Identification of Discriminators of Hepatomaby Gene Expression Profiling

Using a Minimal Dataset ApproachSoek Ying Neo,1 Chon Kar Leow,1,2 Vinsensius B. Vega,1 Philip M. Long,1 Amirul F.M. Islam,1 Paul B.S. Lai,3

Edison T. Liu,1 and Ee Chee Ren1,4

The severity of hepatocellular carcinoma (HCC) and the lack of good diagnostic markers andtreatment strategies have rendered the disease a major challenge. Previous microarray analyses ofHCC were restricted to the selected tissue sample sets without validation on an independentseries of tissue samples. We describe an approach to the identification of a composite discrimi-nator cassette by intersecting different microarray datasets. We studied the global transcriptionalprofiles of matched HCC tumor and nontumor liver samples from 37 patients using cDNA(cDNA) microarrays. Application of nonparametric Wilcoxon statistical analyses (P < 1 � 10�6)and the criteria of 1.5-fold differential gene expression change resulted in the identification of218 genes, including BMI-1, ERBB3, and those involved in the ubiquitin-proteasome pathway.Elevated ERBB2 and epidermal growth factor receptor (EGFR) expression levels were detected inERBB3-expressing tumors, suggesting the presence of ERBB3 cognate partners. Comparison ofour dataset with an earlier study of approximately 150 tissue sets identified multiple overlappingdiscriminator markers, suggesting good concordance of data despite differences in patient pop-ulations and technology platforms. These overlapping discriminator markers could distinguishHCC tumor from nontumor liver samples with reasonable precision and the features wereunlikely to appear by chance, as measured by Monte Carlo simulations. More significantly,validation of the discriminator cassettes on an independent set of 58 liver biopsy specimensyielded greater than 93% prediction accuracy. In conclusion, these data indicate the robustnessof expression profiling in marker discovery using limited patient tissue specimens as well asidentify novel genes that are highly likely to be excellent markers for HCC diagnosis and treat-ment. Supplementary material for this article can be found on the HEPATOLOGY website(http://interscience.wiley.com/jpages/0270-9139/suppmat/index.html). (HEPATOLOGY 2004;39:944–953.)

Hepatocellular carcinoma (HCC) is the mostcommon primary malignant tumor of the liverthat accounts for more than 70% of liver can-

cers worldwide.1 A major risk factor associated with the

development of HCC is hepatitis B virus (HBV) infec-tion. Death is usually due to liver failure associated withcirrhosis and/or rapid outgrowth of multiple nodules. Ap-proximately 0.25 to 1 million new cases of HCC are di-agnosed each year, and the cancer is especially prevalent inSoutheast Asia, China, and sub-Saharan Africa. Althoughsurgical resection is considered to be the main curativetreatment, only 10% to 15% of cases are suitable for sur-gery at the time of presentation. This is because either thedisease is detected at an advanced stage at presentation orthe underlying poor liver functional reserve precludedsurgical intervention.

The lack of molecular markers that characterize tumorformation poses a major problem to effective diagnosisand prognosis of HCC. Current diagnosis of HCC relieson the presence of a liver mass on radiologic investigationsand the detection of an elevated level of serum alpha fe-toprotein (AFP).2 However, an elevated level of AFP is

Abbreviations: HCC, hepatocellular carcinoma; HBV, hepatitis B virus; AFP,alpha fetoprotein; cDNA, complementary DNA; kNN, k-nearest neighbor; RT-PCR, reverse-transcription polymerase chain reaction; IGFBP3, insulin-like growthfactor binding protein 3; EGFR, epidermal growth factor receptor; PBGD, porpho-bilinogen deaminase.

From the 1Genome Institute of Singapore, Singapore; 2Department of Surgery,National University Hospital, Singapore; 3Department of Surgery, Prince of WalesHospital, Hong Kong, People’s Republic of China; and 4Department of Microbiol-ogy, National University of Singapore, Singapore.

Received September 18, 2003; accepted December 20, 2003.Address reprint requests to: E. Chen, Ren, Genome Institute of Singapore, 60

Biopolis Street, The Genome, 02-01, Singapore 138672. E-mail: renec@gis.a-star.edu.sg; fax: 011-65-6478-9052.

Copyright © 2004 by the American Association for the Study of Liver Diseases.Published online in Wiley InterScience (www.interscience.wiley.com).DOI 10.1002/hep.20105

944

not exclusive to HCC and has been observed in benignhepatic disease, such as cirrhosis and other cancers such asgerm cell cancer.3 Treatment of HCC includes the use ofinterferon therapy and antiviral drugs, but the results areunpredictable and the effectiveness may be limited.2,4 Ge-nome-wide analysis by microarray5 offers a systematic ap-proach to uncover comprehensive information about thetranscription profile of HCC. Previous studies have usedmicroarrays to address the changes in gene expression ofHCC.6–11 One major study was published recently byChen et al.12 that investigated more than 100 liver tissuespecimens. However, these reports were restricted to thetissue samples selected for each study and there was anabsence of validation of their findings on an independentseries of tissue samples, limiting the potential significanceand utility of the data.

In the current study, we used complementary DNA(cDNA) microarrays to examine the global cellularchanges in matched pairs of HBV-associated HCC tumorand nontumor liver tissue specimens of 37 patients. Inaddition, gene expression patterns between primary HCCtumors and liver cancer cell lines were examined for pos-sible biologic variation. A comparison was performedwith other independent microarray studies of HCC in anattempt to identify a composite cassette of discriminatorgenes that could potentially serve as tumor markers. Tovalidate the utility of these discriminator cassettes to dis-tinguish tumor from nontumor, prediction accuracy wasassessed on an entirely independent set of 58 liver biopsysamples. These experiments indicated that array-based ex-pression profiling on limited tissue sets generates robustdata and identified novel molecular markers that arehighly likely to be excellent markers for HCC diagnosisand targets for new disease management strategies.

Patients and Methods

RNA Isolation, RNA Amplification, and cDNA Mi-croarray Hybridization. All 37 patients (from whomthe test set of tissue specimens was derived) had HBV-associated HCC and underwent curative liver resection.The paired samples of tumor and corresponding nontu-mor tissue specimens were obtained from the resectedliver specimen. The validation tissue set comprised 58liver biopsy samples from an independent cohort of 29patients who also had HCC associated with HBV andunderwent curative liver resection. Informed consent wasobtained from the patients. The institutional research andethics committee approved the study.

Tissue specimens were snap frozen in liquid nitrogenand stored at �150°C. A small section of each specimenwas sampled and total RNA was isolated using Trizol

reagent (Life Technologies, Bethesda, MD) according tothe manufacturer’s instructions. The integrity of the RNAspecimen was verified by gel electrophoresis. The humanliver cancer cell lines used in this study were PLC/PRF/5,HA22T, Huh1, Huh4, Tong, Hep3B, SNU182, SNU449,SNU475, HepG2, Huh6, Huh7, SKHep1, and Mahlavu.All cell lines were cultured under conditions recommendedby the American Type Culture Collection (VA).

Because the amount of total RNA sample obtained fromthe limited tissue material was insufficient (approximately4-15 �g) to be put directly on the array, RNA was linearlyamplified using a procedure modified from Eberwine et al.13

Human universal reference RNA (Stratagene, La Jolla, CA),which comprised total RNA samples from 10 different hu-man cell lines, was amplified and used as the reference forcDNA microarray analysis. Approximately 9,000 humancDNA features (Incyte Genomics, Palo Alto, CA) were spot-ted onto poly-L-lysine–coated slides using an OmniGridarrayer (GeneMachines). Probes were generated from theamplified RNA material and hybridized to the chip as de-scribed previously.14 To minimize the effects of labeling bi-ases, reciprocal dye swap labeling experiments wereperformed for each sample.

Data Analysis. Raw data were analyzed on GenePixanalysis software version 3.0 (Axon Instruments, Burlin-game, CA) and uploaded to a relational database. ThecDNA clones used for the microarray are represented bytheir UniGene identifiers. For each array, the logarithmicexpression ratios for the spots were normalized. In addition,spots that did not meet our filtering criteria (see supplemen-tary material for this article on the HEPATOLOGY website(http://interscience.wiley.com/jpages/0270-9139/suppmat/index.html)) were excluded, resulting in the inclusion of8,716 features for subsequent analysis. Statistical comparisonof genes between HCC tumor and nontumor specimens wasperformed by the Wilcoxon rank-sum nonparametric test.To evaluate gene expression patterns, hierarchical clusteringusing one minus Pearson’s correlation metric and averagelinkage,15 and multidimensional scaling was performedon normalized data (mean � 0, SD � 1). Functionalcharacterization of genes was based on gene ontology16

and on other published works in the PubMed database(http://www.ncbi.nlm.nih.gov/entrez).

Three different sets of Monte Carlo simulations17 wereperformed to (1) measure the quality of a set of selectedgene features to be used as potential markers (Pa); (2)determine whether the set of genes observed to have agood performance as tumor discriminators, could appearmerely by chance (Pb); and (3) approximate the signifi-cance of the number of observed overlapping genes afterintersection of the important gene lists derived from twoindependent groups (i.e., the current study and the Chen

HEPATOLOGY, Vol. 39, No. 4, 2004 NEO ET AL. 945

et al. study; Pc). To validate the utility of the variousexpression cassettes to distinguish HCC tumor from nontu-mor liver specimens, the prediction accuracy of each discrim-inator cassette was assessed on an independent tissue setcomprising 58 liver clinical biopsy specimens from 29 pa-tients using a k-nearest neighbor (kNN) classification algo-rithm (k � 3) that employs Pearson correlation to measurethe similarity between expression profiles. The algorithmwas trained against the dataset comprising 74 tissue samplesfrom 37 patients before testing against the new tissue set.

Real-Time Semiquantitative Reverse-Transcrip-tion Polymerase Chain Reaction (RT-PCR). TotalRNA samples were analyzed for the expression levels ofselected genes by real-time semiquantitative RT-PCR us-ing the LightCycler RNA amplification kit SYBR Green Ion the LightCycler (Roche, Basel, Switzerland) accordingto the manufacturer’s instructions. Data are presented asthe level of gene expression in each HCC tumor specimenrelative to its corresponding nontumor liver specimen.

Results

Assessment of Global Gene Expression DifferencesBetween HCC Tumor and Nontumor Liver Speci-mens. The gene expression patterns of primary HCC tu-mor specimens and the corresponding nontumor livertissue specimens from 37 patients were examined bycDNA microarray. First, we assessed the overall naturalpatterns of gene expression in the HCC tumor and non-tumor liver tissue specimens based on unsupervised hier-archical clustering. ANOVA in expression levels for eachgene across all the tissue specimens indicated that 500gene features (containing 493 unique UniGenes) showedthe largest variability across both HCC tumor and non-tumor liver tissue specimens (Fig. 1). Included in this listare AFP, an often used prognostic marker for HCC, andother genes associated with HCC such as HGF and MYC.Hierarchical clustering analysis based on these highly vari-ant genes showed two main clusters, one representing theHCC tumor specimens and the other, the nontumor livertissue specimens with only 6 of 37 HCC tumor specimensmisclassified as nontumors (Supplementary Fig. 1B).Thus, the molecular configuration of HCC can be readilydistinguished from that of nontumor liver specimens withminimal data manipulation.

Second, to investigate differential gene expression pat-terns between HCC tumor and nontumor liver specimens,we used the Wilcoxon rank-sum test and identified the top2.5% candidate genes that displayed the smallest (best) Pvalue scores (P � 1 � 10�6) and at least a 1.5-fold change ingene expression. For these 218 genes, false discovery rateanalysis indicates a false-positive error of less than 0.4%.

Multidimensional scaling analysis based on these outliers in-dicated that the HCC tumors were a more heterogeneouspopulation than the nontumor liver tissue specimens (Fig.2). Cancer cell lines derived from the primary tumor havetraditionally been used as in vitro model systems for investi-gating the function of genes in the in vivo tumor environ-ment. We asked how the expression pattern of the same 218genes would appear in 14 established human liver cancer celllines. It is apparent that the cell lines exhibited gene expres-sion profiles that were different from the clinical HCC tu-mor and nontumor liver tissue specimens (Fig. 2), suggestingthat these cell lines may have accumulated additional geneticor epigenetic alterations in culture.

Identification of Gene Clusters Differentially Ex-pressed in HCC Tumor Tissue Specimens. Among the218 significant genes that distinguished HCC tumorsfrom nontumor liver tissue specimens, more genes wereobserved to be overexpressed than underexpressed in themalignant tissue specimens (Supplementary Fig. 1). Map-ping of the chromosomal location of these 218 uniqueoutliers indicated that a disproportionate number ofgenes was located on chromosome 1 (Fig. 3A), particu-larly in the 1q region, and that the majority of these geneswere more highly expressed in the tumor tissue speci-mens. This result correlated well with previous reports offrequent amplification on chromosomal 1q in HCC tu-mor specimens.18 Further characterization of these outliergenes revealed many genes that have yet to be fully under-stood for their roles in HCC. A substantial proportion ofgenes were involved in transport (e.g., PEA15), RNA pro-cessing (e.g., RDBP), and metabolic processes (e.g.,NME1) and showed increased expression in HCC tumorspecimens, possibly indicating accelerated rates of metab-olism (Fig. 3B, Table 1, Supplementary Table 2). Severalgenes (e.g., SMT3H1) were members of the ubiquitin-

Fig. 1. Natural patterns of gene expression differences between HCCtumor and nontumor liver tissue specimens based on unsupervisedclustering. Plot showing the variance of expression value for each of thegene features across all the HCC tumor and nontumor liver tissuespecimens. Dotted line indicates the 500 most variable gene features.

946 NEO ET AL. HEPATOLOGY, April 2004

proteasome pathway, suggesting considerable deregula-tion of this pathway in HCC. Transcription factors (e.g.,ESR1) and genes involved in controlling growth and dif-ferentiation (e.g., GRN), and signal transduction (e.g.,CSTB) formed the other dominant gene groups. Notably,the polycomb group protein BMI-1, which is believed tobe an oncogene, was consistently expressed at muchhigher levels in HCC tumor specimens (Fig. 4). BMI-1expression level is elevated in various tumors includingHodgkin’s disease19 but it has not been previously studiedin HCC. In addition, there were a number of genes (e.g.,MAWBP, AD24) that had no known functions. Genesthat have been studied in HCC previously and were rep-resented in our outlier list included HDGF andGHR.20,21 A number of the differentially expressed genessuch as MDK and CDC23 are consistent with those re-ported for cancer cells.22,23 It is noteworthy that the list ofoutlier genes did not include AFP, which was elevated ina small number of our HCC tumor samples. This agreeswell with previous reports that AFP levels were variablyelevated in HCC12 and that AFP was present in approxi-mately 50% of HCC.24

To validate our microarray data, real-time RT-PCRanalysis was performed for insulin-like growth factorbinding protein 3 (IGFBP3) and ERBB3 in all the 37matched HCC tumor and nontumor liver samples.The results of real-time RT-PCR analyses of IGFBP3and ERBB3 were consistent with previous reports ex-amining these individual markers.25,26 For example,IGFBP3 expression was diminished in 35 of 37 HCCtumors relative to their corresponding nontumor livertissue specimens (Fig. 5A), whereas ERBB3 expression

Fig. 3. Characterization of differentially expressed genes in HCC tumorspecimens (P � 1 � 10�6, approximately1.5-fold change). (A) Chro-mosomal distribution of the 218 outlier genes. The dark-colored and lightshaded bars represent genes that are at least 1.5-fold up-regulated anddown-regulated, respectively, in HCC tumor specimens relative to non-tumor liver samples. (B) Functional characterization of the outlier genesbased on gene ontology and published works.

Fig. 2. Significant gene differential expression be-tween HCC tumor and nontumor liver tissue specimensand comparison with liver cancer cell lines (P � 1 �10�6, approximately 1.5-fold change). Multidimen-sional scaling plot illustrates the ability of these 218outlier genes to separate HCC tumor specimens (redcircles) from nontumor liver tissue specimens (greencircles). The multidimensional plot also shows howdifferent liver cancer cell lines (yellow circles) are fromthe clinical tissue samples.

HEPATOLOGY, Vol. 39, No. 4, 2004 NEO ET AL. 947

Table 1. Selected Genes Significantly Differentially Expressed in HCC Tumors

Function Gene Symbol Gene Name UniGene

ExpressionChange in HCC

Tumor*

Transcription factor ILF2 Interleukin enhancer binding factor 2, 45 kD Hs. 75117 1BMI1 Murine leukemia viral (bmi-1) oncogene homolog Hs. 431 1TAF9 TAF9 RNA polymerase II, TATA box binding protein (TBP)-associated factor, 32 kD Hs. 60679 1ZNF146 Zinc finger protein 146 Hs. 301819 1CHD4 Chromodomain helicase DNA binding protein 4 Hs. 74441 1NR4A1 Nuclear receptor subfamily 4, group A, member 1 Hs. 1119 2ZNF238 Zinc finger protein 238 Hs. 69997 2FOSB FBJ murine osteosarcoma viral oncogene homolog B Hs. 75678 2

RNA processing H2AFY H2A histone family, member Y Hs. 75258 1SNRPB Small nuclear ribonucleoprotein polypeptides B and B1 Hs. 83753 1RPS7 Ribosomal protein S7 Hs. 301547 1MRPS14 Mitochondrial ribosomal protein S14 Hs. 247324 1SNRPD2 Small nuclear ribonucleoprotein D2 polypeptide (16.5 kD) Hs. 53125 1NCL Nucleolin Hs. 79110 1RPS10 Ribosomal protein S10 Hs. 76230 1RPL6 Ribosomal protein L6 Hs. 349961 1SNRPE Small nuclear ribonucleoprotein polypeptide E Hs. 334612 1SF3B4 Splicing factor 3b, subunit 4, 49 kD Hs. 25797 1RDBP RD RNA-binding protein Hs. 106061 1SNRPF Small nuclear ribonucleoprotein polypeptide F Hs. 105465 1RPS10 Ribosomal protein S10 Hs. 76230 1

DNA replication/repair ADPRT ADP-ribosyltransferase (NAD�; poly(ADP-ribose) polymerase) Hs. 177766 1PRKDC Protein kinase, DNA-activated, catalytic polypeptide Hs. 155637 1SMC4L1 SMC4 (structural maintenance of chromosomes 4, yeast)-like 1 Hs. 50758 1FEN1 Flap structure-specific endonuclease 1 Hs. 4756 1MCM2 Minichromosome maintenance deficient (S. cerevisiae) 2 (mitotin) Hs. 57101 1HAT1 Histone acetyltransferase 1 Hs. 13340 1

Cell growth/differentiation GPC3 Glypican 3 Hs. 119651 1MDK Midkine (neurite growth-promoting factor 2) Hs. 82045 1HDGF Hepatoma-derived growth factor (high-mobility group protein 1-like) Hs. 89525 1TP53BP2 Tumor protein p53-binding protein, 2 Hs. 44585 1CDC23 CDC23 (cell division cycle 23, yeast, homolog) Hs. 153546 1IGFBP3 Insulin-like growth factor binding protein 3 Hs. 77326 2

Immune response TMPO Thymopoietin Hs. 11355 1IGKC Immunoglobulin kappa constant Hs. 156110 2IGHG3 Immunoglobulin heavy constant gamma 3 (G3m marker) Hs. 300697 2IGJ Immunoglobulin J polypeptide, linker protein for immunoglobulin alpha and mu

polypeptidesHs. 76325 2

Cell adhesion LAMR1 Laminin receptor 1 (67 kD, ribosomal protein SA) Hs. 181357 1CAPZA2 Capping protein (actin filament) muscle Z-line, alpha 2 Hs. 75546 1ARHE Ras homolog gene family, member E Hs. 6838 2

Signal transduction CAP2 Adenylyl cyclase-associated protein 2 Hs. 296341 1CALM2 Calmodulin 2 (phosphorylase kinase, delta) Hs. 182278 1LASP1 LIM and SH3 protein 1 Hs. 334851 1SHC1 SHC (Src homology 2 domain-containing) transforming protein 1 Hs. 81972 1RGS5 Regulator of G-protein signalling 5 Hs. 24950 1HAX1 HS1 binding protein Hs. 15318 1ERBB3 v-erb-b2 avian erythroblastic leukemia viral oncogene homolog 3 Hs. 199067 1

Ubiquitin-proteasome pathway UBD Diubiquitn Hs. 44532 1USP14 Ubiquitin specific protease 14 (tRNA-guanine transglycosylase) Hs. 75981 1PSMA1 Proteasome (prosome, macropain) subunit, alpha type, 1 Hs. 82159 1PSMB4 Proteasome (prosome, macropain) subunit, beta type, 4 Hs. 89545 1

Transport CCT5 Chaperonin containing TCP1, subunit 5 (epsilon) Hs. 1600 1CCT3 Chaperonin containing TCP1, subunit 3 (gamma) Hs. 1708 1HSPA5 Heat shock 70 kD protein 5 (glucose-regulated protein, 78 kD) Hs. 75410 1XPO1 Exportin 1 (CRM1, yeast, homolog) Hs. 79090 1NUCB2 Nucleobindin 2 Hs. 3164 1ATP6IP1 ATPase, H� transporting, lysosomal interacting protein 1 Hs. 6551 1AP3S1 Adaptor-related protein complex 3, sigma 1 subunit Hs. 80917 1VDAC2 Voltage-dependent anion channel 2 Hs. 78902 1

Metabolism NME1 Non-metastatic cells 1, protein (NM23A) expressed in Hs. 118638 1DPM1 Dolichyl-phosphate mannosyltransferase polypeptide 1, catalytic subunit Hs. 5085 1ACLY ATP citrate lyase Hs. 174140 1GCN1L1 GCN1 (general control of amino-acid synthesis 1, yeast)-like 1 Hs. 75354 1NAT2 N-acetyltransferase 2 (arylamine N-acetyltransferase) Hs. 2 2CYP2C8 Cytochrome P450, subfamily IIC (mephenytoin 4-hydroxylase), polypeptide 8 Hs. 174220 2CYP2E Cytochrome P450, subfamily IIE (ethanol-inducible) Hs. 75183 2

Unknown C20orf24 Chromosome 20 open reading frame 24 Hs. 184062 1C1orf9 Chromosome 1 open reading frame 9 Hs. 108636 1LOC51235 Hypothetical protein Hs. 181444 1FLJ12666 Hypothetical protein FLJ12666 Hs. 23767 2

*Gene expression level showing at least 1.5-fold change in HCC tumor relative to non-tumor liver tissues (P � 1 � 106). (The full gene list with GenBank accession numbers is available in the supplementary data.)

948 NEO ET AL. HEPATOLOGY, April 2004

was elevated in 34 of 37 tumor samples (Fig. 5B). Weobserved good concordance in fold changes betweenmicroarray data and real-time RT-PCR results andthe magnitude of fold change obtained by PCR wassimilar or higher (as much as a fivefold in some cases,given the sensitivity of PCR). Because ERBB3 is defec-tive in tyrosine kinase activity and requires dimeriza-tion with other receptors, possibly another member ofthe ERBB family,27 we tested the hypothesis thatHCC tumors expressing high levels of ERBB3 wereassociated with high expression of ERRB2 or epidermalgrowth factor receptor (EGFR). The expression ofERBB2 was elevated in 12 of 37 tumor specimens,whereas high EGFR expression was found in 15 of 37tumor specimens (Fig. 5B). With the top 50 percentileof high ERBB3-expressing HCC tumors, we found asignificant concomitant increase in ERBB2 expression(t-test P � approximately .0026) but no associationwith high EGFR expression (t-test P � approximately.31). Clearly, the cognate partners of ERRB3 werepresent in tumor specimens expressing high levels ofERBB3.

Validation of HCC Tumor Discriminator Expres-sion Cassettes. To assess the validity of our expressioncassette of genes for distinguishing HCC tumor from

nontumor liver tissue specimens, we explored the inter-section of our data with those published in the literature.In the study by Chen et al.,12 HCC tumor specimensfrom 82 patients and nontumor liver tissue specimens

Fig. 4. Expression of BMI-1 in HCC tumor specimens as determinedby cDNA microarray analysis. Data are presented as the level of expres-sion (log base 2) in each HCC tumor specimen with respect to thecorresponding nontumor liver sample.

™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™™3Fig. 5. Real-time RT-PCR analysis of IGFBP3, ERBB3, ERBB2, and

EGFR in HCC tumor samples. The gene expression patterns for (A)IGFBP3 and (B) ERBB3, ERBB2, and EGFR in all the 37 HCC tumorspecimens and their corresonding nontumor liver tissue specimens wereexamined. All data were normalized to the amount of housekeeping geneporphobilinogen deaminase and are presented as relative fold expressionchange (log base 2) in HCC tumor specimens with respect to itscorresponding nontumor liver counterpart. A positive value depicts ahigher expression level, whereas a negative value depicts a lower ex-pression level in the tumor relative to the nontumor specimen.

HEPATOLOGY, Vol. 39, No. 4, 2004 NEO ET AL. 949

from 74 patients were examined and 1,648 features (con-taining 1,449 unique UniGenes) were reported to dis-criminate tumor from nontumor, of which 600 features(containing 540 unique UniGenes) formed the best dis-criminators. First, we asked whether any of the Chen et al.dataset of 600 most differentially expressed genes was in-cluded in our array of 8,716 features (containing 7,521unique UniGenes) based on UniGene identifiers. A totalof 265 features (containing 245 unique UniGenes) fromour microarray were observed to overlap (SupplementaryTable 3). Hierarchical clustering analysis based on theexpression levels of these 265 “overlap” features separatedour tissue set into two distinct groups of tumor and non-tumor specimens, with five tissue samples misclassified(Fig. 6). Such clustering was significant (Pa � 1 � 10�6)based on random permutation testing of sample labels.The likelihood of a randomly chosen set of 265 featuresproducing five or fewer samples misclassified was low(Pb � 1.5 � 10�3). Therefore, these 265 overlap featurescould distinguish HCC tumor from nontumor liver spec-imens with reasonable precision, and the features wereunlikely to appear by chance.

Second, we asked whether our 218 significant gene list(containing 213 unique UniGenes) was present in theChen et al. microarray of approximately 23,000 features(containing 17,220 unique UniGenes). A total of 230features (containing 166 unique UniGenes) from the

Chen et al. array were observed to overlap (Supplemen-tary Table 4). Hierarchical clustering analysis based onthe expression levels of these 230 overlap features sepa-rated the Chen et al. tissue set into distinct tumor andnontumor groups, with four tissue samples misclassified(Fig. 7). Random permutation of sample labels indicatedthat the clustering was significant (Pa � 1 � 10�6) and itwas unlikely that a randomly chosen set of 230 featurescould produce four or fewer samples misclassified (Pb �1 � 10�4). These 230 overlap features are therefore ableto discern fairly accurately HCC tumor from nontumorliver specimens. We also sort the overlap between our 218significant gene list and the Chen et al. list of 1,648 gooddiscriminating genes. A total of 68 unique UniGenesoverlapped (Supplementary Table 5). The likelihood thatthe overlap would arise by chance if the two gene lists weretotally independent was minuscule (Pc � 1 � 10�8).Therefore, cross-testing the results of each dataset withthe other, we were able to validate tumor-nontumor dis-criminators for HCC diagnosis.

Third, to definitively validate the utility of these probesets to distinguish HCC tumor from nontumor liver tis-sue specimens, we assessed the accuracy of these four discrim-inator cassettes on an independent tissue set consisting of 58liver clinical biopsy specimens from 29 patients. These 58liver samples were processed separately on the same cDNAmicroarray platform as with our test set of tissue samplesfrom 37 patients. Using a kNN prediction algorithm, wefound that all classifier probe cassettes could readily distin-

Fig. 6. Intersection between the best 600 genes from the Chen et al.list of good discriminators and the GIS microarray expression data set(8,716 features) based on UniGene identifiers. The 265 overlap features(containing 245 unique UniGenes) obtained were applied on the GIStissue set. Clustering analysis was performed based on the expressionlevels of these features. The significance of the clustering obtained (Pa)was measured by Monte Carlo sample label permutations. The proba-bility that the observed 265 overlap features could appear merely bychance alone (Pb) was estimated by performing a different series ofMonte Carlo simulations (see supplementary information for details).

Fig. 7. Intersection between the GIS 218 significant gene list and theChen et al. microarray expression data set (approximately 23,000 fea-tures) based on UniGene identifiers. The 230 overlap features (contain-ing 166 unique UniGenes) obtained were applied on the Chen et al.tissue set. Clustering analysis was performed based on the expressionlevels of these features. The significance of the clustering obtained (Pa)was measured by Monte Carlo sample label permutations. The proba-bility that the observed 230 overlap features could appear merely bychance alone (Pb) was estimated by performing a different series ofMonte Carlo simulations (see supplementary information for details).

950 NEO ET AL. HEPATOLOGY, April 2004

guish HCC tumor from nontumor liver specimens (Table2). The GIS discriminator list (N � 218) resulted in a pre-dictive accuracy of 93%, four false-negative and no false-positive results. The GIS genes found in the Chen et al.microarray (N � 166) gave three false-negative and no false-positive results for a predictive accuracy of 95%. The dis-criminators from the Chen et al. study found in the GISmicroarray (N � 265) gave three false-negative and no false-positive results in the validation set for a predictive accuracyof 95%. Lastly, the intersect between the GIS and the Chenet al. discriminator probes comprising only 68 genes gave apredictive accuracy of 96%, one false-negative and one false-positive result. Therefore, the gene discriminators of tumorversus nontumor in HCC derived by the intersect analysis oflimited tissue sets can be validated in an independent man-ner.

DiscussionDue to the severity of the disease and the lack of good

molecular markers for diagnosis and effective treatmentstrategies of the tumor, HCC remains a major cancerchallenge. We present an analysis of gene expression pat-terns on matched HCC tumor and nontumor tissue spec-imens from 37 patients and uncovered 218 genes that candistinguish tumor from nontumor liver specimens. Previ-ous work on HCC to identify unique expression signa-tures uncovered discriminating gene sets. We sought tocapitalize on the progressive standardization of the arrayplatforms by testing the performance of overlapping dis-criminator markers between our array results and those ofChen et al.12 Our results show remarkable concordancewith those of Chen et al., despite the differences in patientpopulations and the technology platforms applied. Wethen validated these intersecting discriminators with anindependent set of 58 HCC tumor and nontumor liverspecimens and showed high predictive accuracy for thesegene probes. Therefore, markers arising from array-basedanalysis of limited tissue sets are surprisingly robust.Moreover, we believe that future approaches combiningin silico and limited experimental validation are sufficientto uncover biologically meaningful data from clinical ex-pression array studies.

In-depth study of the functions of the useful markersshows that a substantial number of the genes are involvedin the ubiquitin-proteasome pathway and are overex-pressed in HCC tumor tissue specimens. Two of thesegenes, UBD and PSMD4, which were highly expressed inthe tumor tissue specimens in the current study, are sim-ilar to the observations made by other independent arraystudies.9,12 In addition to stability, the activities of manyproteins such as the tumor suppressor p53 are modulatedby ubiquitination.28 The molecular components of theproteasome complex are also involved in other cellularevents that do not require proteolysis, such as endocytosis,the localization of certain proteins in the nucleus, andtranscription.29,30 Defects in the ubquitin-proteasomepathway have been observed in a variety of human dis-eases, including neurodegenerative disease, metabolic dis-orders, and cancer.31 Recently, it was demonstrated thatapplication of the proteasome inhibitor PS-341 in ortho-topic human pancreatic tumor xenografts resulted in in-hibition of tumor growth and angiogenesis.32 These datasuggest that the ubiquitination pathways may be a legiti-mate target for HCC therapeutics.

The putative oncogene ERBB3 was consistentlypresent at much higher levels in our set of HCC tumortissue specimens. This is consistent with findings of anindependent study that found high ERRB3 levels inthe majority of HCC tumor specimens by immunohis-tochemistry.26 High ERBB2 expression was found inapproximately 41% of the Chen et al. HCC tumorset,12 similar to our findings that approximately 32% ofour HCC tumor set overexpressed ERBB2. In addi-tion, there was an association between high ERBB3and high ERBB2 expression in the HCC tumor speci-mens in the current study. The physiologic effects ofERBB3 signaling are not completely clear. ERBB3 hasbeen shown to serve as the primary binding site forheregulin in adult rat hepatocytes and whole liver speci-mens.33 In addition, ERBB3 is capable of eliciting bothdifferentiation and growth responses when transactivatedby the other ERBB receptors ERBB2 and EGFR, neitherof which binds heregulin.27 These data, together, suggestthat drugs targeting the kinase domains of ERBB2 and

Table 2. Prediction Accuracy of Gene Classifiers Using k NN Algorithm on 58 Liver Biopsies From 29 Patients

Gene Classifiers

No. ofGene

ClassifiersMisclassification

Rate

No. of FalseNegativeCases*

No. of FalsePositiveCases†

PredictiveAccuracy

GIS discriminating genes 218 4 of 58 4 – 93%GIS genes present in the Chen et al. microarray 166 3 of 58 3 – 95%Chen et al. genes present in the GIS microarray 265 3 of 58 3 – 95%Intersect between GIS and Chen et al. discriminating genes 68 2 of 58 1 1 96%

*False negative cases refer to HCC tumors which were misclassified as non-tumor livers.†False positive cases refer to non-tumor livers which were misclassified as HCC tumors.

HEPATOLOGY, Vol. 39, No. 4, 2004 NEO ET AL. 951

EGFR or the extracellular domain of ERBB3 may be use-ful in HCC treatment strategy.

Our study highlights a sizable number of potentialmarker genes that may be classified as cell adhesion mol-ecules. Alterations in the expression and function of ad-hesion molecules, including integrins and cadherins, areassociated with both tumor suppression and progressionin different diseases.34 Some of these genes, such asITGB1, have been shown to have an essential role inHCC.35 Although little information is available currentlyfor other genes, such as DNCH1, work on other genemembers (e.g., DNLC2A and DNLC2B) indicates thattheir levels are substantially up-regulated and down-reg-ulated, respectively, in HCC tumor specimens comparedwith their adjacent nontumor liver tissue specimens.36

The current development of antagonists, including cilen-gitide,37 against integrins signifies the importance of celladhesion molecules in cancer.

In addition, we showed the polycomb group proteinBMI-1 to be a significant outlier that was highly differen-tially expressed in HCC tumors. Originally identified inDrosophila, polycomb group proteins are evolutionarilyconserved chromatin components that are involved inmaintaining transcriptional “memory” and are vital inembryogenesis and control of cell identity.38 Conceiv-ably, deregulation of this epigenetic silencing would havedeleterious effects on normal development and may resultin the development of cancer. This was demonstrated byVarambally et al.39 in their study of prostate cancer. Theyshowed that the polycomb group protein EZH2 is highlyexpressed in prostate cancer and is involved in the pro-gression of prostate cancer. Recent studies indicate thatBMI-1 has an essential role in determining the prolifera-tive activities of normal and leukemic stem cells.40 Thesefindings suggest that the oncogenic potential of BMI-1may render it an important tumor marker as well as ther-apeutic target for HCC.

IGFBP3 showed consistent low expression levels across ourand other datasets of HCC tumors,6,12 suggesting a role for theIGFaxis inHCCdiseasepresentation.Perturbationsof the IGFaxis are also implicated in the formationof cancer inorgans suchas the colon, prostate, and lung.41 How the actions of IGFBP3are modulated is still unclear, but it is likely to be complex andinvolve multiple players. Besides signaling through the IGF re-ceptor-dependent pathway, IGFBP3 also mediates signals viaother pathways such as transforming growth factor-beta, tumornecrosis factor-alpha, and RXR�.42 GPC3, a member of theheparin sulfate proteoglycans, showed consistent high abun-danceacrossHCCtumor specimens in thecurrent studyand inthe Chen et al.12 datasets. High expression of GPC3 has alsobeen reported in embryonal tumors such as Wilms tumor43

although its expression is reported to be suppressed in breast

cancer.44 Heparin sulfate proteoglycans act as coreceptors forheparin-binding growth factors and ultimately lead to stimula-tion or inhibition of growth factor activity.45 A recent study byMidorikawa et al. determined a role for GPC3 in hepatocarci-nogenesis. They showed that GPC3 interacts with FGF2, in-hibits BMP7 signaling pathway, and modulates the activities ofFGF2 and BMP7.46 The modulation of growth factors byGPC3 suggests a growth-promoting effect for GPC3 in HCCtumorigenesis and the use of GPC3 as a drug target for HCC.

Human liver cancer cell lines established from the pri-mary HCC tumor are traditionally used to study the mech-anisms of tumorigenesis due to the ease of procurement andhandling of these cell lines. How robust cell lines mimic thebiologic behavior of primary tumors has not been well doc-umented. Recently, the genetic profiles of 19 HCC cell lineswere characterized and two subtypes of cell lines that werehighly correlated to the expression of AFP were identified.47

The overexpression of AFP, however, is not consistent acrossprimary HCC tumors, as shown in the current study and byothers.12,24 Our findings reveal that many molecular featurescharacteristic of primary HCC were lacking in the cell lines,whereas other important features exhibited more augmentedor diminished expression levels in cell lines. Comparativegenomic hybridization studies indicate that aberrant regula-tion of gene expression in primary HCC and liver cancer celllines is frequently caused by the gain or loss of chromosomalregions.48 In particular, overrepresentations occur more fre-quently on chromosome 1q.18 This correlates well with ourobservations that a disproportionate number of genes lo-cated on chromosome 1q showed high expression in HCCtumor relative to nontumor tissue specimens.

In conclusion, our findings, taken together with otherarray studies, highlight the consistent biologic associa-tions of HCC tumorigenesis with gene expression pro-files. They suggest candidate markers and pathwaysuncovered by this approach that are likely to be useful inthe diagnosis and treatment of HCC.

Acknowledgment: The authors thank ChandramouliGadisetti and Christos Sotiriou for advice in the initialanalysis; Adaikalavan Ramasamy, Safia S. Rahman, andLinda K. H. Teng for technical assistance; and Lance D.Miller for critical review of the manuscript.

References1. Parkin DM, Pisani P, Ferlay J. Global cancer statistics. CA 1999;49:33–64.2. Yu AS, Keeffe EB. Management of hepatocellular carcinoma. Rev Gastro-

enterol Disord 2003;3:8–24.3. Bosl GJ, Head MD. Serum tumor marker half-life during chemotherapy in

patients with germ cell tumors. Int J Biol Markers 1994;9:25–28.4. Lee WM. Hepatitis B virus infection. N Engl J Med 1997;337:1733–1745.5. Brown PO, Botstein D. Exploring the new world of the genome with DNA

microarrays. Nat Genet 1999;21(Suppl):33–37.6. Okabe H, Satoh S, Kato T, Kitahara O, Yanagawa R, Yamaoka Y, Tsunoda T,

et al. Genome-wide analysis of gene expression in human hepatocellular car-

952 NEO ET AL. HEPATOLOGY, April 2004

cinomas using cDNA microarray: identification of genes involved in viralcarcinogenesis and tumor progression. Cancer Res 2001;61:2129–2137.

7. Honda M, Kaneko S, Kawa H, Shirota Y, Kobayashi K. Differential geneexpression between chronic hepatitis B and C hepatic lesion. Gastroenter-ology 2001;120:955–966.

8. Shirota Y, Kaneko S, Honda M, Kawai HF, Kobayashi K. Identification ofdifferentially expressed genes in hepatocellular carcinoma with cDNA mi-croarrays. HEPATOLOGY 2001;33:832–840.

9. Tackels-Horne D, Goodman MD, Williams AJ, Wilson DJ, Eskandari T,Vogt LM, Boland JF, et al. Identification of differentially expressed genesin hepatocellular carcinoma and metastatic liver tumors by oligonucleotideexpression profiling. Cancer 2001;92:395–405.

10. Xu L, Hui L, Wang S, Gong J, Jin Y, Wang Y, Ji Y, et al. Expressionprofiling suggested a regulatory role of liver-enriched transcription factorsin human hepatocellular carcinoma. Cancer Res 2001;61:3176–3781.

11. Xu X-R, Huang J, Xu Z-G, Qian B-Z, Zhu Z-D, Yan Q, Cai T, et al.Insight into hepatocellular carcinogenesis at transcriptome level by com-paring gene expression profiles of hepatocellular carcinoma with those ofcorresponding noncancerous liver. Proc Natl Acad Sci USA 2001;98:15089–15094.

12. Chen X, Cheung ST, So S, Fan ST, Barry C, Higgins J, Lai K-M, et al.Gene expression patterns in human liver cancers. Mol Biol Cell 2002;13:1929–1939.

13. Eberwine J, Yeh H, Miyashiro K, Cao Y, Nair S, Finnell R, Zettel M, et al.Analysis of gene expression in single live neurons. Proc Natl Acad SciUSA 1992;89:3010–3014.

14. Sotiriou C, Khanna C, Jazaeri AA, Petersen D, Liu ET. Core biopsies canbe used to distinguish differences in expression profiling by cDNA mi-croarrays. J Mol Diag 2002;4:30–36.

15. Eisen MB, Spellman PT, Brown PO, Botstein D. Cluster analysis anddisplay of genome-wide expression patterns. Proc Natl Acad Sci USA1998;95:14863–14688.

16. The Gene Ontology Consortium. Gene ontology: tool for the unificationof biology. Nat Genet 2000; 25: 25-29. (Available: URL: http://www.geneontology.org).

17. Davison AC, Hinkley DV. Bootstrap methods and their application. Cam-bridge: Cambridge University Press, 1997.

18. Marchio A, Meddeb M, Pineau P, Danglot G, Tiollais P, Bernheim A,Dejean A. Recurrent chromosomal abnormalities in hepatocellular carci-noma detected by comparative genomic hybridization. Genes Chrom Can-cer 1997;18:59–65.

19. Raaphorst FM, van Kemenade FJ, Blokzijl T, Fieret E, Hamer KM, SatijnDPE, Otte AP, et al. Coexpression of BMI-1 and EZH2 polycomb groupgenes in Reed-Sternberg cells of Hodgkin’s disease. Am J Pathol 2000;157:709–715.

20. Kishima Y, Yoshida K, Enomoto H, Yamamoto M, Kuroda T, Okuda Y,Uyama H, et al. Antisense oligonucleotides of hepatoma-derived growthfactor (HDGF) suppress the proliferation of hepatoma cells. Hepatogas-troenterology 2002;49:1639–1644.

21. Leung KC, Doyle N, Ballesteros M, Waters MJ, Ho KK. Insulin regulation ofhuman hepatic growth hormone receptors: divergent effects on biosynthesisand surface translocation. J Clin Endocrinol Metab 2000;85:4712–4720.

22. Zhang N, Deuel TF. Pleiotrophin and midkine, a family of mitogenic andangiogenic heparin-binding growth and differentiation factors. Curr OpinHematol 1999;6:44–50.

23. Irniger S, Nasmyth K. The anaphase-promoting complex is required in G1arrested yeast cells to inhibit B-type cyclin accumulation and to preventuncontrolled entry into S-phase. J Cell Sci 1997;110:1523–1531.

24. Johnson PJ. The role of serum alpha-fetoprotein estimation in the diagnosisand management of hepatocellular carcinoma. Clin Liver Dis 2001;5:145–159.

25. Huynh H, Chow PKH, Ooi LLP, Soo K-C. A possible role for insulin-likegrowth factor-binding protein-3 autocrine/paracrine loops in controlling hep-atocellular carcinoma cell proliferation. Cell Growth Diff 2002;13:115–122.

26. Ito Y, Takeda T, Sakon M, Tsujimoto M, Higashiyama S, Noda K, Miyo-shi E, et al. Expression and clinical significance of erb-B receptor family inhepatocellular carcinoma. Br J Cancer 2001;84:1377–1383.

27. Riese DJII, Stern DF. Specificity within the EGF family/ErbB receptorfamily signaling network. Bioessays 1998;20:41–48.

28. Salomoni P, Pandolfi PP. p53 de-ubiquitination: at the edge between lifeand death. Nat Cell Biol 2002;4:E152–E153.

29. Pickart CM. Ubiquitin enters the new millennium. Mol Cell 2001;8:499–504.30. Salghetti SE, Caudy AA, Chenoweth JG, Tansey WP. Regulation of tran-

scriptional activation domain function by ubiquitin. Science 2001;293:1651–1653.

31. Sakamoto KM. Ubiquitin-dependent proteolysis: its role in human dis-eases and the design of therapeutic strategies. Mol Genet Metab 2002;77:44–56.

32. Nawrocki ST, Bruns CJ, Harbison MT, Bold RJ, Gotsch BS, AbbruzzeseJL, Elliott P, et al. Effects of the proteasome inhibitor PS-341 on apoptosisand angiogenesis in orthotopic human pancreatic tumor xenografts. MolCancer Ther 2002;1:1243–1253.

33. Carver RS, Sliwkowski MX, Sitaric C, Russell WE. Insulin regulatesheregulin binding and ErbB3 expression in rat hepatocytes. J Biol Chem1996;271:13491–13496.

34. Christofori G. Changing neighbours, changing behaviour: cell adhesionmolecule-mediated signaling during tumor progression. EMBO J 2003;22:2318–2323.

35. Masumoto A, Arao S, Otsuki M. Role of �1 integrins in adhesion andinvasion of hepatocellular cells. HEPATOLOGY 1999;29:68–74.

36. Jiang J, Yu L, Huang X, Chen X, Li D, Zhang Y, Tang L, et al. Identifi-cation of two novel human dynein light chain genes, DNLC2A andDNLC2B, and their expression changes in hepatocellular carcinoma tis-sues from 68 Chinese patients. Gene 2001;281:103–113.

37. Smith JW. Cilengitide Merck. Curr Opin Invest Drugs 2003;4:741–745.38. Orlando V. Polycomb, epigenomes and control of cell identity. Cell 2003;

112:599–606.39. Varambally S, Dhanasekaran SM, Zhou M, Barrette TR, Kumar-Sinha C,

Sanda MG, Ghosh D, et al. The polycomb group protein EZH2 is in-volved in progression of prostate cancer. Nature 2002;419:624–629.

40. Lessard J, Sauvageau G. Bmi-1 determines the proliferative capacity ofnormal and leukaemic stem cells. Nature 2003;423:255–260.

41. Furstenberger G, Senn H. Insulin-like growth factors and cancer. LancetOncol 2002;3:298–302.

42. Grimberg A, Cohen P. Role of insulin-like growth factors and their bindingproteins in growth control and carcinogenesis. J Cell Physiol 2000;183:1–9.

43. Saikali Z, Sinnett D. Expression of glypican 3 (GPC3) in embryonal tu-mors. Int J Cancer 2000;89:418–422.

44. Xiang YY, Ladeda V, Filmus J. Glypican-3 expression is silenced in humanbreast cancer. Oncogene 2001;20:7408–7412.

45. Filmus J, Selleck SB. Glypicans: proteoglycans with a surprise. J Clin Invest2001;108:497–501.

46. Midorikawa Y, Ishikawa S, Iwanari H, Imamura T, Sakamoto H, Miya-zono K, Kodama T, et al. Glypican-3, overexpressed in hepatocellularcarcinoma, modulates FGF2 and BMP-7 signaling. Int J Cancer 2003;103:455–465.

47. Lee J-S, Thorgeirsson SS. Functional and genomic implications of globalgene expression profiles in cell lines from human hepatocellular cancer.HEPATOLOGY 2002;35:1134–1143.

48. Zimonjic DB, Keck CL, Thorgeirsson SS, Popescu NC. Novel recurrent ge-netic imbalances in human hepatocellular carcinoma cell lines identified bycomparative genomic hybridization. HEPATOLOGY 1999; 29:1208–1214.

HEPATOLOGY, Vol. 39, No. 4, 2004 NEO ET AL. 953