Comparative and integrative functional genomics of HCC - Nature

Oncogene (2006) 25, 3801–3809

& 2006 Nature Publishing Group All rights reserved 0950-9232/06 $30.00 www.nature.com/onc

REVIEW

Comparative and integrative functional genomics of HCC J-S Lee and SS Thorgeirsson Laboratory of Experimental Carcinogenesis, Center for Cancer Research, National Cancer Institute, National Institutes of Health, Bethesda, MD, USA

Global gene expression profiling of hepatocellular carcinoma (HCC) is a promising new technology that has already refined the diagnosis and prognostic predictions of HCC patients. This has been accomplished by identifying genes whose expression pattern is associated with clinicopathological features of HCC tumors. Molecular characterization of HCC from gene expression profiling studies will undoubtedly improve the prediction of treatment responses, selection of treatments for specific molecular subtypes of HCC and ultimately the clinical outcome of HCC patients. The research focus is now shifting toward the identification of genetic determinants that are components of the specific regulatory pathways altered in cancers, and that may constitute novel therapeutic targets. Here we review the recent advances in gene expression profiling of HCC and discuss the future strategies for analysing large and complicated data sets from microarray studies and how to integrate these with diverse genomic data. Oncogene (2006) 25, 3801–3809. doi:10.1038/sj.onc.1209561 Keywords: DNA microarray; gene expression profile; hepatocellular carcinoma (HCC); comparative functional genomics; integrative functional genomics

Introduction Hepatocellular carcinoma (HCC) is one of the most common cancer in the world, accounting for an estimated 600 000 deaths annually (Parkin et al., 2005). Although HCC is common in southeast Asia and subSahara Africa, the incidence rate of HCC has continued to increase in the United States and western Europe over the past 25 years, and the incidence and mortality rates of HCC are expected to double over the next 10–20 years (El Serag and Mason, 1999; Davila et al., 2004; El Serag, 2004). Although much is known about both the cellular changes that lead to HCC and the etiological agents (i.e., HBV, HCV infection and alcohol) responsible for the majority of HCC, the molecular Correspondence: Dr SS Thorgeirsson, Laboratory of Experimental Carcinogenesis, Center for Cancer Research, National Cancer Institute, National Institutes of Health, 37 Convent Drive, Building 37, Room 4146, Bethesda, MD 20892-4262, USA. E-mail: [email protected]

pathogenesis of HCC is not well understood (Thorgeirsson and Grisham, 2002; Llovet et al., 2003; Bruix et al., 2004). Moreover, the severity of HCC, the lack of good diagnostic markers and treatment strategies, and clinical heterogeneity have rendered the disease a major challenge (Thorgeirsson and Grisham, 2002; Lee and Thorgeirsson, 2004). Patients with HCC have a highly variable clinical course (Llovet et al., 1999, 2003), indicating that HCC comprises several biologically distinctive subgroups. Despite the considerable efforts that have been devoted to the rationalization over empirical approaches to establishing classification system (Pugh et al., 1973; Okuda et al., 1985; Calvet et al., 1990; CLIP investigators, 1998; Chevret et al., 1999; Tan et al., 2003), clinical and pathological diagnosis and classification of HCC remain unreliable in predicting patients’ survival and responses to therapy. The prognostic variability likely reflects a molecular heterogeneity that has not been appreciated from methods traditionally used to characterize HCC. Improving the classification of HCC patients into groups with homogeneous prognosis as well as a more comprehensive understanding of the underlying biology of HCC development at the molecular level would at minimum improve the application of currently available treatment modalities and at best offer new treatment strategies. Investigation of various cancers at the molecular level is well underway through functional genomic approaches including DNA microarray technology that can simultaneously detect the expression levels of thousands of genes. In addition to large-scale DNA sequencing projects that have accomplished tremendous success in providing molecular architecture of genes and genomes, high-throughput gene expression profiling projects will continue to be the primary driver behind the comprehensive molecular catalogs of human disease. This new technology has been successfully used to predict the clinical outcome and survival as well as to classify different types of cancer (Alizadeh et al., 2000; Bittner et al., 2000; Beer et al., 2002; Lee and Thorgeirsson, 2002; van’t Veer et al., 2002; Valk et al., 2004; Lee et al., 2004a; Roepman et al., 2005). This review summarizes recent studies on gene expression profiling of HCC aimed at better understanding the molecular pathogenesis of the disease, the promising clinical applications of the gene expression data as well as addressing the potential future use of this new technology.

Functional genomics of HCC J-S Lee and SS Thorgeirsson

3802

Gene expression profiling of HCC Gene expression patterns associated with clinicopathological features of HCC As conventional approaches for the prognostic classification of HCC largely rely on the single or multiple clinicopathological variables such as the severity of the liver function and characteristics of tumor (i.e., size, number of nodules, vascular invasion, distant metastasis and tumor grade), the majority of DNA-microarray studies in human HCC focused on uncovering differences in gene expression patterns among tissues that are clinically or histologically well-defined. As expected, prediction models based on gene expression profiles successfully discriminate HCC from non-tumor livers with expected high accuracy (Kim et al., 2004; Neo et al., 2004; Nam et al., 2005). Many studies have identified differentially expressed gene sets in HCC that differ according to etiological factors (Okabe et al., 2001), mutations of tumor suppressor genes (Chen et al., 2002), rate of recurrence (Iizuka et al., 2003), intrahepatic metastasis (Ye et al., 2003) and different stages (Nam et al., 2005). The results from these studies demonstrate that genomic-scale gene expression profiles recapitulate the well-known morphological distinctions of the tissues and suggest that the differences in gene expression patterns might be biologically relevant. However, most of these studies have identified genes that are associated with limited aspects of the tumor pathogenesis, but do not necessarily reflect the underlying biological properties that likely drive tumor behavior. Therefore, the true value that can be added by genome-wide survey of gene expression is likely to arise from the identification of biologically and clinically homogeneous subgroups of HCC that cannot be recognized by the conventional characterization of tumors and the development of predictive gene expression profiles and selection of specific biomarkers that will guide personalized therapies for HCC patients. Uncovering new subclasses of HCC In a recent study (Lee et al., 2004a), unbiased analytical approach was applied on gene expression data from human HCC to investigate the possibility that variations in gene expression of HCC would permit the identification of distinct subclasses of HCC patients with different prognoses. Two independent but complementary approaches were applied for data analysis to uncover subclasses of HCC and the underlying biological differences at the molecular level between the subclasses. First, hierarchical clustering of the data was applied, and revealed two subclasses of HCC that are strongly associated with survival of the patients (Figure 1a). Second, several independent prediction algorithms were used to determine whether gene expression patterns could be used to predict survival (Figure 1b). Hepatocellular carcinoma patients were randomly divided into two equal groups: a training set that was used to develop the HCC classifiers, and a validation set that was used to evaluate the test. All five models successfully separated Oncogene

poor survival patients from better survival patients. These results demonstrated not only a strong association of gene expression patterns with the survival of the patients but also a robust reproducibility of these gene expression-based predictors. A univariate Cox regression model was used to identify individual genes whose expression is highly correlated with the length of survival. Application of survival-associated genes for subclass prediction was highly accurate as illustrated by the fact that averaged gene expression indices from the selected ‘survival genes’ were sufficient to segregate the two subclasses even without the use of sophisticated prediction models. Information obtained from knowledge-based annotation of the survival genes provided insight into the underlying biological differences between the two subclasses of HCC. Out of several biological groups of the survival genes, the cell proliferation group was the best predictor of an unfavorable outcome of the disease. Expression of typical cell proliferation markers like proliferating cell nuclear antigen (PCNA) and cell cycle regulators such as CDK4, CCNB1, CCNA2 and CKS2 was greater in the poor survival subclass (cluster A). Also, not surprisingly, many genes that are expressed more in poorer survival subclasses are antiapoptotic. Higher expression of genes involved in ubiquitination and sumoylation was also observed in the poor survival subclass. Enhanced activation of ubiquitin-dependent protein degradation may account for deregulation of cell cycle control and faster cell proliferation in the poor survival subclass. The study shows that gene expression profiling analysis can identify previously unrecognized, clinically relevant subclasses of HCC in a robust and reproducible manner. Comparative functional genomics Despite of facts that transgenic and knockout mouse models have significantly enhanced our understanding of human cancers, we do in most cases still rely on casual correlation between human and mouse models owing to lack of methods for direct comparisons (Lee et al., 2005). Aligning orthologous genomic sequence from closely related species allowed the discovery of evolutionarily conserved sequences that might be important functional elements. The neutral theory of molecular evolution provides a framework for the identification of functional DNA sequences in genomes of different species (Kimura, 1968; King and Jukes, 1969). The neutral theory postulates that functionally important elements in genome sequences tend to evolve at a slower rate than do less important elements as we already observed in coding sequences of genes. This difference has permitted the identification of both protein coding sequences and functional noncoding sequences in a genome (Eddy, 2002; Cooper and Sidow, 2003; Ureta-Vidal et al., 2003). If the regulatory elements of evolutionarily related species are conserved, it is reasonable to hypothesize that gene expression signatures reflecting similar phenotypes in different


3803

Figure 1 Two analytical strategies are commonly applied; supervised and unsupervised analysis. Supervised analysis is commonly used to identify groups of genes that have associated biological or pathological features, and if necessary, use these genes to build prediction models based on gene expression profiles. Unsupervised analysis solely relies on gene expression profiles to group or classify tumor tissues. Hierarchical clustering and principal component analysis are the best known methods for unsupervised analysis. These transcriptionally defined groups of tumors can then be explored for the association of biological or clinicopathological features independently obtained from tissues. (a) Hierarchical clustering analysis was applied to gene expression data from hepatocellular carcinoma (HCC), and revealed two distinctive subtypes of gene expression patterns among 89 HCC. Kaplan–Meier plot of overall survival of HCC patients showed strong association of gene expression patterns with patients’ survival. (b) Schematic overview of the strategy used for the construction of prediction models and evaluation of predicted outcomes based on gene expression signatures. The class predictions were carried out using cross-validation approach and several different algorithms, linear discriminator analysis (LDA), support vector machines (SVM), nearest centroid (NC), nearest neighbor (NN) and compound covariate predictor (CCP). Before applying prediction models, HCC tissues were randomly divided into two equal groups, a training set and a validation set, to estimate the reproducibility of the tests. Each predictor independently identified the most differentially expressed genes between two clusters in the training set. The collected genes were then used to build prediction models that can estimate the probability that a given HCC tissue is cluster A or B. In order to minimize the misclassification error, numbers of genes in each predictor were optimized during leave-one-out cross-validation in training set tissues. During cross-validation, one tissue was removed at a time from the training set, and the remaining tissues were used to build prediction models. The identity of the left-out tissue was predicted based on the given algorithm of predictors. Cross-validation was repeated until each tissue in the training set has been left out once. After applying six classifiers (predictors) to the validation set, Kaplan–Meier survival plots and log rank test can be used to access statistical significance of predicted groups in survival.

species could also be conserved. This hypothesis forms the basis for our attempt to cross-compare the gene expression data from human HCC and mouse HCC models to identify aberrant phenotypes reflecting molecular pathways that are evolutionarily conserved. Three recent studies provide clues on how to address the possibility that the comparison of gene

expression patterns from human tumors and mouse tumors would permit the direct identification of common aberrant molecular pathways involved in tumorigenesis (Ellwood-Yen et al., 2003; Lee et al., 2004b; Sweet-Cordero et al., 2005). By comparing gene expression profiles from a prostate cancer mouse model overexpressing Myc and human prostate cancer, Oncogene


3804

Ellwood-Yen et al. (2003) identified a conserved expression module of human genes that corresponded with defined Myc signature genes in the mouse model. Mycspecific gene expression signature in the mouse model permitted the definition of a subset of MYC-like human cancers that is probably driven by MYC amplification or other mechanisms of MYC pathway activation. KRAS mutation is one of the most frequently observed genetic modifications in human lung adenocarcinomas (Wistuba et al., 2001). A recent study reported the detection of a KRAS2 gene expression signature via the comparison of a KRAS2-mediated mouse model of lung cancer with human lung cancer (Sweet-Cordero et al., 2005). Interestingly, when the investigators used gene expression data from human lung adenocarcinomas alone, no statistically significant gene expression correlate of KRAS2 mutation status was identified between wild-type and mutated KRAS2 tumors. Without the human–mouse comparison, one might overlook the existence of a gene expression signature that reflects the KRAS2 mutation in human adenocarcinomas. In an effort to identify the best-fit mouse HCC models that mimic human condition, our lab has recently integrated gene expression data from human and mouse HCC (Lee et al., 2004b). Gene expression patterns of mouse HCC were obtained from seven HCC mouse models. Orthologous human and mouse genes from both data sets were selected, and the gene expression data were integrated after standardizing the relative expression levels for both species. In hierarchical clustering analysis of integrated data (Figure 2a), gene expression patterns of HCC from Myc, E2f1 and Myc/E2f1 mice had the highest similarity with those of the better survival group of human HCC, whereas the expression patterns of Myc/Tgfa and Diethylnitrosamine (DENA)-induced mouse HCC were most similar to those of the poorer survival group of human HCC. These results suggest that these two classes of mouse models might closely recapitulate the molecular patterns of the two subclasses of human HCC. In contrast, gene expression patterns of HCC from Acox1–/– and ciprofibrate-treated mice were least similar to those observed in either subclass of human HCC. The development of HCC in these two models is driven by peroxisome proliferation in the liver (Reddy and Lalwai, 1983; Rao et al., 1984; Fan et al., 1998). These results suggest that the process of hepatocarcinogenesis induced by peroxisome proliferation in mice progresses along molecular pathways that

do not frequently occur in humans (Gonzalez et al., 1998; Cheung et al., 2004). The similarity of gene expression profiles between human and mouse models is in good agreement with the phenotypic characteristics of the tumors (Figure 2b). The human tumors with increased proliferation, decreased apoptosis and worse prognosis are paired with the mouse models with the same characteristics. The gene expression-based prediction of mouse models is highly concordant with the phenotypes of mice. Myc/Tgfa mice have a typically poor prognosis phenotype, such as an earlier and higher incident rate of HCC development, higher mortality, higher genomic instability and higher expression of poor prognostic marker (Sargent et al., 1999; Calvisi et al., 2004). Myc and Myc/E2f1 mice have a relatively higher mutation frequency of and nuclear accumulation of b-catenin (Calvisi et al., 2001, 2004), which in human HCC are indicative of lower genomic instability and better prognosis (Hsu et al., 2000; Laurent-Puig et al., 2001; Mao et al., 2001; Wong et al., 2001). The fact that these findings are first uncovered by using unsupervised methods and validated later using supervised methods indicates that the underlying principles in gene expression changes are conserved between mouse and human HCC. Although the precise molecular mechanism driving hepatocarcinogenesis is yet to be discovered, the relative similarity of Myc/Tgfa mice to the human poor survival group (subclass A) HCC indicates the role of the epidermal growth factor receptor (Yarden and Sliwkowski, 2001), or related signaling pathways in prognosis of human HCC. In a future study, it will be interesting to examine expression level, constitutive phosphorylation and mutations of EGFR in poor prognosis HCC patients to determine the therapeutic potential of EGFR tyrosine kinase inhibitor ‘gefitinib’ in a subclass of HCC patients. These results strongly suggest that well-defined gene expression signatures from experimental conditions or animal models can be used to stratify human cancer patients into more homogeneous groups at the molecular level. In many cancers, specific drugs are indicated for specific subtypes of cancer. Herceptin is only effective to the subpopulation of individuals with breast cancer who express ERBB2 (McNeil, 1998), and Gleevec is for individuals with chronic myeloid leukemia resulting from the BCRABL1 gene fusion (le Coutre et al., 1999). Similarly, two recent studies identified mutations in EGFR in lung adenocarcinomas, which predict response to the tyrosine kinase inhibitor gefitinib (Lynch et al., 2004; Paez et al.,

Figure 2 Comparative functional genomics. (a) Gene expression data of mouse and human hepatocellular carcinoma (HCC) tissues were collected independently and integrated using orthologous genes present in both microarrays. Hierarchical clustering analysis is applied to integrated data 68 mouse and 91 human HCC tumors. Correlation heat map depicting overall similarity and dissimilarity of gene expression pattern among human and mouse HCC tissues is shown. The red and green colors in cells reflect the correlation between gene expression patterns of two HCC tissues compared as indicated in the scale bar. By applying unsupervised and supervised analysis of gene expression patterns, HCC mouse models that best or least mimic human conditions were identified. (b) Phenotypic similarities between HCCs generated in the transgenic mouse models and subclass A and B of human HCC. The ‘best-fit’ HCC mouse models can then be used to test hypotheses on tumor progression that are generated by analysis of cross-species gene expression patterns or from other experimental data. These models should also be extremely valuable for testing both potential therapeutic targets identified in human studies and preclinical trials of drugs. Oncogene


3805

2004). We can anticipate that unique molecular identities of each subclass of HCC uncovered by comparative analysis of a genome-wide survey of gene expression from human and animal models will provide new therapeutic strategies to maximize the efficiency of treatments.

Integrative functional genomics Gene expression profiling studies of various cancers have discovered consistent gene expression patterns associated with a histological or clinical phenotype and discovering subtypes of cancers previously unidentified

Oncogene


3806

with conventional technologies (Alizadeh et al., 2000; van de Vijver et al., 2002; Bullinger et al., 2004; Dave et al., 2004; Lapointe et al., 2004; Valk et al., 2004; Lee et al., 2004a; Roepman et al., 2005). This approach promises to provide diagnostic and prognostic markers that can be clinically used in the near future (Lossos et al., 2004). The research focus has now shifted toward identifying genetic determinants that are components of the specific regulatory pathways altered in cancers, leading to the discovery of novel therapeutic targets. However, it is not easy to select few candidate genes for further studies from the lengthy gene lists generated from gene expression profiling studies owing to the following reasons: first, too many confounding factors are embedded in the gene expression profile data from human cancer tissues. These factors include ages, hospital cares, different protocols of treatments, nonparallel progression of cancer and unspecified environment factors that are irrelevant to HCC development. Second, the magnitude of changes or statistical significance in the altered gene expression profiles does not always reflect functional significance of the genes in HCC development. This is largely owing to the fact that gene expression patterns observed in HCC merely represent the snapshot of gene expression patterns at a single time point (i.e., at tissue collection). It is also important to recognize that differential expression of genes does not necessarily imply causality but might merely reflect the loss of liver function and rapid proliferation of tumor cells. Third, the key regulatory elements that govern the expression patterns of gene sets reflecting clinical phenotypes are often not detectable in gene expression profile studies, because in many cases their activities are regulated at the post-transcriptional level (i.e., localization, modification and interaction with other key regulatory proteins). However, this limitation has been overcome by applying cross-comparison of multiple gene expression data sets from human HCC and animal models, uncovering evolutionarily conserved gene expression patterns during HCC progression (Lee et al., 2004b). The success of new experimental and analytical approach, comparative functional genomics, suggests that more integration of independent data sets will enhance our ability to identify key regulatory elements during HCC development. Cancer cells evolve from normal cells following accumulation of genetic and epigenetic alterations (Hanahan and Weinberg, 2000). Although the gene expression patterns in cancer cells well reflect these alterations, a significant fraction of the gene expression program of the cancer cells is characteristic of the nontransformed cellular lineages from which the cancer originated (Alizadeh et al., 2000). Analysis of gene expression profiles of cancer cell lines indicated that neither physiological adaptation in vivo nor experimental adaptation in vitro was sufficient to overwrite the gene expression programs established during development (Ross et al., 2000). Cancer cells do not invent new pathways, but resurrect the pre-existing but dormant signaling pathways that were active during embryonic development or normal physiological condiOncogene

tion. In recent studies (Chang et al., 2004, 2005), gene expression signatures from experimental condition mimicking wound healing process were used to measure how much tumors may be like wounds, and developed a model to predict cancer progression based on wound healing gene expression signature. Patients carrying the wound healing signature had a significantly increased risk of metastasis when compared with patients who lack the signature, indicating that genomic data from normal physiological condition can help to predict the prognosis of cancer patients. It is therefore important to obtain, in addition to the gene expression data from human HCC and mouse HCC models, gene expression signatures from experimental animals and humans unique for different normal physiological conditions, such as liver development and regeneration as well as hepatic stem cells that can be integrated into the gene expression patterns from human HCC (Figure 3). Whole genome sequencing projects of many species including our own have not only revealed the architectures of genes and genomes, but have provided a ‘bird’s view’ on our own evolution as a species and the evolutionary landscape (Murphy et al., 2004). Crossspecies comparisons of gene expression data from human cancer and mouse models indicated that gene expression signatures reflecting similar phenotypes in two species are conserved as genomic architectures including cis-regulatory elements are conserved (Ellwood-Yen et al., 2003; Lee et al., 2004b; Sweet-Cordero et al., 2005). If gene expression signatures are conserved in evolutionarily related species, it will be reasonable to speculate that the transcription factors governing these signatures would also be conserved. However, it is not easy to identify these transcription factors from gene expression pattern data, as their activities are regulated post-transcriptionally. In order to overcome this shortcoming, it will be necessary to devise new strategy for uncovering key transcription factors governing gene expression patterns that well reflect the clinicopathological features of human cancer. Although current computational in silico prediction methods of cisregulatory elements in genomic sequences are not sophisticated enough to overcome high rate of falsely predicted sites owing to short and degenerated nature of binding sites targeted by transcription factors, comparative genomics has long held the promise for the identification of cis-regulatory elements in mammalian genomes (Hardison et al., 1997). Previous study showed that cross-species comparison in binding sites prediction (‘phylogenetic footprinting’) enhances predictive sensitivity when analysing promoter sequences conserved between humans and mice (Lenhard et al., 2003). Therefore, key transcription factors might be uncovered by identifying enriched transcription factor binding sites that coexist in the promoter regions of genes that show conserved gene expression patterns in human cancer and animal models, even if their expression levels in tumors are not significantly altered (Figure 3). It is important to realize that we can obtain better biologically and clinically significant inference of genomic data by integrating gene expression data with


3807

Figure 3 Integrative functional genomics. In addition to comparative functional genomics approach that integrates gene expression data from primary human hepatocellular carcinoma (HCC) with those from animal HCC models, gene expression signatures unique for different physiological conditions such as liver development and liver regeneration as well as hepatic stem cells will be collected and integrated into the gene expression patterns from human HCC. This approach, integrative functional genomics, will further help to not only stratify patients into more clinically homogeneous groups, but also to uncover the origins of tumor cells and distinct pathways involved in the molecular pathogenesis of HCC. It is expected that unique gene expression patterns from well-defined experimental conditions are regulated by limited number of transcription factors. Therefore, by applying comparative genomics to identify evolutionarily conserved sequences in the promoter regions of genes with conserved gene expression patterns, key transcription factors that govern the clinicopathological features of tumors can be identified.

diverse genomic data such as genomic sequence information in promoters, array-based CGH data and noncoding gene (i.e., microRNA) expression data. Recent advances in translating cancer genomics into clinical oncology strongly indicate that, rather than relying on the current modeling of population risk assessment and empirical treatment of patients, it is

essential to move to predictive personalized models based on molecular classification and targeted therapy. There can be little doubt that the future will bring a shift away from population risk assessment and empirical treatment of patients with HCC to the predictive personalized care based on molecular classification and targeted therapy. Serological molecular markers that are Oncogene


3808

identified by gene expression profile studies and noninvasive imaging probes will be used for screening or early detection. Defined specific gene expression signatures of tumors will be used to identify altered genetic elements or pathways that point to the most beneficial

therapy or combination of therapies. In parallel, new targeted therapies will emerge as gene expression profiling and other genome-wide screening technologies uncover pathways or interactions of pathways that are most vulnerable for therapeutic intervention.

References Alizadeh AA, Eisen MB, Davis RE, Ma C, Lossos IS, Rosenwald A et al. (2000). Nature 403: 503–511. Beer DG, Kardia SL, Huang CC, Giordano TJ, Levin AM, Misek DE et al. (2002). Nat Med 8: 816–824. Bittner M, Meltzer P, Chen Y, Jiang Y, Seftor E, Hendrix M et al. (2000). Nature 406: 536–540. Bruix J, Boix L, Sala M, Llovet JM. (2004). Cancer Cell 5: 215–219. Bullinger L, Dohner K, Bair E, Frohling S, Schlenk RF, Tibshirani R et al. (2004). N Engl J Med 350: 1605–1616. Calvet X, Bruix J, Gines P, Bru C, Sole M, Vilana R et al. (1990). Hepatology 12: 753–760. Calvisi DF, Factor VM, Ladu S, Conner EA, Thorgeirsson SS. (2004). Gastroenterology 126: 1374–1386. Calvisi DF, Factor VM, Loi R, Thorgeirsson SS. (2001). Cancer Res 61: 2085–2091. Chang HY, Nuyten DS, Sneddon JB, Hastie T, Tibshirani R, Sorlie T et al. (2005). Proc Natl Acad Sci USA 102: 3738–3743. Chang HY, Sneddon JB, Alizadeh AA, Sood R, West RB, Montgomery K et al. (2004). PLoS Biol 2: E7. Chen X, Cheung ST, So S, Fan ST, Barry C, Higgins J et al. (2002). Mol Biol Cell 13: 1929–1939. Cheung C, Akiyama TE, Ward JM, Nicol CJ, Feigenbaum L, Vinson C et al. (2004). Cancer Res 64: 3849–3854. Chevret S, Trinchet JC, Mathieu D, Rached AA, Beaugrand M, Chastang C. (1999). J Hepatol 31: 133–141. CLIP investigators (1998). Hepatology 28: 751–755. Cooper GM, Sidow A. (2003). Curr Opin Genet Dev 13: 604–610. Dave SS, Wright G, Tan B, Rosenwald A, Gascoyne RD, Chan WC et al. (2004). N Engl J Med 351: 2159–2169. Davila JA, Morgan RO, Shaib Y, McGlynn KA, El Serag HB. (2004). Gastroenterology 127: 1372–1380. Eddy SR. (2002). Cell 109: 137–140. El Serag HB, Mason AC. (1999). N Engl J Med 340: 745–750. El Serag HB. (2004). Gastroenterology 127: S27–S34. Ellwood-Yen K, Graeber TG, Wongvipat J, Iruela-Arispe ML, Zhang J, Matusik R et al. (2003). Cancer Cell 4: 223–238. Fan CY, Pan J, Usuda N, Yeldandi AV, Rao MS, Reddy JK. (1998). J Biol Chem 273: 15639–15645. Gonzalez FJ, Peters JM, Cattley RC. (1998). J Natl Cancer Inst 90: 1702–1709. Hanahan D, Weinberg RA. (2000). Cell 100: 57–70. Hardison RC, Oeltjen J, Miller W. (1997). Genome Res 7: 959–966. Hsu HC, Jeng YM, Mao TL, Chu JS, Lai PL, Peng SY. (2000). Am J Pathol 157: 763–770. Iizuka N, Oka M, Yamada-Okabe H, Nishida M, Maeda Y, Mori N et al. (2003). Lancet 361: 923–929. Kim BY, Lee JG, Park S, Ahn JY, Ju YJ, Chung JH et al. (2004). Biochim Biophys Acta 1739: 50–61. Kimura M. (1968). Nature 217: 624–626. King JL, Jukes TH. (1969). Science 164: 788–798. Lapointe J, Li C, Higgins JP, Van De RM, Bair E, Montgomery K et al. (2004). Proc Natl Acad Sci USA 101: 811–816. Oncogene

Laurent-Puig P, Legoix P, Bluteau O, Belghiti J, Franco D, Binot F et al. (2001). Gastroenterology 120: 1763–1773. le Coutre P, Mologni L, Cleris L, Marchesi E, Buchdunger E, Giardini R et al. (1999). J Natl Cancer Inst 91: 163–168. Lee JS, Thorgeirsson SS. (2002). Hepatology 35: 1134–1143. Lee JS, Thorgeirsson SS. (2004). Gastroenterology 127: S51–S55. Lee JS, Chu IS, Heo J, Calvisi DF, Sun Z, Roskams T et al. (2004a). Hepatology 40: 667–676. Lee JS, Chu IS, Mikaelyan A, Calvisi DF, Heo J, Reddy JK et al. (2004b). Nat Genet 36: 1306–1311. Lee JS, Grisham JW, Thorgeirsson SS. (2005). Carcinogenesis 26: 1013–1020. Lenhard B, Sandelin A, Mendoza L, Engstrom P, Jareborg N, Wasserman WW. (2003). J Biol 2: 13. Llovet JM, Burroughs A, Bruix J. (2003). Lancet 362: 1907–1917. Llovet JM, Bustamante J, Castells A, Vilana R, Ayuso MC, Sala M et al. (1999). Hepatology 29: 62–67. Lossos IS, Czerwinski DK, Alizadeh AA, Wechser MA, Tibshirani R, Botstein D et al. (2004). N Engl J Med 350: 1828–1837. Lynch TJ, Bell DW, Sordella R, Gurubhagavatula S, Okimoto RA, Brannigan BW et al. (2004). N Engl J Med 350: 2129–2139. Mao TL, Chu JS, Jeng YM, Lai PL, Hsu HC. (2001). J Pathol 193: 95–101. McNeil C. (1998). J Natl Cancer Inst 90: 882–883. Murphy WJ, Pevzner PA, O’Brien SJ. (2004). Trends Genet 20: 631–639. Nam SW, Park JY, Ramasamy A, Shevade S, Islam A, Long PM et al. (2005). Hepatology 42: 809–818. Neo SY, Leow CK, Vega VB, Long PM, Islam AF, Lai PB et al. (2004). Hepatology 39: 944–953. Okabe H, Satoh S, Kato T, Kitahara O, Yanagawa R, Yamaoka Y et al. (2001). Cancer Res 61: 2129–2137. Okuda K, Ohtsuki T, Obata H, Tomimatsu M, Okazaki N, Hasegawa H et al. (1985). Cancer 56: 918–928. Paez JG, Janne PA, Lee JC, Tracy S, Greulich H, Gabriel S et al. (2004). Science 304: 1497–1500. Parkin DM, Bray F, Ferlay J, Pisani P. (2005). CA Cancer J Clin 55: 74–108. Pugh RN, Murray-Lyon IM, Dawson JL, Pietroni MC, Williams R. (1973). Br J Surg 60: 646–649. Rao MS, Lalwani ND, Watanabe TK, Reddy JK. (1984). Cancer Res 44: 1072–1076. Reddy JK, Lalwai ND. (1983). Crit Rev Toxicol 12: 1–58. Roepman P, Wessels LF, Kettelarij N, Kemmeren P, Miles AJ, Lijnzaad P et al. (2005). Nat Genet 37: 182–186. Ross DT, Scherf U, Eisen MB, Perou CM, Rees C, Spellman P et al. (2000). Nat Genet 24: 227–235. Sargent LM, Zhou X, Keck CL, Sanderson ND, Zimonjic DB, Popescu NC et al. (1999). Am J Pathol 154: 1047–1055. Sweet-Cordero A, Mukherjee S, Subramanian A, You H, Roix JJ, Ladd-Acosta C et al. (2005). Nat Genet 37: 48–55. Tan CK, Law NM, Ng HS, Machin D. (2003). J Clin Oncol 21: 2294–2298. Thorgeirsson SS, Grisham JW. (2002). Nat Genet 31: 339–346.


3809 Ureta-Vidal A, Ettwiller L, Birney E. (2003). Nat Rev Genet 4: 251–262. Valk PJ, Verhaak RG, Beijen MA, Erpelinck CA, DoornKhosrovani SBW, Boer JM et al. (2004). N Engl J Med 350: 1617–1628. van de Vijver MJ, He YD, van’t Veer LJ, Dai H, Hart AA, Voskuil DW et al. (2002). N Engl J Med 347: 1999–2009.

van’t Veer LJ, Dai H, van de Vijver MJ, He YD, Hart AA, Mao M et al. (2002). Nature 415: 530–536. Wistuba II, Gazdar AF, Minna JD. (2001). Semin Oncol 28: 3–13. Wong CM, Fan ST, Ng IO. (2001). Cancer 92: 136–145. Yarden Y, Sliwkowski MX. (2001). Nat Rev Mol Cell Biol 2: 127–137. Ye QH, Qin LX, Forgues M, He P, Kim JW, Peng AC et al. (2003). Nat Med 9: 416–423.

Oncogene