Gene expression signatures modulated by epidermal growth factor receptor activation and their relationship to cetuximab resistance in head and neck squamous cell carcinoma
© Fertig et al.; licensee BioMed Central Ltd. 2012
Received: 24 October 2011
Accepted: 22 March 2012
Published: 1 May 2012
Aberrant activation of signaling pathways downstream of epidermal growth factor receptor (EGFR) has been hypothesized to be one of the mechanisms of cetuximab (a monoclonal antibody against EGFR) resistance in head and neck squamous cell carcinoma (HNSCC). To infer relevant and specific pathway activation downstream of EGFR from gene expression in HNSCC, we generated gene expression signatures using immortalized keratinocytes (HaCaT) subjected to ligand stimulation and transfected with EGFR, RELA/p65, or HRASVal12D.
The gene expression patterns that distinguished the HaCaT variants and conditions were inferred using the Markov chain Monte Carlo (MCMC) matrix factorization algorithm Coordinated Gene Activity in Pattern Sets (CoGAPS). This approach inferred gene expression signatures with greater relevance to cell signaling pathway activation than the expression signatures inferred with standard linear models. Furthermore, the pathway signature generated using HaCaT-HRASVal12D further associated with the cetuximab treatment response in isogenic cetuximab-sensitive (UMSCC1) and -resistant (1CC8) cell lines.
Our data suggest that the CoGAPS algorithm can generate gene expression signatures that are pertinent to downstream effects of receptor signaling pathway activation and potentially be useful in modeling resistance mechanisms to targeted therapies.
Aberrant signal transduction pathways induce and maintain many cancers [1–3]. Therefore, targeted therapeutics blocking this aberrant cellular signaling activity can impede malignant progression. However, the clinical success of targeted agents relies on accurate identification of the relative contribution of the targeted pathway to the malignancy prior to treatment. For example, EGFR overexpression is associated with poor prognosis of head and neck squamous cell carcinoma (HNSCC) leading to adoption of EGFR-targeted agents including cetuximab, a monoclonal antibody against EGFR, for clinical management [4–7]. Although many HNSCC patients benefit from cetuximab treatment, a majority of patients do not respond or eventually develop acquired resistance after the initial clinical response. Previous studies have implicated activation of epistatic signaling intermediaries downstream of EGFR activation in cetuximab resistance . Thus, inference of aberrant pathway activity controlled by EGFR activation may shed light on molecular underpinnings of acquired cetuximab resistance in patients with HNSCC.
Gene expression profiles constitute an important tool to investigate and predict biochemical network activity in complex cellular systems such as human tumors. Standard class discovery techniques, such as hierarchical clustering  and singular value decomposition (SVD; ), can implicate gene expression activity common to subsets of samples. However, clustering algorithms are unable to account for the reuse of genes in diverse biological processes as is common in eukaryotic systems [11, 12]. Moreover, algorithms such as SVD infer complex combinations of responses across all measured genes, without regard for the biochemical structure of cellular signaling networks. As a result, these inference techniques often obscure the relationship of the resulting genetic signatures to the specific cell signaling processes that are active in the measured system .
On the other hand, coordinated expression changes in a priori sets of downstream targets of pathway-activated transcription factors can implicate activity in specific signaling pathways [14–16]. These pathway-inference techniques predominately use statistical techniques that quantify the magnitude of class comparison statistics, notably t-statistics, in a priori gene sets relating to cell signaling relative to a background, null distribution (reviewed in ). Similar to clustering algorithms, co-regulation of individual genes by multiple pathways and transcription factors will bias these gene set-based analysis techniques . Previous studies have shown that Markov chain Monte Carlo (MCMC) matrix factorization techniques, such as Bayesian Decomposition (BD; ) and Coordinated Gene Activity in Pattern Sets (CoGAPS; ), robustly infer gene expression patterns relating to transcription factor activity in cancers .
To generate pathway signatures for HNSCC that are well characterized at the molecular level, we use HaCaT keratinocytes and isogenic variants thereof. HaCaT cells are well characterized in reference to their biological and malignant properties [20–23]. Therefore, these HaCaT models provide a potential to identify oncogenic signatures related to signaling responses that are unencumbered by ‘background noise’ inherent to tumor tissues. The HaCaT cell lines have the further advantage as a model system because their genetic aberrations closely parallel early oncogenic events seen in HNSCC. Specifically, like HNSCC , the HaCaT lines represent a spontaneously immortalized aneuploid human keratinocyte cell line of monoclonal origin with increased telomerase activity , two mutant p53 alleles , and absence of p16INK4a expression due to promoter hypermethylation .
Here we show that the MCMC matrix factorization algorithm CoGAPS infers gene expression patterns in HaCaT keratinocytes associated with modulation of the EGFR activity by forced expression, ligand stimulation, and pharmacological inhibition . We relate these patterns directly to activity at the pathway level, in contrast to previous studies that have focused on transcription factor-level activity [13, 29]. In addition, the CoGAPS inferred patterns predict molecular response to cetuximab treatment in an isogenic pair of HNSCC cell lines with varying cetuximab sensitivities.
Compilation of protein-protein interaction data in the signaling pathways downstream of EGFR in HNSCC for pathway-level gene set inference
CoGAPS reveals transcriptional responses to EGFR pathway modulation in the isogenic HaCaT model system reflected in their gene expression changes
HaCaT experimental design
where N represents the normal distribution on each element of the matrix multiplication AP and the ith row and jth column of the matrix Σ is the standard deviation of gene expression for the ith gene and jth sample. The resulting signals that are common to subsets of the samples are summarized in the rows of the pattern matrix P, with related gene expression patterns in the corresponding columns of A. In contrast to standard techniques class-comparison algorithms that strictly infer differential expression between sets of samples, the CoGAPS analysis infers predominant signals from the gene expression data in an unsupervised analysis. As a result, CoGAPS can capture degrees of gene expression activity that are common to various sample classes. A further advantage of CoGAPS over standard class-comparison or clustering algorithms is its ability to infer activity in subsets of genes concurrently affected by the diverse biological processes in each sample type [12, 19, 29].
In analyzing the CoGAPS estimated pathway activity (Figure 3), we observe the expected global upregulation of STAT, RAS, and AKT pathways in HaCaT-EGFRWT, upregulation of RAS in HaCaT-HRASVal12D, and upregulation of AKT resulting in HaCaT-p65WT. We also observe unexpected weak activation of TGF-β in HaCaT-EGFRWT and of RAS and TGF-β in HaCaT-p65WT. These unexpected signals are likely due to pathway cross talk and the observed weak upregulation of RAS and TGF-β pathways in HaCaT-vector control. The presence of serum also enhances this upregulation of RAS and TGF-β. The final pattern is consistent across samples and reveals a global upregulation of Notch and STAT pathways in the HaCaT cell lines.
Comparison of gene expression signatures inferred in CoGAPS and linear models
We note that the six CoGAPS patterns are inferred without any prior information about the experimental conditions in Table 1, whereas standard linear models of pathway response require information about the pertinent experimental conditions. As a result, linear models formed based upon the experimental conditions in the HaCaT forced expression experiments may overfit expression changes due to treatment and stimulation conditions. Moreover, they would be unable to account for relative changes in the magnitude of expression across comparable samples encoded in the relative magnitude of rows of P (Figure 2).
Figure 4b and 4c show the pathway statistics from the linear model gene expression signatures using standard gene set tests and a test analogous to CoGAPS, respectively, as described in the Methods section. In contrast to CoGAPS (Figure 3), the gene set statistics from the linear model (Figure 4b) show strong upregulation of only the AKT and Notch pathways in both the EGFR and RAS pathways, with only weak upregulation of RAS, which would be predicted from the experimental conditions. While the linear model reveals expected upregulation of the AKT pathway in HaCaT-p65WT, it also infers unexpected upregulation of Notch and strong upregulation of the STAT pathway. This latter STAT upregulation due to p65 overexpression is inconsistent with the structure of EGFR protein-protein interactions (Figure 1), in which p65 is far downstream of STAT. In further contrast to the CoGAPS gene expression patterns, the linear model does not infer expected STAT pathway upregulation in the HaCaT-HRASVal12D or HaCaT-EGFRWT data. Similar to CoGAPS, the linear model infers a strong upregulation of TGF-β from the introduction of serum. Unlike CoGAPS, the linear model also predicts that serum strongly upregulates signaling in the RAS pathway. The signaling patterns are largely similar using the CoGAPS-based permutation statistics from eq. 4 in Figure 4c. In this case, the linear model statistics do infer a weak upregulation of RAS in HaCaT-HRASVal12D. However, unlike the gene expression patterns from CoGAPS, the linear model for HaCaT-EGFRWT still does not infer the expected upregulation of the STAT or RAS pathways.
Gene expression signatures from CoGAPS distinguish pathway-level response in isogenic cetuximab sensitive and resistant HNSCC cell lines
CoGAPS demonstrated improved ability to correctly infer patterns with even subtle transcriptional differences because this algorithm can accurately account for gene re-use [12, 19, 39]. In contrast to linear models, CoGAPS seeks gene expression signatures with minimal structure. As a result, the signatures inferred with CoGAPS are more similar than those inferred with linear models. Therefore, these CoGAPS expression signatures reveal degrees of activation of the pathways downstream of EGFR due to pathway cross-talk, unlike the standard gene set test or permutation gene set analysis for linear models. For example, only the CoGAPS analysis can detect the modest RAS and STAT signals associated with forced expression of EGFR and HRAS in HaCaT cells. Moreover, the sparsity in the CoGAPS prior ensures that gene expression amplitude is assigned only to pertinent patterns. As a result, while the linear model assigns Notch activity to each of the forced expression signatures, CoGAPS correctly assigns this global signature to the background term.
This type of statistical modeling can delineate the key pathway activity of relevance to the development of targeted agents in genetically heterogeneous HNSCC. In our previous molecular characterization of HNSCC based on gene expression profiling, we have shown that there are at least four subtypes of HNSCC and these subtypes have prognostic implications reflecting the biological heterogeneity . In addition, recent completion of whole exome sequencing of HNSCC has shown that tobacco-related HNSCC contain average of 20.6 mutations per tumor and many of the mutated genes are tumor suppressors that cannot be easily targeted [34, 35]. Furthermore, many of these mutated genes are regulated in contextual manner meaning that one mutation of a gene in a tumor may result in a different phenotype depending on the genetic context determined by other co-existing mutations or deregulations. The analyses of global gene expression signature will potentially yield the dominant pathways that result from lack of tumor suppressor or contextual regulated genes, and allow exploitation as therapeutic targets or biomarkers of clinical outcome. It is also a potentially valuable tool of determining the on- and off-target effects of targeted agents and hypothesis generating approach to unravel the mechanism of resistance by unbiased examination of global changes in the inter-connected pathways induced by inhibition of the targeted pathway. Due to the small sample size, we could not infer accurate expression signatures from the pharmacological inhibition of EGFR, MEK or PI3K in our model system. However, additional studies to optimize the experimental design and to further validate the model for utilization in experimental therapeutics are in progress.
In this study, the inferred CoGAPS gene expression signatures implicate signaling processes in the HNSCC system that they model. For example, the gene expression signature related to constitutive activation of the RAS pathway in the HaCaT-HRASVal12D distinguishes the transcriptional profile of UMSCC1 and 1CC8 and their relative transcriptional response to cetuximab treatment. This observation suggests that over-activation of the RAS pathway cannot be repressed by cetuximab in resistant HNSCC cell lines. While HRAS mutations are found in HNSCC , the UMSCC1 and 1CC8 cell lines are HRAS wild type. Therefore, this aberrant activity in the RAS pathway that our algorithm inferred for 1CC8 is consistent with activation of the wild type RAS pathway in cetuximab resistance. One possible mechanism for this activation would be the compensatory pathway activity proposed in . Further studies to validate the mechanisms underlying these are currently ongoing.
This work demonstrates the versatility of the CoGAPS matrix factorization algorithm to infer biological signaling nodes and intermediaries as they relate to specific gene expression in immortalized HaCaT and transformed variants thereof. For example, upon stimulation/deregulation of EGFR activity, the algorithm successfully identified gene expression signatures consistent with known elements of the EGFR signaling network (Figure 1). In contrast to linear models, the CoGAPS algorithm performs pathway inference without a priori knowledge of the experimental conditions listed in Table 1. Pathway inference by CoGAPS was predictive across a heterogeneous set of experimental pathway manipulations, i.e., transcriptional responses triggered either by overexpression of EGFR, NF-kB/p65 or mutant HRAS, or those induced by addition of serum to culture media.
Similarly to the UMSCC1/1CC8 model, we hypothesize that the pathway-level gene expression signatures inferred from the HaCaT model with CoGAPS will implicate relevant molecular mechanisms in gene expression data from HNSCC patients in future studies. For example, the gene expression signature relating to constitutive RAS activation estimated from the HaCaT-HRASVal12D provides a potential biomarker to infer patient-specific cetuximab sensitivity and resistance prior to clinical treatment. Moreover, previous studies have also linked cetuximab resistance to increased AKT pathway activity in HNSCC cell lines with cetuximab resistance . We hypothesize that application of the CoGAPS algorithm to future samples of HaCaT with forced expression of RAF and PI3K with activating mutations will distinguish whether the activation of the RAS pathway induces AKT pathway activity preferentially over MAPK pathway (Figure 1) in cetuximab resistance. When applied to tumor data, these additional CoGAPS inferred gene expression signatures would also provide candidate molecular targets for patients with predicted cetuximab resistance.
Inhibition and stimulation of EGFR signaling pathway in the HaCaT model system and HNSCC cell lines
HaCaT cells overexpressing EGFR (HaCaT-EGFR) were maintained in cell culture media (W489) as previously described [20, 22]. The cells were seeded in 100 mm tissue culture plates in regular media until they reach 70-80 % confluency. After incubation with serum-free media for 12 hours, EGF or TNF-α (10 ng/ml, Sigma-Aldrich, St. Louis, MO) were then added for 4 or 8 hours before cells were harvested for total RNA isolation. The culture conditions, total RNA isolation and microarray experiments of SCC1 and 1CC8 HNSCC cell lines used in the study were previously published .
Microarray data preprocessing
To ensure that the microarray measurements from the HaCaT samples and the UMSCC1/1CC8 cell lines are comparable, we normalize all microarray measurements using fRMA . After fRMA normalization, clustering reveals an apparent batch effect from processing date in the HaCaT expression data (Additional file 3: Figure S 3a). We, therefore, fit a linear model with date and experimental conditions summarized in Table 1 using the lmFit function in the R Bioconductor package limma (Linear Models for Microarray Data; ). Additional file 3: Figure S 3b shows that samples correctly cluster by experimental condition after removing the modeled date effect using the clustering diagram generated with . This batch corrected, normalized data will be used for subsequent analyses of the HaCaT cell lines. For robustness, the clustering and linear models were performed on all HaCaT samples, including eight samples that were treated with pharmacological agents. These latter samples were excluded from subsequent analyses due to an inability to infer accurate expression signatures from their small samples size. On the other hand, the UMSCC1 and 1CC8 samples are processed in the same batch, requiring no correction after fRMA normalization. All the data are MIAME compliant and the raw data for all samples have been deposited in Gene Expression Omnibus (HaCaT data GSE32975; SCC1 and 1CC8 data in GSE21483). All analyses are performed with software R using scripts provided as Additional file 7: Files S 1, Additional file 8: File S 2, Additional file 9: File S 3, Additional file 10: File S 4 and Additional file 11: File S 5.
Pathway and transcription factor targets
We identify candidate transcription factor regulators for each probe of the Affymetrix U133 Plus 2.0 array from TRANSFAC using the Automated Sequence Annotation Pipelines (ASAP; ). The list used in these analyses was obtained and frozen on December 8, 2010 and provided as a supplemental file containing the ASAP results (Additional file 7: File S 2). Gene sets related to each pathway listed in Additional file 6: Table S 1are defined as the targets of each transcription factor identified as downstream to the pathway from Additional file 6: Table S 1.
We limit all subsequent analyses of HaCaT and UMSCC1/1CC8 cell line expression to only those probes that are annotated in TRANSFAC. We select a single probe for each gene to further avoid biases in gene set tests from using multiple probes for a single gene. Specifically, we retain the probe for each gene with the smallest p-value resulting from comparisons of the HaCaT experimental conditions using t-statistics moderated with empirical Bayes from the limma package .
CoGAPS pattern inference and pathway analysis
CoGAPS factors the expression matrix D into amplitude (A) and pattern (P) matrices with p patterns according to the distribution in eq. 1. The posterior distribution for elements of each of these matrices are computed with an MCMC Gibbs sampler based upon the atomic prior of  and implemented in the Bioconductor package CoGAPS . Here, the standard deviation in eq. 1 is given by for each gene i and sample j, based upon the established, microarray error-model in  and previous applications of [13, 29].
The quality of the CoGAPS fit is assessed through the χ2 fit of the posterior mean of A and P and identifiability of these matrices across MCMC simulations. Using these criterion, the optimal number of patterns p for the matrix factorization is the minimum number of patterns for which the χ2 reaches a minimal value and the inferred patterns persist across simulations. For the HaCaT expression data analyzed in this paper, the χ2value of the fit begins to plateau for 6 patterns (Additional file 4: Figure S 4), at which point multiple CoGAPS simulations obtain the same A and P matrices (Additional file 1: Figures S 1 and Additional file 2: S 2).
where and are the sample mean and variance, respectively, from the CoGAPS MCMC samples for gene (row) g and pattern (column) p amplitude. We compute p-values through a permutation test that compares the computed value of to a null distribution obtained by the values of the statistic in eq. 2 resulting from selecting random sets of R genes. In contrast to , we apply this statistic to both targets of individual transcription factors and targets of sets of transcription factors regulated in the pathways (Additional file 6: Table S 1) to infer pathway level activity. For visualization, p-values computed from the statistic in eq. 2 are transformed as follows
Therefore, a transformed value of −1 indicates underrepresentation of set G and +1 indicates over-representation. The associated pathway-level statistics in Figure 3 represent the mean of the statistic from eq. (3) across three CoGAPS simulations, with error bars representing the minimum and maximum values in each of these simulations.
Projecting the gene expression signatures in the columns of A onto additional samples can implicate the relative activity of inferred patterns in those samples. In this paper, the projection is implemented by solving the factorization in eq. 1 for the new data matrix where A is fixed as the average of the CoGAPS posterior mean for each of the three CoGAPS simulations performed. We estimate the patterns P associated with this amplitude matrix using the least-squares fit to the new data implemented with the lmFit function in the limma package . Applying this projection to the original, HaCaT data reveals that the projection provides similar, albeit slightly nosier, estimates when compared to the CoGAPS posterior mean for P (Additional file 5: Figure S 5). This linear projection is, therefore, used to project the gene signatures inferred from the HaCaT data onto gene expression data from the HNSCC UMSCC1 and 1CC8 cell lines. Differences between the values of the pattern matrix P inferred through CoGAPS or these subsequent projections are quantified using p-values from a t-test between the groups of experimental conditions being compared.
Linear models and gene set pathway-level analysis
where Pdown and Pup are the p-values resulting from the geneSetTest function if the alternative hypothesis is specified as down or up regulated, respectively. The heatmap in Figure 4b then rescales these statistics across columns are 1 at the maximum value of and −1 at the minimum value. For further comparison to the CoGAPS results, Figure 4c plots the transformed p-values resulting from the permutation-based CoGAPS gene set test in eqs. 2 and 3. To reflect the statistics of the linear model, the posterior estimate for the ratio in eq. 2 is replaced with the estimated, empirical Bayes moderated t-statistics for each of the six conditions specified in the linear model.
EJF, MC, and MFO were responsible for data analyses and manuscript preparation. QR, HH, and HC performed data collection and assisted with manuscript preparation. UR and AD were involved with study design, data analysis, and manuscript preparation. CHC was responsible for study design, data collection, data analysis, and manuscript preparation.
Acknowledgements and Funding
The authors would like to acknowledge funding from the following sources: EJF: K25 (CA141053), pilot project from the Sidney Kimmel Cancer Center of Johns Hopkins University Head and Neck SPORE; AD: funding from RTOG; UR: Commonwealth of Pennsylvania through the American College of Radiology and seed grant from TRP-RTOG; MFO: R21 (LM009382); CHC: R01 (DE017982) and Damon Runyon Clinical Investigator Award (CI-28-05).
- Cox AD, Der CJ: Ras family signaling: therapeutic targeting. Cancer Biol Ther. 2002, 1: 599-606.View ArticlePubMedGoogle Scholar
- Downward J: Targeting RAS signalling pathways in cancer therapy. Nat Rev Cancer. 2003, 3: 11-22. 10.1038/nrc969.View ArticlePubMedGoogle Scholar
- Hanahan D, Weinberg RA: Hallmarks of cancer: the next generation. Cell. 2011, 144: 646-74. 10.1016/j.cell.2011.02.013.View ArticlePubMedGoogle Scholar
- Cohen EEW, Rosen F, Stadler WM, Recant W, Stenson K, Huo D: Phase II trial of ZD1839 in recurrent or metastatic squamous cell carcinoma of the head and neck. J Clin Oncol. 2003, 21: 1980-7. 10.1200/JCO.2003.10.051.View ArticlePubMedGoogle Scholar
- Soulieres D, Senzer NN, Vokes EE, Hidalgo M, Agarwala SS, Siu LL: Multicenter phase II study of erlotinib, an oral epidermal growth factor receptor tyrosine kinase inhibitor, in patients with recurrent or metastatic squamous cell cancer of the head and neck. J Clin Oncol. 2004, 22: 77-85.View ArticlePubMedGoogle Scholar
- Bonner JA, Harari PM, Giralt J, Azarnia N, Shin DM, Cohen RB: Radiotherapy plus cetuximab for squamous-cell carcinoma of the head and neck. N Engl J Med. 2006, 354: 567-78. 10.1056/NEJMoa053422.View ArticlePubMedGoogle Scholar
- Vermorken JB, Mesia R, Rivera F, Remenar E, Kawecki A, Rottey S: Platinum-based chemotherapy plus cetuximab in head and neck cancer. N Engl J Med. 2008, 359: 1116-27. 10.1056/NEJMoa0802656.View ArticlePubMedGoogle Scholar
- Hatakeyama H, Cheng H, Wirth P, Counsell A, Marcrom SR, Wood CB: Regulation of heparin-binding EGF-like growth factor by miR-212 and acquired cetuximab-resistance in head and neck squamous cell carcinoma. PLoS One. 2010, 5: e12702-10.1371/journal.pone.0012702.PubMed CentralView ArticlePubMedGoogle Scholar
- Eisen MB, Spellman PT, Brown PO, Botstein D: Cluster analysis and display of genome-wide expression patterns. Proc Natl Acad Sci USA. 1998, 95: 14863-8. 10.1073/pnas.95.25.14863.PubMed CentralView ArticlePubMedGoogle Scholar
- Alter O, Brown PO, Botstein D: Singular value decomposition for genome-wide expression data processing and modeling. Proc Natl Acad Sci USA. 2000, 97: 10101-6. 10.1073/pnas.97.18.10101.PubMed CentralView ArticlePubMedGoogle Scholar
- Moloshok TD, Klevecz RR, Grant JD, Manion FJ, Speier WF: Ochs MF: Application of Bayesian decomposition for analysing microarray data. Bioinformatics. 2002, 18: 566-75. 10.1093/bioinformatics/18.4.566.View ArticlePubMedGoogle Scholar
- Bidaut G, Suhre K, Claverie J, Ochs MF: Determination of strongly overlapping signaling activity from microarray data. BMC Bioinformatics. 2006, 7: 99-10.1186/1471-2105-7-99.PubMed CentralView ArticlePubMedGoogle Scholar
- Ochs MF, Rink L, Tarn C, Mburu S, Taguchi T, Eisenberg B: Detection of treatment-induced changes in signaling pathways in gastrointestinal stromal tumors using transcriptomic data. Cancer Res. 2009, 69: 9125-32. 10.1158/0008-5472.CAN-09-1709.PubMed CentralView ArticlePubMedGoogle Scholar
- Efroni S, Schaefer CF, Buetow KH: Identification of key processes underlying cancer phenotypes using biologic pathway analysis. PLoS One. 2007, 2: e425-10.1371/journal.pone.0000425.PubMed CentralView ArticlePubMedGoogle Scholar
- Tarca AL, Draghici S, Khatri P, Hassan SS, Mittal P, Kim J: A novel signaling pathway impact analysis. Bioinformatics. 2009, 25: 75-82. 10.1093/bioinformatics/btn577.PubMed CentralView ArticlePubMedGoogle Scholar
- Vaske CJ, Benz SC, Sanborn JZ, Earl D, Szeto C, Zhu J: Inference of patient-specific pathway activities from multi-dimensional cancer genomics data using PARADIGM. Bioinformatics. 2010, 26: i237-45. 10.1093/bioinformatics/btq182.PubMed CentralView ArticlePubMedGoogle Scholar
- Emmert-Streib F, Glazko GV: Pathway analysis of expression data: deciphering functional building blocks of complex diseases. PLoS Comput Biol. 2011, 7: e1002053-10.1371/journal.pcbi.1002053.PubMed CentralView ArticlePubMedGoogle Scholar
- Kossenkov AV, Ochs MF: Matrix factorization for recovery of biological processes from microarray data. Methods Enzymol. 2009, 467: 59-77.PubMed CentralView ArticlePubMedGoogle Scholar
- Fertig EJ, Ding J, Favorov AV, Parmigiani G, Ochs MF: CoGAPS: an R/C++ package to identify patterns and biological process activity in transcriptomic data. Bioinformatics. 2010, 26: 2792-3. 10.1093/bioinformatics/btq503.PubMed CentralView ArticlePubMedGoogle Scholar
- Boukamp P, Petrussevska RT, Breitkreutz D, Hornung J, Markham A, Fusenig NE: Normal keratinization in a spontaneously immortalized aneuploid human keratinocyte cell line. J Cell Biol. 1988, 106: 761-71. 10.1083/jcb.106.3.761.View ArticlePubMedGoogle Scholar
- Jost M, Huggett TM, Kari C, Boise LH, Rodeck U: Epidermal growth factor receptor-dependent control of keratinocyte survival and Bcl-xL expression through a MEK-dependent pathway. J Biol Chem. 2001, 276: 6320-6. 10.1074/jbc.M008210200.View ArticlePubMedGoogle Scholar
- Quadros MRD, Peruzzi F, Kari C, Rodeck U: Complex regulation of signal transducers and activators of transcription 3 activation in normal and malignant keratinocytes. Cancer Res. 2004, 64: 3934-9. 10.1158/0008-5472.CAN-04-0214.View ArticlePubMedGoogle Scholar
- Ren Q, Kari C, Quadros MRD, Burd R, McCue P, Dicker AP: Malignant transformation of immortalized HaCaT keratinocytes through deregulated nuclear factor kappaB signaling. Cancer Res. 2006, 66: 5209-15. 10.1158/0008-5472.CAN-05-4158.View ArticlePubMedGoogle Scholar
- Forastiere A, Koch W, Trotti A, Sidransky D: Head and neck cancer. N Engl J Med. 2001, 345: 1890-900. 10.1056/NEJMra001375.View ArticlePubMedGoogle Scholar
- Kallassy M, Martel N, Damour O, Yamasaki H, Nakazawa H: Growth arrest of immortalized human keratinocytes and suppression of telomerase activity by p21WAF1 gene expression. Mol Carcinog. 1998, 21: 26-36. 10.1002/(SICI)1098-2744(199801)21:1<26::AID-MC5>3.0.CO;2-N.View ArticlePubMedGoogle Scholar
- Lehman TA, Modali R, Boukamp P, Stanek J, Bennett WP, Welsh JA: p53 mutations in human immortalized epithelial cell lines. Carcinogenesis. 1993, 14: 833-9. 10.1093/carcin/14.5.833.View ArticlePubMedGoogle Scholar
- Harvat BL, Jetten AM: Decreased growth inhibitory responses of squamous carcinoma cells to interferon-gamma involve failure to recruit cki proteins into cdk2 complexes. J Invest Dermatol. 2001, 117: 1274-81. 10.1046/j.0022-202x.2001.01495.x.View ArticlePubMedGoogle Scholar
- Chung CH, Parker J, Levy S, Slebos RJ, Dicker AP, Rodeck U: Gene expression profiles as markers of aggressive disease-EGFR as a factor. Int J Radiat Oncol Biol Phys. 2007, 69: S102-5.PubMed CentralView ArticlePubMedGoogle Scholar
- Kossenkov AV, Peterson AJ, Ochs MF: Determining transcription factor activity from microarray data using Bayesian Markov chain Monte Carlo sampling. Stud Health Technol Inform. 2007, 129: 1250-1254.PubMedGoogle Scholar
- Morgan S, Grandis JR: ErbB receptors in the biology and pathology of the aerodigestive tract. Exp Cell Res. 2009, 315: 572-582. 10.1016/j.yexcr.2008.08.009. Feb 15PubMed CentralView ArticlePubMedGoogle Scholar
- Ratushny V, Astsaturov I, Burtness BA, Golemis EA, Silverman JS: Targeting EGFR resistance networks in head and neck cancer. Cell Signal. 2009 Aug, 21: 1255-1268. 10.1016/j.cellsig.2009.02.021.PubMed CentralView ArticlePubMedGoogle Scholar
- Koch U, Radtke F: Notch signaling in solid tumors. Curr Top Dev Biol. 2010, 92: 411-55.View ArticlePubMedGoogle Scholar
- White RA, Malkoski SP, Wang X: TGFβ signaling in head and neck squamous cell carcinoma. Oncogene. 2010, 29: 5437-46. 10.1038/onc.2010.306.PubMed CentralView ArticlePubMedGoogle Scholar
- Agrawal N, Frederick MJ, Pickering CR, Bettegowda C, Chang K, Li RJ: Exome sequencing of head and neck squamous cell carcinoma reveals inactivating mutations in NOTCH1. Science. 2011, 333: 1154-7. 10.1126/science.1206923.PubMed CentralView ArticlePubMedGoogle Scholar
- Stransky N, Egloff AM, Tward AD, Kostic AD, Cibulskis K, Sivachenko A: The Mutational Landscape of Head and Neck Squamous Cell Carcinoma. Science. 2011, 333: 1157-1160. 10.1126/science.1208130.PubMed CentralView ArticlePubMedGoogle Scholar
- Smyth GK: Linear models and empirical bayes methods for assessing differential expression in microarray experiments. Stat Appl Genet Mol Biol. 2004, 3: Article3Google Scholar
- Boukamp P, Stanbridge EJ, Foo DY, Cerutti PA, Fusenig NE: c-Ha-ras oncogene expression in immortalized human keratinocytes (HaCaT) alters growth potential in vivo but lacks correlation with malignancy. Cancer Res. 1990, 50: 2840-7.PubMedGoogle Scholar
- Breitkreutz D, Boukamp P, Ryle CM, Stark HJ, Roop DR, Fusenig NE: Epidermal morphogenesis and keratin expression in c-Ha-ras-transfected tumorigenic clones of the human HaCaT cell line. Cancer Res. 1991, 51: 4402-9.PubMedGoogle Scholar
- Ochs MF, Moloshok TD, Bidaut G, Toby G: Bayesian decomposition: analyzing microarray data within a biological context. Ann N Y Acad Sci. 2004, 1020: 212-26. 10.1196/annals.1310.018.View ArticlePubMedGoogle Scholar
- Chung CH, Parker JS, Karaca G, Wu J, Funkhouser WK, Moore D: Molecular classification of head and neck squamous cell carcinomas using patterns of gene expression. Cancer Cell. 2004, 5: 489-500. 10.1016/S1535-6108(04)00112-6.View ArticlePubMedGoogle Scholar
- McCall MN, Bolstad BM, Irizarry RA: Frozen robust multiarray analysis (fRMA). Biostatistics. 2010, 11: 242-53. 10.1093/biostatistics/kxp059.PubMed CentralView ArticlePubMedGoogle Scholar
- Coombes KR: ClassDiscovery: Classes and methods for "class discovery" with microarrays or proteomics. [http://bioinformatics.mdanderson.org/Software/OOMPA]
- Kossenkov A, Manion FJ, Korotkov E, Moloshok TD, Ochs MF: ASAP: automated sequence annotation pipeline for web-based updating of sequence information with a local dynamic database. Bioinformatics. 2003, 19: 675-6. 10.1093/bioinformatics/btg056.View ArticlePubMedGoogle Scholar
- Sibisi S, Skilling J: Prior distributions on measure space. Journal of the Royal Statistical Society, B. 1997, 59: 217-235. 10.1111/1467-9868.00065.View ArticleGoogle Scholar
- Rocke DM, Durbin B: A model for measurement error for gene expression arrays. J Comput Biol. 2001, 8: 557-69. 10.1089/106652701753307485.View ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.