MicroRNA-21 as a Potential Biomarker for Colorectal Cancer Diagnosis: A Meta-analysis

Aims: Colorectal cancer (CRC) occupies an important position in the morbidity and mortality constitution of malignancies. In recent years, mounting literature has reported about the upregulating expression of microRNA-21 in blood and stool of CRC patients, which suggested that microRNA-21 may become a novel potential biomarker for CRC. Consequently, this meta-analysis was designed to systematically review the values of microRNA-21 in CRC diagnosis. but the specificity is favorable in predicting CRC patients.


INTRODUCTION
In recent years, morbidity and mortality constitution of tumors published worldwide has demonstrated that [1][2][3] colorectal cancer is increasingly threatening to human health. Although early detection, early diagnosis, and early treatment are emphasized in clinical practice, limitations still exist in the present imaging technology for early diagnosis of colorectal cancer (CRC) [4,5]. For example, nuclear magnetic resonance imaging, computed tomography, and positron emission computed tomography show limitations of high cost, selective sizes of tumor tissues, and disconfirmation; air barium double-contrast radiography examination, transrectal ultrasound, and colonoscopy are usually rejected because of discomfort during operation [6]. Furthermore, serum tumor markers [7] and fecal occult-blood testing (FOBT) [8] exhibit low sensitivity and unsatisfactory specificity.

MicroRNAs
(miRNAs) are evolutionarily conserved, endogenous, and non-coding RNA molecules, which consist of 17-25 nucleotides and can identify specific target mRNAs to act as negative gene regulators at post-transcriptional level. Considerable research has found that abnormal miRNA expression emerges in the occurrence and development of CRC, and tumorspecific miRNA exists in CRC patients' serum, plasma [9,10], feces [11] or tissue [12,13]; all of which can be stably detected and can predict tumor status; such discoveries ensure a promising application prospect for miRNAs as novel tumor biomarkers.
As a widely studied oncogenic miRNA, miR-21 has been notably upregulated in CRC patients. Previous studies proposed that miR-21 is positively correlated with pathological staging of CRC and negatively correlated with disease-free interval [14,15]. Additionally, miR-21 is considered a powerful prognostic marker applicable to CRC patients from various racial populations, and it can identify the progression of high-risk diseases in the early stage of cancer [16,17]. This meta-analysis was conducted to evaluate the clinical values of miR-21 in tissues, feces, serum, and plasma for systematic diagnosis of CRC.

Search Strategy
All relevant articles were searched via the following electronic databases: Cochrane library, PubMed, EMbase, Google Scholar, and Chinese National Knowledge Infrastructure until January 10, 2015. Brief search strategies were as follows: ("Subject headings" or colorectal cancer or colon cancer or rectal cancer), ("Subject headings" or microRNA-21 or miRNA-21), ("Subject headings" or sensitivity or specificity), and other precise search strategies of PubMed (Table 1).

Inclusion criteria
(1) all CRC patients were pathologically verified (gold standard); (2) CRC patients did not receive radiotherapy, chemotherapy, surgery nor any other treatment before sample collection; (3) studies had to contain evaluating indicators of microRNA-21 used alone for CRC diagnosis: sensitivity or specificity; (4) normal controls were set; (5) for similar literature published by the same author or research center, those with more recent publication time or higher influence factor were included，and resemble studies focusing on different indicators should all be involved; (6) cases of each group should be ≥10; (7) no restriction in age, gender, nationality, and race, and original articles published in either English or Chinese.

Data Extraction and Quality Assessment
On the basis of the above mentioned inclusion and exclusion criteria, two investigators (Chaohui ZHEN and Guojun YAO) independently conducted literature selection, quality assessment, and data extraction. Any disagreement would be fully discussed between the former two researchers, or under the assistance of the third senior investigator to reach a consensus. The final enrolled articles were assessed by referring to quality assessment of diagnosis accuracy studies-2 (QUADAS-2) [18]. In diagnostic tests, QUADAS-2 was generally recognized as a quality assessing tool comprising four domains: (1) patient selection; (2) index test; (3) reference standard; and (4) flow and timing. The signaling questions of each domain are rated as "high," "unclear," and "low" to judge on risk of bias and applicability.

Statistical Analysis
All statistical analyses were performed by Meta-DiSc and STATA12.0 statistical software. By extracting all relevant data (true positives, false positives, true negatives, and false negatives) from the enrolled studies, the pooled sensitivity (SEN), specificity (SPE), positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnostic odds ratio (DOR), and 95% confidence interval (95% CI) were obtained. Simultaneously, the investigators generated the summary receiver operating characteristic (SROC) curve and calculated the area under the curve (AUC). The analysis steps were as follows: (1) Heterogeneity test: a. The threshold effect was assessed using Spearman correlation coefficient [19]; a P value < 0.05 suggested the presence of the threshold effect requiring no meta-analysis; b. The nonthreshold effect was detected by Cochran-Q method and test of inconsistency index (I2) [20]; a low P value of <0.05 or a high I2 value of >50% suggested the existence of non-threshold effect, and meta-regression analysis was consequently used to determine the sources; (2) Metaregression analysis: The research-related covariates were compared, which were eliminated one-by-one based on the P values (from high to low); the covariates causing heterogeneity were determined, and subgroup analysis was performed; (3) Quantitative synthesis: a. Fixed-effects model (Mantel-Haenszel method) was used to combine the data if heterogeneity did not exist; b. Subgroup analysis or random-effects model (DerSimonian-Laird method) was applied to combine the data if non-threshold effect existed; (4) Publication bias was assessed by Begg's test and Egger's test [21]; a P value ≥ 0.05 indicated no publication bias.

Characteristics of Eligible Studies
After selection based on PRISMA statement [22], about 15 studies with a total of 1268 CRC patients and 910 healthy controls were eventually eligible. Fig. 1 shows the study selection process, and Table 2 lists the main characteristics of the eligible studies.

Quality Assessment
The quality of the included studies was assessed using QUADAS-2 quality assessment. All 15 inclusions reached middle to high quality (Figs. 2 and 3). However, no blinding method was set in the detection of gold standard. Hence, a major bias was concentrated upon "index test"; bias also existed in "patient selection" because most of the research did not report nor recruit patients randomly.

Heterogeneity Test
The heterogeneity among the studies is a critical factor influencing the accuracy of the synthesis.

Meta-regression
Meta-regression was employed to explain the research-related covariates causing nonthreshold effect. These covariates included race, internal control, assay type and study quality. Unfortunately, no satisfactory clues were found.

Quantitative Synthesis
Research without non-threshold effect should use FEM to merge data. The potential heterogeneity caused by the non-threshold effect in studies of some sample types should also

Publication bias
Publication bias is another crucial factor influencing the estimating accuracy in metaanalysis of diagnostic tests. Begg's test and Egger's test were used in this meta-analysis. The P values for Begg's test and Egger's test were 0.404 and 0.187, respectively, which suggest that no publication bias existed among the included studies focusing on blood samples (Fig. 9). The number of studies on fecal samples was very small for use in Begg's test and Egger's test.

DISCUSSION
To explore novel tumor markers for clinical applications, we performed this meta-analysis to review the values of miR-21 in different samples for systematic diagnosis of CRC. In this metaanalysis, discrepant expression levels of miR-21 in the blood, feces, and tissues showed obvious statistical significance between the CRC patients and the control individuals. For miR-21 in blood samples, the pooled SEN, SPE, PLR,

Fig. 1. The flow chart of data identification and selection
and NLR were 0.71 (95% CI: 0.67-0.74), 0.78 (95% CI: 0.74-0.83), 3.53 (95% CI: 2.35-5.30), and 0.37 (95% CI: 0.27-0.49), respectively, thus indicating prominent advantage in sensitivity and specificity of miR-21 in blood samples for the diagnosis of CRC. DOR [38] reflects the correlation between the diagnostic results and diseases. When DOR is >1, a larger value suggests better diagnostic performance. AUC [39] is used to evaluate the overall performance, and its optimal value tends to be 1. The pooled DOR and AUC of miR-21 in blood samples were 10 4.77 (95% CI: 1.64-13.87), and 0.69 (95% CI: 0.51-0.93), respectively, indicating that fecal miR-21 exhibits favorable SPE, even though its SEN is unsatisfactory. When combined with multiple miRNAs, fecal miR-21 can decline the misdiagnosis rate of CRC. There were only 2 researches focusing on tissue miR-21 in the present study, and meta analysis was not recommended because the result of the combined data is not persuasive. Nevertheless, both single-center studies suggested high expression of miR-21 in colorectal cancer tissue, which will provide theoretical basis for future researches on other specimen types.    [27,31]. All in all, the results indicated that miR-21 has a potential diagnostic value with moderate sensitivity and good specificity for CRC. Combined with our research, the diagnostic performance of miR-21 in blood samples is better than that of fecal samples, and serum samples are more beneficial than plasma for miRNA extraction and detection.
For any meta-analysis, completely avoiding the heterogeneity existing in the eligible studies [40] is indispensable; even publishers themselves prefer positive results, and publication bias thus emerges [41]. Fortunately, no threshold effect or publication bias was observed in our metaanalysis. Examination showed that non-threshold effect existed in the selected articles; however, meta-regression did not determine the responsible covariates. Hence, we combined the data by using REM to improve the accuracy of the combined results.
The prognosis of CRC patients is inversely related to tumor staging, early diagnosis, and early treatment, which can significantly decrease its mortality and recurrence. Undoubtedly, colonoscopy is a method with the maximum diagnostic performance, but it is rejected by most of patients because of its invasive nature [6]. Moreover, noninvasive examinations, including carcinoembryonic antigen [7] and FOBT [8] cannot work efficiently because of poor sensitivity and specificity. The occurrence, staging, metastasis, and recurrence of CRC are closely related to the abnormal expression of miRNAs, and their stable existence in serum, plasma [9,10], and feces [11] has facilitated clinical applications of miRNAs. Furthermore, our meta-analysis demonstrated that miR-21 in the blood yields preferable performance in the diagnosis of CRC; fecal miR-21 exhibits satisfactory specificity, and it can reduce the misdiagnosis rate in cooperation with other indicators. This evidence suggested that miR-21 is a potential powerful biomarker in CRC diagnosis.
Although its results are promising, this metaanalysis still has limitations. First, miR-21 has only recently emerged as a novel marker in CRC diagnosis, thus, our meta-analysis only includes a small sample size. Therefore, further validations of miR-21 in large cohort and independent studies are needed. Second, many other cancer types dysregulate miR-21. However, studies that distinguish CRC from other cancer types are lacking, which may limit the application of miR-21 in clinical settings. Further comprehensive studies are needed to solve this problem. Furthermore, the detecting methods for miRNAs are all based on qRT-PCR. Unified primers and reference miRNAs for qPCR analysis remain unavailable to date.

CONCLUSION
In conclusion, our research demonstrated that miR-21 in blood samples has better diagnostic performance than that in feces, and serum samples are even better for the detection of miR-21. Fecal miR-21 exhibits favorable specificity but with poor sensitivity. However, several limitations exist in this meta-analysis. First, the implementation of the gold standard did not meet the requirements of the blind method; second, case selection did not strictly abide the continuous random inclusion; finally, small sample size was used in some selected studies. Consequently, blind-designed random studies in large scale should be strictly conducted by multicenter medical institutions locally and internationally to achieve more authoritative results.

CONSENT
It is not applicable.

ETHICAL APPROVAL
It is not applicable.