CD147/EMMPRIN overexpression and prognosis in cancer: A systematic review and meta-analysis

CD147/EMMPRIN (extracellular matrix metalloproteinase inducer) plays an important role in tumor progression and a number of studies have suggested that it is an indicator of tumor prognosis. This current meta-analysis systematically reevaluated the predictive potential of CD147/EMMPRIN in various cancers. We searched PubMed and Embase databases to screen the literature. Fixed-effect and random-effect meta-analytical techniques were used to correlate CD147 expression with outcome measures. A total of 53 studies that included 68 datasets were eligible for inclusion in the final analysis. We found a significant association between CD147/EMMPRIN overexpression and adverse tumor outcomes, such as overall survival, disease-specific survival, progression-free survival, metastasis-free survival or recurrence-free survival, irrespective of the model analysis. In addition, CD147/EMMPRIN overexpression predicted a high risk for chemotherapy drugs resistance. CD147/EMMPRIN is a central player in tumor progression and predicts a poor prognosis, including in patients who have received chemo-radiotherapy. Our results provide the evidence that CD147/EMMPRIN could be a potential therapeutic target for cancers.


Results
Our systematic literature search of CD147 and its correlation with different outcomes identified 910 articles. Among these, 836 articles were excluded and only 82 studies satisfied the inclusion eligibility for the meta-analysis. Upon further review of the full text articles, eight additional articles from reference sources were included. An additional 29 studies were excluded due to the following reasons: two articles were reviews; three were duplicate reports; four had insufficient data; four included only a few cases; nine explored the prognostic value of CD147/EMMPRIN in combination with other biomarkers, such as VEGF, MMP-2, CD44s, MCT-1; and seven studies were determined to be too complicated for subgroup analysis. The remaining 53 studies [5][6][7][9][10][11][12][13][14][15][16][17] containing 68 datasets that met our inclusion criteria were included in this review (Fig. 2). Among the eligible studies, 33 provided survival information about the correlation of CD147/EMMPRIN expression with tumor prognosis using a multivariate model and 20 presented the same information using a univariate model. The characteristics of these two models are summarized in Tables 1 and 2, respectively.
The 53 eligible studies represented 26 different carcinomas or sarcomas and a median number of 204.5 patients (range, . CD147/EMMPRIN expression in these studies was mainly detected by immunohistochemistry (IHC). One publication had three groups of subjects from different cancers and was therefore considered as three different studies 49 . Ten publications [12][13][14]16,25,29,38,39,51,52 presented two different prognostic results, while one publication 31 presented three prognostic results. Some studies also investigated the link between CD147/EMMPRIN and chemotherapy [5][6][7] or radiation resistance 57 . In terms of clinicopathological variables, most of the studies suggested that increased CD147/EMMPRIN expression correlated significantly with higher clinical grade, tumor size, invasion depth, lymphatic invasion, histological grade, and some additional parameters (Table 3).
Similarly, the subgroup analysis of tumors stratified by cancer type again demonstrated a significant association between CD147/EMMPRIN overexpression and adverse outcome of PFS/MFS/RFS in breast carcinoma (meta-HR = 2.50; 95% CI: 1.63-3.83) and ovarian cancer (meta-HR = 1.72; 95% CI: 1. 23-2.40) in the multivariate model and renal cell carcinoma (meta-HR = 1.58; 95% CI: 1.34-1.85) in the univariate model. Similar associations were observed in hepatocellular carcinoma, bladder cancer, prostate cancer, colorectal cancer, osteosarcoma, gastric cancer and hypopharyngeal squamous cell carcinoma. However, we did not observe an association in cases of endometrial carcinoma, and the results were not consistent for cervical carcinoma (Supplementary Figure 2A

Heterogeneity analysis.
There was evidence of significant heterogeneity (I 2 > 50%) between OS and PFS studies, but not among DSS studies (Table 4). Therefore, the random-effect model was used in all analyses except for the DSS. For the OS and PFS studies, we conducted a meta-regression analysis using publication year, cancer type, sample size, and country as covariates. All covariates were entered into the meta-regression model simultaneously, and the covariates with the highest p values were omitted one at a time to identify sources of heterogeneity. The meta-regression did not identify any of these covariates as a significant source of heterogeneity for OS and PFS studies in multivariate, and indicated that cancer type may be the source of heterogeneity for OS studies in univariate analysis (Coef = 0.093, p = 0.042) (Supplementary Figure 4). Meta-regression was not used for univariate analysis of the PFS since these studies numbered less than 10.

Sensitivity analysis.
We also performed a sensitivity analysis to evaluate the effects of individual studies on the pooled HR estimates by omitting one study at a time. The HR estimates for the DSS and PFS/MFS/RFS in the multivariate model were not altered. However, the HR estimates for the OS in the multivariate model, and OS and PFS/MFS/ RFS in the univariate model were altered when one, one and two studies were excluded, respectively (data not shown).

Publication bias.
To assess confidence in our study, we performed a publication bias analysis using the funnel plot and Egger's and Begg's rank correlation tests. There was no significant publication bias in both models for the DSS and PFS/MFS/RFS groups (Table 4; Fig. 6D-F). In the case of the OS group, the univariate model (Table 4; Fig. 6C) suggested no publication bias, but in the multivariate model, results were inconsistent based on the p value obtained by Begg's rank test (0.05) and Egger's test (0.007) ( Table 4; Fig. 6A). These test results were nonparametric and therefore the Trim and Fill method was used to further verify this analysis. After filling the deleted studies (square dots), we found no obvious asymmetry in the funnel plot (Fig. 6B). Thus, the HR estimates for the prognostic value of CD147/EMMPRIN were not notably altered (data not shown). This suggests that there was no publication bias even in the OS group in the multivariate model.

Discussion
CD147/EMMPRIN is a glycosylated, multifunctional molecule that participates in tumor progression 61 . In the present study, we quantitatively analyzed the data from 53 studies, including 68 datasets, to examine the associations between CD147/EMMPRIN expression and its prognosis predictive value in cancer. Previous studies have    Table 4. Meta-analysis of association between CD147/EMMPRIN expression and tumor prognosis. P* present P for HR, P# present P for I2. by cancer type. Meanwhile, the predictive role of CD147/EMMPRIN in cervical cancer and hepatocellular carcinoma prognosis has been controversial and its result in cervical carcinoma, endometrial carcinoma,  pancreatobiliary adenocarcinoma and some additional tumors did not consistently reach significance. However, the sample size and studies in our stratified analysis were small, and our findings should be further verified.
Previous meta-analyses didn't explore any significant association between CD147/EMMPRIN and susceptibility of radio-chemotherapy. In the view of more studies are warranted to validate CD147/EMMPRIN association with chemotherapy and radiation resistance prediction [5][6][7]57 . It has been suggested that chemotherapy drugs combined with CD147-targeted therapy may increase the sensitivity of tumor cells to several different chemotherapeutics, and can result in more effective inhibition of tumor proliferation and recurrence 7,67 . Our study also established that increased CD147/EMMPRIN expression is linked to high risk of drug resistance and our preliminary analysis linked CD147 to radiation resistance. It follows that CD147/EMMPRIN has been proposed to be an important potential therapeutic target.
The results of the present study must be interpreted with caution due to the presence of substantial heterogeneity. In addition, there were several limitations in this study. First, the composition of the cancer type or stage varied between studies, and the detection and corresponding cut-offs varied. For instance, some studies only included pediatric patients 26 while others included older patients 10 . Also, in non-small-cell lung cancer, CD147/ EMMPRIN was associated with poor survival in patients with adenocarcinoma only, but not with squamous cell carcinoma 22 . The scoring criteria were also inconsistent. Second, the follow-up time varied across studies, which may have contributed to the non-homogeneity of prognostic information. Sensitivity analysis results also showed instability within individual articles. Third, this study was based on published articles only and, since negative data are hard to publish, there could be publication bias, which is an inherent limitation of all meta-analyses, irrespective of outcomes from the Egger's linear regression test and Begg's rank correlation test.

Conclusion
In summary, this meta-analysis indicated that higher expression of CD147/EMMPRIN potentially may be a prognostic marker for most cancers, and thus can serve as a potential therapeutic target. We further verified that CD147/ EMMPRIN had a complex role in tumor progression by crosstalk with numerous factors. However, additional multicenter prospective studies are warranted to confirm these findings, especially in various types of tumors.

Methods
Study identification. We searched PubMed and Embase databases through March 2015 to identify relevant studies for inclusion in our meta-analysis. The following keywords were used in the literature searches; "CD147", "extracellular matrix metalloproteinase inducer", "EMMPRIN", "basigin", "survival", "prognosis", "tumor", "cancer", "carcinoma", "neoplasm", or their combinations. Eligible articles were selected based on title, abstract and full text. If the same patient cohort was reported in multiple publications, only the most complete and most recent publication was selected. We also searched of the reference lists from electronically identified articles.

Inclusion and exclusion criteria. Based on the Reporting Recommendations for Tumor Marker Prognostic
Studies (REMARK) guidelines, we included studies that met all of the following criteria: (1) written in English; (2) reported quantitative outcomes from prognostic association studies of tumor and CD147/EMMPRIN; (3) described outcomes as overall survival (OS), disease-specific survival (DSS), progression-free survival (PFS), metastasis-free survival (MFS) or recurrence-free survival (RFS), depending upon the study; (4) had a minimum of 40 case numbers describing prognosis; (5) provided a detailed protocol, including the source of raw materials, methodology, quantification methods, and scoring criteria; (6) precisely defined the time-to-event outcome, time to follow-up, and the median follow-up time; and (8) data were presented as the estimated hazard ratios (HRs) with 95% confidence intervals (CIs) or in the case where HRs or 95% CIs were not reported directly, a calculation was used to determine if the conditions of the study were suitable for inclusion 68 . Multivariate analyses were used for statistical analysis biomarkers that had independent prognostic factors for cancers after adjusting for one or more additional standard clinical prognostic variables like age, pathology, stage, grade, lymphatic metastasis or other biological marker variables. Studies were excluded if they were case reports, case-only studies, letters, reviews, reported insufficient data, lacked statistical analysis, combined with other factors, or were duplicate studies. All studies were independently reviewed by two authors, and in a case of conflict a third author resolved the issue after thorough discussion.
Data collection and analysis. We assessed heterogeneity using the Chi 2 test and I 2 test 69 . If heterogeneity was present, meta-regression was used to determine the source. We combined data from different trials using a fixed-effect model when there was no significant heterogeneity in populations (I 2 < 50%) and a random-effect model when there was considerable heterogeneity. If heterogeneity was present, meta-regression was used to determine the source. Variables were synthesized using HR/OR. By convention, an overall HR (OR) > 1 with a 95% CI implied a poor outcome (high risk) for the group with either positive or negative biomarker expression. The high HR value corresponded with poor survival. To evaluate the effects of individual studies on the pooled HR estimates, we performed a sensitivity analysis omitting one study at a time. The statistical significance was set at 0.05. We used funnel plot asymmetry using Egger's linear regression test and Begg's rank correlation test to assess the publication bias. If both test results were inconsistent, Nonparametric Trim and Fill method was used to verify the results 70 . A P value of < 0.05 suggested significant publication bias. All statistical analyses were performed using STATA 12.0 (StataCorp, College Station, TX) statistical software.