Gene Expression Profiling Reveals Potential Biomarkers of Human Hepatocellular Carcinoma

Purpose: Hepatocellular carcinoma (HCC), a common cancer worldwide, has a dismal outcome partly due to the poor identification of early-stage HCC. Currently, one third of HCC patients present with low serum α-fetoprotein (AFP) levels, the only clinically available diagnostic marker for HCC. The aim of this study was to identify new diagnostic molecular markers for HCC, especially for individuals with low serum AFP.

Experimental Design: We used the microarray technique to determine the expression profiles of 218 HCC specimens from patients with either high or low serum AFP. From the microarray study, we selected five candidate genes (i.e., GPC3, PEG10, MDK, SERPINI1, and QP-C), which were overexpressed in HCCs. Using quantitative real-time PCR analyses, we validated the expression of these five genes in 50 AFP-normal and 8 AFP-positive HCC specimens and 36 cirrhotic noncancerous hepatic specimens, which include 52 independent specimens not used in microarray analysis.

Results: A significant increase in the expression of the five candidate genes could be detected in most of the HCC samples, including those with normal serum AFP and small tumors. GPC3, MDK, and SERPINI1 encode known serum proteins. Consistently, a significant increase in serum midkine, encoded by MDK, was associated with HCC patients, including those with normal serum AFP. Using prediction analysis of microarray, we showed that a combined score of these five genes can accurately classify noncancerous hepatic tissues (100%) and HCC (71%).

Conclusions: We suggest that a diagnostic signature approach using a combined score of these five biomarkers rather than a single marker may improve the prediction accuracy of HCC patients, including those with normal serum AFP and smaller-sized tumors.

Hepatocellular carcinoma (HCC), endemic to Asia and Africa with a rising incidence in Western countries (1–3), is one of the most common and aggressive cancers worldwide. It has been the third cancer killer worldwide and the second cancer killer in China since 1990s (4, 5). Globally, the 5-year survival rate of HCC is <5% and ∼598,000 HCC patients die each year (6). The high mortality associated with this disease is mainly attributed to the inability to diagnose HCC patients at an early stage. In fact, most symptomatic HCC patients are diagnosed at an advanced stage, thus precluding their chance for surgical intervention (7). In contrast, HCC patients who were diagnosed at an early stage and received curative resection had a significantly improved survival time (8–10). Thus, early detection and resection have been generally recognized as the most important factors to achieve long-term survival for HCC patients.

Since its detection in the serum of HCC patients in 1970s, α-fetoprotein (AFP) has been the only serologic marker widely used for diagnosing HCC patients. This marker allows the identification of a small set of HCC patients with smaller tumors, and these patients have a relatively long-term survival rate following curative treatment (8–10). Presently, the only approach to screen for the presence of HCC in high-risk populations is the combination of serum AFP and ultrasonography (11–13). However, elevated serum AFP is only observed in about 60% to 70% of HCC patients and, to a lesser extent (33-65%), in patients with smaller HCCs (14). Moreover, nonspecific elevation of serum AFP has been found in 15% to 58% of patients with chronic hepatitis and 11% to 47% of patients with liver cirrhosis (14). Therefore, it is necessary to identify new serologic HCC biomarkers that have a sufficient sensitivity and specificity for the diagnosis of HCC patients, especially in AFP-normal and/or smaller HCC cases.

This study was designed to identify new biomarkers for HCC. In this study, we first conducted microarray analyses to identify genes that were differentially expressed in HCC samples among patients with high serum AFP to those with low serum AFP. We then selected five candidate genes that were highly expressed in HCC samples, including those with low serum AFP and whose protein products could potentially be identified in sera. We did validation analysis by quantitative real-time PCR analyses (qRT-PCR). We show that these five genes separate HCC specimens from cirrhotic and noncancerous liver specimens. A significant increased expression can be found in most of the HCC specimens, including cases with normal serum AFP and small tumor size, suggesting that the products of these five genes may serve as biomarkers to aid HCC diagnosis. Serum midkine was significantly elevated in HCC patients, including those who had normal serum AFP and smaller tumors. These results warrant further efforts in the development of serum-based detection assays for GPC3 and SERPINI1, and possibly PEG10 and QP-C, and an algorithm by combining these five markers as a potential early indicator for HCC.

Clinical specimens. Surgical tissue specimens from Chinese HCC patients were collected with informed consent and approved by the Institutional Review Board of the Liver Cancer Institute (Shanghai, China). A total of 206 HCC tumor tissue and paired adjacent nontumor liver tissue samples [Fig. 1, cohort B: 185 cases used in oligo microarray study and additional 21 cases used in qRT-PCR (Fig. 1, cohort B + C)] was collected from patients undergoing surgery during the period August 2002 to November 2003 at the Liver Cancer Institute. An additional 33 HCC samples used in cDNA microarray analysis (Fig. 1, cohort A) and 26 cirrhotic liver specimens from non-HCC chronic liver disease patients (Fig. 1, part of cohort C) were previously described (15, 16). Normal liver tissue samples were obtained from seven disease-free liver donors. Pathologic diagnosis was independently done by two pathologists. In addition, independently archived serum samples from a total of 64 HCC patients (53 males from Liver Cancer Institute) and 26 normal blood donors (24 males from the Department of Transfusion Medicine, NIH, Bethesda, MD) were also included in the study (Fig. 1, cohort D). The median ages for cases and controls are 54 (range, 25-71) and 45 (range, 18-65), respectively. An additional set of serum samples collected at the Shanghai Public Health Center of Fudan University (Shanghai, China) was independently tested by MDK ELISA, which included 32 HCC patients, 20 hospital controls, and 49 cirrhosis patients (Fig. 1, cohort E). Measurements of hepatitis viral status and AFP level were standardized between studies. The clinical characteristics of these samples are included in Table 1 and in Supplementary Tables S1, S2, and S3.

AFP grouping. The clinically acceptable normal serum AFP was defined as ≤20 ng/mL. Patients with serum AFP of >300 ng/mL were considered as true positive cases because of a higher specificity by such a cutoff (11, 17). We therefore grouped HCC cases with AFP of ≤20, <300, and ≥300 ng/mL as AFP-normal HCC, low AFP HCC, and high AFP HCC, respectively.

Microarrays, RNA isolation, qRT-PCR, MDK ELISA, and statistical analyses. The microarray slides were generated at the Advanced Technology Center of the National Cancer Institute (Gaithersburg, MD). The Human Array-Ready Oligo Set (version 2.0) containing 70-mer probe of 21,329 genes was obtained from Qiagen, Inc. (Valencia, CA). Total RNA samples were extracted from snap-frozen tissue sections using Trizol reagent (Invitrogen, Carlsbad, CA) according to the manufacturer's protocol. The microarray methodology was similar to that previously described (16), except an indirect labeling approach was used. The BRB ArrayTools software V3.2.2, developed by Simon et al. at the National Cancer Institute Biometric Research Branch,⁵

was used for supervised class comparison between high and low AFP groups. The class comparison algorithm with random permutation tests was proven to be superior in identifying significant discriminatory genes with a control for false discovery (15, 16). This method was used to search for genes that were differentially expressed at a significance level of P < 0.002 between low and high AFP HCC.

qRT-PCR was used to validate the microarray results. The RT-PCR products collected at the last PCR cycle were analyzed by agarose gel electrophoresis to ensure the unique amplification of targeted genes (Supplementary Fig. S1). Human 18S rRNA labeled with VIC reporter dye was used as an endogenous control for each sample. A RNA pool from seven disease-free normal liver donors was used as a reference. Only AFP-normal HCC cases (≤20 ng/mL of serum AFP) and high AFP HCC cases (≥300 ng/mL of serum AFP) were analyzed by qRT-PCR. The relative mRNA expression levels were analyzed by the 2-ΔΔC_t method (18). Prediction analysis of microarray developed by Tibshirani et al. (19) was used to classify HCC samples based on the qRT-PCR data. Multidimensional scaling analysis based on Euclidean distance was used to visualize the classification outcome. The positive predictive value and negative predictive value were calculated as follows: positive predictive value = n11 / (n11 + n21) and negative predictive value = n22 / (n12 + n22), respectively, where n11 = number of HCC predicted as HCC, n12 = number of HCC predicted as non-HCC, n21 = number of non-HCC predicted as HCC, and n22 = number of non-HCC predicted as non-HCC.

Serum midkine concentrations were determined by ELISA using a commercial kit (BioVendor, LLC, Candler, NC). The assay was done according to the manufacturer's instructions.

We used the microarray technique to search for potential biomarkers that can aid the diagnosis of HCC patients, including those with low serum AFP. Our strategy used a primary search for differentially expressed HCC-associated genes among AFP-low and AFP-high groups and a secondary selection of those with a ≥2-fold expression ratio in AFP-low HCC cases. We first analyzed our recent cDNA microarray expression data (containing ∼9,000 human genes) of 33 HCC specimens (Fig. 1, cohort A; ref. 16). Among them, 20 cases (61%) had a high AFP serum level (≥300 ng/mL), whereas 13 cases (39%) had a low serum AFP (<300 ng/mL). Supervised class comparison with 1,000 random permutation tests and a false discovery rate control resulted in 90 significant genes (P < 0.002) that could discriminate these two groups. However, only three genes (MDK, GPC3, and QP-C) had an average expression ratio of ≥2-fold in low AFP cases (Supplementary Table S4). Due to this low success rate in identifying enough candidates for further screening and validation, we analyzed the gene expression profiles of an additional 185 HCC samples compared with their matched noncancerous liver tissues using a microarray platform consisting of 21,329 expression features (Fig. 1, cohort B). Among them, 88 (48%) were high AFP and 97 (52%) were low AFP. Similar supervised analysis resulted in a total of 1,317 significant genes (P < 0.002), of which 35 had an average expression ratio of ≥2-fold in AFP-low cases (Supplementary Table S4). Notably, the GPC3 gene was commonly identified from both microarray studies.

Although each of the 37 candidates may serve as a potential diagnostic biomarker, initially, we subjectively selected five genes (i.e., MDK, GPC3, QP-C, PEG10, and SERPINI1) for further study by qRT-PCR based on the following considerations: (a) all three genes identified by cDNA array with a relatively high expression ratio in AFP-low HCC compared with the matched noncancerous liver tissues (MDK, GPC3, and QP-C) and (b) an additional three genes from the oligo array analysis, two of which had the highest AFP-HCC expression fold change [GPC3 (also identified by cDNA array) and PEG10], and an additional gene whose encoded protein is secreted and could potentially be detected in serum or plasma samples (SERPINI1). For example, 92% of low AFP HCC cases had a ≥2-fold expression in the MDK gene, whereas 69% to 81% of low AFP HCC cases had a ≥2-fold expression in the GPC3 gene (Supplementary Table S4). It seemed that most low AFP HCC samples had an increased expression of MDK, GPC3, QP-C, and PEG10.

Next, we quantified the expression of these five genes by qRT-PCR on 36 noncancerous liver tissues and 58 HCC tumors, of which 21 HCC tumors and 31 noncancerous liver tissues were not used in the microarray studies described above (Table 1; Fig. 1, cohort B + C). Fifty HCC samples had normal serum AFP (≤20 ng/mL of serum AFP) and most of them were derived from cirrhotic livers (95%). The average serum levels of AFP were 6 ng/mL (range, 1-18) in the AFP-normal group and 1,758 ng/mL (range, 329-4,620) in the high AFP group (Table 1). Among 36 noncancerous liver tissues, 10 [8 AFP normal, 5 ng/mL (1–8); 2 high AFP HCC, 1,061 ng/mL (range, 329-1,794)] were randomly selected from the 58 HCC patients described above, and 26 were from cirrhotic and noncancer patients [i.e., primary biliary cirrhosis (n = 15), autoimmune hepatitis (n = 7), and alcoholic liver disease (n = 4)] described previously (15). The average size of primary tumor lesions from the 58 HCC patients was 4.6 cm in diameter (range, 1.1-15.3). The relative mRNA abundance of each gene was obtained by normalization to the 18S ribosomal gene and then to a pool of seven disease-free normal liver tissues. Although a total of 94 specimens was used in qRT-PCR to validate the microarray results, only 51 independent samples (21 HCC samples and 31 noncancerous hepatic samples) not used in microarray analyses were included in computing the prediction accuracy. A “normal” threshold (noncancer) for each gene was established as the mean level of each gene in 31 independent noncancerous liver samples plus two SDs of the variance. Thus, a HCC sample was considered to have an elevated expression of a given gene when its lowest level in a triplicate measurement was above the normal threshold. As shown in Fig. 2, most HCC cases had an elevated expression of these five genes above the defined normal threshold and their increased expressions seemed to be specific to HCC samples, including those with AFP normal (<20 ng/mL) and smaller tumors (<3 or 5 cm in diameter; Table 2; Fig. 2). Furthermore, the expression levels of MDK and SERPINI1 were significantly higher in AFP-high samples than that of AFP-normal samples, whereas the level of QP-C was significantly lower in AFP-high cases (Supplementary Fig. S2). Only a few samples had an expression below the threshold, regardless of whether they were from new HCC samples (Fig. 2, black circles) or those originally analyzed by microarray (Fig. 2, red circles). Similar to the microarray data, the expression levels of GPC3 and PEG10 were also higher in AFP-positive than that of AFP-normal samples, although the differences were statistically insignificant (Supplementary Fig. S2). It seems that the qRT-PCR data largely correlate with the microarray data.

To further test if these five genes provided a sufficient combined expression score to separate HCC from nontumor liver specimens, we used qRT-PCR expression data to determine the probability of correctly classifying only 21 independent HCC samples and 31 independent nontumor liver specimens by the prediction analysis of microarray algorithm (Fig. 1, only cohort C). Prediction analysis of microarray analysis, using nearest shrunken centroid classification with 10-fold cross-validation, resulted in 100% correct classification of nontumor samples and 71% correct classification of independent HCC samples (Fig. 3). Similarly, we did prediction analysis of all 94 samples and found that the five-gene signature could provide 84% sensitivity (58 HCC) and 100% specificity (36 non-HCC; data not shown). In addition, similar results were obtained when all 94 samples were included and randomly partitioned into 50% training and 50% testing samples, a process that was repeated five times (Supplementary Fig. S3A). It seemed that GPC3, PEG10, and MDK provided the most weight to discriminate HCC from nontumor liver specimens (Supplementary Fig. S3B and C). Consistently, the same HCC samples were misclassified with these two approaches, suggesting that misclassification did not occur by random chance. Thus, this five-gene signature provided high sensitivity and specificity to classify new HCC samples. The positive predictive value and negative predictive value were 100% and 80%, respectively.

Three of the five genes encode known proteins found in sera. The level of serum midkine (MDK) was shown to be elevated in several types of human cancers, such as esophageal cancer, gastric cancer, and neuroblastoma (20). In addition, increased MDK expression was also reported in HCC (21). Because a commercial MDK ELISA assay was available, we determined serum MDK level in 64 additional HCC patients (cases) and 26 normal blood donors (controls; Fig. 1, cohort D). The medians (25th and 75th percentiles) of serum MDK were 442.8 (288.0 and 666.5) pg/mL in the cases and 11.7 (3.8 and 42.3) pg/mL in the controls. A statistically significant difference (P < 0.0001, two tailed) was observed between cases and controls as determined by unpaired t test with Welch's correction (Fig. 4A). No significant difference was observed in the cases when stratified by overall AFP (P = 0.96) or tumor size (P = 0.22) as determined by one-way ANOVA. However, when using the median tumor size (5 cm in diameter) as the cutoff, we found that larger tumors (≥5 cm) had significantly higher serum AFP than that of smaller tumors (<5 cm; P = 0.02; Fig. 4B). In contrast, the difference in serum MDK between small and large tumors (Fig. 4C) was insignificant (P > 0.05). These results suggest that serum AFP level may reflect the degree of tumor burden, whereas serum MDK is independent of the tumor size. Similar results were obtained in a separate cohort (Fig. 1, cohort E). We found that HCC patients had a statistically significant higher serum MDK than patients with cirrhosis or hospital controls (P < 0.0001 and P = 0.0002, respectively; Supplementary Fig. S4) regardless of the AFP level. Thus, it is feasible to diagnose AFP-normal and small HCC with serum samples.

HCC is a common and aggressive malignant tumor worldwide with a dismal outcome. Early detection and resection may offer an opportunity to improve the long-term survival for HCC patients. Unfortunately, with current diagnostic approaches, only about 10% to 20% of HCC patients are eligible for resection (22). Currently, AFP is the only serologic biomarker for HCC in clinical practice. With its low sensitivity and nonspecific elevation in nonmalignant hepatic diseases, there is an urgent need to identify additional diagnostic biomarkers that can assist HCC early diagnosis, especially in AFP-normal cases and in tumors with smaller sizes.

In this study, we first did microarray analyses of HCC samples from patients with different levels of serum AFP. The class comparison analysis used in this study successfully identified AFP, the current HCC diagnostic biomarker, as a significant differentially expressed gene, further validating the utility of this approach. The microarray analysis allows us to identify 37 potential diagnostic biomarkers that are differentially expressed in AFP-positive and AFP-negative HCC samples. Five selected genes were validated by qRT-PCR as biomarkers to distinguish tumor samples from noncancerous hepatic tissues (Supplementary Table S5). A traditional HCC diagnosis has relied on the use of a single biomarker approach (e.g., AFP). We propose that the use of multiple markers in combination may improve the accuracy of identifying HCC cases. Among the five candidate genes, GPC3, MDK, and SERPINI1 encode known extracellular proteins (i.e., glypican-3, midkine, and serine/cysteine proteinase inhibitor 12), which can be found in sera. We have tested serum MDK and have found it can distinguish normal and cirrhosis individuals from HCC patients, including those with normal AFP and small tumors from two independent cohorts, and could therefore be used as a diagnostic marker. In principle, such an approach can also be used to classify HCC by combining readings from several serum markers. Thus, this study warrants further development of simple methodologies, such as an ELISA-based assay, to quantify these proteins in patient serum samples. It should be noted that most of HCC cases from this study are only positive for hepatitis B. Thus, the applicability of these markers for HCC with other etiologies, such as hepatitis C virus, remains to be determined.

Our qRT-PCR analyses indicate that GPC3 is the lead gene in distinguishing normal from HCC patients, followed by PEG10 and MDK. GPC3 encodes glypican-3, a member of the heparin sulfate proteoglycan family, which binds to the cell membrane through glycosylphosphatidylinositol anchors (23). Recent studies indicate that the serum GPC3 level is elevated in HCC patients and support its use for HCC patients (24–26). In addition, PEG10 is a novel imprinted gene located on human chromosome 7q21, whereas MDK encodes a novel heparin-binding growth factor originally identified in embryonal carcinoma cells that is involved in the early stage of retinoic acid–induced differentiation (27). Analogous to AFP, MDK mRNA is highly expressed during embryogenesis, but is undetectable in adult tissues, except kidney (28). The serum MDK level has been reported to be elevated in patients with various types of carcinomas, but not in normal individuals (20). Similarly, an increased expression of MDK and PEG10 has been reported to be associated with HCC (21, 29). Moreover, the expression of GPC3 is independent of the differentiation status and size of HCC (26). We have found that an increased expression of GPC3, PEG10, MDK, SERPINI1, and QP-C is associated with most HCC samples, including those with normal serum AFP and small tumor size. A combined expression score of AFP and these five genes may prove useful in the clinical setting and will be the subject of future studies. These findings warrant extensive validation of serum GPC3, MDK, and SERPINI1 as well as the potential presence of serum PEG10 and QP-C as clinical diagnostic tools by further investigation in a large case-control study with matched clinical variables, followed by a prospective study. Taken together, our results and published data indicate that these five genes may serve as diagnostic aids of HCC, including those with normal AFP and small HCC.

We thank Kei-Feng Lei, Yan-Fang Wang, Hong-Hai Wang, Xin Wu, Nicholas Younes, and Xiao-Qun Zhu for their technical support and Jia Fan, Zeng-Chen Ma, Zheng-Gang Ren, Zhi-Quan Wu, Sheng-Long Ye, and Xin-da Zhou for their help in collecting clinical specimens. We also thank Dorothea Dudek-Creaven for editorial assistance.