funding, established procedures for tumorgraft generation and processing, directed the project, and wrote the manuscript with input from other authors.

Funding

Supported by National Institutes of Health grants R21DE019712 (to A.J.), R01CA149456 (to A.J.), P30‐CA046934 (to the University of Colorado Cancer Center), and DE020649 (to X.J.W., J.I.S. and A.J.). Oncothyreon Inc. provided financial support (to A.J.) for the animal and molecular studies related to the investigational compound. The Vora Family Foundation provided support (to A.J.) for equipment purchasing and molecular testing. The Janet Mordecai Family Foundation gave support (to A.J.) for molecular testing.

Conflicts of interest

S.P. and D.F.H. are employees and hold stock of Oncothyreon Inc. A.J. has filed intellectual property interests on ideas/materials described herein. The rest of the authors declare that they have no competing interests.

Data and materials availability, Materials will be shared according to the University of Colorado's Office for Technology Transfer policies and Institutional Review Board. Gene expression arrays will be deposited in NCBI Gene Expression Omnibus.

Supporting information

The following are the supplementary data related to this article:

Supplementary Table 1 Primers for mutation analyses.

Supplementary Table 2 Significant genes from the 5 multivariate analyses.

Supplementary Table 3 Genesets enriched by BiNGO based on the significant genes and their neighboring genes in the 5 multivariate analyses.

Supplementary Table 4 GSEA outputs of the Cetuximab and PX‐866 sensitive vs. resistant samples based on KEGG pathways definition.

Supplementary Figure 1A. Comparison of the gene expression platforms Affymetrix microarray and Illumina RNAseq for five samples. RNAseq produces a greater dynamic range than the array. Fold change was calculated by dividing the gene expression for HPV positive by HPV negative. Most genes show similar trend for both platforms, however some genes show dynamic fold change by RNAseq but not by array. B. Comparison of RNAseq coverage. The same sample, CUHN022, was sequenced several times with different levels of coverage. The reads from the run with 19 million reads were shortened to match the read length for GAIIx. More genes and isoforms are detected by increasing the number of reads from 10 to 40 million reads and from read length from 40 to 100 bp.

Supplementary Figure 2Description of the cases with mutations in PI3KCA, Tp53, NOTCH1 or NOTCH2. Ninety‐three percent of the patients harboring mutations in these genes reported a history of smoking. Ninety percent of the non‐synonymous mutations identified in these tumors were the result of G:C > T:A substitutions.

Supplementary Figure 3Networks analysis of differentially expressed genes identified by Limma analysis. The top panel represents the gene networks of Limma‐generated differentially expressed genes and their neighboring relationship, and the bottom panel represents an example of a significantly enriched module/pathway. For each binary comparison, neighboring genes of the identified differentially expressed genes were queried using STRING database to find enriched Gene Ontology (GO) Biological Processes within those genesets.

Supplementary data 1.

1.1.

Supplementary data related to this article can be found at http://dx.doi.org/10.1016/j.molonc.2013.03.004.

Keysar Stephen B., Astling David P., Anderson Ryan T., Vogler Brian W., Bowles Daniel W., Morton J. Jason, Paylor Jeramiah J., Glogowska Magdalena J., Le Phuong N., Eagles-Soukup Justin R., Kako Severine L., Takimoto Sarah M., Sehrt Daniel B., Umpierrez Adrian, Pittman Morgan A., Macfadden Sarah M., Helber Ryan M., Peterson Scott, Hausman Diana F., Said Sherif, Leem Ted H., Goddard Julie A., Arcaroli John J., Messersmith Wells A., Robinson William A., Hirsch Fred R., Varella-Garcia Marileila, Raben David, Wang Xiao-Jing, Song John I., Tan Aik-Choon, Jimeno Antonio, (2013), A patient tumor transplant model of squamous cell cancer identifies PI3K inhibitors as candidate therapeutics in defined molecular bins, Molecular Oncology, 7, doi: 10.1016/j.molonc.2013.03.004.