Introduction

MircoRNAs (miRNAs) are small and single-stranded noncoding RNAs consist of approximately 18–22 evolutionarily conserved nucleotides in length. By binding complementary sequences in the 3′ untranslated region (3′-UTR) of mRNAs, miRNAs either mediate translational suppression or direct mRNAs degradation. Consequently, the mRNAs translation repression or destabilization will lead to down-regulated expression of the encoded proteins1, 2. Meanwhile, as more than half of the sequences encoding miRNAs are located in tumor-associated genomic regions or fragile sites and accumulated evidence has revealed that miRNAs may participate in various cancer-related biological processes including apoptosis, differentiation, proliferation, stress response and metabolism3,4,5, it is widely-accepted that deregulated expression of miRNAs might be used as a novel kind of biomarkers for early cancer diagnosis or prognosis prediction6.

The miR-17-92 cluster is a typical and most extensively studied example of miRNAs, which located at chromosomal locus 13q31.3 and encoded the miR-17, miR-18a, miR-19a/b, miR-20a, and miR-92a. Recent studies have reported that miR-17-92 cluster was frequently overexpressed in various cancer types, and played critically suppresive role in degradation or inhibition of its target genes7,8,9. Despite the miR-17-92 cluster has shown great potential in prediction of cancer prognosis, the concretely prognostic value of highly-expressed miR-17-92 cluster in various cancer types remains controversial. However, meta-analysis can explore the authentic and comprehensive results through incorporating all available evidences to get a relatively precise and accurate estimation using statistical software10. Thus, we conduct a meta-analysis to assess the possible correlations between high-expression of miR-17-92 cluster with cancer prognosis, which efforts may hold great promise in exploring some novel potential biomarkers for monitoring therapeutic efficacy and prognosis of various cancers.

Methods

Ethics statement

The PRISMA statement was used to performed the current meta-analysis11. No patient’s privacy or clinical samples were involved in this study, hence the ethical approval was not required.

Search strategy

Literature resources including PubMed, Cochrane Library, Embase, CBM and CNKI were searched for eligible literatures, using the terms (“microRNA OR miRNA OR miR-17 OR miR-18 OR miR-19 OR miR-20 OR miR-92 OR miR-17-92 cluster”), (“survival OR prognosis OR prognostic”) and (“cancer OR tumor OR tumour OR neoplasm OR neoplasma OR neoplasia OR carcinoma OR cancers OR tumors OR tumours OR neoplasms OR neoplasmas OR neoplasias OR carcinomas”). Last search of current investigation was updated on November 25, 2016. Additionally, the publication language was only limited to English and Chinese. In case of omission, we identified the reference lists of the relevant articles and review articles to seek for the potentially relevant studies. We did not contract the corresponding authors if the relevant data were unavailable.

Inclusion and exclusion criteria

Studies met the following criteria could be identified: (1) clinical study about the association of high-expression of miR-17-92 cluster with cancer prognostic value; (2) relevant available data of the hazard ratio (HRs) and their corresponding 95% confidence interval (CIs) to evaluate its associations could be obtained; (3) patients prognostic outcomes including overall survival (OS), cancer-specific survival (CSS), relapse-free survival (RFS), progression-free sur-vival (PFS), disease-free survival (DFS). Studies met the following four criteria were excluded: (1) the available data regarding about associations were absent; (2) similar or duplicate study (when the same or similar cohort was applied, after careful examination, the most complete information was included); (3) other types of articles including reviews or abstracts; (4) studies were involved with cells lines or animal models.

Data extraction

Based on the inclusion and exclusion criteria, we extracted the relevant information from each eligible publication. If disagreements were noticed, we are clearly open to discussion by each other (Kaiping Zhang and Li Zhang), or reviewed by a third author (Min Chao). The data on first author, publication year, study country, age, cancer type, stage range, miRNAs category, sample source, follow-up time, test method, sample size, survival outcome, analysis method, HR and 95%CI and the cut-off value were extracted. We have not contacted any author of the original researches even though the essential information could not be available. Besides, study country came from China and others. Sample source was stratified into tissue, blood, formalin-fixed and paraffin-embedded (FFPE) and tissue microarray (TMA). Test method included immunohistochemistry (IHC), in situ hybridization (ISH) and reverse transcription-polymerase chain reaction (RT-PCR); sample sizes were separated into ≥100 and <100 and cancer types included solid cancer and others. Analysis methods were divided into univariate analysis and multivariate analysis.

Statistical analysis

We explored the association of high-expression of miR-17-92 cluster with cancer prognostic value by applying Review Manager software (RevMan 5, The Cochrane Collaboration, Oxford, UK) and STATA software (Version 12.0, Stata Corpotation, College Station, TX). HR and 95% CI were collected for assessing the prognostic value of highly-expressed miR-17-92 cluster in various cancers. Meanwhile, the heterogeneity has been assessed via chi-square-based Q and I2 test across studies (no heterogeneity I2<25%, moderate heterogeneity I2 = 25–50%, extreme heterogeneity I2 > 50%)12. In case of extreme heterogeneity (I2 > 50% or P < 0.01 for Q test), we used random-effects (DerSimonian and Laird method) model13. Otherwise, fixed-effects (Mantel-Haenszel method) model was introduced14. One-way sensitivity analyses individually removed publications in meta-analysis were conducted to assess results’ stability. It mainly explored the impact of specific study upon mixed HR. The Begg’s funnel plot was performed to evaluated the publication bias. P value less than 0.05 indicated that there was a bias of study15. Additionally, different subgroups consisted of country, test method, sample source, miR-17-92 component, sample size and cancer type were conducted.

Results

Characteristics of the studies

As a result, 36 studies consisted of 4965 samples satisfied the eligible studies16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51 (Fig. 1). The principal characteristics of the included studies were summarized in Table 1. Of these studies, Li’s18 study was involved with two different cohorts of Tianjin cohort and Xiangya cohort. Wu et al.27 designed a multiphase study to identify tissue and serum miRNAs expression, respectively. Fang et al.30 studied miRNAs expression profiles in colorectal cancer (CRC) patients, comparing chemoresistant and chemosensitive groups. Patients were randomly divided into the training set, internal testing set and independent validation set to search for prognostic value of highly-expressed miR-17-92 cluster in stage II colon cancer by Zhang et al.31. Ma et al.43 conducted a study to detect the prognostic value of high-expression of miR-17-5p in CRC using RT-PCR and ISH methods. The expression of miRNAs was measured by RT-PCR in tissues from non small cell lung cancer (NSCLC) patients that originated from Maryland, Norway and Japan by Saito et al.46. As mentioned above, we enrolled them independently into meta-analysis. Eventually, this meta-analysis was established based on 44 studies (Table 2).

Figure 1
figure 1

Flow diagram of the study selection process in the meta-analysis.

Full size image
Table 1 Main characteristics of the eligible studies.
Full size table
Table 2 miR-17-92 cluster evaluation and survival data of the selected studies.
Full size table

Of 44 studies, 41 were written in English and 3 were published in Chinese. The sample sizes ranged from 30 to 376. The included tumor types were as follows: 1 burkitt lym phoma (BL), 4 gastric cancer (GC), 1 T-cell lymphoblastic lymphoma (T-LBL), 8 CRC, 1 renal cell carcinoma (RCC), 2 esophageal squamous cell carcinoma (ESCC), 1 multiple myeloma (MM), 7 NSCLC, 5 colon cancer (CC), 1 adenoid cystic carcinoma (ACC), 1 breast cancer (BCa), 4 hepatocellular carcinoma (HCC), 1 gastrointestinal cancer (GIC), 1 epithelial ovarian cancer (EOC), 2 pancreatic cancer (PC), 1 glioblastoma (GBM), 1 lung cancer (LC), 2 prostate cancer (PCa). Meanwhile, 7 ISH, 35 RT-PCR, 1 IHC and 1 HiSeq in test methods were applied. According to the sample source, there were 12 formalin-fixed and paraffin-embedded (FFPE); 16 tissue, 13 blood and 3 tissue microarray (TMA). As for the survival outcomes, 44 eligible studies were divided into 51 datasets: 31 for OS, 3 for PFS, 4 for RFS, 10 for DFS and 3 for CSS. However, the cut-off value for the high-expression of miR-17-92 cluster was inconsistent among these included studies (Table 2).

Meta-analysis of OS

In univariate analysis, 21 studies were involved in current meta-analysis to assess the prognostic value of highly-expressed miR-17-92 cluster in tumors. High-expression of miR-17-92 cluster in various tumors was associated with unfavorable OS (HR = 2.05, 95%CI: 1.58-2.65, P < 0.001) (Fig. 2A). Besides, it seemed that there were certain associations via sub-analyses regarding country, test method, sample source, miR-17-92 component, sample size and cancer type. (Table 3)

Figure 2
figure 2

Forest plot of the association between high-expression of miR-17-92 cluster in various tumors and OS under different types of analysis. (A) Univariate analysis; (B) multivariate analysis). The squares and horizontal lines correspond to the study-specific HR and 95% CI. The area of the squares reflects the weight. The diamond represents the summary HR and 95% CI. CI = confidence interval, HR = hazard ratio.

Full size image
Table 3 Stratifi ed analysis of the high-expression of miR-17-92 cluster and overall survival.
Full size table

In multivariate analysis, 27 studies were included to assess the prognostic value of miR-17-92 cluster. Consequently, high-expression of miR-17-92 cluster in various tumors was associated with unfavorable OS (HR = 2.14, 95%CI: 1.75–2.61, P < 0.001) (Fig. 2B). Likewise, a similar result was found in different subgroups (Table 3).

Meta-analysis of DFS

7 studies and 6 studies were included in univariate analysis and multivariate analysis, respectively. Ultimately, we found that high-expression of miR-17-92 cluster was linked with poor DFS both in univariate analysis (HR = 1.96, 95%CI: 1.55-2.48, P < 0.001) (Fig. 3A) and multivariate analysis (HR = 2.18, 95%CI: 1.63–2.91, P < 0.001) (Fig. 3B).

Figure 3
figure 3

Forest plot of the association between high-expression of miR-17-92 cluster in various tumors and DFS under different types of analysis. (A) Univariate analysis; (B) multivariate analysis). The squares and horizontal lines correspond to the study-specific HR and 95% CI. The area of the squares reflects the weight. The diamond represents the summary HR and 95% CI. CI = confidence interval, HR = hazard ratio.

Full size image

Meta-analysis of RFS/PFS/CSS

In univariate analysis, there were 5studies, 1 study and 3 studies involved with RFS, PFS and CSS, respectively. Correspondingly, 3 studies, 3 studies and 1 study were collected in multivariate analysis. No association of high-expression of miR-17-92 cluster was detected with RFS (Univariate: HR = 1.24, 95%CI: 0.56–2.73, P = 0.597; Multivariate: HR = 2.21, 95%CI: 0.24–20.33, P = 0.482) (Fig. 4A and B). We also explored that high-expression of miR-17-92 cluster was associated with favorable PFS (Univariate: HR = 0.36, 95%CI: 0.16–0.80, P = 0.012; Multivariate: HR = 1.55, 95%CI: 0.79–3.05, P = 0.201), and poor CSS in the univariate (HR = 1.77, 95%CI: 1.21–2.60, P = 0.004) rather than multivariate analysis (HR = 1.77, 95%CI: 0.80–3.92, P = 0.160).

Figure 4
figure 4

Forest plot of the association between high-expression of miR-17-92 cluster in various tumors and RFS under different types of analysis. (A) Univariate analysis; (B) multivariate analysis). The squares and horizontal lines correspond to the study-specific HR and 95% CI. The area of the squares reflects the weight. The diamond represents the summary HR and 95% CI. CI = confidence interval, HR = hazard ratio.

Full size image

Sensitivity analysis

Each single study here was deleted at a time to assess the specific effect of the individual data on the pooled HRs, and one-way sensitivity analysis suggested pooled results were relatively stable. Among them, the pooled results of OS, DFS and RFS in both univariate and multivariate analyses were shown in Fig. 5A and B, Fig. 6A and B, Fig. 7A and B, respectively.

Figure 5
figure 5

One-way sensitivity analysis of high-expression of miR-17-92 cluster in various tumors with OS under different types of analysis. (A) Univariate analysis; (B) multivariate analysis). Individually removed the studies and suggested that the results of this meta-analysis were stable.

Full size image
Figure 6
figure 6

One-way sensitivity analysis of high-expression of miR-17-92 cluster in various tumors with DFS under different types of analysis. (A) Univariate analysis; (B) multivariate analysis). Individually removed the studies and suggested that the results of this meta-analysis were stable.

Full size image
Figure 7
figure 7

One-way sensitivity analysis of high-expression of miR-17-92 cluster in various tumors with RFS under different types of analysis. (A) Univariate analysis; (B) multivariate analysis). Individually removed the studies and suggested that the results of this meta-analysis were stable.

Full size image

Publication bias evaluation

In univariate analysis, Begg’s funnel plot indicated that publication bias was not found in meta-analysis of OS (P = 0.822 > 0.05, Fig. 8A), DFS (P = 0.764 > 0.05, Fig. 8C), RFS (P = 0.462 > 0.05, Fig. 8E). Meanwhile, in multivariate analysis, there was no publication bias of OS (P = 0.059 > 0.05, Fig. 8B), DFS (P = 0.348 > 0.05, Fig. 8D), RFS (P = 1.000 > 0.05, Fig. 8F) from Begg’s funnel plot. Moreover, no publication bias was found in each subgroup of mata-analysis of OS. However, we did not evaluated the publication bias for the CSS/PFS meta-analysis because of fewer datasets for meta-analysis.

Figure 8
figure 8

Begg’s funnel plot for publication bias test. (A) OS of high-expression of miR-17-92 cluster in various tumors under univariate analysis; (B) OS of high-expression of miR-17-92 cluster in various tumors under multivariate analysis; (C) DFS of high-expression of miR-17-92 cluster in various tumors under univariate analysis; (D) DFS of high-expression of miR-17-92 cluster in various tumors under multivariate analysis; (E) RFS of high-expression of miR-17-92 cluster in various tumors under univariate analysis; (F) RFS of high-expression of miR-17-92 cluster in various tumors under multivariate analysis;). The x-axis is log (HR), and the y-axis is natural logarithm of HR. The horizontal line in the figure represents the overall estimated log (HR). The two diagonal lines indicate the pseudo 95% confidence limits of the effect estimate. Log (HR) = log-transformed HR, HR = hazard ratio.

Full size image

Discussion

In recent decades, to explore the clinically useful cancer signatures remains to be the focus of research due to the complexity of cancer. Fortunately, considerable progresses have been achieved to identify the combinatory cancer hallmark-based gene signature sets (CSS sets) for prognostic indicators and therapeutic strategy design. For instance, a model of seven-gene signatures (NHLRC3, ZDHHC21, PRR14L, CCBL1, PTPRB, PNPO and PPIP5K2) was applied to predict OS by dividing colorectal cancer (CRC) patients into low-risk and high-risk groups. Consequently, the poorer OS was detected in high-risk group compared with low-risk CRC patients52. Additionally, the benefit from adjuvant chemotherapy for patients with stage II CRC after surgery remains a matter of debate53. Gao et al. analyzed data from approximately 1000 patients with stage II CRC from 13 independent cohorts and explored eight CSS sets for determining prognosis of patients. The CSS sets accurately stratified patients into low-, intermediate-, high-risk groups, and predicted five-year RFS rates were 94%, 78% and 45%, respectively for 60%, 28% and 12% of patients with stage II disease. Meanwhile, they have addressed that CSS sets-defined high-risk patients with stage II CRC could gain survival benefit from fluorouracil-based adjuvant chemotherapy54.

Apart from CSS sets, miRNAs have also added some novel insights into cancer diagnosis and therapy. Meanwhile, as miRNAs mainly regulated gene expression by targeting the 3′-UTR of mRNAs, the post-transcriptional effects could consequently make the target genes potential cancer hallmarks in clinical trials. For instance, the miR-17-92 cluster consisted of miR-17, miR-18a, miR-19a/b, miR-20a, and miR-92a, has been proven to play significantly regulatory roles in development, progression and prognosis of several cancer types. It was previously reported that miR-18-mediated low-expression of target gene TGF-β intimately contributed to prolonged survival of glioblastoma multiforme (GBM) patients55. Similarly, miR-19 could exhibit modulatory effect on GSK-3β/β-catenin axis, down-regulated target gene GSK-3β could promote the metastatic potential of lung cancer cells, revealing a poor survival outcome for cancer patients56, 57. Besides, miR-20a was able to suppress the hepatocellular cancer cell proliferation and migration by directly targeted RUNX3, while clinical evidences have proved that RUNX3 was negatively associated with tumor progression, lymph node metastasis and poor prognosis58, 59. Collectively, the gene targets of miR-17-92 cluster could potentially serve as different cancer hallmarks. These cancer hallmarks constituted an organizing principle that provided a logical framework for understanding the remarkable diversities of neoplastic diseases60.

Although it appears reasonable to identify the clinically prognostic value of the dysregulated expression of miR-17-92 cluster itself in various cancers, the investigations focused on the clinical correlation between high-expression of miR-17-92 cluster with cancer prognosis were relatively rare and inconclusive. Moreover, small sample-sized studies lacking statistical power often have resulted in apparently contradicting conclusions. Meta-analysis is a useful tool for providing convincing evidence as it could present inconsistent results from different investigations to get a relatively precise estimation. As far as we know, the current meta-analysis is the first try to comprehensively assess the correlation of high-expression of miR-17-92 cluster with cancer prognosis. Meanwhile, the potential associations were explored in different subgroups. Consequently, the finding of significant correlation between high-expression of miR-17-92 cluster and unfavorable OS/DFS/CSS in various tumors by two different statistical methods is of particular interest. Likewise, similar results were found in different subgroups. However, no association of high-expression of miR-17-92 cluster was detected with RFS. Additionally, we demonstrated that high-expression of miR-17-92 cluster was associated with favorable PFS by two different statistical methods. Probably due to relatively fewer studies of RFS/PFS/CSS, these results remain inconclusive and require further investigation.

Due to significant heterogeneity of the current meta-analysis, careful interpretation and search for influencing factors were required. Firstly, impact of ethnicity on prognosis in patients was considerable, which should be take into consideration when evaluating the prognosis of cancer for patients61. It is a well-established fact that formalin-fixation and/or prolonged storage can elicit damage to nucleic acids, further conferring considerable limitation on results62, 63. Accordingly, differences in the detection and quantification methods, types and numbers of miRNAs evaluated and sample source should be also considered as potential sources of heterogeneity. We performed further subgroup analyses according to country, test method, sample source, miR-17-92 cluster component and sample size. All of the subgroup analyses also indicated that high-expression of miR-17-92 cluster was associated with poor OS. As for PFS/RFS/DFS/CSS, we did not perform subgroup analyses due to relatively fewer eligible studies.

Actually, our meta-analysis has its limitations. Firstly, only published studies might not provide sufficient evidences. Additionally, studies regarding various tumors without a consistent cut-off value may influence the ultimate results and two eligible studies20, 25 did not clearly illustrate the hypotype of miR-19. Meanwhile, the heterogeneity suggested that potential or undiscovered factors including adjustment for surgery, radiation, chemotherapy, socioeconomic status, tumor characteristics, and so on might be ignored. Whereas, in spite of aforementioned limitations, a certain relationship of high-expression of miR-17-92 in prognostic value was found in current meta-analysis.

In conclusion, the current study is the first original meta-analysis to address the correlation between the miR-17-92 cluster expression and prognostic value for patients. A marginally significant association was explored in overall population as well as the corresponding subgroups. Concretely, it presented that high-expression of miR-17-92 cluster might be associated with poor OS/DFS/CSS and favorable CSS to some extent, while no association was detected between high miR-17-92 expression with RFS. Notably, due to relatively fewer studies of RFS/PFS/CSS, these results still require further verification in the near future.