The potential of microRNAs as human prostate cancer biomarkers: A meta‐analysis of related studies

Abstract Prostate cancer (PC) is a very important kind of male malignancies. When PC evolves into a stage of hormone resistance or metastasis, the fatality rate is very high. Currently, discoveries and advances in miRNAs as biomarkers have opened the potential for the diagnosis of PC, especially early diagnosis. miRNAs not only can noninvasively or minimally invasively identify PC, but also can provide the data for optimization and personalization of therapy. Moreover, miRNAs have been shown to play an important role to predict prognosis of PC. The purpose of this meta‐analysis is to integrate the currently published expression profile data of miRNAs in PC, and evaluate the value of miRNAs as biomarkers for PC. All of relevant records were selected via electronic databases: Pubmed, Embase, Cochrane, and CNKI based on the assessment of title, abstract, and full text. we extracted mean ± SD or fold change of miRNAs expression levels in PC versus BPH or normal controls. Pooled hazard ratios (HRs) with 95% confidence intervals (CI) for overall survival (OS) and recurrence‐free survival (RFS), were also calculated to detect the relationship between high miRNAs expression and PC prognosis. Selected 104 articles were published in 2007‐2017. According to the inclusion criteria, 104 records were included for this meta‐analysis. The pooled or stratified analyze showed 10 up‐regulated miRNAs (miR‐18a, miR‐34a, miR‐106b, miR‐141, miR‐182, miR‐183, miR‐200a/b, miR‐301a, and miR‐375) and 14 down‐regulated miRNAs (miR‐1, miR‐23b/27b, miR‐30c, miR‐99b, miR‐139‐5p, miR‐152, miR‐187, miR‐204, miR‐205, miR‐224, miR‐452, miR‐505, and let‐7c) had relatively good diagnostic and predictive potential to discriminate PC from BPH/normal controls. Furthermore, high expression of miR‐32 and low expression of let‐7c could be used to differentiate metastatic PC from local/primary PC. Additional interesting findings were that the expression profiles of five miRNAs (miR‐21, miR‐30c, miR‐129, miR‐145, and let‐7c) could predict poor RFS of PC, while the evaluation of miR‐375 was associated with worse OS. miRNAs are important regulators in PC progression. Our results indicate that miRNAs are suitable for predicting the different stages of PC. The detection of miRNAs is an effective way to control patient's prognosis and evaluate therapeutic efficacy. However, large‐scale detections based on common clinical guidelines are still necessary to further validate our conclusions, due to the bias induced by molecular heterogeneity and differences in study design and detection methods.


| Search strategy
We performed a detailed literature search in PubMed, Embase, Cochrane, and Chinese National Knowledge Infrastructure databases to obtain relevant articles for this meta-analysis. Relevant studies were selected according to a combination of keywords and Medical Subject Headings (MeSH): ("prostate cancer" or "prostate neoplasm" or "prostate tumor") and ("microRNAs" or "miRNAs" or "miR-") and ("marker or "biomarker"). All selected studies in English or Chinese were viewed, and their reference lists were also examined for other eligible publications. Most studies were published between 2007and 2017. The last search update was finished on July 8, 2017. These studies regarding miRNAs and PC are performed in clinical samples or PC cell lines. Published data are subject to the limitation of small sample size and selection bias.

| Inclusion and exclusion criteria
More than 1300 articles were retrieved, and 104 publications were included and reviewed in the meta-analysis ( Figure 1). Eligible studies had to fit the following inclusion criteria: (i) a kind of miRNA was involved in the studies; (ii) patients with PC were studied, and gold standard test (eg, histological examination) was used for the PC diagnosis; (iii) prostate tissue or serum or urine samples were used from PC patients or non-PC patients for miRNA expression comparison; and (iv) validation method and enough patients' information were reported. Eligible studies that met above mentioned criteria were further evaluated and excluded according to a selection process showed in Figure 1. Exclusion criteria were as follows: (i) reviews, letters, commentary, or erratum; (ii) non-English or non-Chinese studies; (iii) data was obtained from PC cell lines; (iv) no sufficient data to extract; and (v) duplicate records.

| Data extraction
We assessed the data quality of each publication and extracted the following information: (i) basic features, such as first author, publication year, case region, study design, sample number, validation method, and detected miRNAs, as showed in Table 1; (ii) expression levels or fold-change of detected miRNAs and predictive data, including OS and RFS; and (iii) information needed for quality assessment. If there were no data that could be extracted directly, we used the computer of revman 5.3 software to calculate and generate the relevant data.

| Statistical analysis
We drew forest plots to estimate miRNAs expression levels in PC and control patients' samples, and their effects on PC patients' OS and RFS. Publication bias was explored by funnel plots. 12,13 The fixed-effects model was used to calculate HR and 95% CI in all enrolled studies. 14 We used Chisquared and the inconsistency index (I 2 ) tests to assess the heterogeneities (P value ≤0.1 and I 2 value ≥50 %). To avoid the influence of heterogeneity, subgroup analyses were performed based on the characteristics of included studies, such as patients' ethnicities, pathological types, and detected sample types, etc. All P values were two-tailed and a P value <0.05 was considered to be statistically significant.

| Summary of included studies
A total of 1336 primary literatures were searched in PubMed, Embase, Cochrane, and CNKI. As shown in the selection process ( Figure 1), we firstly removed 49 studies due to duplication. Then, we excluded 980 and 202 studies, respectively, after abstracts and full texts were reviewed. Ultimately, only 104 articles were considered eligible for the meta-analysis. The characteristics of 104 included studies were summarized in Table 1 in alphabetical order of the miRNAs. The publication years of these records ranged from 2007 to 2017. In these 104 studies, some were divided into several parts because of multiple miRNAs. Data of enrolled records were collected from the United States, China, Germany, Greece, Italy, Austria, Korea, and Brazil, etc. The dominant ethnicity was Caucasian in more than half of studies, while 38-2 studies were executed in Asians. Most studies were prospective in design. The expression level of miRNA was usually detected by quantitative real-time polymerase chain reaction (qRT-PCR) and microarray in tissue samples, while 6 + 2 studies were in serum or plasma samples, 6 + 2 studies were in urine (Table 1). Among these studies, 71 records were associated with Mean ± SD and fold-change of miRNA expression level in tumor or control samples (Table 2 and Figures 2-5). A 29 focused on RFS (Table 3 and Figure 6A-E), and 11 focused on OS (Table 4 and Figure 6F). In the analysis of RFS and OS, 26, and 9 records directly reported HRs and 95% CIs, respectively, while in other studies we extrapolated these necessary variables by available original data (Tables 3, 4 and Figure 6).

| miRNAs and PC diagnosis
miRNAs may regulate the wide range of biologic processes, and their deregulation are associated with PC onset, progression, and metastasis. More and more studies investigated differentially expressed miRNA as PC diagnostic and prognostic markers by comparing the expression levels of miRNAs in tumor tissues to that in BPH or normal controls. But there were the high variability in the data obtained from the different records. These could be caused by several factors as follows: (i) different sample groups; (ii) different detecting and verifying methods; (iii) small sample size. Nonetheless, these studies depicted a starting point, and some of the included records screened the same miRNA which was found with the same trend in multiple studies with different methods, as shown in Table 1. However, a confirmed diagnostic miRNA which could be translated into the clinic was not arised. Further confirmed experiments are needed in additional large patient cohorts.

| Publication bias and subgroup analysis
The high heterogeneity between the data from the included records could be associated with several factors: different study design, different races of patients, different methods of sample collection and detection, incomplete information, and small sample size. There were also many difficultly statistical factors: proportion of contaminating cells, limited tumor size and differences in miRNAs stability and processing. In addition, different control samples (BPH or adjacent normal or unmatched normal) and the different characteristics of PC (low/high-risk or metastasis or recurrence) could explain, at least in part, the different results. Significant heterogeneities (P < 0.05, I 2 > 50%) were found in most miRNAs expression profiles, we performed the subgroup analyses to seek the source of heterogeneity, which include ethnicity, sources of control (BPH or N), and sample types (serum/plasma or urine), etc.
To assess publication bias of 11 studies on miR-21, the funnel plot was drawed. As shown in the Figure 3C, significant publication bias was found in the pooled analysis of miR-21 (P < 0.00001, I 2 = 95%), most of the research data was distributed on the edge line. In order to avoid the effect of heterogeneity, we performed four subgroup analyses divided by ethnicity, and sample categories, including: China, Brazil, local versus meta, and PC versus control. Unfortunately, the heterogeneities were significantly reduced in Brazil subgroup   and local versus meta subgroup, while the expression of miR-21 in other subgroups still had obvious heterogeneity. In Brazil subgroup, I2 value was less than 50%, but SD value from the first study by Betina Katz 15 was too large, and covered the scope of the other two data. Therefore, we believe that the study on miR-21 still needs to be further expanded. We did not find corresponding increased miR-21 in Chinese by merging four studies [16][17][18][19] ( Figure 3C). When stratified by the category of detected samples, increased expression of miR-21 showed consistency in local versus meta subgroup, but no statistically significant result was observed in PC versus control subgroup. Five studies on miR-100 had obvious heterogeneity, as shown in the Figure 3F. After carefully reviewing the five full-texts, they were divided into three subgroups, including: Brazil, urine and recurrence and non-recurrence. Among them, heterogeneity in the Brazil subgroup was significantly reduced (P = 0.24, I 2 = 26%). Subgroup analysis of miR-141 expression showed that miR-141 expression was consistently up-regulated in four studies of PC versus control subgroup and more obviously up-regulated in serum samples data from Heather H. Cheng. 20 Four studies on miR-200c had also obvious heterogeneity, as shown in the Figure 3K Six studies on miR-221 were divided into four subgroups: local versus meta, aggressive versus non-aggressive, PC versus control and urine, and the heterogeneity in PC versus control subgroup was significantly reduced to 0%. In addition, existing data showed that the expression of miR-221 in primary PC was less than that in normal tissues, but miR-221 was significantly increased when PC progressed to more malignant stages (metastasis or recurrence or hormone resistance). Among them, Tong's research data were divided into two parts, which were included in local versus meta subgroup and PC versus control subgroup, respectively. The studies on miR-15a and miR-16 were divided into two subgroups: PC versus meta, PC versus control. Results showed that both of miR-15a and miR-16 were up-regulated in metastasis PC, while their expression levels were lower in PC tissues than in non-cancerous tissues. Inconsistently expression of let-7c was reported. The three studies on let-7c were divided into two subgroups: high-risk versus meta and high-risk versus control, and the research data from Katia R. M. Leite 2013 22 were separately counted in the two subgroups because two sets of data were involved. The heterogeneity was significantly reduced to 0% and 12%. The study of miR-143, 145 191, −25-32 was divided into two subgroups, PC versus meta, PC versus control. Moreover, miR-222 and miR-375 were inconsistently expressed in prostate tumor tissues and matched normal tissues ( Figures 3M and 3N). So it was essential to conduct subgroup analyses on miR-222 and miR-375 expression. Five studies on miR-222 could be divided into three subgroups: PC versus control, China, and urine. The study of D Lin was a comparative study on the malignant and non-malignant PC in China. The heterogeneity in PC versus control subgroup significantly decreased to 27%. Five studies on miR-375 were divided into three subgroups: PC versus control, serum, and urine. The heterogeneities in PC versus control and serum subgroups were significantly reduced to 23% and 0%, respectively. The analyses of the above-mentioned subgroups showed that the expression of miR-375 in the urine samples were widely different, and also deviated from the expression profiles of tissues and plasma samples. In addition to the above mentioned miRNAs expression data, there were also significant heterogeneities in (V) miR-505. Squares and horizontal lines correspond to study-specific HRs and 95% CIs; respectively. The area of the squares correlates the weight of each enrolled study and the diamonds represent the summary HRs and 95% CIs the studies on seven miRNAs (Figure 4). Among them, studies on miR-10b, miR-18a, miR-30c, and miR-206, research data from Beatriz A. Walter 23 deviated significantly from other research data. The heterogeneity decreased significantly when we rejected the deviant data. Moreover, in several studies on miR-139-5p and miR-182, Cheng Pang 24 and Fan Feng 21 detected miRNAs expression profiles in whole blood and urine, respectively, which could explain the causes of heterogeneity. Finally, in three studies about miR-146 a, Bin Xu 25 collected ADPC and AIPC patients' samples in China, which were obviously different from the other two studies by Katia R. M. Leite 26,27 in Brazil.

| miRNA expression and recurrence-free survival
Biochemical recurrence (BCR) was considered as the first key point to estimate treatment success after RP. BCR can predate SONG ET AL. the development of metastases and other signs of clinical progression, or ultimately death. Recently, a lot of studies attempted to find miRNAs to be potential predictors for patients with biochemical failure. We summarized previous data in the meta-analysis, miR-30c, miR-129, miR-145, and let-7c were found to have the same trend to predict BCR in eight articles ( Figure 6B-E). While the relationship of miR-21 and BCR were studied in four articles with significant heterogeneity ( Figure 6A). After reviewing the four full texts, we found that Ernest K Amankwah 28 examined the effect of the interaction between obesity and miR-21 expression on PC recurrence. Obese patients were included in the study. Removing the data from Ernest K Amankwah, miR-21 could distinguish biochemical failure patients from non-recurrence ( Figure 6A).

| miRNA expression and overall survival
A total of 11 records comprised OS analysis involving 15 miRNAs (Table 4 and Figure 6F). Among three articles on miR-375, significant heterogeneity was observed (P = 0.03, I 2 = 70%). After reviewing three full texts, we found plasma samples were used in the studies of Hui-ming Lin 29 and Xiaoyi Huang, 30 while serum samples were used in the study of Sven Wach. 31 Removing the data from Sven Wach, the heterogeneity was markedly decreased (P = 0.67, I 2 = 0%) ( Figure 6F). Hence, a fixed model was applied to calculate a pooled RR and 95%CI, and we found that patients with high miR-375 expression had significantly poorer OS compared to low miR-375 expression (RR = 2.93, 95%CI, 1.96-4.40) ( Figure 6F).
In the other eight studies involving 14 miRNAs (Table 4), eight miRNAs (miR-132, miR-150, miR-200a/b/c, miR-429, miR-708, and miR-1290) were showed that increased expression predicted significantly worse OS, and low expression of four miRNAs (miR-23a, miR-23b, miR-221, and miR-224) were associated with poorer OS. Moreover, in the analyses on miR-205 and miR-1207-3p, no statistically significant results were observed. It was worth noting that miR-200a, miR-200b, miR-200c, and miR-429 were the members of the same family, their change trends were consistent in different studies, and all of them were associated with poorer OS.

| DISCUSSION
The major challenges for PC clinical management were its accurate diagnosis and dynamic monitoring after RP, chemotherapy or radiotherapy, etc. Although PSA routinely screening improved the ratio of early detection, its levels was poorly associated with tumor aggressiveness, and had a little help to predict PC patients' prognosis. Moreover, biopsies were not only invasive but also not conclusive, for example, sampling errors could lead to missed diagnosis and wrong therapies in clinic, especially in the cases with multifocal PC.
Recently, miRNAs had been found to be closely associated with a variety of tumors by regulating their target genes to affect carcinogenesis and progression. And a number of researches showed a significant correlation between the expression levels of miRNAs and the diagnosis and prognosis of PC. These study data would be helpful miRNAs as biomarkers to be transfer into the clinical application for diagnosis and prognosis of PC. Moreover, Compared to mRNAs, clinical samples containing miRNAs are more likely to be collected and detested because miRNAs are stable not to be easily degraded. The expression profiles of miRNAs are (C) miR-30c; (D) miR-139-5p; (E) miR-146a; (F) miR-182; (G) miR-206. Squares and horizontal lines correspond to study-specific HRs and 95% CIs; respectively. The area of the squares correlates the weight of each enrolled study and the diamonds represent the summary HRs and 95% CIs SONG ET AL. special in various cancer or normal tissues. And they can be accurately quantified by microarray, qRT-PCR, and RNA sequencing in not only frozen or fresh or formalin-fixed paraffin-embedded tissues, but also serum or plasma samples, even in urine or saliva samples. However, these results on the clinical value of miRNAs were inconsistent and even contradictory due to the clinical complexity of PC. Therefore, it is necessary to conduct stratified and systematic analyses to confirm their expression pattern and application scope.
By meta-analyses of included studies, we successfully come to some valuable conclusions for future applications in clinic. The most studied miRNA was miR-21, with 11 articles providing the data of its expression level in clinical PC samples. Secondly, the expression profile data of miR-221 and miR-205 were clearly reported in seven and six studies, respectively. And the expression levels of 7 miRNAs (miR-25, miR-32, miR-100, miR-125b, miR-141, miR-222, miR-375) were reported in five literatures. In addition, the most obviously increased miRNAs were the members of the miR-200 family: miR-200a and miR-200b, their HR and 95% CI were 5.17 (3.22-7.13) and 4.08 (2.91-5.24), respectively. The most significantly decreased miRNA was miR-199a, its pooled HR and 95%CI was −4.23 (−16.22-7.76).
In order to remove the interference of genetic backgrounds due to patients' ethnic groups, the included studies were classified into China subgroup and Brazil subgroup, etc. We found increased miR-21 expression could distinguish PC patients from normal controls, and could predict a significantly poor RFS. The expression of miR-100 in the Brazilian population was significantly reduced, and HR and 95%CI was Second, we conducted subgroup analyses depending on the pathological types of PC to classify the enrolled studies into subgroups of cancer categories: normal controls/BPH, primary/local PC, metastatic PC, high-risk PC, and (N) miR-548c. Squares and horizontal lines correspond to study-specific HRs and 95% CIs; respectively. The area of the squares correlates the weight of each enrolled study and the diamonds represent the summary HRs and 95% CIs recurrence PC/non-recurrence PC subgroups, etc. In the comparisons of the expression profiles of miRNAs in primary/local PC versus metastatic PC subgroups, we found that miR-21 and miR-32 were up-regulated in metastatic PC tissues, while miR-25, miR-143, miR-145, miR-191, and let-7c were down-regulated. In subgroup analyses of PC versus control, we found that miR-141, miR-200c, and miR-375 were increased, while miR-30 c, miR-143, miR-145, miR-191, miR-221, miR-222, and let-7c were reduced. Among them, low expression of three miRNAs (miR-30c, miR-45, and let-7c) predicted worse RFS, the HR 95%CI were 0.32 (0.15-0.66), 3.86 (1.85-8.03), and 3.14 (1.49-6.60), respectively. In addition, the expression model of miR-15a and miR-16 was special, both of them were lower expressed in PC tissues than that in normal controls, and their expression levels were increased again when PC progressed to malignant metastatic stages.
Finally, we performed subgroup analyses to clarify the diagnostic values of miRNAs based on the data of serum/ plasma and urine samples, etc. We found that high-expression of miR-375 was significantly associated with a worse OS (HR = 2.93, 95%CI 1. 96-4.40) in serum/plasma subgroup, and its high-expression was also shown in tissue subgroup (HR = 7.41, 95%CI 6.49-8.33) and urine subgroup (HR = 1799.29, 95%CI 1796. 45-1802.13). In addition, we processed subgroup analyses of the expression levels of other five miRNAs in serum or urine samples. Among them, the members of miR-200 family: miR-141 and miR-200c were up-regulated, while the expression levels of miR-221 and miR-221 were decreased. The expression of miR-100 was increased in the urine samples, which was contrary to its expression in patients' tissues. Although the detection of miRNAs in tissues was widely accepted by researchers and doctors to diagnose and predict PC progression, the detection in serum or urine samples was more convenient and uninjurious, which could dynamically monitor the therapeutic effects and patients' prognosis at any time point of the lifetime of PC patients.  The meta-analysis has some merits. First, we strictly followed the literature inclusion criteria and the quality of enrolled literatures was satisfactory. Second, we conducted subgroup analyses to effectively minimize the influence of heterogeneity among the enrolled studies, and to further explore the scope of application for miRNAs as a prognostic biomarker of malignant tumors. All of these have increased the statistical power of the meta-analysis. But there are also many shortcomings in the meta-analysis. First, only a few articles are eligible for a kind of miRNA leading to the relative shortage in subgroup analyses. Secondly, after data integration and subgroup analyses, some miRNAs data still lack statistical significance, such as miR-15a (P = 0.45), miR-16 (P = 0.69), miR-21 (P = 0.49), miR-25 (P = 0.83), miR-191 (P = 0.49), and miR-200c (P = 0.63), etc. Besides, no study is carried out in Africa, which blocks the integrated investigation of the association between miRNAs expression and PC diagnosis and prognosis. Finally, because of the lack of unified cut-off value of miRNAs expression in different researches, which would reduce the potency of miRNAs as predictive biomarkers. Therefore, the application value of miRNAs as prognostic factors for PC is still controversial, requiring more researches to verify.

| CONCLUSIONS
The potential use of miRNAs as diagnosis and prognosis factors for PC in the clinic was based on a growing body of investigations in the last decades. Currently, ongoing researches were still controversial that delayed the transformation from bench to bedside. Nevertheless, the potential value of miRNAs used in clinical practice had been generally accepted, which represented not only promising biomarkers for PC but also candidated therapeutic targets. Besides, detecting the expression levels of miRNAs in serum or plasma or urine samples was more exciting than detecting miRNAs in tissues, because of low cost, rapid test, and noninvasion, etc. However, in this meta-analysis, we found that the expression profiles of miRNAs in the blood samples were different from that of the tissues, and the deviation in the urine samples was more obvious.
Due to the lack of relevant data, further studies in larger sample sizes are needed to conduct more precise stratification between miRNAs expression levels and different progression stages of PC. We will also continue to evaluate and report the clinical value of miRNAs detection when larger studies further verify the validity of miRNAs. The guidelines on study design and sample collection still need to be further improved, in particular, the detecting platforms should be clearly defined. Taken together, the meta-analysis underline that the use of miRNAs as biomarkers for diagnosis and prediction of PC is promising, though not yet a reality in clinical practice.