Journal of Clinical & Experimental OncologyISSN: 2324-9110

All submissions of the EM system will be redirected to Online Manuscript Submission System. Authors are requested to submit articles directly to Online Manuscript Submission System of respective journal.

Research Article, J Clin Exp Oncol Vol: 7 Issue: 2

Prognostic Model of Stage II/III Colon Cancer Constructed using Gene Expression Subtypes and KRAS Mutation Status

Kengo Gotoh1#, Eiji Shinto2#, Yuichiro Yoshida1*, Hideki Ueno2, Yoshiki Kajiwara2, Masato Yamadera2, Ken Nagata2, Hitoshi Tsuda3, Junji Yamamoto2 and Kazuo Hase2

1Central Research Laboratories, Sysmex Corporation, Kobe, Japan

2Surgery, National Defense Medical College, Saitama, Japan

3Pathology, National Defense Medical College, Saitama, Japan

#These authors contributed equally to this work

*Corresponding Author : Yuichiro Yoshida
Central Research Laboratories, Sysmex Corporation, 4-4-4 Takatsukadai, Nishi-ku, Kobe 651-2271, Japan
Tel: 81-78-991-2367

Received: March 06, 2018 Accepted: March 19, 2018 Published: March 26, 2018

Citation: Gotoh K, Shinto E, Yoshida Y, Ueno H, Kajiwara Y, et al. (2018) Prognostic Model of Stage II/III Colon Cancer Constructed using Gene Expression Subtypes and KRAS Mutation Status. J Clin Exp Oncol 7:2. doi: 10.4172/2324-9110.1000214


Objectives: Cancer subtypes classified according to DNA microarray data predict prognosis with high accuracy. Here we constructed a new colon cancer (CC) subtype classification based on only of genes with known biological functions with the aim of establishing a new prognostic model for clinical use.

Methods: We performed an expression correlation analysis using data for 73 primary CC cases in the public dataset (learning set), focusing on genes located on the long arms of chromosomes 18 and 20 and stromal-related genes. We determined the representation of each gene in the modules with closely correlated expression levels in the same module. Mutations in KRAS, BRAF and TP53 were assessed using direct sequencing. Microsatellite instability (MSI) was analyzed using the Bethesda reference panel.

Results: We constructed a discriminant model with a view to classifying CC into three subtypes (“stromal”, “chromosomal instability [CIN]-like”, “MSI-like”) based on the expression levels of 55 genes of the Learning set. When we applied this predictor to microarray data from other patients with stage II/III colon cancer (n=258, test set), we discovered a significant difference in diseasefree survival between the stromal subtype and the other subtypes (p=1.25e-03). Accordingly, we created an integrated prognostic model for classifying the patients into high- and low-risk groups according to the expression levels of the 55 genes and KRAS mutations (p=1.56e-06). Analysis of independent specimens from patients with stage II/III colon cancer who underwent radical resection (n=59, validation set) confirmed the prognostic value of our model (p=4.75e-02).

Conclusion: The model produced a biologically discriminatory classifier that associated MSI status with the risk of recurrence that may be clinically applicable to the selection of patients with Stage II/ III CC for adjuvant therapy.

Keywords: Colon cancer; Prognosis; DNA microarray; Chromosomal instability; Microsatellite instability; Stroma


Colorectal cancer (CRC) is the second and third most common cancer of women and men, with nearly 1.4 million new cases diagnosed worldwide in 2012 [1]. In the same year, there were 693,933 CRC-related deaths (8.5% of all cancer deaths) [1]. The postoperative recurrence rate of Stage II CRC is 10%-20% [2]. Postoperative adjuvant chemotherapy is recommended for patients with Stage II CRC with high-risk pathological factors [3]. Nevertheless, risk factors have not been determined with precision, and this recommendation is not justified through randomized control trials. Large-scale studies of patients with stage III CRC show that postoperative adjuvant chemotherapy effectively prolongs survival [4]. However, prognosis is better for patients with Stage IIIA compared with Stage IIB/IIC disease [2]. Therefore, precise prediction of prognosis is urgently required for patients with Stage III CRC as well. CRC subtypes classified according to DNA microarray data predict prognosis with high accuracy [5-9]. However, there is little consistency among reports regarding the genes selected for subtype classification. This diversity is explained by the employment of different methodologies. For example, some researchers determine the expression levels of genes with known functions (candidate-gene approach), while others select genes through unbiased microarray analysis (data-driven approach) [10]. The former strategy risks missing potentially important but poorly annotated genes, while the latter is associated with a serious risk of over fitting to the training set.

Most colon cancers (CCs) can be separated into at least two types, the Chromosomal instability (CIN) type or the Microsatellite instability (MSI) type [11]. The prognosis of MSI-type patients with CC is more favorable compared with that of those with CIN-type [12-15]. Chromosomal abnormalities in CC frequently occur on the long arms of chromosomes 18 and 20 [16,17]. Therefore, we hypothesized that we could distinguish CIN-type from MSI-type by analyzing the expression levels of genes located in these chromosomal regions. Otherwise, the rate of the MSI-type would be higher than the rates of the other subtypes. Moreover, convincing evidence reveals an association between the number of infiltrating stroma cells in tumor tissue and the prognosis of CC [5,18-20]. We therefore reasoned that the analysis of expression levels of stroma-related genes might aid in the definition of cancer subtypes.

Here we developed a new model of classifying CC subtypes using a methodology that combines candidate-gene and data-driven approaches to overcome the disadvantages of each. The method involves analysis of the unique combination of genes (referred to as modules) located on 18q and 20q as well as those expressed by the tumor stroma. First, we extracted these gene modules through correlating gene expression levels and selected representative genes from each module. Using a 55 gene-expression signature, we subtyped CCs into CIN-like, MSI-like, and stromal subtypes and assessed their prognostic significance. Moreover, we evaluated the relationship between common gene mutations in CC and clinical outcomes of patients with each CC subtype. This allowed us to establish a new prognostic model based on the assessment of a novel combination of the three gene modules and KRAS mutations. The clinical significance of the proposed prognostic model was tested using an independent public dataset, and its clinical utility was demonstrated using an inhouse dataset and clinical data.

Materials and Methods

Public microarray datasets

We used two public datasets to construct the model. Gene expression data from the H. Lee Moffitt Cancer Center [21], which is part of the GSE14333 (learning set; n=73, Dukes stages A, B and C CC, without adjuvant chemotherapy), were used for gene cluster discovery. Gene expression data from The French Cartes d’Identite´ des Tumeurs (CIT) program, which is a part of GSE39582 [9] (test set; n=258, Stages II and III CC, without adjuvant chemotherapy), were used to construct our prognostic model. The datasets are available from the National Center for Biotechnology Information Gene Expression Omnibus database (

Tissue samples

We collected tumor tissue specimens from patients with Stage II/III CC who underwent curative surgery at the National Defense Medical College (validation set; n=60, without adjuvant chemotherapy, 2002- 2003) according to a protocol approved by the Institutional Review Board. Archived primary tumor and adjacent non-neoplastic colon tissues were collected during surgery. Frozen sections with a tumor component >60% were prepared from each specimen and subjected to DNA microarray and gene mutation analyses. Patients and tumor characteristics of the three datasets are presented in Table 1.

Characteristics Learning set Test set Validation set P value
(n=73) (n=258) (n=59)
Age at diagnosis Mean, years (SD) 68.9(12.0) 70.9(12.8) 69.4(8.3) 0.308
<40, n (%) 1(1.4%) 6(2.3%) 0(0%)
40-49, n (%) 2(2.7%) 13(5.0%) 1(1.7%)
50-59, n (%) 11(15.1%) 22(8.5%) 7(11.9%)
60-69, n (%) 22(30.1%) 61(23.6%) 20(33.9%)
>70, n (%) 37(50.7%) 156(60.5%) 31(52.5%)
Sex, n (%) Male 37(50.7%) 147(57.0%) 32(54.2%) 6.00e-03
Female 36(49.3%) 111(43.0%) 27(45.8%)
Stage, n (%) I 21(28.8%) 0 (0%) 0 (0%) 2.42E-22
II 37(50.7%) 203(78.7%) 43(72.9%)
III 15(20.5%) 55(21.3%) 16(27.1%)
Location, n (%) Right 35(47.9%) 118(45.7%) 25(42.4%) 0.814
Left 38(52.1%) 140(54.3%) 34(57.6%)
Follow up, n (%) Mean 4.02(0.08-11.9) 4.41(0-16.8) 5.92(0.05-11.5) 2.00e-03
Relapse, n (%) + 11(15.1%) 62(24.0%) 6(10.2%) 0.027
- 62(84.9%) 196(76.0%) 53(89.8%)
KRAS mutant, n (%) + - 86(34.5%) 23(39.0%) 0.624
- - 163(65.5%) 36(61.0%)
N/A 73 9 0
BRAF mutant, n (%) + - 33(13.8%) 2(3.4%) 0.046
- - 206(86.2%) 57(96.6%)
N/A 73 19 0
TP53 mutant, n (%) + - 79(49.7%) 35(59.3%) 0.266
- - 80(50.3%) 24(40.7%)
N/A 73 99 0
MSI, n (%) + - 47(20.7%) 5(8.6%) 0.053
- - 180(79.3%) 53(91.4%)
N/A 73 31 1
Recurrence, n (%) + 11(15.1%) 62(24.0%) 6(10.2%) 0.272
- 62(84.9%) 196(76.0%) 53(89.8%)
5-year survival rate, % 83.9% 72.5% 89.5% -

Table 1: Patient characteristics.

Construction of a molecular subtype predictor using a learning set

To discover functional gene modules, we performed co-expression analysis using R software. We filtered out low intensity probes (average signal intensity<350) and low deviation probes (CV<0.18) from all probes in GSE14333 microarray data. Then we calculated Pearson’s correlation coefficient of gene expression between the remaining 10,449 genes, and we extracted gene modules from gene clusters. After we identified three gene modules representing the long arms of chromosomes 18 and 20 and stromal-related genes, representative genes were selected according to the significant associations of their expression levels within each module as well as to information about their biological functions. We were limited to 73 samples because of our relatively narrow selection criteria.

Accordingly, we next constructed a subtype predictor with reference to Hu et al. [22]. First, we performed hierarchical clustering of the learning dataset using the expression profiles. We next averaged the gene expression profiles (centroid) for samples assigned to each subtype using the hierarchical clustering dendrogram as a guide. New samples were assigned to the nearest subtype/centroid as determined using the Spearman correlation. This predictor was then used to classify tumors into subtypes within the test and validation datasets.

Gene expression analysis

Total RNA and DNA were extracted using an RNeasy Mini Kit and a QIAamp DNA Mini Kit (Qiagen Valencia, CA). The RNA quality of samples was then assessed using a 2100 Bio analyzer (Agilent Technologies, Palo Alto, CA). The preparations were of optimal quality for oligonucleotide microarray analysis, as shown by the integrity of the 18S and 28S ribosomal RNA bands. Gene expression data were generated using an Affymetrix Human Genome U133 Plus 2.0 Gene Chip Array (Affymetrix, Santa Clara, CA). Each dataset was normalized using the Microarray Suite (MAS) 5.0 in the Expression console (Affymetrix). Data are available from the NCBI Gene Expression Omnibus (; accession number GSE 92921).

Gene mutation and MSI analysis

The mutations in exon 2 and 3 of KRAS [23], exon 15 of BRAF [24] and exons 5 to 9 of TP53 [25] were assessed using direct sequencing as previously described. MSI was analyzed using a panel of five microsatellite loci of the Bethesda reference panel [26]. MSIhigh tumors were classified as deficient in DNA mismatch repair (dMMR), and MSI-low and microsatellite-stable (MSS) tumors as proficient MMR (pMMR).

Survival analysis

We used Kaplan-Meier analysis and the log-rank test to estimate prognosis. Recurrence-free survival (RFS) was defined as the time from surgery until primary tumor recurrence. P<0.05 indicates statistical significance, and all statistical tests were two-tailed. All statistical analyses were conducted using the statistical software EZR [27].


Identification of three CC subtypes and development of a subtype predictor

We performed an expression correlation analysis using data for 73 primary CC cases in GSE14333 (learning set, Dukes Stage A/B/C CC). We identified several gene clusters, which included gene groups located on chromosomes 18q, 20q, and 13q as well as the gene groups associated with the immune response, cell proliferation, and stromal related genes (Figure 1).

Figure 1: Correlation heat map. The heat map shows pairwise Pearson correlation coefficients between genes. The Pearson correlation coefficients from low to high are indicated in blue to yellow, respectively. The bars on the right side of heat map indicate gene clusters.

Among these gene groups, we focused on genes located on the long arms of chromosomes 18 and 20 and stromal-related genes. We determined the representation of each gene in the modules with closely correlated expression levels in the same module. Finally, we extracted 55 genes according to their representation among the three gene groups and information about their biological functions (Table 2).

Affymetrix Probe Set ID Representative Public ID UniGene ID Gene Symbol Gene Title Chromosomal Location
200027_at NM_004539 Hs.465224 NARS asparaginyl-tRNA synthetase chr18q21.31
201588_at NM_004786 Hs.114412 TXNL1 thioredoxin-like 1 chr18q21.31
202484_s_at AF072242 Hs.25674 MBD2 methyl-CpG binding domain protein 2 chr18q21
202836_s_at NM_006701 Hs.465498 TXNL4A thioredoxin-like 4A chr18q23
209366_x_at M22865 Hs.465413 CYB5A cytochrome b5 type A (microsomal) chr18q23
217752_s_at NM_018235 Hs.149185 CNDP2 CNDP dipeptidase 2 (metallopeptidase M20 family) chr18q22.3
218171_at AF195514 Hs.126550 VPS4B vacuolar protein sorting 4 homolog B (S. cerevisiae) chr18q21.33
218208_at NM_025078 Hs.288284 PQLC1 PQ loop repeat containing 1 chr18q23
219419_at NM_024805 Hs.593610 RBFA ribosome binding factor A (putative) chr18q23
220180_at NM_025214 Hs.120790 CCDC68 coiled-coil domain containing 68 chr18q21
223180_s_at BC000892 Hs.532835 C18orf55 chromosome 18 open reading frame 55 chr18q22.3
223299_at AF212233 Hs.45107 SEC11C SEC11 homolog C (S. cerevisiae) chr18q21.32
225301_s_at AI991160 Hs.720076 MYO5B myosin VB chr18q21
225407_at N37023 Hs.551713 MBP myelin basic protein chr18q23
226924_at AI016355 Hs.61508 LOC400657 hypothetical LOC400657 chr18q22.3
227072_at BG167480 Hs.654809 RTTN Rotatin chr18q22.2
227542_at AU157543 Hs.44439 SOCS6 suppressor of cytokine signaling 6 chr18q22.2
227768_at AB051490 Hs.536490 ZNF407 zinc finger protein 407 chr18q23
232594_at AK001829 Hs.191582 HSBP1L1 heat shock factor binding protein 1-like 1 chr18q23
238846_at AW026379 Hs.204044 TNFRSF11A tumor necrosis factor receptor superfamily, member 11a, NFKB activator chr18q22.1
202716_at NM_002827 Hs.417549 PTPN1 protein tyrosine phosphatase, non-receptor type 1 chr20q13.1-q13.2
202925_s_at NM_002657 Hs.154104 PLAGL2 pleomorphic adenoma gene-like 2 chr20q11.21
203943_at NM_004798 Hs.369670 KIF3B kinesin family member 3B chr20q11.21
204554_at AL109928 Hs.42215 PPP1R3D protein phosphatase 1, regulatory (inhibitor) subunit 3D chr20q13.3
204849_at NM_006602 Hs.126248 TCFL5 transcription factor-like 5 (basic helix-loop-helix) chr20q13.3-qter
206918_s_at NM_003915 Hs.246413 CPNE1 copine I chr20q11.22
207941_s_at NM_004902 Hs.282901 RBM39 RNA binding motif protein 39 chr20q11.22
208979_at AF128458 Hs.368971 NCOA6 nuclear receptor coactivator 6 chr20q11
209422_at AL109965 Hs.517044 PHF20 PHD finger protein 20 chr20q11.22-q11.23
212237_at N64780 Hs.374043 ASXL1 additional sex combs like 1 (Drosophila) chr20q11.1
212349_at AL045513 Hs.472409 POFUT1 protein O-fucosyltransferase 1 chr20q11
213090_s_at AI744029 Hs.18857 TAF4 TAF4 RNA polymerase II, TATA box binding protein (TBP)-associated factor, 135kDa chr20q13.33
218325_s_at NM_022105 Hs.517172 DIDO1 death inducer-obliterator 1 chr20q13.33
224611_s_at AL118506 Hs.164419 DNAJC5 DnaJ (Hsp40) homolog, subfamily C, member 5 chr20q13.33
224815_at AA148301 Hs.408427 COMMD7 COMM domain containing 7 chr20q11.21
225224_at AL034550 Hs.516978 C20orf112 chromosome 20 open reading frame 112 chr20q11.1-q11.23
225313_at AI627538 Hs.504920 C20orf177 chromosome 20 open reading frame 177 chr20q13.2-q13.33
225498_at AV713673 Hs.472471 CHMP4B chromatin modifying protein 4B chr20q11.22
225903_at AL118520 Hs.253319 PIGU phosphatidylinositol glycan anchor biosynthesis, class U chr20q11.22
1555630_a_at AF327350 Hs.301853 RAB34 RAB34, member RAS oncogene family chr17q11.2
201069_at NM_004530 Hs.513617 MMP2 matrix metallopeptidase 2 (gelatinase A, 72kDa gelatinase, 72kDa type IV collagenase) chr16q13-q21
201426_s_at AI922599 Hs.455493 VIM Vimentin chr10p13
201893_x_at AF138300 Hs.156316 DCN Decorin chr12q21.33
202351_at AI093579 Hs.436873 ITGAV integrin, alpha V (vitronectin receptor, alpha polypeptide, antigen CD51) chr2q31-q32
202450_s_at NM_000396 Hs.632466 CTSK cathepsin K chr1q21
202686_s_at NM_021913 Hs.590970 AXL AXL receptor tyrosine kinase chr19q13.1
203729_at NM_001425 Hs.9999 EMP3 epithelial membrane protein 3 chr19q13.3
208747_s_at M18767 Hs.458355 C1S complement component 1, s subcomponent chr12p13
212067_s_at AL573058 Hs.524224 C1R complement component 1, r subcomponent chr12p13
213001_at AF007150 Hs.653262 ANGPTL2 angiopoietin-like 2 chr9q34
218469_at NM_013372 Hs.40098 GREM1 gremlin 1 chr15q13.3
226001_at AK002174 Hs.272251 KLHL5 kelch-like 5 (Drosophila) chr4p14
226051_at BF973568 Hs.55940 SELM selenoprotein M chr22q12.2
229067_at BF977829 Hs.523529 SRGAP2P1 SLIT-ROBO Rho GTPases activating protein 2 pseudo gene 1 chr1p11.2
57715_at W72694 Hs.241545 CALHM2 calcium homeostasis modulator 2 chr10pter-q26.12

Table 2: Reference genes.

Unsupervised clustering using the 55 genes revealed that patients with CC (learning set) could be classified into the subtypes as follows: Subtype 1 (n=32; 43.2%), characterized by high expression of stromal genes; Subtype 2 (N=24; 33.8%), associated with the 18q-lower/20qhigher expression than average; and Subtype 3 (n=17; 23.0%), associated with the 18q-higher/20q-lower expression than average (Figure 2).

Figure 2: Heat map of unsupervised hierarchal clustering. Unsupervised hierarchical clustering of the learning dataset comprising 55 genes expressed at detectable levels in all samples. Each column represents a sample and each row a gene. Gene expression above and below the median is depicted in red and blue, respectively. The blue and yellow squares in the second upper bar indicate patients with no recurrence and recurrence, respectively.

Subtype 1 was named the ‘stromal subtype’ because of the high expression of stroma-related genes.

Subtype 2 was named the ‘CIN-like subtype’, because CIN cancers frequently express the 18q-low/20q-high pattern [16,17]. Subtype 3 was named the ‘MSI-like subtype’, because we suspect that MSI cases are infrequently categorized as Subtype 2. Thus, Subtype 3 is enriched with MSI cases. The 5-year disease free survival (DFS) rates in the CIN-like, MSI-like, and stromal subtypes were 88.2%, 86.4%, and 72.8%, respectively. We found no statistically significant difference, though patients with the stromal subtype tended to have a higher risk of recurrence compared with the other subtypes (Figure 3 and Table 3). For general use, we developed a subtype classifier ‘55-gene centroid predictor’ to divide CCs into the three subtypes, according to a previous report [22].

Characteristic CIN-like subtype MSI-like subtype Stromal subtype P value
(n=32, 43.8%) (n=24, 32.9%)  (n=17, 23.3%)
Age at diagnosis Mean, Year (SD) 72.9(10.5) 67.2(12.7) 63.8(11.9%) 0.2856
<40, n (%) 0(0%) 0(0%) 1(5.9%)
40-49, n (%) 0(0%) 2(8.3%) 0(0%)
50-59, n (%) 4(12.5%) 4(16.7%) 3(17.6%)
60-69, n (%) 8(25.0%) 8(33.3%) 6(35.3%)
>70, n (%) 20(62.5%) 10(41.7%) 7(41.2%)
Sex, n (%) Male 18(56.3%) 11(45.8%) 8(47.1%) 0.724
Female 14(43.8%) 13(54.2%) 9(52.9%)
Stage at diagnosis, n (%) Dukes A 13(40.6%) 5(20.8%) 3(17.6%) 0.315
Dukes B 13(40.6%) 15(62.5%) 9(52.9%)
Dukes C 6(18.8%) 4(16.7%) 5(29.4%)
Location, n (%) Right 12(37.5%) 16(66.7%) 7(41.2%) 9.19e-02
Left 20(62.5%) 8(33.3%) 10(58.8%)
Recurrence, n (%) + 3(9.4%) 3(12.5%) 5(29.4%) 0.226
- 29(90.6%) 21(87.5%) 12(70.6%)
5-year DFS,%   88.2% 86.4% 72.8% 0.222

Table 3: Correlation of subtypes with clinical parameters of the learning set.

Figure 3: Survival of subtypes in the learning dataset. Kaplan–Meier analysis of disease-free survival times of patients in each subtype. The orange, blue and pink lines indicate the CIN-like, MSI-like and Stromal subtypes, respectively. Censored data are marked at each line. P values were calculated using the log-rank test.

Prognostic significance of the subtype model and its biological importance for the test set

We applied the subtype predictor to an independent public data set of 258 stage II/III CC samples (GSE39582). The CIN-like, MSIlike, and stromal subtypes accounted for 39.9% (103 specimens), 39.5% (102 specimens), and 20.5% (53 specimens), respectively. The 5-year RFS of patients with the CIN-like, MSI-like, and stromal subtype cancers were 72.2%, 81.2%, and 56.0%, respectively (Figure 4). The log-rank test revealed that there was a significantly higher incidence of recurrent events in the stromal subtype (stromal subtype vs. other subtypes, p=3.62e−03).

Figure 4: Survival curves corresponding to each subtypes in the test dataset. Kaplan–Meier analysis of RFS of patients in each subtype. The orange, blue and pink lines indicate CIN-like, MSI-like and Stromal subtypes, respectively. Censored data are marked at each line. P values were calculated using the log-rank test.

The relationships between each subtype and factor in the test dataset are summarized in Table 4. The results supported our assumption that CIN was significantly enriched in CIN-like subtype cases (p=3.86e-15), MSI was significantly enriched in MSI-like subtype cases (p=2.20e-16), and CpG island methylated phenotype (CIMP) was significantly enriched in MSI-like subtype cases (p=2.38e-13). There was a high incidence of BRAF mutations (p=1.22e-10) in MSI-like subtype cases and a low incidence of TP53 mutations (p=1.09e-06).

Characteristic CIN-like subtype MSI-like subtype Stromal subtype P value
(n=103, 39.9%)  (n=102, 39.5%) (n=53, 20.5%)
Age at diagnosis Mean, Year (SD) 70.5(11.9) 71.3(14.0) 71(12.3) 0.198
<40, n (%) 1(1.0%) 4(3.9%) 1(1.9%)
40-49, n (%) 4(3.9%) 7(6.9%) 2(3.8%)
50-59, n (%) 13(12.6%) 3(2.9%) 6(11.3%)
60-69, n (%) 25(24.3%) 21(20.6%) 15(28.3%)
>70, n (%) 60(58.3%) 67(65.7%) 29(54.7%)
Sex, n (%) Male 63(61.2%) 56(54.9%) 28(52.8%) 0.525
Female 40(38.8%) 46(45.1%) 25(47.2%)
Stage at diagnosis, n (%) II 87(84.5%) 79(77.5%) 37(69.8%) 0.096
III 48(46.6%) 23(22.5%) 16(30.2%)
Location, n (%) distal 80(77.7%) 30(29.4%) 30(56.6%) 3.35e-11
proximal 23(22.3%) 72(70.6%) 23(43.4%)
MSI status, n (%) MSI 0(0%) 47(51.6%) 0(0%) <2.2e-16
MSS 93(100%) 44(48.4%) 43(100%)
N/A 10 11 10
CIN status, n (%) + 82(95.3%) 39(41.9%) 37(84.1%) 3.86e-15
- 4(4.7%) 54(58.1%) 7(15.9%)
N/A 17 9 9
CIMP status, n (%) + 2(2.2%) 45(47.4%) 5(11.4%) 2.38e-13
- 89(97.8%) 50(52.6%) 39(88.6%)
N/A 12 7 9
KRAS mutation status, n (%) + 24(24.7%) 36(36.0%) 26(50%) 7.79e-03
- 73(75.3%) 64(64.0%) 26(50%)
N/A 6 2 1
BRAF mutation status, n (%) + 0(0%) 31(32.0%) 2(4.2%) 1.22e-10
- 94(100%) 66(68.0%) 46(95.8%)
N/A 9 5 5
TP53 mutation status, n (%) + 47(67.1%) 13(22.4%) 19(61.3%) 1.09e-06
- 23(32.9%) 45(77.6%) 12(38.7%)
N/A 33 44 22
Recurrence, n (%) + 25(24.3%) 16(15.7%) 21(39.6%) 4.19e-03
- 78(75.7%) 86(84.3%) 32(60.4%)
5-year RFS, %   72.2% 81.2% 56.0% 5.84e-03

Table 4: Correlation of subtypes with clinical parameters in the test set.

Modification of the subtype model incorporating gene mutation status

When we evaluated the prognostic values of common gene mutations in each subtype, we found that KRAS mutations represented a significant risk factor for recurrence (p=2.74e-04) in the CIN-like subtype cases (Figure 5), while BRAF and TP53 mutations did not have significant prognostic value. Accordingly, we created a prognostic model that classified patients with CC into the good or poor prognosis group according to the molecular subtype and KRAS mutation status (Figure 6). Our prognostic model classified patients with stage II/III CC into two groups with significantly different risks of recurrence (p=1.56e-06) in the test set GSE39582 (Figure 7).

Figure 5: Prognostic values of common gene mutations in the CIN subtype. Kaplan–Meyer analysis of RFS of CIN-like subtype patients in the test dataset compared with patients with wild-type (blue line) or mutated (red line) KRAS (A), BRAF (B), or TP53 (C). Censored data are marked at each line.

Figure 6: Flow chart of the prognosis model.

Figure 7: Kaplan-Meier analysis of RFS of patients in the test data predicted with a high versus low risk of recurrence. The blue and red lines indicate low-risk and high-risk patients, respectively. Censored data are marked at each line. P values were calculated using the log-rank test.

The results of these analyses are summarized in Figure 6. The relative risks of the subtypes associated with 5-year RFS are as follows: MSI-like, 81.2%; CIN-like, wild-type KRAS, 78.9%; CIN-like, mutant KRAS, 46.6%; and stromal, 56.0%.

Validation of the prognostic model with in-house data set

We next validated the prognostic model with the newly acquired microarray datasets of tissue specimens from 59 patients treated at the National Defense Medical College with resected stage II/III CC who had not received postoperative adjuvant chemotherapy (inhouse validation dataset). Patients were classified into the CIN-like (32.2%, 19 specimens), MSI-like (32.2%, 19 specimens), and stromal (35.6%, 21 specimens) subtypes using the 55-gene centroid predictor. Finally, we applied our prognostic model to the independent dataset and demonstrated that it identified high- and low-risk patient groups (p=0.0475) (Figure 8 and Table 5).

Figure 8: Kaplan-Meier analysis of RFS of patients in the test data predicted with a high versus low risk of recurrence. The blue and red lines indicate low-risk and high-risk patients, respectively. Censored data are marked at each line. P values were calculated using the log-rank test.

Variable CIN-like subtype MSI-like subtype Stromal subtype P value
(n=19, 32.2%) (n=19, 32.2%) (n=21, 35.6%)
Age at diagnosis Mean, Year (SD) 68.7 (8.3) 71.3 (9.0) 68.3 (7.8) 0.220
<40, n (%) 0 (0%) 0 (0%) 0 (0%)
40-49, n (%) 0 (0%) 1 (5.3%) 0 (0%)
50-59, n (%) 3 (15.8%) 1 (5.3%) 3 (14.3%)
60-69, n (%) 8 (42.1%) 3 (15.8%) 9 (42.9%)
>70, n (%) 8 (42.1%) 14 (73.7%) 9 (42.9%)
Sex, n (%) Male 9 (47.4%) 12 (63.2%) 11 (52.4%) 0.648
Female 10 (52.6%) 7 (36.8%) 10 (47.6%)  
Stage at diagnosis, n (%) II 12 (63.2%) 16 (84.2%) 15 (71.4%) 0.356
III 7 (36.8%) 3 (15.8%) 6 (28.6%)
Location, n (%) distal 7 (36.8%) 10 (52.6%) 8 (38.1%) 0.545
proximal 12 (63.2%) 9 (47.4%) 13 (61.9%)
MSI status, n (%) dMMR 0 (0%) 5 (27.8%) 0 (0%) 1.87e-03
pMMR 19 (100%) 13 (72.2%) 21 (100%)
N/A 0 1 0
KRAS mutation status, n (%) + 5 (26.3%) 12 (63.2%) 6 (28.6%) 0.0472
- 14 (73.7%) 7 (36.8%) 15 (71.4%)
N/A 0 0 0
BRAF mutation status, n (%) + 0 (0%) 2 (10.5%) 0 (0%) 0.200
- 19 (100%) 17 (89.5%) 21 (100%)
N/A 0 0 0
TP53 mutation status, n (%) + 14 (73.7%) 7 (36.8%) 14 (66.7%) 0.0561
- 5 (26.3%) 12 (63.2%) 7 (33.3)
N/A 0 0 0
Recurrence, n (%) + 1 (5.3%) 1 (5.3%) 4 (19.0%) 0.346
- 18 (94.7%) 18 (94.7%) 17 (81.0%)
5-year DFS,%   94.4% 94.4% 80.7% 0.267

Table 5: Correlation of subtypes with clinical parameters of the validation set.


To develop a commercially available prognostic test for patients with CC that offers unique advantages over existing assays [28-34], we focused on selecting genes with known biological functions that reside on specific chromosomal loci as well as those that are expressed by the tumor stroma known to influence tumor phenotypes (i.e. “rational” vs. “shotgun” approach”). Here we describe such a test that importantly does not infringe on patents for others designed to predict prognosis of patients with CC (Our model described in this paper is patent pending). In particular, selection of differentially expressed genes using bioinformatics without considering function may diminish the test’s specificity and sensitivity to levels that might disqualify it for routine clinical use. Specifically, we estimated the performance of the prognostic model using 4,818 genes in three gene clusters (Stromal/Chromosome 20q, 13q/Chromosome 18q (Figure 1 and Table 6). The prognostic performance of this model was diminished in the test and validation sets. This result indicates that our gene selection methods effectively help to construct a highly reproducible discriminative model comprising multiple genes.

  Learning set Test set Validation set
n 5 year DFS n 5 year RFS n 5 year DFS
55 genes CIN 32 88.2% 103 72.2% 19 94.4%
MSI 24 86.4% 102 81.2% 19 94.4%
Stromal 17 72.8% 53 56.0% 21 80.7%
P value 0.222 5.84e-03 0.267
High risk - - 77 52.9% 26 80.6%
Low risk - - 181 81.0% 33 96.8%
P value   1.56e-06 4.75e-02
4,818 genes CIN 35 82.9% 110 68.7% 20 90.0%
MSI 27 96.0% 100 80.8% 33 87.3%
Stromal 11 56.0% 48 63.3% 6 100.0%
P value 3.23e-03 2.00e-02 0.666
High risk - - 80 57.6% 9 100.0%
Low risk - - 171 78.7% 50 87.5%
P value   7.64e-05 0.276

Table 6: Comparison of 5-year survival rates of subtypes using 55 and 4,818 genes.

For this purpose, here we performed unsupervised clustering of gene expression in CC samples, focusing on genes located on the long arms of chromosomes 18 and 20 as well as stroma-related genes. We selected the gene modules “Chromosomal 18q Loss”, “Chromosomal 20q Amplification,” and “Stromal” to construct a classifier comprising 55 genes. This strategy enabled us to categorize CCs into the subtypes “Stromal,” “CIN-like,” and “MSI-like.” Next, we discovered that KRAS mutations were associated with shorter RFS of patients with the CIN-subtype but not those with the MSI- and Stromal-types. We incorporated these findings to construct a new prognostic model that incorporated the combination of subtype classifications and KRAS mutations.

Adjuvant chemotherapy is frequently considered as the treatment option of choice for high-risk patients with stage II CC. The National Comprehensive Cancer Network [3] and the European Society for Medical Oncology [35] recommend considering adjuvant treatment for patients with stage II CC if specific features are present as follows: clinical obstruction, perforation, T4 tumor, suboptimal lymph node sampling (<12 lymph nodes), presence of lymphovascular or perineural invasion, or poor differentiation. Kumar et al. [36] determined the effectiveness of adjuvant chemotherapy for high-risk patients with these features. Further, ≥ 76% of Stage II patients are categorized as high-risk [36], although the recurrence rate of Stage II patients ranges from 10% to 20% [2].

Patients with stage III CC are generally advised to undergo adjuvant chemotherapy; however, we doubt that patients with Stage IIIA CC require this type of therapy, because their prognosis is better compared with those with Stage IIB/IIC. Further investigations are therefore necessary to develop a robust standard to select candidate adjuvant chemotherapeutic strategies according to cancer stage and subtyping performed using gene-expression profiles.

The RFS of low-risk patients with Stage II/III CC judged using our prognostic model was 81.0% in the test set and 96.8% in the validation set, indicating that adjuvant therapy is not required for such patients. Thus, we used our model to assign 77% and 50% of the test and validation sets, respectively, to the low-risk category. These frequencies are significantly higher compared with the proportion of patients classified as low-risk [36]. Thus, the developed model was designed to identify CC patients with good and poor prognosis, if used clinically, would spare numerous patients from unnecessary and potentially harmful adjuvant therapy and reduce medical costs.

MSI status is a key feature of CC, and evidence [37,38] indicates that the benefits of adjuvant chemotherapy differ between patients with MSI and MSS. In the present study, all MSI cancers were included in MSI-like subtype, not only in the test set, but in the validation set as well. Consequently, all CIN-like and stromal subtypes exhibited the MSS phenotype, indicating that additional screening for MSI screening of patients with CIN-like and stromal subtype cancers is not required. The 55-gene centroid predictor was significant for predicting prognosis and was biologically significant by its ability to identify MSI subtypes.

Gene expression profiling is increasingly utilized for cancer diagnosis, prognosis, and treatment decisions. The Mamma Print assay [39-41], introduced by Agendia in 2007, and was the first FDAapproved molecular test that profiles gene expression levels to predict the risk of early-stage breast cancer to guide treatment decisions. The OncotypeDX Colon Cancer Assay manufactured by Genomic Health improves prediction of recurrence risk for patients with stage II CC [28-30,32,33,36]. The ColoPrint, produced by Agendia, significantly improves prognostic accuracy in predicting the survival of patients with stages II and III colorectal cancer [31,34]. Subsequently, several clinical trials [28-30] validated the performance of these two classifiers. However, the OncotypeDX Colon Cancer Assay categorized 26%– 31% of these patients as “intermediate risk” [33]. This high incidence of “intermediate risk” patients complicates clinical decision-making. Moreover, these assays required an MSI test to improve their quality, because a good prognosis was predicted for patients with MSI-type cancer regardless of the results of the OncotypeDX Colon Cancer Assay. Thus, our prognostic model has a distinct advantage, because it divides patients with CC into two risk groups without performing the MSI test.

Although these findings seem promising, the present study has some limitations. First, our models were structured using data acquired from frozen specimens. For practical use, the quality of our model must be confirmed using formalin-fixed paraffin-embedded specimens. Second, we were unable to obtain statistically significant data for Stage II or Stage III patients because few patient samples were available. We believe that a larger study will aid in determining the potential of the subtype classifier and the prognostic model described here.


We developed a new prognostic model for patients with CC based on analysis of the expression of 55 genes and KRAS mutations status. This novel model was developed by analyzing a combination of 18q, 20q, and stromal gene modules. The model produced a biologically discriminatory classifier that associated MSI status with the risk of recurrence. Finally, we demonstrated the prognostic significance of the proposed model that may be clinically applicable for the selection of patients with Stage II/III CC for adjuvant therapy.


We thank Dr. Igor Kurochkin for his helpful comments on earlier drafts of this manuscript and Aya Matsuoka and Mayuko Sueoka for their technical assistance. We thank Steven R. Tronick, PhD of DMC Corp. ( <>) for editing drafts of this manuscript.

Grant Numbers and Source of Support

K.G. and Y.Y. are employees of Sysmex Corporation. The other authors have no conflicts of interest to declare. The microarray measurement, gene mutation, MSI, and statistical analyses were funded by the research and development fund of Sysmex Corporation. The funder did not have a role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript.


international publisher, scitechnol, subscription journals, subscription, international, publisher, science

Track Your Manuscript

Awards Nomination

Recommended Conferences

International Conference on Clinical Nephrology

Las Vegas, USA

Media Partners