Exploring Genomic Profiles of Hepatocellular Carcinoma

Abstract

Gene expression profiling using microarray technologies provides a powerful approach to understand complex biological systems and the pathogenesis of diseases. In the field of liver cancer research, a number of genome-wide profiling studies have been published. These studies have provided gene sets, that is, signature, which could classify tumors and predict clinical outcomes such as survival, recurrence, and metastasis. More recently, the application of genomic profiling has been extended to identify molecular targets, pathways, and the cellular origins of the tumors. Systemic and integrative analyses of multiple data sets and emerging new technologies also accelerate the progress of the cancer genomic studies. Here, we review the genomic signatures identified from the genomic profiling studies of hepatocellular carcinoma (HCC), and categorize and characterize them into prediction, phenotype, function, and molecular target signatures according to their utilities and properties. Our classification of the signatures would be helpful to understand and design studies with extended application of genomic profiles.

INTRODUCTION

Hepatocellular carcinoma (HCC) is one of the most common cancers with dismal clinical outcome accounting for the third cause of cancer-related death in the world [1], and its incidence has continued to increase in the United States and Western Europe [2]. A variety of etiological and risk factors including hepatitis virus, dietary exposure of aflatoxin B1, alcohol, diabetes, or parasites have associated with the development of HCC. Moreover, multi-step processes including genetic and epigenetic alterations are thought to be accumulated during progression of HCC [3,4]. Such complexity of the pathophysiology leads to a daunting heterogeneity of the tumors precluding unveiling disease mechanism.

In the last decade, numerous genome-wide profiling studies have been performed to delineate the complexity of the cancers. Based on gene expression patterns, heterogeneous HCC has been classified into homogeneous tumors with correlation of clinical outcomes. Molecular-based subgrouping of patients allowed the development of new targeted therapies [5,6]. This success could be accomplished by substantial progress on data analysis. Using gene set (i.e., signature)-based rather than single gene-based approach, data normalization, clustering, classification, and validation workflows are now becoming standardized [7–9]. Those developments of analysis strategies and tools have been accelerated by publicly open statistical computing resources such as bioconductor (http://www.bioconductor.-org), gene set enrichment analysis (GSEA) [10], genetic network analysis, and other analogous approaches.

More recently, research interest has moved toward identifying molecular determinants of disease pathogenesis and novel therapeutic targets. For this purpose, systematic and integrative analyses using multiple and multi-level data resources are thought as a state-of-art strategy to identify pivotal aberrations of genetic events or signaling pathways [11]. Accordingly, the study design including analytical strategy and reporting might be different according to the study purposes. With this reason, the gene sets identified by genome-wide profiling studies may have different biological meanings and usage.

Here, we categorized the previously identified gene signatures from gene expression profiling studies of liver cancer into four signature classes of prediction, phenotype, function, and molecular targets (Table 1), and discuss their distinct properties in study design, analysis, and validation strategies.

Table 1.

Classification of Gene Signatures Identified by Gene Expression Profiling

Category	Description	Clinical correlations	External validation methoda
Prediction signature: predict clinical outcomes
Iizuka et al. [12]	Predict early recurrence by 12 gene signature	Recurrence	Same platform
Ye et al. [13]	Predict metastasis in HBV-related HCC	Metastasis	Same platform
Lee et al. [14]	Predict survival by 406 gene signature	Survival	Same platform
Budhu et al. [15]	Predict metastasis by stromal tissue signature	Metastasis, survival	Same platform
Wang et al. [16]	Predict recurrence by 57 gene signature	Recurrence	Same platform
Woo et al. [17]	Predict early recurrence in HBV-related HCC CD24 was identified as biomarker for early recurrence	Recurrence	Cross platform
Hoshida et al. [18]	Predict late recurrence using formalin-fixed, paraffin-embedded (FFPE) stromal tissues	Recurrence	Same platform
Budhu et al. [19]	Predict metastasis by microRNA signature	Metastasis, survival	Same platform
Ji et al. [20]	Identify mir-26 as predictor for survival and response to IFN therapy	Survival, response to IFN therapy	Same platform
Poon et al. [21]	Predict survival by DNA copy numbers	Survival	Not determined
Hernandez-Vargas et al. [22]	Predict survival by 58 DNA methylation	Survival	Not determined
Phenotype signature: define characteristics of cancer subtype
Nam et al. [31]	Identify classifiers for low-to high-grade dysplasia and HCCs	Not determined	Not determined
Wurmbach et al. [32]	Identify dysregulated pathways for 4 neoplastic stages of HCV-related HCC	Not determined	Not determined
Ura et al. [33]	Identify a microRNA signature which is differentially expressed between HBV- and HCV-related HCC	Not determined	siRNA-mediated silencing
Chiang et al. [34]	Classify phenotypes of proliferation, CTNNB1, IFN-related, polysomy 7, and unannotated	Recurrence	Not determined
Katoh et al. [35]	Identify subclasses, e.g., c-Myc-induced HCC, 6p/1q-amplified HCC, and 17q-amplified HCC	Survival	Rapamycin sensitivity in 17q gained tumor cell lines
Boyault et al. [36]	Identify 6 classes based on gene expression, mutation, methylation, and LOH	Survival	Not determined
Lee et al. [37]	Identify hepatoblast-like class	Survival, recurrence	Same platform
Yamashita et al. [38]	Identify progenitor-like class using EpCAM and AFP	Survival	Same platform, immunohistochemistry
Andersen et al. [39]	Identify progenitor-like class using CK19 expression	Survival, recurrence	Not determined
Cairo et al. [40]	Identify child hepatoblastoma-derived signature	Survival	qPCR
Woo et al. [41]	Identify cholangiocarcinoma-like class	Survival, recurrence	Cross-platform
Function signature: functionally related gene sets (pathways)
Kaposi-Novak et al. [44]	Met knock-out signature	Survival	Cross platform
Coulouarn et al. [45]	TGF-β knock-out signature	Survival	Cross platform
Kaposi-Novak et al. [46]	Myc signature expression during malignant conversion	Not determined	Cross platform
Mayhew et al. [47]	RB knock-out signature	Survival	Not determined
Lee et al. [48]	Signatures from mouse transgenics of E2F, MYC, MYC-E2F1, and MYC-TGFA	Survival group	Not determined
Molecular target signature: candidate gene sets for molecular therapies
Zender et al. [55]	cIAP1 and Yap are identified in DNA amplicons	Not determined	In vivo overexpression
Zender et al. [56]	13 genes including XPO4 are identified by shRNA screening of the overlapped deleted genomic regions in human and mouse tumors	Not determined	shRNA-mediated silencing
Woo et al. [57]	50 potential driver genes are identified by combined analysis of array CGH and gene expression profiles	Survival	siRNA-mediated silencing

^aFunctional validations by experiments and external validations using independent data sets were considered.

CLASSIFICATION OF SIGNATURES

Prediction Signature

The majority of the gene signatures in previous studies are the prediction signature which predicts clinical outcomes. It includes the signatures generated from gene expression [12–18], microRNA [19,20], DNA copy numbers [21], and epigenetic regulation [22]. The gene signatures are usually developed and trained by the clinical information (e.g., survival, recurrence, or metastasis) to be predicted.

Major concern for the prediction signatures is its clinical applicability. For clinical use, the prediction signatures should be validated in a concrete manner. Because the predictability of the signatures could be affected by many confounding factors including experimental conditions (e.g., sample preparation and hybridization conditions) or population characteristics (e.g., postoperative treatment, patient’s health, and etiological differences), rigorous validation with appropriate sample size is required. External validation with independent data sets is also indispensable. Simon and coworkers have suggested helpful guidelines for study design, analysis, and validation methods for the development of predictors [23,24]. In general, the prediction signatures do not necessarily have functional roles [25], because they merely reflect the general condition of liver functions or host responses without causal relations. Therefore, the prediction signature genes may not appropriate candidates for therapeutic targets.

There are several issues to be resolved for the clinical use of the prediction signatures. Cross-platform difference of microarray is thought to be one of major obstacle for the clinical use of microarrays. Although the Microarray Quality Control (MAQC) project launched by US Food and Drug Administration (FDA) has addressed the technical reproducibility of microarray data across different sites and platforms [26], many of microarray studies generated only a small number of overlapping genes [24,27]. This discrepant result might be derived from the use of different platforms, probe sets, data normalization, and processing as well as differences in the study populations [28]. To resolve this issue, we have shown cross-platform consistency and robustness of tumor recurrence predictors by comparing two independent data sets [17]. Such validation by cross-platform comparison would be a strong evidence for the robustness of the clinical utility of prediction signatures.

Sample preparation is also one of critical drawback for the clinical use of microarray, because handling of the RNA samples from frozen tissues is not easy in clinical setting. Recently, Hoshida et al. [18] have shown the clinical utility of RNAs prepared from formalin fixed, paraffin-embedded (FFPE) tissues for the prediction of prognostic outcomes using a cDNA-mediated annealing selection extension and ligation (DASL) assay. This study has brought the microarray technology one step closer to clinical use, because the FFPE samples can be easily assessed in clinic [29]. They showed prognostic significance of the prediction signature generated from nontumoral tissues, but failed to show prognostic significance of the signatures from tumoral tissues. This finding is discrepant to the previous studies [12–14,16,17] which have consistently demonstrated the prognostic values of the signatures from tumoral tissue, therefore, further elaboration might be required [30].

Phenotype Signature

While the prediction signatures are generally obtained from the parameters to be predicted, the phenotype signatures represent distinct class phenotype regardless of the prediction parameters. For example, clinical and pathological features such as tumor grade [31], clinical stage [32], etiology (e.g., HBV or HCV) [33], drug responses, or the presence of certain biomarkers can be used for sample classification labels. Also, genetic events such as gene mutations or DNA copy number variations can be used as phenotype classifiers [34–36].

Remarkably, recent identification of the expression traits of tumor cellular origins could represent distinct tumor phenotypes. The identification of hepatoblast-like signature has stratified a novel HCC subtype harboring hepatic progenitor cell-like trait [37]. This study has demonstrated a strong correlation of the cell-type traits with patients’ prognosis, suggesting that the cellular origin of the tumors is a critical determinant for cancer progression. EpCam [38] and CK19 [39] were also identified as classifier for progenitor-derived HCC. Similarly, the hepato-blastoma-derived signature has been identified from childhood hepatoblastomas predicting clinical outcome [40]. However, the progenitors might be heterogeneous according to their differentiation status. With this concern, we have recently identified a cholangiocarcinoma (CC)-like gene expression trait (i.e., CC signature) representing a biliary differentiation trait in HCC [41]. In this study, we have shown functional and prognostic implications of the CC signature in HCC progression, suggesting distinct property of differentiation trait compared with those of stem-like trait. Taken these findings together, we could suggest that the different cellular origins from distinct developmental stage of primitive hepatic progenitor cells to differentiated hepatic or biliary precursors can give rise to heterogeneous HCC phenotypes (Figure 1).

Figure 1.

Postulated diagram of the cellular origin of HCC. The postulated cellular origins of HCC based on the expression status of cholangiocarcinoma (CC), embryonic stem (ES) cell, and hepatoblast (HB) signatures were illustrated. Details for cellular origin of CC were omitted. ⁺ The expression of cellular origin signatures. CHC represents combined hepatocellular cholangiocarcinoma. Reprinted from Ref. [41].

As like the prediction signature, the phenotype signatures have correlated with clinical outcomes (Table 1) [37,40,41]. However, these phenotype signatures do not necessarily have strong prediction power for clinical outcomes, because their primary goal is to address the functional characteristics of a certain tumor subtype. For example, the HCC class identified by a certain genetic phenotypes such as 17q amplification was not strongly associated with patient’s survival, but functionally liked to drug sensitivity to rapamycin [35]. Thus, the major concern for the phenotype signature is class characterization rather than prediction of clinical outcomes. Proper classification with well-defined phenotype signatures and linking them with functional and clinical utilities such as chemical or drug responses would be a promising strategy to develop novel cancer treatment modality for a certain tumor subtype, opening new era of personalized medicine [42,43].

Function Signature

The gene signatures representing distinct cellular processes or molecular pathways are referred as function signature. For example, the target gene signatures for MET [44], TGFB [45], MYC [46], and RB [47] have been identified from knock-out or transgenic animal models. As an application of function signatures, comparative analysis using different mouse models has demonstrated that the MYC-TGFA transgenics had similar expression patterns with HCCs of the poorer survival group, suggesting the regulatory role of TGFA signaling in aggressive progression of HCC [48]. In addition to such experimental models, computationally predicted or manually curated databases can be categorized as function signatures such as Gene Ontology (http://www.geneontology.org), microRNA targets (e.g., MirBase, http://microrna.sanger.ac.uk, or TargetScan, http://www.targetscan.org), protein–protein interactions (e.g., IntAct, http://www.ebi. ac.uk/intact), or transcription factor-binding (e.g., TRASNFAC, http://www.biobase-international.com).

As like phenotype signature, function signatures can represent phenotypes, but they have more causal or functional relevance. Thus, they are useful in functional interpretation of tumor classification. However, since gene expression profile is only a “snap-shot” of gene-to-gene interaction, they inevitably include many of nonspecific “by-stander” genes which are not functionally associated. Therefore, functional roles of the genes in function signature need to be addressed more carefully in detail. Particularly, proliferation- or cell cycle-related genes can easily be co-founded as by-stander genes, because the expression of these genes might be altered as an end result of disease progression. With this concern, a previous study has tested the possible effect of those genes by simply subtracting them from embryonic stem (ES) cell signatures [49]. This would be one of solutions for the signature purification to remove unwanted effects of the by-stander genes, although further accurate solutions might be required.

The pathogenesis of HCC is thought to be processed by multi-step pathologic states of chronic cirrhotic lesions, dysplastic nodules, and progression to early and advanced HCC [3,4]. The function signatures as well as phenotype signatures could address key regulators and genetic events during step-by-step progression of tumors (Figure 2). In earlier study, Nam et al. [31] have shown the discrete expression patterns (i.e., phenotype signature) across the multi-step hepatocarcinogenesis from low- to high-grade dysplastic nodules and HCCs. In addition, by performing pathway analysis, Wurmbach et al. [32] have further revealed the dysregulation of the Notch and Toll-like receptor pathways in cirrhosis, followed by dysregulation of Jak/STAT in early carcinogenesis. The application of MYC signature has also revealed its regulatory role during malignant conversion from dysplasia to early HCC [46].

Figure 2.

Genetic aberrations during multi-step hepatocarcinogenesis identified by genome-wide high-throughput profiling studies.

Besides these findings, genetic events such as gene mutation (e.g., TP53, CTNNB1, and KRAS), DNA methylation [50], and other gene expressions (e.g., IGF, VEGFR, CD24) have been noticed to involve the multi-step hepatocarcinogenesis, and their targeted therapies are now under investigation [51]. Signaling pathways such as proliferation (e.g., TGFA, EGF, IGF2, and HGF), differentiation (e.g., Wnt and TGFB), inflammation (e.g., IL6, IFN, and TNFA), and angiogenesis (e.g., VEGF, FGF, PDGF, and angiopoietin) were frequently deregulated resulting in high complexity and heterogeneity of HCC [see details in review [52,53]]. Such multiple and multistep complexity of genetic interaction impedes complete understanding of the genetic pathogenesis of hepatocarcinogenesis. Further extended studies defining phenotype or function signatures with multi-layered platform data including microRNAs, DNA copy numbers, or epigenetic changes are required to further delineate the multi-step hepato-carcinogenesis.

Molecular Targets

In cancer genomics studies, many of researchers are now interested in identifying molecular determinants of disease pathogenesis and novel therapeutic targets. These could be identified from prediction, function, or phenotype signatures. However, since lengthy gene list of those signatures would preclude the further evaluation of target genes, it would be important to select the most promising candidates as minimum as possible. To prioritize the most probable candidate genes, systemic integration of the gene signatures from multiple layers of data sets is thought to be “state-of-art” strategy. Data integration with multiple-layered platforms may have advantage to overcome the limitation of each platform data by reducing biased observation in each data set. For example, recent integrative analysis of microRNA target signatures from different databases could successfully identify novel oncogenic function of mir-101 in prostate cancer [54]. In HCC, cross-species comparative analysis of array-based comparative genomic hybridization (CGH) has revealed novel cancer genes, cIAP1 and Yap [55]. The same group also performed an efficient high-throughput screening of siRNA libraries by analyzing array CGH data, which discovered 13 novel tumor suppressor genes including XPO4 [56]. In addition, we have shown a successful integration of array CGH and gene expression profiling data [57]. Guiding the gene selection strategy by assessing the prognostic impact of the selected genes, we identified 50 potential driver genes including NCSTN and SCRIB. Then, further interrogation of the driver genes by using Connectivity Map database [58] could identify molecular targets (i.e., EGFR, mTOR, and AMPK) of the driver genes suggesting their mechanistic implications. These studies of systematic and integrative analyses of genomic profiles to select candidate target genes can speed up the screening and validation process of cancer-gene discovery and therapeutic targets.

INTEGRATIVE ANALYSIS UNVEILS COMPLEXITY

As shown in previous studies, appropriate integration of independent genomic data had a great advantage to identify novel molecular targets. Huge data repositories of microarray data, Gene Expression Omnibus (GEO) [59] and ArrayExpress [60] has enabled the researchers to perform integrative data analysis. In addition, as one of pioneering resources for integrative genomic profiling, National Cancer Institute 60 (NCI60) cell line panel has been established [61]. NCI60 had multi-dimensional data sets for 60 cancer cell lines including profiles of mRNAs, proteins, microRNAs, DNA copy numbers, gene mutations, and drug responses. These multilayered data integration could unveil molecular complexity by linking phenotypes and gene expression signatures [61–63]. Similar approaches for the cell-line based screen of candidate drugs have shown an efficient improvement of the evaluation process for anti-cancer agents [see review in [64]]. In addition, recent developments of huge data repositories such as The Cancer Genome Atlas (TCGA) [65], Connectivity Map, and Oncomine [66] will also be useful resources for integrative analysis of cancer genome, providing more qualified molecular targets [67,68].

Interaction of signatures is one of the challenging issues in the future integrative analyses. Recently, Chinnaiyan’s group has succeeded integrative data analyses using the gene signatures from multidimensional data sets. They developed a newer technique, “molecular concept map” (MCM) to address higher-level patterns across multiple and disparate microarray experiments [69]. Differing from the previous studies linking one-to-one or one-to-many association between signatures and phenotypes, MCM has linked many-to-many association among function and phenotype signatures by integrating different data types of pathways, proteins, and networks in an unbiased approach. The same group also developed an algorithm, that is, Cancer Outlier Profile Analysis (COPA) [70]. By identifying outlier expression across cancer samples, it could successfully reveal novel and uncharacterized oncogenic alterations. In another study, network analysis by constructing a BRCA-centered network (BCN) could identify novel susceptibility genes in breast cancer [71]. Such attempts on higher-level systematic approaches by integrating multiple and multi-level data resources would be promising to delineate genomic determinants in cancer development and progression (summarized in Figure 3); however, only a few of such integrative approaches have been applied to liver cancers.

Figure 3.

Strategies and resources for integrative analysis of genomic profiling data. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

CONCLUSION

We have reviewed the current status of the HCC gene expression profiling studies particularly the signature outcomes. Our classification of gene expression signatures as prediction, phenotype, function, and molecular targets would be helpful to address the distinction and the current trends of the application of genomic data to cancer research.

The accumulation of genome-wide “omics” data has provided us new opportunities to understand disease mechanisms at systems level. Moreover, emerging new technologies such as massively parallel sequencing and their applications (e.g., RNA-seq and ChiP-Seq) enabled researchers to investigate new parameters, such as mutations, alternative splicing, epigenetic silencing, and DNA-protein interactions which have not been resolved by current microarray technologies [72–74]. However, as data become complex and the study purposes are diverse, more careful attention on appropriate study design, reporting, and analytical strategies become essential in genomics studies. We believe that the systems view of cancers by constructing integrative atlas of genomic profiles with appropriate study design and analysis strategies would provide great potential for improving cancer management as well as understanding the cancer pathobiology.

Abbreviations

HCC

hepatocellular carcinoma

CC

cholangio-carcinoma

CGH

comparative genomic hybridization