Gene expression signature for recurrence in stage III colorectal cancers

Abstract

BACKGROUND:

Colorectal cancer patients with lymph node metastases (stage III) show poorer prognosis than those without. Predicting development of recurrence may guide the need for intensive follow-up and/or adjuvant chemotherapy in such patients. The authors' objective was to identify a set of discriminating genes that could predict recurrence in stage III colorectal cancer.

METHODS:

Thirty-six stage III colorectal cancer patients were studied. Tumor samples were obtained from surgically resected specimens. Thirteen patients developed recurrence, whereas 23 patients did not. Gene expression profiles were determined using human HG-U133 Plus 2.0 Gene Chip (Affymetrix, Santa Clara, Calif).

RESULTS:

The authors identified 45 discriminating genes between patients with and without recurrence. By using this gene set, they established a new model to predict recurrence with an accuracy of 90.9%. The discriminating genes included calcineurin-binding protein 1 (CABIN1), whose expression differed remarkably between patients with and without recurrence (P = .0073). The authors further examined the DNA copy number of CABIN1 and were able to show a significant relation with recurrence (P < .012). Patients having CABIN1 gene loss demonstrated a higher risk of recurrence (odds ratio, 18.8). DNA copy number of CABIN1 alone could predict recurrence with an accuracy of 80.0%.

CONCLUSIONS:

The results of the current study demonstrated that gene expression profiling is useful in predicting recurrence in stage III colorectal cancer. The authors identified CABIN1 among discriminating genes that may play a key role in the development of recurrence. These results may help to establish an individualized therapy for stage III colorectal cancer. Cancer 2009. © 2009 American Cancer Society.

Colorectal cancer remains 1 of the leading causes of cancer mortality. The presence of lymph node metastases is considered an important factor in determining the outcome of colorectal cancer. Patients with lymph node metastases (stage III patients) show poorer prognosis as compared with those without. To improve the postoperative outcome of stage III colorectal cancer, we have previously demonstrated that molecular markers could select patients who would benefit from adjuvant chemotherapy.1 However, even after chemotherapy, some patients still develop recurrence, which may determine the outcome of the patients. Therefore, if we can predict which patients will develop recurrence, it may provide us with an indication for intensive follow-up and/or adjuvant chemotherapy in such patients.

Advances in expression genomics by DNA microarray have made it possible to analyze tens of thousands of genes at a time and have demonstrated that the expression profiles of cancer cells may be used for predicting the response to drugs, as well as their toxicity and adverse effects.2, 3 Also, we have recently reported that gene expression profiles could predict response to radiotherapy, development of colorectal cancer, and/or dysplasia in ulcerative colitis or classify molecularly different phenotypes of microsatellite instability colorectal cancers.4–6 In the present study, applying the same method, we examined the gene expression profiles of stage III colorectal cancers and were able to build a predictive model of postoperative recurrence with a high accuracy rate. Among discriminating genes between patients with and without recurrence, we identified calcineurin-binding protein 1 (CABIN1) as showing significantly lower expression among patients with recurrence. A previous study showed that alteration of the DNA copy number in a region near CABIN1 was associated with numerous cancer types.7 Therefore, we further examined the DNA copy number of CABIN1 and could show a significant relation between copy number and recurrence. To the best of our knowledge, CABIN1 has not been reported in the context of recurrence in malignancy. Our objective here was to identify a set of discriminating genes that could predict recurrence in stage III colorectal cancer, and furthermore, to demonstrate the significance of CABIN1 expression in relation to recurrence.

MATERIALS AND METHODS

Patient Samples

Informed consent was obtained from colorectal cancer patients for the collection of specimens, and the study protocol was approved by the local ethics committee. Thirty-six colorectal cancer patients who had undergone curative surgical resection of colorectal cancer were studied. No patients had any treatment before surgery or distant metastasis at the time of surgery, and in all patients lymph node metastasis was detected by pathologic examination. Tumor samples were taken from surgically resected specimens, snap-frozen in liquid nitrogen, and stored at −80°C until use. Paralleled tumor specimens were formalin fixed and paraffin embedded for histological examination. Samples were used for RNA extraction when microscopic examination verified that specimens contained at least 70% tumor cells, as described previously.5

RNA Isolation and Microarray Procedures

Total RNA was isolated from each of the frozen samples using RNeasy Mini Kit (QIAGEN, Chatswort, Calif) for gene expression analysis, as described previously.5 Gene expression profiles were determined using Affymetrix HG-U133 Plus2.0 GeneChips (Affymetrix, Santa Clara, Calif) according to the manufacturer's recommendations.

Microarray Analysis

Class prediction

Expression analysis was performed on GeneSpring GX software version 7.3 (Agilent Technologies, Santa Clara, Calif). Gene expression data, when classified as either flag-P or flag-M in >30% of all samples, were loaded into the software, and then normalized in 2 ways: “per chip normalization” and “per gene normalization.” For per chip normalization, all expression data on a chip were normalized to the 50th percentile of all values on that chip. For per gene normalization, the data for a given gene were normalized to the median expression level of that gene across all samples. To identify discriminating genes markers, the expression profiles were compared between patients with and without recurrence, using the Student t test for unpaired data (with Welch correction for unequal variances, controlling the false discovery rate using Benjamini–Hochberg multiple comparisons procedure) and fold-change. Two-dimensional hierarchical clustering was then applied to the log-transformed data with average-linkage clustering with standard correlation as the similarity metric for the discriminating genes that we identified as differentially expressed between patients with and without recurrence. Variation of multigene expression between patients with and without recurrence was also compared by principal component analysis (PCA). We then performed supervised class prediction using the k-nearest-neighbor method and a 9-fold cross-validation with the discriminating genes.8 The significance of classification accuracy was calculated by the following procedure. At first, the average classification accuracy was measured by a 100× repeated 9-fold cross-validation using a 45-gene predictor. Then the class labels of the samples were permutated 4000 times, obtaining a new signature and calculating the 9-fold cross-validation accuracy for each permutated data set. Finally we measured the random chance of getting a signature that has higher accuracy than the 45-gene signature. The above data analysis was conducted in the R computing environment. We performed survival analysis on the basis of class predictive results. Rates of recurrence-free survival were estimated by the Kaplan-Meier method and compared by the log-rank test.

Gene functional category analysis

Gene ontology categories were analyzed by the BioScript Library tool on GeneSpring GX 7.3. Genes were classified according to their annotated role in biological processes, molecular function, and cellular components from Gene Ontology (The Gene Ontology Consortium). This program identifies genes belonging to different Gene Ontology categories and also calculates the statistical significance of nonrandom representation, that is, overlapping P value. By using hypergeometric probability, the overlapping P value was calculated by

• 

in which the probability of overlap corresponding to k or more genes between a gene list of n genes is compared against a gene list of m genes when randomly sampled from a universe of u genes.

Genomic DNA Preparation and Quantitative Real-time Polymerase Chain Reaction for CABIN1 Gene Copy Number Determination

Genomic DNA was extracted from samples using the QIAamp DNA mini kit (Qiagen Inc., Valencia, Calif), according to the manufacturer's instructions. Quantitative real-time polymerase chain reaction (PCR) was performed on a PRISM 7300 sequence detector (Applied Biosystems, Foster City, Calif) by using a QuantiTect SYBR Green kit (Qiagen, Inc., Valencia, Calif). We quantified the DNA of each sample by comparing the target locus to the reference Line-1, a repetitive element for which copy numbers per haploid genome are similar among all of the human normal and neoplastic cells. The quantification is based on standard curves from a serial dilution of human genomic DNA. The relative target copy number level was also normalized to normal human genomic DNA as calibrator. Copy number change of the target gene relative to the Line-1 and the calibrator were determined by using the formula (T_target/T_Line-1)/(C_target/C_Line-1), where T_target and T_Line-1 are quantities from tumor DNA by using target and Line-1, and C_target and C_Line-1 are quantities from calibrator by using target and Line-1.9, 10 PCR for each primer set was performed in at least triplicate, and the mean values and standard deviation were reported. Conditions for quantitative PCR reaction were as follows, 1 cycle at 94°C for 15 minutes, 45 cycles at 94°C for 20 seconds, 56°C for 20 seconds, and 70°C for 30 seconds. At the end of the PCR reaction, samples were subjected to a melting analysis to confirm specificity of the amplicon. Primer Express software (version 2.0; Applied Biosystems) was used to design the primers to span a 200-base pair (bp) nonrepetitive region, which were then synthesized by Operon Biotechnologies Inc. (Tokyo, Japan). The primer set was subsequently compared with the human genome using the basic local alignment search tool algorithm to determine its uniqueness. Primer sequences for CABIN1 used in this study are described below.

Forward primer was 5′-GGGAGCACGCCTTGTTGTCA-3′ and reverse primer was 5′-CCCCCAGCTTCCTCACGTAA-3′ The forward primer was designed to anneal to the 5′-upstream region from the transcription start site to perfectly eliminate the hybrid formation by the forward primer with the first strand reversely transcribed from retained CABIN1 mRNA with reverse primer.

Western Blot Analysis for CABIN1 Protein Expression

We further evaluated CABIN1 protein expression by Western blot analysis. Total protein from fresh frozen tumor tissue was extracted using a protein extraction reagent (T-PER, Pierce Biotechnology, Rockford, Ill) supplemented with protease inhibitors (Halt Protease Inhibitor Cocktail kit, Pierce Biotechnology). Tissue lysates (20 μg protein/sample) were loaded into 15% sodium dodecyl sulfate polyacrylamide gels. After electrophoresis, the separated proteins were electrotransblotted onto polyvinylidene difluoride membranes (Immobilon-P membrane, Millipore, Billerica, Mass). After blocking, membranes were probed with antihuman CABIN1 polyclonal goat antibody (Santa Cruz Biotechnology, Santa Cruz, Calif). The proteins were observed using horseradish peroxidase–conjugated antibodies (Pierce) followed by enhanced chemiluminescence (Pierce). The intensity of luminescence was quantified using an image analysis system (LAS-3000, Fuji Film, Tokyo, Japan).

RESULTS

Gene Expression Profiling: Class Comparison Between Patients With and Without Disease Recurrence

Gene expression profiling was established using DNA array. The patients were observed with a median follow-up period of 3.2 years (range, 0.9-7.4 years). Among 36 patients, 13 developed disease recurrence, whereas 23 did not. Five patients developed both lung and liver metastases, 2 patients developed both lung and local recurrence, 2 patients developed local recurrence alone, 1 patient developed both lung recurrence and peritoneal dissemination, 1 patient developed peritoneal dissemination alone, 1 developed patient liver metastasis alone, and 1 patient developed lymph node recurrence alone. There was no significant difference between patients with and without disease recurrence in clinicopathologic factors such as sex, age, and tumor location. By using class-comparison analysis, we identified a list of 45 genes that were differentially expressed at significant levels (P <.05) between patients with and without recurrence (Table 1). Twenty-one genes demonstrated higher and 24 genes lower expression in patients with recurrence as compared with those without. By using 45 discriminating genes, we performed a hierarchical cluster analysis (Fig. 1). Patients with and without recurrence were clustered into 2 distinct groups. We also used 45 discriminating genes to generate a 3-dimensional (from 45-dimensional) plot of the data (Fig. 2). The 3 axes are the first 3 principal components fitted to the patient's molecular profile data. PCA-based multidimensional scaling visualization separated patients with lymph node recurrence and those without into a linearly separable gene expression data space.

Table 1. List of 45 Genes That Discriminate Between Patients With and Patients Without Disease Recurrence

Probe ID	Fold Change (Recurrence Negative/ Recurrence Positive)	P	Gene Symbol	GenBank	Gene Name
TNF indicates tumor necrosis factor; kD, kilodaltons.
230760_at	5.77	.0496	ZFY	BF592062	Zinc finger protein, Y-linked
1556381_at	3.903	.0277	TBDN100	AK091308	Transcriptional coactivator tubedown-100
205001_s_at	3.8	.0496	DDX3Y	AF000985	DEAD (Asp-Glu-Ala-Asp) box polypeptide 3, Y-linked
234082_at	3.457	.0477	ZNF236	AU146017	Zinc finger protein 236
1560469_at	3.138	.0496	NR5A2	BC027893	Nuclear receptor subfamily 5, group A, member 2
241970_at	2.67	.0277		C14898	Homo sapiens transcribed sequences
220048_at	2.557	.0277	EDAR	NM_022336	Ectodysplasin 1, anhidrotic receptor
235535_x_at	2.528	.0443		AI369073	qw29g11.x1 NCI_CGAP_Ut4 H. sapiens cDNA clone IMAGE:1992548 3′, mRNA sequence.
202554_s_at	2.513	.0304	GSTM3	AL527430	Glutathione S-transferase M3 (brain)
216628_at	2.347	.035		AL117447	H. sapiens mRNA; cDNA DKFZp586A0617 (from clone DKFZp586A0617)
209750_at	2.323	.0496	NR1D2	N32859	Nuclear receptor subfamily 1, group D, member 2
1557581_x_at	2.302	.0499	CABIN1	BC027347	Calcineurin binding protein 1
206336_at	2.251	.0304	CXCL6	NM_002993	Chemokine (C-X-C motif) ligand 6 (granulocyte chemotactic protein 2)
205032_at	2.244	.0438	ITGA2	NM_002203	Integrin, alpha-2 (CD49B, alpha-2 subunit of VLA-2 receptor)
236882_at	2.216	.0227	PON2	AA922154	Paraoxonase 2
229765_at	2.204	.0496		AW511239	H. sapiens transcribed sequences
240044_x_at	2.199	.0227		AI864078	H. sapiens cDNA clone IMAGE:6195280, partial cds
238909_at	2.198	.0277	S100A10	BF126155	S-100 calcium binding protein A10 (annexin II ligand, calpactin I, light polypeptide [p11])
230733_at	2.175	.0227		H98113	H. sapiens transcribed sequences
1552516_a_at	2.148	.0277	HIPK1	NM_152696	Homeodomain interacting protein kinase 1
216607_s_at	2.14	.0304	CYP51P2	U40053	Human lanosterol 14-alpha demethylase (CYP51P2) processed pseudogene, complete cds
242473_at	2.097	.0413	TRAF4	BF000155	TNF receptor-associated factor 4
242420_at	2.07	.0496		AI084326	H. sapiens cDNA FLJ37573 fis, clone BRCOC2002835
223170_at	2.014	.0477	DKFZP564K1964	AF132000	DKFZP564K1964 protein
219870_at	0.493	.0438	ATF7IP2	NM_024997	Activating transcription factor 7 interacting protein 2
1559882_at	0.492	.0496	SAMHD1	AF147427	SAM domain and HD domain 1
242517_at	0.488	.0496	GPR54	AI819198	G protein-coupled receptor 54
239646_at	0.487	.0277		BF003148	H. sapiens transcribed sequence with weak similarity to protein ref:NP_060265.1 (H. sapiens) hypothetical protein FLJ20378 (H. sapiens)
223878_at	0.476	.0329	INPP4B	BC005273	Inositol polyphosphate-4-phosphatase, type II, 105 kD
225696_at	0.476	.0227	COPS7B	AK024273	COP9 constitutive photomorphogenic homolog subunit 7B (Arabidopsis)
AFFX-r2-Bs-thr-5_s_at	0.475	.0334		AFFX-r2-Bs-thr-5	Bacillus subtilis/GEN=thrC /DB_XREF=gb:X04603.1/NOTE=SIF corresponding to nucleotides 290-888 of gb:X04603.1/DEF=B. subtilis thrB and thrC genes for homoserine kinase and threonine synthase.
1563792_at	0.472	.0496	AMN	AK092824	Amnionless homolog (mouse)
235414_at	0.463	.0496	FLJ35863	BF432571	Hypothetical protein FLJ35863
228737_at	0.447	.0496	C20orf100	AA211909	Chromosome 20 open reading frame 100
244297_at	0.438	.0496	FLJ35740	AI806762	FLJ35740 protein
242723_at	0.435	.0443		AI001880	H. sapiens transcribed sequences
210128_s_at	0.36	.0257	LTB4R	U41070	Leukotriene B4 receptor
1555894_s_at	0.354	.0277	LOC92154	AA829283	Hypothetical protein BC002770
241416_at	0.323	.0189	REPIN1	BE672607	Replication initiator 1
204425_at	0.313	.0227	ARHGAP4	NM_001666	Rho GTPase activating protein 4
224589_at	0.11	.0438	LOC139202	BF223193	H. sapiens cDNA: FLJ21545 fis, clone COL06195
221728_x_at	0.0788	.0438	LOC139202	AA628440	H. sapiens cDNA: FLJ21545 fis, clone COL06195
214218_s_at	0.0676	.0277	LOC139202	AV699347	H. sapiens cDNA: FLJ21545 fis, clone COL06195
224590_at	0.0497	.0438	LOC139202	BE644917	H. sapiens cDNA: FLJ21545 fis, clone COL06195
224588_at	0.0354	.0498	LOC139202	AA167449	H. sapiens cDNA: FLJ21545 fis, clone COL06195

Figure 1. Supervised clustering of colorectal cancer and 45 genes is shown. Two-way hierarchical clustering was used to order samples (columns) and array targets (rows). Red indicates overexpression; green, underexpression. At the bottom of the figure, yellow indicates patients with disease recurrence, and red indicates those without. Samples were correctly classified except for 4 cases with recurrence and 3 without.

Figure 2. Discriminating genes were used to generate a 3-dimensional (from 45-dimensional) plot of the data. The 3 axes are the first 3 principal components fitted to the patient's molecular profile data. The cumulative proportion of the variance captured by each principal component axis is: (x) principal component axis 54.14%; (y) principal component axis 11.98%; and (z) principal component axis 5.685%. Principal component analysis–based multidimensional scaling visualization separated patients with disease recurrence (red) and those without (yellow) into a linearly separable gene expression data space.

Gene Functional Category Analysis

To investigate the biologic functions involved in discriminating genes, we performed Gene Ontology category analysis. When selected discriminating genes were compared with all genes whose expression profile could be evaluated, the categories signal transducer activity (9 genes, P = .00293), receptor activity (9 genes, P = .00293), receptor binding (6 genes, P = .0024), cell communication (11 genes, P = .0365), and signal transduction (9 genes, P = .00724) were found to demonstrate a significantly higher proportion among the 45 selected genes.

Gene Expression Profiling: Class Prediction of Patients With and Without Disease Recurrence

We next examined whether the expression profiling is useful in predicting recurrence among stage III patients. By using all samples, we performed supervised class prediction using the k-nearest-neighbor method and a 9-fold cross-validation with the 45 discriminating genes. After repeating the procedure 100 times, the average prediction accuracy was determined to be 90.9%.

The average sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of the present model were 91.6%, 82.6%, 91.2%, and 83.3%, respectively. The predictive performance of the model was evaluated by a permutation-based procedure. A false-finding rate was 0.0015; ie, 6 among 4000 permutations exhibited higher prediction accuracy than it did in the actual data. We performed recurrence-free survival analysis depending on the 9-fold cross-validation results. Among 36 patients, 12 patients were classified as recurrence positive and the remaining 24 patients as recurrence negative. Between the 2 groups, we examined the recurrence-free survival rate. There was a significant difference in survival rate between the 2 groups (P < .001) (Fig. 3).

Figure 3. Recurrence-free survival analysis is based on the predictive results. In the predictive model, 12 patients were classified as recurrence (Rec.) positive (+), whereas the remaining 24 patients were recurrence negative (−). Between the 2 groups, there was a significant difference in recurrence-free survival rate (P < .001).

Determination of CABIN1 Gene Copy Number

Among 45 discriminating genes, the percentage of the “Present-call” flag value of CABIN1 genes was drastically low in patients with recurrence (84.6% [11 of 13] vs 60.9% [14 of 23]; P = .0073). To examine whether genetic loss in CABIN1 locus may lead to a depression in mRNA expression of this gene, we designed the specific PCR primers for genomic sequence for CABIN1 genes and determined the gene copy number using a real-time PCR–based technique. There was a strong correlation between the copy number and the level of CABIN1 mRNA (corresponding P value, P = .014; Pearson coefficient, r = .47) (Fig. 4). The average copy number and standard deviation of CABIN1 genes determined by PCR in non-neoplastic colorectal tissue samples were 1.93 and 0.34, respectively. Patients with a copy number below the average value minus twice the value of the standard deviation were defined as “Loss” in CABIN1 gene, and others as “No Loss.” We could determine the copy number of CABIN1 in 25 patients, 6 patients with Loss and 19 patients with No Loss.

Figure 4. The relation is shown between CABIN1 DNA copy number and mRNA expression level. CABIN1 DNA copy number was determined by a real-time polymerase chain reaction–based technique. There was a strong correlation noted between the copy number and the level of CABIN1 mRNA by microarray (corresponding P value of .014 and Pearson coefficient of 0.47).

CABIN1 Gene Copy Number and Disease Recurrence

We next examined Loss status of CABIN1 gene with respect to recurrence. Fisher exact test revealed a significant relation between the Loss status and disease recurrence. Namely, the percentage of Loss was significantly higher in patients with recurrence (55.6% [5 of 9] vs 6.3% [1 of 16]; P = .012) (Table 2). Patients with CABIN1 gene loss showed a 18.8× higher risk ratio in recurrence than those without (odds ratio, 18.8). When Loss status is used as a marker of recurrence, the sensitivity, specificity, PPV, and NPV were 55.6%, 93.8%, 83.3%, and 78.9%, respectively. There were 4 misclassifications in patients with recurrence, and 1 in those without. The accuracy of prediction was 80.0%.

Table 2. Loss Status of CABIN1 Copy Number and Disease Recurrence

Case No.	CABIN1 Copy Number		Recurrence	GeneChip		Loss Status of CABIN1 Copy Number
	Mean Value	SD		mRNA Level	Flag Call
SD indicates standard deviation.
1	1.71	0.34	No	0.863	Absent	No loss
2	1.01	0.26	Yes	1.497	Absent	Loss
3	1.83	0.38	No	1.352	Present	No loss
4	2.54	0.98	Yes	0.888	Absent	No loss
5	2.10	0.45	No	0.753	Present	No loss
6	1.09	0.16	Yes	0.927	Absent	Loss
7	1.81	0.16	Yes	0.574	Absent	No loss
8	2.02	0.23	Yes	0.445	Absent	No loss
9	1.01	0.05	No	1.244	Present	Loss
10	1.81	0.41	No	1.028	Present	No loss
11	0.89	0.07	Yes	0.186	Absent	Loss
12	1.55	0.47	Yes	0.736	Absent	No loss
13	2.35	0.16	No	1.205	Absent	No loss
14	1.71	0.19	No	0.985	Absent	No loss
15	2.34	0.44	No	1.842	Present	No loss
16	1.65	0.52	No	2.248	Present	No loss
17	1.54	0.31	No	0.91	Absent	No loss
18	2.22	0.51	No	1.334	Absent	No loss
19	1.04	0.09	Yes	0.215	Absent	Loss
20	1.64	0.30	No	0.725	Absent	No loss
21	0.82	0.21	Yes	0.596	Absent	Loss
22	1.82	0.64	No	3.012	Present	No loss
23	1.34	0.40	No	1.669	Present	No loss
24	3.58	0.32	No	3.54	Present	No loss
25	2.48	0.54	No	0.842	Absent	No loss

Western Blot Analysis for CABIN1 Protein Expression

We performed western blot analysis with anti-CABIN1 polyclonal antibody to examine CABIN1 protein expression level in respect to recurrence. In 10 patients, 4 patients with recurrence and 6 without recurrence, samples were available for western blotting analysis. As shown in Figure 5, the protein expression of CABIN1 was relatively low in patients with recurrence compared with that in patients without recurrence. Densitometric analysis indicated a significant difference between patients with and without recurrence (P = .008).

Figure 5. Protein expression levels of CABIN1 are shown by Western blot analysis. β-actin served as an internal control. The protein expression of CABIN1 was relatively low in patients with disease recurrence compared with that in patients without disease recurrence. Densitometric analysis indicated a significant difference between patients with and without disease recurrence (P = .008).

DISCUSSION

By using global gene expression profiling, we established a new model to predict the development of recurrence in stage III colorectal cancer. Patients who developed recurrence could be predicted with an accuracy rate of 90.9%. The present study examined a larger number of patients and the predictive accuracy was higher in comparison with a previous report.11 Prediction of recurrence is important in determining the treatment modality in colorectal cancer patients after surgery. Patients at high risk for developing recurrence may need intensive follow-up as well as adjuvant chemotherapy. To our knowledge to date, several markers, such as TP53 or K-ras mutational, have been reported as possible markers that identify patients who would develop recurrence after surgery.12–14 In addition, we have demonstrated that loss of heterozygosity in chromosome 18q could be a predictive marker for prognosis in stage III colorectal cancer.1 However, because the accuracy of these markers is limited, they have not become routinely used in the clinical setting. In the present study, however, we could establish a new model, which can predict recurrence with a high accuracy rate, and this may be used for individualized treatment in stage III colorectal cancer.

By comparing gene expression profiling, we identified 45 genes whose expression differed significantly between patients with and without recurrence. Among the 45 genes, 21 genes demonstrated higher and 24 genes lower expression in patients with recurrence. Gene Ontology category analysis for discriminating genes identified nonrandom enrichment of a variety of biologic process categories, including signal transducer activity (P = .00293), receptor activity (P = .00293), receptor binding (P = .0024), cell communication (P = .0365), and signal transduction (P = .00724) transcription. This suggested that these categories might be important in the development of recurrence in stage III colorectal cancer.

Among the 45 discriminating genes, we further focused on the expression of CABIN1, which showed a remarkable difference between patients with and without recurrence. Alteration in DNA copy number is known to modify gene expression and function, and gene amplification is closely associated with its overexpression.15, 16 The present study also showed a close correlation between DNA copy number and expression level of CABIN1 by microarray. We further evaluated CABIN1 protein expression by western blotting and could show that the protein expression of CABIN1 was low in patients with recurrence compared with that in patients without recurrence. A previous study demonstrated that alteration of DNA copy number in the region near CABIN1 was associated with various cancers.17, 18CABIN1 gene is located immediately downstream of the GSTT1, and increased expression of GSTT1 is known to prevent some cancers.7, 16 Conversely, GSTT1 gene coding is polymorphic, and null genotype shows significantly poorer differentiation of colorectal cancer.19 In the present study as well, GSTT1 showed lower expression in patients with recurrence. However, this did not reach statistical significance (3.35-fold change; P = .0573), and there was no significant correlation between GSTT1 gene copy number and tumor recurrence (data not shown). As for CABIN1, patients with recurrence showed significantly lower expression than those without (2.30-fold change; P = .0499), and we could confirm the same difference by an analysis of DNA copy number of CABIN1. Patients with recurrence showed significantly higher frequency of CABIN1 loss (P = .011; odds ratio, 18.8). By using the DNA copy number of CABIN1 alone, we could predict the recurrence status with an accuracy of 80.0%. Normal function of the CABIN1 protein is involved in protein-protein interactions with histone deacetylases, and CABIN1 plays a role in the maintenance of normal chromatin structure.20 Chromatin modification regulates gene transcription and is an important epigenetic mechanism that is altered in cancer cells. Patients with disease recurrence demonstrated significantly lower expression and less DNA copy number of CABIN1 than those without. Taken together, these results suggested that CABIN1 may play a key role in the development of recurrence in stage III colorectal cancer.

The results of the current study demonstrated that gene expression profiling could predict the development of recurrence in stage III colorectal cancer with a high accuracy rate (90.9%). Furthermore, we identified CABIN1 among the discriminating genes, which demonstrated significantly lower expression and decreased DNA copy number in patients with recurrence. CABIN1 alone could predict recurrence with an accuracy of 80.0%. These results may help to establish an individualized tailored therapy for stage III colorectal cancer. Furthermore, global expression profiles discriminating genes including CABIN1 may provide insights into the development of novel therapeutic targets.

Acknowledgements

We thank Eriko Hashimoto and Riyo Kakimoto for technical and secretarial support.

Conflict of Interest Disclosures

This study was supported by a Grant-in-Aid for Scientific Research from the Ministry of Education, Culture, Sports, Science, and Technology of Japan and a grant from the Ministry of Health, Labor and Welfare of Japan.