Classification Based on the Combination of Molecular and Pathologic Predictors is Superior to Molecular Classification on Prognosis in Colorectal Carcinoma

Materials and Methods

Case materials. All 221 patients (66 cases died of colorectal carcinoma) with colorectal carcinoma were inhabitants of Xiaoshan District, Zhejiang Province, China. There were 121 males and 100 females. The age range of the patients was from 26 to 85 years (median, 59 years). All archival paraffin-embedded tissue blocks were collected from the Department of Pathology, Zhejiang University, and the People's No. 1 Hospital of Xiaoshan, and the Zhejiang Cancer Hospital from January 1990 to December 2000. Patient and treatment data were collected from patient records. The 221 patients had not received chemotherapy or radiotherapy prior to surgery. Pathologic predictors and immunohistochemical scores of each section were independently scored by two pathologists who were blinded to the patients' outcomes. Any interobserver divergences were resolved with a multiheaded microscope. Information about follow-ups was provided by the Xiaoshan Center of Disease Control, with a median follow-up period of 51 months (range, 1-152 months), censored to the end of 2002. This study was approved by the Ethics Committee of Biomedicine, Zhejiang University, China.

Pathology. Paraffin-embedded tissue blocks were processed according to standard histologic procedures and stained with H&E. The following pathologic predictors were derived from histology: (a) the type of histology: tubular adenocarcinoma (n = 164), papillary adenocarcinoma (n = 26), mucinous adenocarcinoma (n = 27), ring cell carcinoma, and undifferentiated carcinoma (n = 4); (b) histologic grade: low grade (n = 148; gland formation >50% in tubular adenocarcinoma and papillary carcinoma) and high grade (n = 73; gland formation <50% in tubular adenocarcinoma, mucinous adenocarcinoma, signet ring cell carcinoma, and undifferentiated carcinoma); (c) depth of infiltration: inside serosa (n = 199) and outside serosa (n = 22); (d) metastasis in lymph node (negative, n = 122; positive, n = 99); (e) distant metastasis (negative, n = 199; positive, n = 22); (f) TNM stages (stage I, n = 35; stage II, n = 83; stage III, n = 81; stage IV, n = 22).

Immunohistochemistry. Archival paraffin-embedded tissue blocks were consecutively cut (4 μm) and mounted on 3-aminopropyl-triethoxysilane–coated slides. Thirteen immunohistochemical predictors were investigated in this study, including P53, Bcl2, Ki67, NF-κB, syndecan-1, β-catenin, CD68, CXCR4, PPARγ, IGF-IR, IGFBP7, insulin receptor, and thymidylate synthase. Information regarding the antibodies used are summarized in Table 1 , along with the staining pattern and patient number for each marker used. Missed data were caused by a shortage of sections.

Immunostains by EnVision method were done according to the conditions summarized in Table 1. For negative controls, the primary antibodies were replaced with a PBS solution (100 mmol/L; pH 7.4). A definite brown granular stain was defined as positive. The percentages of positive cells were scored using the following scale: 0, no staining or <5%; 1, 5% to 25%; 2, 26% to 50%; 3, 51% to 75%; 4, >75%, except IGFBP7, Bcl2, CXCR4, and NF-κB, which were scored as 0 (<10%) or 1 (≥10%). The number of macrophages (CD68-positive cells) in the invasive margin and in the center of tumor were counted separately in four fields (40×), and mean values were calculated and then classified into two grades according to the median. Macrophage counts ranged from 0 to 164 (median, 16) in the invasive margin, and from 0 to 70 (median, 13) in the center of tumor. We noticed that the expression of CXCR4 (positive percentage or intensity) in the invasive margin of the tumor was significantly increased compared with that in the central regions of the tumor in 18.6% (34 of 183) cases, so we defined the phenomenon as CXCR4 margin.

Statistical analysis. Univariate survival analyses were done and survival curves were drawn using the Kaplan-Meier method. The differences between curves were tested by the log-rank test. Cumulative survival rate was calculated by a life-table method. Multivariate analysis was done using the COX proportional hazard model and a forward stepwise method was used to bring variables into the model. A significant difference was identified at P < 0.05, and a potential significance was identified at P < 0.1. Hierarchical cluster analyses were done by Cluster 3.0 (Stanford University), and the results were visualized by TreeView (Stanford University). Complete linkage's method was used as the cluster method, using Spearman rank correlation as an interval measure. Comparison of the resultant clusters was made by χ² test or Fisher's exact test using SPSS 13.0 statistical software (SPSS Inc.).

Results

Univariate survival analyses of pathologic predictors and molecular predictors. Unfavorable prognosis was significantly correlated with pathologic predictors, i.e., high histologic grade, deeper infiltration, lymph node metastasis, distant metastasis, and higher TNM stage. Papillary adenocarcinoma had a better prognosis than other histologic types. Among the molecular predictors, the significant predictors included P53, IGFBP7, CD68 in the invasive margin and CXCR4 margin (Fig. 1 ), and a trend toward significant predictor was PPARγ (P = 0.085). More specifically, patients had the worst prognosis when P53 scored 2, IGFBP7 scored 0, CD68 scored 0, and CXCR4 margin scored 1, respectively. The rest of the molecular predictors, including Ki67, Bcl2, thymidylate synthase, syndecan-1, NF-κB, IGF-IR, insulin receptor, and β-catenin were not correlated with prognosis.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Survival analysis of molecular factors in colorectal carcinoma based on P53 (A), IGFBP7 (B), CXCR4 margin (C), and CD68 in the invasive margin (D).

Hierarchical clustering analysis. Unsupervised hierarchical clustering analysis was done using complete linkage method with Spearman rank test to group the data with different predictor profiles. First, we clustered data with all 13 molecular predictors (13m), two main groups (group 1, n = 50; group 2, n = 171) resulted from this process, showing no correlation with prognosis (P = 0.4238). Second, we clustered data with five molecular predictors (5m) which were significant or potentially significant in the univariate survival analyses, the resulting two groups (group 1, n = 95; group 2, n = 126) correlated with prognosis (P = 0.0532). Twenty-four patients in group 1 passed away at the census point (median survival, 61 months), and 42 patients in group 2 passed away at the census point (median survival, 45 months). Third, we clustered data with five pathologic predictors (5p), excluding TNM stage, no well-defined group was produced.

Furthermore, we combined pathologic and molecular predictors to cluster the data. Two distinct groups (group 1, n = 149; group 2, n = 72) resulted from the process of clustering 5 pathologic predictors and 13 molecular predictors (5p13m), showing no significant correlation with prognosis (P = 0.0775). Three distinct groups (group 1, n = 130; group 2, n = 67; group 3, n = 24) resulted from the process of clustering five pathologic predictors and five molecular predictors (5p5m), showing significant correlation with prognosis when comparing group 2 (median survival, 33 months) versus group 1 (median survival, 62 months; P < 0.0001) and when comparing group 3 (median survival, 38 months) versus group 1 (P = 0.0342), but there were no significant differences between groups 2 and 3 (P = 0.3918; Figs. 2 and 3 ). The 1-year, 3-year, and 5-year cumulative survival rates of group 1 were 86%, 83%, and 79%, respectively, and that of group 2 were 69%, 54%, and 47%, and that of group 3 were 75%, 61%, and 61%, respectively. We merged group 2 with group 3 in χ² test and Fisher's exact test, and multivariate analysis.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Clustergrams of unsupervised hierarchical clustering analysis with five molecular markers (A), and five pathologic and five molecular markers (B). The gradient was scored from low (green) to high (red); black cubes, missing values. Clustered groups are identified by differently colored branches; group 1 (black), group 2 (red), and group 3 (blue).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Survival analysis of classification based on five molecular predictors (A), five pathologic and five molecular predictors (B), and five molecular predictors in TNM stage (C) by Kaplan-Meier method with log-rank test.

The efficacy of the individual clustering predictors was evaluated by χ² test or Fisher's exact test. We found that histologic type, histologic grade, depth of infiltration, metastasis in lymph node, distant metastasis, CD68 in the invasive margin, IGFBP7, CXCR4 margin, and PPARγ showed significant differences between the two groups, whereas P53 did not show any significant difference (Table 2 ). Then we clustered the data of 5p and 4m, excluding P53, the resulting three groups showed no significant correlation with prognosis (P = 0.0919).

Furthermore, we processed Kaplan-Meier analysis in TNM stage II (n = 83) and III (n = 81), respectively, but not in TNM stage I (n = 35) and IV (n = 22) due to an insufficient numbers of patients. Cluster 5m grouped the patients of TNM stage II into two groups (group 1, n = 36; group 2, n = 48) with different survival outcomes (P = 0.0518; Fig. 3). Three patients from group 1 passed away at the census point, and 11 patients from group 2 passed away at the census point. There was a significant difference in positive rate of PPARγ (group 1,100%, 33 of 33; group 2, 43%, 20 of 47; χ² = 28.615; P < 0.0001) rather than other molecular predictors between the two groups. Five-year cumulative survival rates were 90% and 73% for groups 1 and 2, respectively.

Multivariate analysis. Multivariate analysis was done to determine the independent prognostic significance of predictors. Age, cluster 5p5m, cluster 5m and TNM stage, histologic type, and histologic grade were analyzed by COX proportional hazard model. The results showed that only cluster 5p5m and TNM stage were independent prognostic significant markers (Table 3 ). At the first step, TNM stage was included in the model, and at the second step, cluster 5p5m was included. Other variables were excluded from the model. The relative risk of death was 1.912 (group 2 versus group 1) in classification of 5p5m. The relative risks of death were 1.449 (stage II versus stage I), 2.988 (stage III versus stage I), and 8.394 (stage IV versus stage I), respectively, according to TNM stage.

Discussion

There have been many reports published on molecular classification since the term was first introduced in 1999 (6). During the last decade, research about cancer classification had always focused on molecular predictors but ignored the robust clinical significance of pathologic predictors. In this study, we have evaluated the prognostic significance of molecular classification on 221 cases of colorectal carcinoma versus that of the TNM stage system and classification based on the combination of molecular and pathologic predictors. Hierarchical clustering analysis has been applied based on 13 molecular predictors, 5 pathologic predictors, and the combination of molecular and pathologic predictors. We found that clusters based on molecular and pathologic predictors are superior to clusters based on molecular predictors in prognosis, and molecular predictors are valuable in substaging patients with TNM stage II disease. To our knowledge, this is the first study to compare molecular predictors, pathologic predictors, and their combination in the prognostic evaluation of colorectal carcinoma.

Among the 13 molecular predictors which are involved in carcinogenesis, development, and metastasis in colorectal carcinoma, we have identified five molecular predictors which are correlated with prognosis by univariate survival analysis. To find the simplest and the most accurate classification, we clustered the data by different combinations. Defined groups were produced excluding 5p, but only cluster 5m and cluster 5p5m were significant according to prognosis by the Kaplan-Meier method. Furthermore, only cluster 5p5m was proven to be an independent factor in the COX proportional hazard model, suggesting that the combined system of pathologic and molecular predictors would perhaps be a preferred staging system, and that more molecules and more patients should be included to establish a reliable system. However, to date, the TNM stage system, including the local extent of the untreated primary tumor, the status of the regional lymph nodes and distant metastasis, is still the most valuable criterion on the prognosis of colorectal carcinoma. It is noteworthy that a sufficient number of lymph nodes and diligent search are necessary. For this reason, the American Joint Committee on Cancer has recommended the examination of at least 12 lymph nodes to diagnose patients with stage III disease (7).

There has been a heated debate on the question of adjuvant chemotherapy in patients with stage II disease (8, 9). Clinical guidelines for the management of stage II colonic cancer state that the standard care is surgical resection alone (10). However, these guidelines also point out that adjuvant chemotherapy may be considered in some cases, such as in T₄. In our study, cluster 5m identified a high-risk subgroup of patients with TNM stage II disease with lower positive rates of PPARγ. PPARγ, a member of the nuclear hormone receptors, could modulate the Wnt pathway, NF-κB, cell cycle control, and expression of proapoptotic and antiapoptotic proteins. Thus, it is a potential target of adjuvant therapy in TNM stage II disease.

Among those molecular predictors, the prognostic significance of IGFBP7 is still unclear. Our study suggested that IGFBP7 positivity was associated with a better prognosis, which contradicts the study by Adachi et al. (11). Although IGFBP7 did not correlate with the macrophage in the invasive margin (data not shown), the results of cluster analysis showed that IGFBP7 was close to macrophages in the invasive margin. Macrophages play a complex role in the progression of tumors (12–16). In our study, the macrophages in the invasive margin, rather than that in the central region, were a favorable prognostic marker, and negatively correlated with either lymph node metastasis (Spearman correlation coefficient, −0.190; P = 0.006) or distant metastasis (Spearman correlation coefficient, −0.174; P = 0.011). Hence, macrophages in different regions might have different prognostic values and it is necessary to distinguish the macrophage in the invasive margin from that in the central region in future studies.

Chemokines are small chemoattractive proteins that regulate leukocyte migration and have been implicated in cancer progression and metastasis (17–19). We analyzed the correlation between CXCR4 and other pathologic and molecular predictors according to Spearman rank correlation. NF-κB could regulate cancer cells by directly up-regulating the expression of CXCR4 (20), but in our investigation, there was no correlation between CXCR4 with either NF-κB or metastasis (data not shown). Previous studies by Kato et al. indicated that tumors with focal CXCR4 staining have a higher potential for lymph node metastasis than tumors with diffuse CXCR4 staining (21), but the phenomenon of the CXCR4 margin has never been reported. Our results suggest that the CXCR4 margin, rather than CXCR4, is a predictor of worse prognosis, which is different from the results of other studies.

In conclusion, despite the number of molecular predictors which have been identified as prognostic predictors, and several molecular phenotypes which have been reported in colorectal cancer, our results indicate that molecular predictors should serve as a supplement rather than a replacement for pathologic classification. Classification based on molecular predictors could identify additional different prognoses in subtypes in stage II disease. A more accurate classification should combine pathologic and molecular predictors together.