Prognostic value of high-expression of miR-17-92 cluster in various tumors: evidence from a meta-analysis

The prognostic value of miR-17-92 cluster high-expression in various tumors remains controversial. Therefore, we conducted this meta-analysis by searching literatures in PubMed, Embase, Cochrane Library, China Biology Medicine disc, China National Knowledge Infrastructure to identify eligible studies. Eventually, we analyzed 36 articles that examined 17 tumor types from 4965 patients. Consequently, high-expression of miR-17-92 cluster in various tumors was associated with unfavorable overall survival in both univariate (HR = 2.05, 95%CI: 1.58–2.65, P<0.001) and multivariate (HR = 2.14, 95%CI: 1.75–2.61, P<0.001) analyses. Likewise, similar results were found in different subgroups of country, test method, miR-17-92 cluster component, sample source and size. Additionally, high-expression of miR-17-92 cluster was linked with poor disease-free survival (Univariate: HR = 1.96, 95%CI: 1.55–2.48, P<0.001; Multivariate: HR = 2.18, 95%CI: 1.63–2.91, P<0.001), favorable progression-free survival (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 cancer specific survival in univariate rather than multivariate analyses (Univariate: HR = 1.77, 95%CI: 1.21–2.60, P = 0.004; Multivariate: HR = 1.77, 95%CI: 0.80–3.92, P = 0.160). However, no association of miR-17-92 cluster high-expression was detected with recurrence or relapse-free survival. In summary, this meta-analysis towards high-expression of miR-17-92 cluster has indicated poor prognosis of various cancers. Notably, future studies comprising large cohort size from multicenter are required to confirm our conclusions.

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 I 2 test across studies (no heterogeneity I 2 <25%, moderate heterogeneity I 2 = 25-50%, extreme heterogeneity I 2 > 50%) 12 . In case of extreme heterogeneity (I 2 > 50% or P < 0.01 for Q test), we used random-effects (DerSimonian and Laird method) model 13 . Otherwise, fixed-effects (Mantel-Haenszel method) model was introduced 14 . 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 study 15

Results
Characteristics of the studies. As a result, 36 studies consisted of 4965 samples satisfied the eligible studies   (Fig. 1). The principal characteristics of the included studies were summarized in Table 1. Of these studies, Li's 18 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) 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 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  (Fig. 4A and B). We also explored that high-expression of miR-17-92 cluster was associated with favorable PFS (Univariate:   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.  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.

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 patients 52 . Additionally, the benefit from adjuvant chemotherapy for patients with stage II CRC after surgery remains a matter of debate 53 . 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 chemotherapy 54 . 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) patients 55 . 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 patients 56,57 . 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 prognosis 58,59 . 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 diseases 60 .
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 patients 61 . It is a well-established fact that formalin-fixation and/or prolonged storage can elicit damage to nucleic acids, further conferring considerable limitation on results 62,63 . 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 studies 20, 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.