Identification of prognostic biomarkers in gastric cancer using endoscopic biopsy samples

Abstract

Endoscopic biopsy prior to chemotherapy provides an opportunity for studying biomarkers to predict the overall survival in gastric cancer patients. This prospective study was performed to identify prognostic biomarkers in patients with unresected gastric cancer. Fifty‐nine cases of chemotherapy‐naive metastatic gastric cancer were enrolled in this study. A microarray analysis was performed using 40 biopsy samples to identify candidate genes whose expressions might be correlated with the overall survival. After adjusting for clinical covariates based on a multivariate analysis, the identified genes were validated using real‐time reverse transcription polymerase chain reaction (RT‐PCR) analysis in 19 independent validation samples. Ninety‐eight candidate genes whose expression levels were significantly correlated with the overall survival were identified using a microarray analysis based on a proportional hazards model (P < 0.005). Multivariate analysis was performed to assess 10 of these genes, and the results yielded a statistical significance level for DACH1 and PDCD6. We further evaluated these two genes in independent samples using real‐time RT‐PCR and found that lower mRNA expression levels of PDCD6 were correlated significantly with a poor overall survival. We identified PDCD6 as a prognostic biomarker in patients with unresected gastric cancer using endoscopic biopsy samples. Our PCR‐based single gene prediction strategy successfully predicted the overall survival and may lead to a better understanding of this disease subgroup. (Cancer Sci 2008; 99: 2193–2199)

Over the past two decades, various anticancer agents have been examined for their efficacy against gastric cancer, including 5‐fluorouracil (5‐FU) and 5‐FU‐based drugs, taxanes, CPT‐11 and cisplatin, all administered either as monotherapy or in combination regimens;^{( 1 )} however, the median survival time (MST) of these patients remains at only approximately 7 months.^{( 2} , ^{3 )} In a recent randomized phase III trial examining oral S‐1 monotherapy and cisplatin plus irinotecan combination therapy, the response rates to both S‐1 and to the cisplatin plus irinotecan combination therapy were approximately 50%, indicating that around half of the patients did not respond to chemotherapy,^{( 4} , ⁵ , ⁶ , ^{7 )} and the MST in both the arms was less than 1 year.^{( 8 )} Thus, the prognosis of patients with gastric cancer remains poor.

The commonly recognized prognostic factors in cases of unresectable gastric cancer are the performance status, presence/absence of liver metastases, presence/absence of peritoneal metastases and the serum levels of alkaline phosphatase.^{( 9 )} Many molecular biomarkers have been also investigated for their potential to predict the outcome in hypothesis‐based studies. Several studies have shown that the mRNA levels and immunohistochemical staining intensity of thymidylate synthase (TS) in gastric cancers treated with fluorouracil are associated with the response and survival; in addition, the excision repair cross‐complementing (ERCC)1 gene expression level has been shown to be associated with the clinical outcome in patients treated with cisplatin.^{( 10} , ^{11 )} HER2 expression has also been reported to be a prognostic marker in cases of differentiated gastric cancer.^{( 12} , ^{13 )} Mutation of p53 and high p53 protein expression, and high expression levels of urokinase‐plasminogen activator, xanthine oxidoreductase, claudin‐4, vascular endothelial growth factor, interleukin‐8 and cyclin E have all been correlated with poor survival.^{( 13} , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ^{19 )} In terms of epigenetic alterations, reduced expression of acetylated histone H4 or DNA methylation of CDH1 and RAR‐β have been shown to be correlated with tumor invasiveness and the tumor metastasizing potential.^{( 20} , ^{21 )}

On the other hand, the recent introduction of the microarray technology has enabled significant genes to be identified almost throughout the genome using a hypothesis‐free approach. The possibility of performing genome‐wide searches is a major advantage, and such searches may be the only way to discover genes that would otherwise be unlikely to even be suggested as candidates. In gastric cancer, biopsy samples of the primary lesions can be easily obtained by endoscopy prior to treatment; however, few prospective biomarker studies using endoscopic biopsy samples to predict patient outcome have been performed to date. Therefore, we conducted a prospective study to identify biomarkers for predicting survival in patients with unresected metastatic gastric cancer.

Materials and Methods

Patients and samples.  The eligible subjects in this study were patients with histologically confirmed, untreated and metastatic stage IV gastric cancer between 20 and 75 years of age. Additional inclusion criteria included an Eastern Cooperative Oncology Group performance status of 0–2. The exclusion criteria included history of prior chemotherapy or major surgery. All patients received chemotherapy using a 5‐FU‐based regimen (5‐FU alone, S1 alone, 5‐FU + methotrexate, 5‐FU + cisplatin, or S1 + cisplatin) or a CPT‐11 plus cisplatin regimen. Sixty‐five gastric cancer patients were enrolled in the study. Of these, two were excluded because of insufficient RNA quantities extracted from their biopsy specimens, and four were excluded because of the poor RNA quality. Thus, samples from the remaining 59 patients were analyzed. The survival time was followed after the patients were initiated on chemotherapy. This study was approved by the Institutional Review Board of the National Cancer Center Hospital, and written informed consent was obtained from all the patients.

The endoscopic biopsy samples collected were immediately placed in an RNA stabilization solution (Isogen; Nippongene, Tokyo, Japan) and stored at –80°C. Other biopsy samples obtained from the same location were reviewed by a pathologist to confirm the presence of tumor cells. The RNA extraction method and the quality check protocol have been described previously.^{( 22 )}

Study design.  This prospective study was started in July 2003 and enrollment was completed in November 2006 at the National Cancer Center Hospital. Fifty‐nine gastric cancer samples were evaluated in this study. The samples were divided into a training set (n = 40) and a validation set (n = 19; 2:1) using computer‐generated randomization (Microsoft Office Excel, Microsoft, Redmond, WA, USA). A microarray analysis was performed using the training set of 40 samples, and candidate genes whose expressions were correlated with the overall survival were identified. Multivariate analysis was performed to adjust the expression of 10 of these candidate genes for clinical features. Finally, the significant genes were evaluated in an independent set of 19 samples and survival was predicted using the results of real‐time reverse transcription polymerase chain reaction (RT‐PCR) analyses.

Real‐time RT‐PCR.  Real‐time RT‐PCR was performed for 10 genes: DACH1 (dachshund homolog 1, NM_004392); EGFR (epidermal growth factor receptor, NM_005228); MT1X (metallothionein 1X, NM_005952); YWHAE (tyrosine 3‐monooxygenase/tryptophan 5‐monooxygenase activation protein, epsilon polypeptide, NM_006761); GPX3 (glutathione peroxidase 3, NM_002084); PDCD6 (programmed cell death 6, NM_013232); WDR33 (WD repeat domain 33, NM_018383); C14orf43 (chromosome 14 open reading frame 43, NM_194278); MYLIP (myosin regulatory light chain interacting protein, NM_013262); and GKAP1 (G kinase anchoring protein 1, NM_025211). Glyceraldehyde 3 phosphate dehydrogenase (GAPD, NM_002046) was used to normalize the expression levels in the subsequent quantitative analyses. RNA was converted to cDNA using a GeneAmp RNA PCR Core kit (Applied Biosystems, Foster City, CA). The transcripts were quantified using the Power SYBR Green PCR Master Mix (Applied Biosystems) and 7900HT Fast Real‐time PCR system (Applied Biosystems) and reported relative to the GAPD expression levels. The PCR conditions were as follows: one cycle of denaturation at 95°C for 10 min, followed by 40 cycles at 95°C for 15 s and 60°C for 60 s. To amplify the target genes, the following primers were purchased from Takara (Yotsukaichi, Japan): DACH1‐FW, 5′‐AAG GGC TGC TAA AGC AAT CAG G‐3′, and DACH1‐RW, 5′‐CTT TGT GGC AAA GCG ACA TTA GG‐3′; EGFR‐FW, 5′‐GGT GCG AAT GAC AGT AGC ATT ATG A‐3′, and EGFR‐RW, 5′‐AAA TGG GCT CCT AAC TAG CTG AAT C‐3′; MT1X‐FW, 5′‐TTG ATC GGG AAC TCC TGC TTC T‐3′, and MT1X‐RW, 5′‐ACA CTT GGC ACA GCC GAC A‐3′; GPX3‐FW, 5′‐ATG CCT ACA GGT ATG CGT GAT TG‐3′, and GPX3‐RW, 5′‐TGC AGG CAC ACA GAT GGT ACA‐3′; PDCD6‐FW, 5′‐TCA AGG CCA GAC TAG ATC AGC CTA A‐3′, and PDCD6‐RW, 5′‐GCT GGG ATG AGG CAC ATG AC‐3′; YWHAE‐FW, 5′‐GGC AGA ATT TGC CAC AGG AA‐3′, and YWHAE‐RW, 5′‐ACC TAA GCG AAT AGG ATG CGT TG‐3′; WDR33‐FW, 5′‐ATG CAT GGG CTC TGT CAG TTT C‐3′, and WDR33‐RW, 5′‐GGC TGA TAC CGG GAC AAC ACT AC‐3′; C14orf43‐FW, 5′‐CAG ACT GGC AAG CCT AAC TCC ATA‐3′, and C14orf43‐RW, 5′‐CAA GGC TGT TCC TGT GCT CTG‐3′; MYLIP‐FW, 5′‐ACG TCT ATC TGC CAA CGC ACA C‐3′, and MYLIP‐RW, 5′‐CAG TTC ATG GAA ACA TGC CAA GTC‐3′; GKAP1‐FW, 5′‐TTG CGA ATA AGT TTC GGA GCA TC‐3′, and GKAP1‐RW, 5′‐GCC ACT GCC ACT ATC CAC TTG TAA‐3′; GAPD‐FW, 5′‐GCA CCG TCA AGG CTG AGA AC‐3′, and GAPD‐RW, 5′‐ATG GTG GTG AAG ACG CCA GT‐3′.

Oligonucleotide microarray study.  The microarray procedure was performed according to the Affymetrix protocols (Santa Clara, CA). In brief, the total RNA extracted from the tumor samples was analyzed using an Agilent 2100 Bioanalyzer (Agilent Technologies, Waldbronn, Germany) for quality check, and cRNA was synthesized using the GeneChip 3′‐Amplification Reagents One‐Cycle cDNA Synthesis Kit (Affymetrix). The labeled cRNA were then purified and used for construction of the probes. Hybridization was performed using the Affymetrix GeneChip HG‐U133 Plus 2.0 array for 16 h at 45°C. The signal intensities were measured using a GeneChip Scanner3000 (Affymetrix) and converted to numerical data using the GeneChip Operating Software, ver. 1 (Affymetrix).

Statistical analysis.  The microarray analysis was performed using the BRB Array Tools software ver. 3.3.0 (http://linus.nci.nih.gov/BRB‐ArrayTools.html) developed by Dr Richard Simon and Dr Amy Peng. In brief, a log base 2 transformation was applied to the raw microarray data, and global normalization was used to calculate the median over the entire array. Genes were excluded if the percentage of data missing or filtered out exceeded 20%. Genes that passed the filtering criteria were then considered for further analysis. We computed a statistical significance level (P < 0.005) for each gene based on a univariate proportional hazards model.

To adjust the expression of 10 genes (DACH1, EGFR, MT1X, YWHAE, GPX3, PDCD6, WDR33, C14orf43, MYLIP and GKAP1) for clinical features (age, sex, performance status [PS], number of metastatic sites, received chemotherapy), clinical data and the normalized microarray expression data of the 10 genes were imported into SAS software ver. 9.1.3 (SAS Institute, Cary, NC, USA) and a Cox regression model was constructed for multivariate analysis against each of the variables. The study groups were divided into two groups based on each of the clinical features: age (<65 or ≥65 years), sex (male or female), PS (0 or ≥1), number of metastatic sites (<3 or ≥3), chemotherapy (5‐FU‐based or CPT11 + CDDP) and expression levels of 10 genes). P < 0.05 was considered significant.

Results

Identification of 98 candidate prognosis‐related genes using a microarray analysis.  The univariate analysis of clinical features including age (<65 or ≥65 years), sex, PS (0 or ≥1), number of metastatic sites (1, 2 or ≥3) and received chemotherapy (5‐FU‐based or CPT11 + CDDP) were performed for 40 microarray samples (Table 1). There were no significant differences between any of the two groups divided according to age, sex, number of metastatic sites or received chemotherapy; however, significant differences were noted between the two groups divided according to PS (P = 0.048).

Table 1.

Univariate analysis of clinical features

Variable	No. of patients	MST (days)	P‐value (log–rank test)
Age (years)
≥65	16	235	0.454
<65	24	250
Sex
Male	29	243	0.926
Female	11	267
PS
≥1	24	182	0.048
0	16	309
Metastasis
1, 2	10	137	0.102
≥3	30	261
Chemotherapy
5‐FU‐based	26	245	0.594
CPT11 + CDDP	14	240

MST, median survival time; PS, performance status.

To identify the candidate prognosis‐related genes from amongst over 47 000 transcripts, a microarray analysis was performed for a training set of 40 samples. A total of 21 308 genes passed the filtering criteria and were further analyzed. Ninety‐eight genes were significantly correlated with survival, according to a Cox proportional hazards model (P < 0.005) (Table 2). Fifty‐nine genes were protective genes (hazard ratio, <1), and 39 were risk genes (hazard ratio >1).

Table 2.

Prognosis‐related genes identified using microarray analysis

P‐value	Hazard ratio	Description	Gene	Probe set	Pass	PCR
0.0002	1.8	Epidermal growth factor receptor	EGFR	201984_s_at	2	PCR	1	0.1
0.0005	0.1	DEAD (Asp‐Glu‐Ala‐Asp) box polypeptide 54	DDX54	219111_s_at			2	0.1
0.0005	0.5	Chimerin (chimaerin) 2	CHN2	213385_at			3	0.1
0.0005	6.1	Ubiquitin‐like domain containing CTD phosphatase 1	UBLCP1	227413_at			4	0.2
0.0006	0.5	PTK2 protein tyrosine kinase 2	PTK2	241387_at			5	0.2
0.0008	3.4	Der1‐like domain family, member 2	DERL2	218333_at			6	0.2
0.0008	0.5	Leucine rich repeat containing 14	LRRC14	32062_at			7	0.2
0.0009	4.5	WD repeat domain 33	WDR33	222763_s_at		PCR	8	0.2
0.0009	0.1	Rhomboid domain containing 3	RHBDD3	217622_at			9	0.2
0.001	0.3	Myosin regulatory light chain interacting protein	MYLIP	228098_s_at	3	PCR	10	0.2
0.0013	4.7	Chromosome 14 open reading frame 43	C14orf43	225980_at		PCR	11	0.2
0.0013	0.2	BCL6 co‐repressor	BCOR	223915_at			12	0.2
0.0013	0.5	MAD1 mitotic arrest deficient‐like 1 (yeast)	MAD1L1	233921_s_at			13	0.2
0.0013	4.9	Chromosome 14 open reading frame 109	C14orf109	213246_at			14	0.2
0.0014	4.2	Hypothetical protein LOC124512	LOC124512	225808_at			15	0.2
0.0014	5.0	Ring finger protein 167	RNF167	212047_s_at			16	0.2
0.0014	0.6	Hypothetical LOC25845	LOC25845	225457_s_at			17	0.2
0.0014	4.2	General transcription factor II, i	GTF2I	232710_at			18	0.3
0.0014	0.2	Rho guanine nucleotide exchange factor (GEF) 10‐like	ARHGEF10L	1570511_at			19	0.3
0.0014	0.3	G kinase anchoring protein 1	GKAP1	229312_s_at		PCR	20	0.3
0.0015	1.9	Glutathione peroxidase 3 (plasma)	GPX3	214091_s_at	2	PCR	21	0.3
0.0016	0.5	Dachshund homolog 1 (Drosophila)	DACH1	1567101_at	2	PCR	22	0.3
0.0016	0.3	Diacylglycerol kinase, theta 110kDa	DGKQ	226605_at			23	0.3
0.0017	0.6	Hepatocellular carcinoma‐associated antigen 112	HCA112	218345_at			24	0.3
0.0018	3.5	Mediator of RNA polymerase II transcription, subunit 31 homolog	MED31	222867_s_at			25	0.3
0.0018	6.9	Tyrosine 3‐monooxygenase/tryptophan 5‐monooxygenase activation protein, epsilon polypeptide	YWHAE	210317_s_at		PCR	26	0.3
0.0018	0.1	KH domain containing, RNA binding, signal transduction associated 1	KHDRBS1	201488_x_at			27	0.3
0.0019	0.3	Solute carrier family 25 (mitochondrial carrier; Graves disease autoantigen), member 16	SLC25A16	210686_x_at			28	0.3
0.0019	4.9	Hypothetical protein LOC51255	LOC51255	223064_at			29	0.3
0.002	0.2	Cyclin L2 /// similar to Aurora kinase A‐interacting protein	CCNL2 /// LOC643556	222999_s_at			30	0.3
0.002	7.4	Lectin, mannose‐binding, 1	LMAN1	224629_at			31	0.3
0.002	0.2	Erythrocyte membrane protein band 4.1 like 4A	EPB41L4A	228259_s_at			32	0.3
0.0022	0.2	KIAA0999 protein	KIAA0999	204155_s_at			33	0.3
0.0022	0.5	ELOVL family member 7	ELOVL7	227180_at			34	0.3
0.0023	4.0	Churchill domain containing 1	CHURC1	233268_s_at			35	0.4
0.0024	4.0	Yippee‐like 2 (Drosophila)	YPEL2	227020_at			36	0.4
0.0024	5.9	Hermansky–Pudlak syndrome 1	HPS1	210112_at			37	0.4
0.0025	0.3	Hypothetical protein LOC285831	LOC285831	228857_at			38	0.4
0.0026	3.5	CDC37 cell division cycle 37 homolog (Saccharomyces cerevisiae)‐like 1	CDC37L1	219343_at			39	0.4
0.0026	2.1	Ankyrin repeat and SOCS box‐containing 9	ASB9	205673_s_at			40	0.4
0.0026	0.2	Hypothetical gene supported by AK125149	LOC401577	239247_at			41	0.5
0.0026	0.3	TBC1 domain family, member 23	TBC1D23	236755_at			42	0.5
0.0026	0.3	MRNA full length insert cDNA clone EUROIMAGE 2362292		235505_s_at			43	0.5
0.0026	0.4	Dehydrogenase/reductase (SDR family) member 8	DHRS8	217989_at			44	0.5
0.0026	0.4	Nuclear receptor coactivator 2	NCOA2	242369_x_at			45	0.5
0.0026	0.2	MRNA; cDNA DKFZp667E0114 (from clone DKFZp667E0114)		235660_at			46	0.5
0.0027	0.4	Transforming, acidic coiled‐coil containing protein 1	TACC1	242290_at			47	0.5
0.0027	0.2	POU domain, class 2, transcription factor 1	POU2F1	1562280_at			48	0.5
0.0027	2.9	p21(CDKN1A)‐activated kinase 6	PAK6	1555310_a_at				0.5
0.0027	0.5	Mannosyl (alpha‐1,3‐)‐glycoprotein β‐1,4‐N‐acetylglucosaminyltransferase, isozyme A	MGAT4A	226039_at			50	0.5
0.0027	5.1	Zinc finger CCCH‐type containing 14	ZC3H14	204216_s_at			51	0.5
0.0028	0.5	Acyl‐CoA synthetase short‐chain family member 2	ACSS2	235805_at			52	0.5
0.0028	0.3	Programmed cell death 6	PDCD6	222380_s_at		PCR	53	0.6
0.0029	3.8	ERGIC and golgi 2	ERGIC2	226422_at			54	0.6
0.0029	0.4	Erythrocyte membrane protein band 4.1 like 5	EPB41L5	225855_at			55	0.6
0.003	6.5	Chromosome 14 open reading frame 32	C14orf32	212644_s_at			56	0.6
0.0031	0.2	Transcribed locus		239437_at			57	1.8
0.0031	0.3	DOT1‐like, histone H3 methyltransferase (S. cerevisiae)	DOT1L	231297_at			58	1.9
0.0031	2.2	Transcription elongation factor A (SII)‐like 8	TCEAL8	224819_at			59	1.9
0.0031	0.3	Laminin, β 1	LAMB1	236437_at			60	2.0
0.0032	2.7	FK506 binding protein 5	FKBP5	224840_at			61	2.0
0.0033	0.5	Integrin, α 6	ITGA6	244665_at			62	2.1
0.0034	2.7	COMM domain containing 9	COMMD9	218072_at			63	2.2
0.0034	0.2	Eukaryotic translation initiation factor 4 γ, 3	EIF4G3	201936_s_at			64	2.3
0.0035	0.5	235616_at	235616_at	235616_at			65	2.6
0.0036	1.9	Metallothionein 1X	MT1X	204326_x_at		PCR	66	2.6
0.0036	2.7	Peroxiredoxin 5	PRDX5	1560587_s_at			67	2.7
0.0037	0.3	Core‐binding factor, runt domain, α subunit 2; translocated to, 2	CBFA2T2	207625_s_at			68	2.7
0.0037	0.4	Transcribed locus, moderately similar to XP_531878.2		230168_at			69	2.7
0.0038	0.3	Zinc finger protein 346	ZNF346	236267_at			70	2.8
0.0038	2.0	Metallothionein 1H‐like protein /// hypothetical protein LOC650610	LOC645745 /// LOC650610	211456_x_at			71	2.9
0.0039	0.2	Hypothetical protein DKFZp586I1420	DKFZp586I1420	213546_at			72	3.4
0.0039	2.0	Adrenergic, β‐2‐, receptor, surface	ADRB2	206170_at			73	3.5
0.0039	0.3	CTD‐binding SR‐like protein rA9	KIAA1542	234952_s_at			74	3.5
0.0039	2.6	Peroxiredoxin 5	PRDX5	222994_at			75	3.6
0.004	0.2	ATPase, H+ transporting, lysosomal 42kDa, V1 subunit C1	ATP6V1C1	226463_at			76	3.8
0.004	8.0	XK, Kell blood group complex subunit‐related family, member 8	XKR8	218753_at			77	3.8
0.004	0.3	Caspase 6, apoptosis‐related cystein peptidase	CASP6	242323_at			78	4.0
0.0041	0.4	Coagulation factor XII (Hageman factor)	F12	205774_at			79	4.0
0.0041	0.3	Centaurin, γ 2	CENTG2	240758_at			80	4.2
0.0042	0.6	LR8 protein	LR8	220532_s_at			81	4.2
0.0042	0.2	WD repeat domain 42A	WDR42A	243318_at			82	4.5
0.0042	2.6	Potassium channel tetramerisation domain containing 14	KCTD14	219545_at			83	4.7
0.0043	2.8	6‐Phosphogluconolactonase	PGLS	218388_at			84	4.9
0.0044	3.8	Bruno‐like 6, RNA binding protein (Drosophila)	BRUNOL6	227775_at			85	4.9
0.0044	2.3	Zinc finger protein 415	ZNF415	205514_at			86	5.0
0.0045	0.5	HIR histone cell cycle regulation defective homolog A (S. cerevisiae)	HIRA	240451_at			87	5.1
0.0046	0.5	Cardiolipin synthase 1	CRLS1	241741_at			88	5.9
0.0046	0.3	c‐mer proto‐oncogene tyrosine kinase	MERTK	233079_at			89	6.1
0.0047	0.2	Additional sex combs like 2 (Drosophila)	ASXL2	218659_at			90	6.5
0.0047	3.6	Platelet endothelial aggregation receptor 1	PEAR1	228618_at			91	6.9
0.0047	0.3	Core‐binding factor, runt domain, α subunit 2; translocated to, 2	CBFA2T2	238549_at			92	7.4
0.005	0.6	Lysosomal associated protein transmembrane 4 β	LAPTM4B	208029_s_at			93	8.0

Pass, number of overlapped probes; PCR, the genes that were subsequently examined using real‐time RT‐PCR.

A heat‐map of the expression values of the 98 selected genes comparing the unfavorable prognosis group (survival time, <180 days) and favorable prognosis group (survival time, ≥180 days) is shown in Fig. 1. Genes are plotted via hierarchical clustering.

Figure 1.

Heat map of expression values for microarray identifying 98 genes whose expressions were correlated with survival. The hierarchical clustering of the 98 genes comparing the unfavorable prognosis group (survival time, <180 days) and favorable prognosis group (survival time, ≥180 days) is shown. The blue or red colors of each block represent the normalized gene expression levels. Each row represents a sample, and each column represents a gene. The 10 genes included in the multivariate analysis (Table 3) are shown.

Multivariate analysis of prognosis‐related genes.  Of the 98 candidate genes, we prioritized those that: (i) were selected by overlapping probes; (ii) were novel genes; or (iii) had a lower P‐value according to a Cox proportional hazards model. We selected the following 10 genes of interest for real‐time RT‐PCR analysis: DACH1, EGFR, MT1X, YWHAE, GPX3, PDCD6, WDR33, C14orf43, MYLIP and GKAP1.

To adjust for relevant clinical covariates against these 10 genes, we performed a multivariate analysis (Table 3). The results of the multivariate analysis revealed that high DACH1 expression and high PDCD6 expression were significantly correlated with the favorable outcome (P = 0.0134 and P = 0.0015, respectively). We therefore considered that the DACH1 and PDCD6 expressions were independent prognostic markers from the results of the multivariate analysis. Results of microarray data and patient survival in the training set of 40 patients are shown in Fig. 2. The Kaplan–Meier method was used for DACH1 and PDCD6. The low PDCD6 and DACH1 expression groups had significantly poorer outcomes (P < 0.0001 and P = 0.0045).

Table 3.

Multivariate analysis of prognosis‐related genes

Variable	Hazard ratio	95% confidence interval	P‐value
Age (≥65)	1.78	0.570–5.559	0.3212
Sex (male)	3.26	0.732–14.489	0.1210
Performance status (≥1)	2.36	0.687–8.078	0.1728
Metastasis (≥3)	1.58	0.450–5.561	0.4739
Chemotherapy (5‐FU)	1.48	0.402–5.475	0.5541
DACH1	0.38	0.175–0.817	0.0134
EGFR	1.41	0.992–2.001	0.0553
MT1X	0.71	0.317–1.600	0.4111
YWHAE	1.91	0.401–9.061	0.4169
GPX3	1.62	0.869–3.007	0.1293
PDCD6	0.06	0.010–0.334	0.0015
WDR33	1.38	0.268–7.067	0.7017
C14orf43	0.64	0.122–3.407	0.6045
MYLIP	0.67	0.221–2.042	0.4826
GKAP1	2.31	0.751–7.106	0.1440

Cox regression model was performed for multivariate analysis against each of the variables.

Figure 2.

Results of microarray data and patient survival in the training set of 40 patients. The Kaplan–Meier method was used for DACH1 and PDCD6. The patients were divided into high and low expression groups by median values. The low PDCD6 and DACH1 expression groups had significantly poorer outcomes (P < 0.0001 and P = 0.0045). High‐Array, group with high expression levels as determined by signal intensity of microarray data. Low‐Array, group with low expression levels as determined by signal intensity of microarray data.

Validation using real‐time RT‐PCR in independent samples.  The mRNA expression levels of DACH1 and PDCD6 were quantified using real‐time RT‐PCR in 19 independent samples to validate the results of the microarray. While the expression levels of DACH1 were not correlated with survival, those of PDCD6 in independent samples were significantly correlated with the survival (P = 0.007) (Table 4). The Kaplan–Meier method was used to estimate the overall survival using the median value (Fig. 3a). All quantified expression levels of real time RT‐PCR data are shown as Fig. 3(b). The mRNA expressions of PDCD6 varied by approximately 25 fold (range, 0.98–25.1). The low PDCD6 expression groups had significantly poorer outcomes (P = 0.0018). We concluded that PDCD6 was a valuable gene for predicting the survival in patients with gastric cancer. These results indicate that our PCR‐based single gene prediction strategy using endoscopic biopsy samples could successfully predict the overall patient survival.

Table 4.

Results of real‐time RT‐PCR for PDCD6 and DACH1 in an independent valiation set

Genes	Hazard ratio	95% confidence limits		P‐value
		Upper	Lower
PDCD6 *	0.29	0.12	0.71	0.007
DACH1	0.79	0.56	1.13	0.199

^* , P < 0.05.

Figure 3.

Results of real‐time reverse transcription polymerase chain reaction (RT‐PCR) analysis and patient survival in the independent validation set of 19 samples. (a) The Kaplan–Meier method was used to estimate the overall survival. The low PDCD6 expression groups had significantly poorer outcomes (P = 0.0018). High‐PCR, group with high expression levels as determined by PCR. Low‐PCR, group with low expression levels as determined by PCR. (b) All quantified expression levels of real time RT‐PCR data are shown. The mRNA expressions of PDCD6 were significantly lower in unfavorable group (P = 0.003) and varied ~25 fold (range, 0.98–25.1). Favorable, the patients with survival time over 180 days. Unfavorable, the patients with a survival time less than 180 days.

Discussion

Several studies have identified prognostic biomarkers in cases of gastric cancer using microarray analysis. Hasegawa et al. identified 12 genes that were associated with lymph node metastasis.^{( 23 )} Hippo et al. identified several genes associated with lymph node metastasis, including Oct‐2, and genes associated with the histological type, including liver‐intestine cadherin.^{( 24 )} These studies introduced a novel direction in which microarray analysis could be used to predict postoperative recurrences. Inoue et al. selected 78 genes that were differentially expressed between aggressive and non‐aggressive cancers and constructed a prognostic scoring system.^{( 25 )} Leung et al. found that high CCL18 expression levels were associated with prolonged overall and disease‐free survival.^{( 26 )} They also found that phospholipase A2 group IIA expression in gastric adenocarcinoma was associated with prolonged survival and less frequent metastasis.^{( 27 )} Chen et al. demonstrated a survival prediction model consisting of three genes (CD36, SLAM, PIM‐1) that was capable of predicting poor or good survival in 23 (76.7%) of 30 newly enrolled patients.^{( 28 )} Most of these studies used surgical specimens to predict postsurgical survival and were conducted retrospectively. Thus, we think that our present prospective study is unique in that we used endoscopic biopsy samples to predict the survival time in patients with unresectable gastric cancer. In patients with unresectable cancer, endoscopic biopsy samples may be the most appropriate specimens available non‐invasively for microarray analysis. Although tumor heterogeneity may pose problems when biopsy samples are used as representative tissue specimens and further investigation is required, we believe that endoscopic biopsy samples should continue to be used for microarray analyses. Current clinical study has been confronted with a number of obstacles. Microarray analysis for clinical studies, in particular, has been hampered with bottlenecks such as RNA quality, the extremely large number of genes to be analyzed, an immature analytical tool or methodology and so on. There are two types of obstacles: controllable obstacles and uncontrollable ones. One uncontrollable obstacle is a complex chemotherapy regimen. It is easy to say that a clinical biomarker study should be performed in one particular regimen. Chemotherapy regimen has, however, progressed and become more sophisticated in a short range of time. This study was prospective clinical study and was largely followed by a guideline, Recommendations for Tumor Marker Prognostic Studies (REMARK). To minimize the uncontrollable factors, we aimed to avoid controllable factors with our best efforts. In this sense, we believe that the present study has succeeded in stratifying potential controllable variables.

Based on the results of the series of analyses conducted in the current study, we validated PDCD6 as a molecular biomarker of the prognosis in gastric cancer.

PDCD6, also known as ALG‐2 (apoptosis‐linked gene‐2), was first identified in a study on T‐cell apoptosis conducted by Vito et al.^{( 29 )} PDCD6 encodes a calcium‐binding protein that belongs to the penta‐EF‐hand protein family. The gene product participates in T‐cell receptor‐, Fas‐ and glucocorticoid‐induced programmed cell death and cell proliferation. The stimulation of cells to enter the cell cycle is thought to drive the cellular apoptotic program, and the presence of additional survival or pro‐apototic signals determines whether a cell proliferates or commits suicide. Krebs et al. indicated that the deregulation of such an obviously delicate balance could lead to pathological developments, such as cancer.^{( 30 )} Detailed biological function of PDCD6 genes in gastric cancer is still unclear. The speculated function may lead us to hypothesize that the expression is generally downregulated in cancer.

Our ultimate goal is to use real‐time RT‐PCR or immunohistochemical examination to identify patients with a poor prognosis prior to undertaking chemotherapy. We are now planning a large‐scale prospective study based on the evidence obtained in the current study.

In conclusion, we identified prognostic biomarkers in patients with unresected gastric cancer, and our PCR‐based single gene prediction strategy successfully predicted the overall survival of patients with gastric cancer. Our findings may provide a novel insight into the treatment of gastric cancer and may lead to a better understanding of this disease subgroup.