Protein expression associated with early intrahepatic recurrence of hepatocellular carcinoma after curative surgery

Abstract

The poor prognosis of patients with hepatocellular carcinoma (HCC) is attributed to intrahepatic recurrence. To understand the molecular background of early intrahepatic recurrence, we conducted a global protein expression study. We compared the protein expression profiles of the primary HCC tissues of 12 patients who showed intrahepatic recurrence within 6 months post surgery with those of 15 patients who had no recurrence 2 years post surgery. Two‐dimensional difference gel electrophoresis identified 23 protein spots, the intensity of which was highly associated with early intrahepatic recurrence. To validate the prediction performance of the identified proteins, we examined additional HCC tissues from 13 HCC patients; six with early intrahepatic recurrence and seven without recurrence. We found that all but one of the 13 patients were grouped according to their recurrence status based on the intensity of the 23 protein spots. Mass spectrometry identified 23 proteins corresponding to the spots. Although 13 of 23 have been previously reported to be correlated with HCC, their association with early intrahepatic recurrence had not been established. The identified proteins are involved in signal transduction pathways, glucose metabolism, cytoskeletal structure, cell adhesion, or function as antioxidants and chaperones. The identified proteins may be candidates for prognostic markers and contribute to the improvement of existing therapeutic strategies. (Cancer Sci 2007; 98: 665–673)

Despiterecent progress in early diagnostic modalities and therapeutic management, the prognosis of hepatocellular carcinoma (HCC) patients remains poor, mainly due to intrahepatic recurrence.^{( 1 )} The incidence of intrahepatic recurrence in the liver remnant ranges from 50% to 100% of HCC patients who undergo curative resection as a result of either intrahepatic metastasis from the primary tumor or multicentric recurrence^{( 2} , ³ , ⁴ , ⁵ , ^{6 )} and the median survival after recurrence is only 11 months.^{( 7 )} Although various molecular alterations have been correlated with early recurrence of HCC^{( 8} , ⁹ , ¹⁰ , ¹¹ , ¹² , ^{13 )} and studies on such events furthered our understanding of HCC biology, the underlying mechanisms remain obscure. Studies focusing on individual candidate genes may be insufficient for precisely understanding the background of the malignant behavior of tumor cells, because the features of cancer cells are defined by multiple genetic alterations in a functionally coordinated manner. With this notion in mind, global mRNA expression analyses have been conducted to identify the gene network associated with early intrahepatic recurrence.^{( 14} , ¹⁵ , ¹⁶ , ^{17 )} Such molecular pathogenesis studies may lead to the development of gene‐based biomarkers and novel therapeutic strategies. However, mRNA expression reflects the protein expression of only a small proportion of genes, probably because of post‐translational control of protein quantity.^{( 18} , ¹⁹ , ²⁰ , ^{21 )} To complement gene expression studies, global protein expression profiling, namely a proteomic approach, has been carried out for the study of HCC.^{( 22} , ²³ , ²⁴ , ²⁵ , ²⁶ , ²⁷ , ^{28 )} Conceptually, proteomics reflects the molecular background of cancer phenotypes more directly, as the final product of genetic and epigenetic events is the proteins contained in the cells.

In this study we investigated the protein expression of HCC tissues from patients with early intrahepatic recurrence and from patients without recurrence. We used two‐dimensional fluorescence difference gel electrophoresis (2D‐DIGE) to generate quantitative protein expression profiles, and data‐mining methods and mass spectrometry to determine the proteins associated with early intrahepatic recurrence.

Materials and Methods

Patients.  HCC patients who had undergone surgery at the National Cancer Center Hospital from January 1998 to April 2000 were evaluated for inclusion in this study. We assessed 40 patients who showed intrahepatic recurrence within 6 months or no recurrence within 2 years post surgery. The absence of intrahepatic metastasis in the residual liver was monitored by ultrasonography (US), computer tomography (CT) scans and angiography preoperatively, and was further confirmed by intraoperative US. CT scans were also carried out to rule out distant metastasis. Surgical margins were assessed post surgery by experienced pathologists. The patients were monitored for HCC recurrence at least once every 4 months after surgery using US and CT scans. To identify the protein spots associated with early intrahepatic recurrence, we validated the prediction performance of the selected spots using a newly enrolled sample set. The clinico‐pathological characteristics of the patient groups are summarized in Table 1. There was no significant difference of virus type between the early recurrence group and non‐recurrence group in both the training and test set. The number of primary tumor, venous invasion, and tumor‐node‐metastasis (TNM) staging were significantly different between the early recurrence and non‐recurrence groups in the training set (Table 1). The majority of cases were infected with hepatitis C virus (Table 1). This project was approved by the institutional review board for the use of human subjects of the National Cancer Center.

Table 1.

Clinical information of the patients

	Training set			Test set
	Early recurrence group (n = 12)	Non‐recurrence group (n = 15)	P‐value	Early recurrence group (n = 6)	Non‐recurrence group (n = 7)	P‐value
Age (mean ± SD)	63.2 ± 5.5	67.3 ± 3.6	0.06	60.5 ± 7.2	61 ± 9.1	0.91
Gender (%)
Male	7 (58.3)	12 (80)	0.06	5 (83.3)	7 (100)	0.46
Female	5 (41.7)	3 (20)		1 (16.7)	0 (0)
Viral infection (%)
HBV(+)	1 (8.3)	3 (20)	0.43	3 (50)	2 (28.6)	0.1
HCV(+)	7 (58.4)	10 (66.7)		1 (16.7)	4 (57.1)
HBV(+)HCV(+)	1 (8.3)	0 (0)		0 (0)	1 (14.3)
HBV(–)HCV(–)	3 (25)	2 (13.3)		2 (33.3)	0 (0)
Child‐Pugh classification (%)
A	12 (100)	15 (100)		6 (100)	7 (100)
B	0 (0)	0 (0)	–	0 (0)	0 (0)	–
C	0 (0)	0 (0)		0 (0)	0 (0)
ICG‐15 (mean ± SD)	19.7 ± 10.5	15.4 ± 9	0.26	22.1 ± 15.3	18.5 ± 9.8	0.63
AFP (ng/mL)
> 400	4 (33.3)	1 (6.7)	0.13	0 (0)	2 (28.6)	0.46
≤ 400	8 (66.7)	14 (93.3)		6 (100)	5 (71.4)
Tumor size (mean ± SD)	6.6 ± 5.2	4.1 ± 1.5	0.14	9.3 ± 7.3	4.5 ± 1.3	0.18
Primary lesion (%)
Single	2 (16.7)	14 (93.3)	< 0.001	1 (16.7)	7 (100)	0.004
Multiple	10 (63.3)	1 (6.7)		5 (83.3)	0 (0)
Gross type (%)
Type 1	3 (25)	7 (46.7)		0 (0)	4 (57.1)
Type 2	4 (33.3)	6 (40)	0.2	2 (33.3)	3 (42.9)	0.06
Type 3	5 (41.7)	2 (13.3)		4 (66.7)	0 (0)
Tumor differentiation (%)
Well	0 (0)	3 (20)		0 (0)	2 (28.6)
Moderately	7 (58.3)	10 (66.7)	0.18	4 (66.7)	4 (57.1)	0.62
Poorly	5 (41.7)	2 (13.3)		2 (33.3)	1 (14.3)
Venous invasion (%)
Presence	12 (100)	6 (40)	0.001	5 (83.3)	4 (57.1)	0.31
Absence	0 (0)	9 (60)		1 (16.7)	3 (42.9)
UICC pTNM stage (%)
Stage I	0 (0)	0 (0)	< 0.001	0 (0)	0 (0)	0.03
Stage II	0 (0)	12 (80)		0 (0)	4 (57.1)
Stage IIIA	10 (83.3)	3 (20)		3 (50)	3 (42.9)
Stage IVA	2 (16.7)	0 (0)		3 (50)	0 (0)
Background liver (%)
Normal	0 (0)	1 (6.7)		1 (16.7)	0 (0)
Chronic hepatitis	9 (75)	12 (80)	0.49	4 (66.6)	6 (85.7)	0.79
Cirrhosis	3 (25)	2 (13.3)		1 (16.7)	1 (14.3)

AFP, alpha‐fetoprotein; HBV, hepatitis B virus; HCV, hepatitis C virus; ICG, indocyanine green; pTNM, pathological TNM; UICC, Union Internationale Centre le Cancer

Protein extraction, fluorescence labeling and separation by two‐dimensional polyacrylamide gel electrophoresis.  The tissue specimens obtained from the surgically resected tumors were stored in vapor nitrogen until use. The frozen tissues were homogenized in urea lysis buffer (7 M urea, 2 M thiourea, 3% CHAPS, 1% Triton X‐100) and incubated on ice for 30 min. After centrifugation at 15 000 g for 30 min, the supernatant was recovered. Fluorescence labeling of proteins was carried out as described previously (Fig. 1A).^{( 22 )} In brief, a portion of all samples was mixed together to make an internal control sample, aliquoted and stored at −80°C until use. Fifty micrograms of both the internal control sample and each experimental sample were labeled with 200 nmol of 1‐(5‐carboxypentyl)‐1′‐propylindocarbocyamine halide (Cy3) and 1‐(5‐carboxypentyl)‐1′‐methylindocarbocyamine halide (Cy5) (GE Healthcare Bio‐sciences, Little Chalfont, Buckinghamshire, UK), respectively. After the labeling reaction was terminated with 0.2 mM lysine, equal amounts of the Cy3‐ or Cy5‐labeled samples were mixed and the sample volume was made up to 420 µL with urea lysis buffer containing 65 mM dithiothreitol (DTT) and 1% ampholine (Amersham Biosciences, Upsala, Sweden).

Figure 1.

(A) Fluorescence labeling and separation by two‐dimensional polyacrylamide gel electrophoresis (2D‐PAGE). A protein sample from a cell line and the internal control sample were labeled with Cy5 and Cy3, respectively. The Cy5‐labeled sample was mixed with the Cy3‐labeled internal control sample and coseparated in the gel. The Cy5 images such as the one shown here represent the protein expression profile of individual samples while the Cy3 image represents that of the common internal control sample. (B) Scatter plot analysis of three protein expression profiles obtained from the same sample. The intensity values obtained for each of the protein spots ranged within a twofold difference between the experiments for more than 90% of the spots.

Two‐dimensional gel electrophoresis was carried out as in our previous report (Fig. 1A).^{( 22 )} In brief, the first separation was achieved using IPG DryStrip gels (24 cm long with a pI range between 3.0 and 10.0, Amersham Biosciences) and IPGphor (GE Healthcare Bio‐sciences), and the second‐dimension separation was achieved using Ettan Dalt II and 9–15% polyacrylamide gradient gels. Identical samples were run in triplicate gels. After electrophoresis, gels were scanned with 2920 2D‐Master Imager (GE Healthcare Bio‐sciences). The Cy5‐image of each sample was normalized by the Cy3‐image of the internal standard sample using the DeCyder software (GE Healthcare Bio‐sciences).

We analyzed the acquired proteome data, which consisted of the intensity signal ratios of approximately 1400 protein spots from 40 HCC samples, using the Impressionist software (Gene Data, Basel, Switzerland). We assessed the reproducibility of the acquired protein profiles by running an identical sample three times and examining the similarities between the protein expression profiles with scatter plot analysis. The intensity values obtained for each of the protein spots ranged within a twofold difference between the experiments for more than 90% of the spots, while the correlation coefficient for the overall protein expression was at least 0.8274 (Fig. 1B). These results indicated that the protein expression profiles obtained by 2D‐DIGE were highly reproducible both in a quantitative and in a qualitative way.

Mass spectrometric identification of proteins.  Protein identification was achieved as reported previously.^{( 22 )} In brief, a 500‐µg protein sample was separated by two‐dimensional polyacrylamide gel electrophoresis (2D‐PAGE) and stained with SYPRO Ruby (Molecular Probes, Eugene, OR, USA). SpotPicker (GE Healthcare Bio‐sciences) collected the spots of interest. The tryptic peptides generated by in‐gel digestion^{( 22 )} were subjected to matrix‐assisted laser desorption/ionization time‐of‐flight mass spectrometry (MALDI‐TOF MS) using a Q‐Star Pulser‐i equipped with an orthogonal injection/MALDI ion source (Applied Biosystems, Framingham, CA, USA). Mass spectra were processed with the Analyst QS program (Applied Biosystems) and Mascot program (Matrix Science, Boston, MA, USA) using the Swiss‐Prot database for protein identification.

Results

We created a training set of 27 samples, consisting of 12 samples from patients who had early intrahepatic recurrence within 6 months and 15 samples from patients who had no recurrence within 2 years post surgery. We first examined the similarity between the 2D profiles of the samples in the training set. The correlation efficiency of all paired samples was calculated and the results were summarized in a correlation matrix (Fig. 2A). We found that the protein expression patterns of samples within each group (the non‐recurrence and the early recurrence group) were very similar to each other, while they were different to those of the other group, suggesting the presence of a certain proteomic background corresponding to the events leading to the recurrence. This observation allowed us to construct an early recurrence prediction model using spot intensity. We used a support‐vector‐machine algorithm to build a class prediction model. A support‐vector‐machine creates the hyper‐plane in a multidimensional space to best‐distinguish two sample groups.^{( 29 )} We identified the minimal number of protein spots that can be used to classify the samples with the minimal misclassification error rate. First, we constructed a classifier using a certain number of spots and calculated the misclassification error rate by a leave‐one‐out cross validation. Then, we ranked the protein spots according to their contribution to the classification model using support vector machine weight. The classification model was then constructed again after the lower‐ranked protein spots were deleted, and its misclassification rate was calculated. This process was repeated and the misclassification error rates were plotted as a function of the number of protein spots (Fig. 2B). Although the classifier that included all protein spots had a discriminatory power of 100% accuracy (0% misclassification error rate), a classifier with a smaller spot number (23 spots) also showed 0% misclassification error rate and was considered as the optimal for the classification model to distinguish the samples according to their recurrence status. These 23 protein spots are the minimal protein set to distinguish these two groups based on the support‐vector machine algorithm, and the reduction of the number of spots resulted in the increased misclassification error rate.

Figure 2.

Protein spots associated with early intrahepatic recurrence. (A) Correlation matrix showing the similarity of the overall feature of proteome. (B) A leave‐one‐out cross‐validation error rate was plotted as a function of spot number, showing that the minimal number of spots to make the error rate 0% was 23. Principal component analysis (C) and hierarchical clustering analysis (D) clearly separated the samples according to their recurrence status using the 23 selected protein spots. (E) All but one of the samples in the training and test set were correctly classified on the basis of the expression pattern of the 23 protein spots. The patients were assigned as those that had early recurrence within 6 months (R1‐R18) and those that did not have recurrence for 2 years post surgery (NR1‐NR22).

The classification performance of the 23 selected protein spots was validated using different classification methods. With principal component analysis, which visualizes the correlation of samples in a three‐dimensional space created by the intensity of the 23 protein spots, the 27 samples were clearly separated according to their recurrence status (Fig. 2C). A hierarchical clustering study based on similarities in the expression profiles of the 23 protein spots categorized the early intrahepatic recurrence and non‐recurrence groups into distinct branches (Fig. 2D). As the separation was ambiguous when all spots were used (data not shown), these 23 protein spots characterized the proteome of the tumors with or without recurrence.

The predictive performance of the selected set of spots was validated using a test set, which consisted of early intrahepatic recurrence (n = 6) and non‐recurrence cases (n = 7). We determined the affinity value of each sample by calculating the minimal distance between each sample and the hyper‐plane (Fig. 2E). The distance was defined so that the samples classified in the recurrence group had a positive distance value while the samples categorized in the non‐recurrence group had a negative value. Figure 2(E) shows that all samples except NR20 were correctly classified into the groups. The expression pattern of the 23 proteins in sample NR20 was more similar to that of the recurrence group, although there were no clinico‐pathological features unique to this case.

The portal vein is the main route for intrahepatic metastasis, and portal invasion is correlated with early recurrence.^{( 30 )} We therefore examined the proteins associated with portal invasion and the number of primary lesions. The training sample set was divided to portal invasion positive (n = 18) and negative (n = 9) cases (Table 1). We did not find a protein spot set with a misclassification rate of less than 13%. Using the selected protein spots, principal component analysis and hierarchical clustering analysis did not separate the samples according to their portal invasion status (data not shown). The number of primary lesions (single or multiple) has been also associated with early recurrence.^{( 31 )} We separated the samples in the training set into two groups: cases with single primary tumors (n = 16) and multiple primary tumors (n = 11) (Table 1). We did not identify any spots based on which the HCC samples could be separated according to the number of primary tumors (data not shown).

The location of the 23 protein spots on the 2D image is demonstrated in Figure 3(A). Figure 3(B) shows the 2D differential expression images of the 23 spots. We used a standardized spot intensity, which was calculated by dividing the intensity of the Cy5 signal by that of the Cy3 signal for each protein spot in the same gel, and the apparent spot intensity in the Cy5 image did not necessarily reflect the normalized intensity (Fig. 3B). Thus, we demonstrate the standardized spot intensity across the samples used to compare the expression level among the samples (Fig. 4). The mean ratios of the identified proteins in the recurrence and non‐recurrence groups are summarized in Table 2.

Figure 3.

(A) Localization of the 23 selected spots on the 2D image. (B) The enlarged images are shown to compare the typical image of a sample from the early recurrence cases (R1) and one from the non‐recurrence cases (NR1).

Figure 4.

Quantitative differential display of the 23 selected spots. The standardized spot intensities of all hepatocellular carcinoma (HCC) samples in the training set are displayed. Assignment of patients corresponds to Figure 2.

Table 2.

List of the identified 23 proteins

Ranking†	Spot no. ‡	Identification§	Acct.no.¶	Theoretic††		Observed‡‡		Fold diff§§	MS score†††	Coverage	MS/MS score¶¶ (%)‡‡‡	Function	Known connection to HCC	References
				Mr	pI	Mr	pI	R /NR
1	1400	Phosphoserine aminotransferase	Q9Y617	40.4	7.7	38.4	7.9	3.8	642	37.6	41	Serine biosynthesis	–	–
2	1033	IFP53/Tryptophanyl‐Trna synthetase	P23381	53.2	5.8	48.8	5.8	0.27	212	13.4	66	DNA/RNA synthesis	–	–
3	1877	Peroxiredoxin 3	P30048	27.7	7.7	25.7	5.9	1.82	171	21.9	51	Antioxidant	yes	40
4	485	GRP94 (Tumor rejection antigen 1)	P14625	92.5	4.8	93.5	4.9	0.59	1574	35.1	–	Chaperone	yes	41
5	1558	Annexin A1 (Phospholipase A2 inhibitory protein)	P04083	38.6	6.6	34.1	6.8	0.55	382	14.7	–	Signal transduction	yes	42
6	2003	Superoxide dismutase [Cu‐Zn]	P00441	15.8	5.7	21.4	5.8	0.44	306	22.7	31	Antioxidant	yes	45
7	1523	Annexin A2	P07355	38.5	7.6	34.7	7.7	0.42	548	18.9	–	Signal transduction	–	–
8	1629	Annexin A4	P09525	35.8	5.9	31.8	5.8	1.6	647	25.7	–	Signal transduction	–	–
9	1528	L‐lactate dehydrogenase B chain	P07195	36.5	5.7	34.4	5.8	1.61	427	10.8	–	Glucose metabolism	yes	44
10	1985	Cyclophilin B (Peptidyl‐prolyl cis‐trans isomerase B)	P23284	22.7	9.3	22.3	10	0.53	726	43.8	–	Signal ransduction	yes	47
11	345	ORP150	Q9Y4L1	111.3	5.2	115	5.2	0.59	671	23.9	–	Chaperone	–	–
12	412	Vinculin	P12003	124.4	5.4	99.8	5.9	1.82	597	49.2	–	Cytoskeletal structure	yes	13, 14
13	1870	Flavin reductase	P30043	21.9	7.3	25.8	8.1	0.52	106	14.1	48	Antioxidant	–	–
14	2025	Stathmin	P16949	17.2	5.8	20.9	5.8	0.56	269	34.9	79	Signal transduction	yes	–
15	1719	Protein kinase C inhibitor protein‐1 (14‐3‐3 zeta/delta)	P63104	27.8	4.7	29.2	4.6	2.7	1306	57.1	–	Signal transduction	–	–
16	1586	Heterogeneous nuclear ribonucleoprotein A2/B1	P22626	37.4	9	33.4	9.5	0.55	597	20.1	–	Pre‐mRNA processing	–	–
17	1723	26S proteasome non‐ATPase regulatory subunit 8	P48556	30	6.9	29.1	6.9	0.64	104	8.9	38	Protein degradation	–	–
18	711	Glutamine amidotransferase	P49915	76.7	6.4	74.6	6.7	1.31	945	51.2	53	DNA and RNA synthesis	–	–
19	1230	Phosphoglycerate kinase 1	P00558	44.6	8.3	40.6	9.1	0.69	234	34.9	–	Glucose metabolism	–	–
20	2107	TCP‐1 chaperonin cofactor A	O75347	12.7	5.3	18.1	5.3	1.57	72	26.9	44	Chaperonine	yes	47
21	841	TCP‐1‐gamma	P49368	60.4	6.1	65.9	6.6	0.64	783	27.8	–	Chaperonine	yes	48
22	1329	Multidrug resistance‐associ ated protein MGr1‐Ag	P08865	32.9	4.8	39	4.7	1.51	405	15.3	74	Cell adhesion	yes	49
23	1675	Nuclear chloride ion channel 27	O00299	26.9	5.1	30.5	5.2	1.64	1041	44	–	Ion channel	yes	38

^† Ranking orders are determined by support vector machine weight.

^‡ Spot numbers refer to those in 3, 4.

^§ §Identifications are according to Swiss‐Prot and NCBI database.

^¶ ¶Accetion numbers are derived from Swiss‐Prot database.

^†† Theoretical molecular weight (Mr) and isoelectric point (pI) are determined by Swiss‐Prot and the ExPASy databas (http://au.expasy.org).

^‡‡ ‡‡Theoretical molecular weight (Mr) and isoelectric point (pI) are determined by the location of the spots on the 2D image.

^§§ §§Fold differences indicate the ratio between the mean intensity of the recurrence and non‐recurrence sample groups.

^¶¶ ¶¶MS scores are calculated by MasCot program.

^††† †††Coverages indicate the percentage of the length of the identified peptide against the full length protein.

^‡‡‡ MS/MS scores are calculated by MasCot program. The blank (–) indicate that MS/MS analysis was not successfully achieved.

The mass spectrum of tryptic peptides in spot 1400 is shown in Figure 5A as an example of mass spectrometric protein identification. Eleven peptides were matched to those of phosphoserine aminotransferase with an Analyst QS score of 945. One ion peak (m/z 1184.6646) was further processed for tandem mass spectrometry analysis, resulting in identification of the same protein (Fig. 5B–D). The peptide sequence used for the identification is shown in Figure 5(E). The other identified proteins are summarized in Table 2.

Figure 5.

Mass spectrometric protein identification. (A) Mass spectrogram of spot 1400. Peptide ions with m/z‐values were used for protein identification. The ion peak shown with an arrow was subjected to MS/MS study. (B) MS/MS spectrogram of the peptide with m/z 1184.65. (C, D) Database search by the Mascot program using the MS/MS data. (E) Sequence of phosphoserine aminotransferase. The sequences used for the protein identification are underlined.

Discussion

We identified a set of 23 proteins whose expression pattern is significantly associated with early intrahepatic recurrence of hepatocellular carcinoma. Based on the expression profile of this set and using hierarchical clustering and principal component analysis, the HCC tumor samples were accurately classified into two groups according to their recurrence status, while they were ambiguously classified when all protein spots were used as the basis for their classification. As the prediction value of these proteins was validated by external validation samples, these proteins may be good candidates for tumor markers to predict early recurrence of HCC. The identification of patients at high risk for recurrence would be useful for further developing therapeutic strategies to possibly include prophylactic liver transplantation, an idea that would require the study of a larger number of HCC samples to be confirmed.

The proteins we identified in this study are involved in a wide range of biological processes, suggesting that various factors determine the malignant behavior of HCC cells. Phosphoserine aminotransferase is involved in the L‐serine biosynthesis pathway^{( 32 )} and enzymic imbalance of serine metabolism in rat hepatoma has been reported previously.^{( 33 )} As phosphoserine aminotransferase is highly expressed in tissues with a high rate of cell turnover^{( 34 )} its up‐regulation in HCC tissue in cases showing early recurrence may reflect the proliferation potential of the tumor cells. We also identified proteins involved in DNA/RNA synthesis, such as IFP53 and glutamine amidotransferase, premRNA processing, such as hnRNP A2/B1, chaperones such as GRP94, ORP150 and peptidyl‐prolyl cis‐trans isomerase B, chaperonins, such as TCP‐1‐chaperonin cofactor A and TCP‐1 gamma, and proteins in the proteolytic pathway, such as the 26S proteasome subunit. As these proteins play a key role in fundamental biological processes, from protein synthesis to degradation, they may be responsible for the overall proteomic differences observed between cases with and without recurrence (Fig. 2A). We identified three antioxidant proteins, namely peroxiredocin 3, superoxide dismutase and flavin reductase. Oxidative stress has been considered as a modulator of cancer‐relevant signaling pathways leading to the accumulation of oncogenic mutations in HCC.^{( 35 )} Our findings may either reflect the defensive response to the oxidative stress or the fact that tumor cells were clonally selected by the oxidative stress. Indeed, the proteins involved in the signal transduction pathways of the annexin families (annexin A1, 2, and 4) and protein kinase C (14‐3‐3 zeta/delta), and stathmin, a proliferation regulator, showed different expression levels between the early recurrence and non‐recurrence groups. Stathmin depolymerizes microtubules and plays a key role in signal transduction and cellular proliferation. Stathmin has been implicated in carcinogenesis and cancer progression in many types of malignant tumors, and was recently linked with carcinogenesis in HCC.^{( 36 )} Two enzymes involved in glucose metabolism, phosphoglycerate kinase 1 and L‐lactate dehydrogenase, were associated with early recurrence. Tumor formation is generally linked with increased activity of glycolytic enzymes, and a previous proteomic study on liver cancer reported the aberrant regulation of several glycolytic enzymes.^{( 27 )} In lung cancer patients, a proteomic study revealed that higher expression of PGK1 was also associated with reduced survival^{( 37 )} suggesting the importance of aberrant glycolysis in cancer progression. Nuclear chloride ion channel 27 was increased during the course of carcinogenesis in HCC^{( 38 )} and was also up‐regulated in the early recurrence group in our study. We observed up‐regulation of the resistance‐associated protein MGr1‐Ag in the early recurrence group. Recently, MGr1‐Ag turned out to be a ligand for PCK3145, an antimetastatic synthetic peptide for hormone‐refractory prostate cancer^{( 39 )} possibly presenting a novel molecular therapy target against early recurrence HCC. The early recurrence and non‐recurrence groups in the training set had the different number of primary tumor, venous invasion and TNM staging in the training set (Table 1). Therefore, these proteins might be involved in the mechanisms of these clinical features in a certain coordinated manner. However, as these clinical parameters were not significantly different between the two groups in the test set and these 23 proteins could distinguish them, the developed classifier may have more prediction value than these clinical parameters.

We may need an extensive validation study using a large‐scale sample set from the multiple hospitals to apply the present results for clinical applications because various clinical backgrounds should influence the prediction performance of the biomarkers. 2D‐DIGE is a powerful tool to discover the biomarker candidate, because it can provide quantitative protein expression data in a reproducible way across multiple samples in a relatively short time. However, it may be difficult to use 2D‐DIGE as a routine clinical examination tool in local hospitals because of its technical complexity and expensive initial and running costs. A more popular, convenient and less expensive examination tool, such as ELISA should be considered for screening purposes. Array technology has already enabled us to monitor the expression level of multiple proteins across a large number of samples in a less labor‐intense, high‐throughput and less expensive way, and the development of the specific antibody against the identified protein variants should be the key step to applying the proteomics results in the clinical setting.

Literature validation showed that many of the identified proteins were associated with HCC (Table 2).^{( 38} , ⁴⁰ , ⁴¹ , ⁴² , ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁶ , ⁴⁷ , ⁴⁸ , ^{49 )} However, with the exception of vinculin^{( 14} , ^{15 )} the genes reported to be associated with early recurrence of HCC in transcriptomic studies were not identified in this study, a fact possibly due to the current detection limitations of 2D‐DIGE using the CyDye DIGE Fluor minimal dye. To improve the sensitivity of our proteomic studies, we are now considering the use of prefractionation methods, narrow range IPG gels, more sensitive fluorescent dye such as the CyDye DIGE Fluor saturation dye, and larger format 2D gels.^{( 50 )} We found that the large format 2D gel with CyDye DIGE Fluor saturation dye could increase the number of spots up to 5000.^{( 51 )} The extended version of 2D‐DIGE linking to the modern mass spectrometry should be considered as a key modality of cancer proteomics. In addition, the combination of multiple proteomic modalities such as the above, mass spectrometry and array technology will act in a complementary way and result in a comprehensive view of liver cancer proteomics.