Appraising circular RNAs as novel biomarkers for the diagnosis and prognosis of gastric cancer: A pair‐wise meta‐analysis

Abstract Background Circular RNAs (circRNAs), proven as single‐stranded closed RNA molecules, have been implicated in the onset and development of multiple cancers. This study aimed to summarize existing evidences regarding the clinicopathologic, diagnostic, and prognostic significances of circRNAs in gastric cancer (GC). Methods Eligible studies were identified using online databases. The quality of the included studies was judged, and patients' clinical characteristics, diagnostic data, and overall survival (OS) were extracted from the electronic medical record. Fisher's method was adopted to determine P values for clinicopathologic features. The diagnostic and prognostic data from all included studies were merged. Results Thirty eligible studies were comprised of 2687 GC patients were enrolled in the meta‐analyses. Altered expressions of circRNAs in GC tissues were significantly associated with worse clinicopathologic features. Abnormally expressed circRNAs yielded a pooled sensitivity of 0.76 (95% CI: 0.69‐0.81) and a specificity of 0.77 (95% CI: 0.70‐0.83) in distinguishing GC from noncancerous controls, which corresponded to an area under the curve (AUC) of 0.83. The survival analysis showed that the oncogenic circRNA signature could be an independent risk factor of OS (HR = 2.11, 95% CI: 1.60‐2.78, P = .000). Patients with down‐regulated circRNAs (tumor suppressor genes) presented a significantly shorter OS time than those with high‐level circRNAs (HR = 0.33, 95% CI: 0.27‐0.42, P = .000). Stratified analyses based on sample type, control source, circRNA expression status, and cutoff setting also produced robust results. Conclusions CircRNAs may play an important role as potential diagnostic and prognostic biomarkers of GC.


| Data extraction
Two authors independently retrieved the name of the first author, year of publication, country, study design, case numbers, sample types, control sources, circRNA signatures, test methods, cutoff value settings, reference genes, values of sensitivity and specificity, HR values with 95% CIs, and follow-up time. Any disagreement was resolved by group discussion until consensus was reached.

| Study bias and quality assessment
We first used the Quality Assessment for Studies of Diagnostic Accuracy 2 (QUADAS-2) checklist to judge the quality and bias of the eligible studies that evaluated diagnostic performances of circRNA(s) in GC. 40 The QUADAS-2 checklist was composed of two parts, "risk of bias" and "applicability concerns," and contained seven items categorized into patient selection, index test, reference standard, flow, and timing. Each item could be rated as low risk, high risk, or unclear risk, and an answer of "low risk" merely received 1 point, while that of either "high risk" or "unclear risk" did not receive any point. In addition, guidelines from the Newcastle-Ottawa Quality Assessment Scale (NOS) checklist were used to determine the bias of prognostic studies, 41 in which eight items regarding study selection, comparability, and outcome were addressed. Risk of bias was judged as low risk, high risk, or unclear risk, corresponding to quantitative scores of 1, 0, and 0 points.

| Statistical analysis
Statistical analyses were conducted using STATA (version 12.0) and Meta-DiSc software (version 1.4). The estimated I 2 and Chi-square statistics were used to assess the heterogeneity among studies. A P-value of <0.1 in the Chi-square test with I 2 of >50% indicated significant heterogeneity. Fisher's method was used to combine the P values for clinicopathologic features. Pooled estimates of sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NRL), diagnostic odds ratio (DOR), and HRs with 95% CIs were calculated using a random effect model when significant heterogeneity was observed. Otherwise, a fixed-effect model was used. Influence and meta-regression tests were performed to trace the underlying causes of study heterogeneity. Deek's funnel plot, and Begg's and Egger's tests were adopted to analyze qualitative publication bias, and a P-value of <.05 was considered statistically significant. When publication bias was observed, the trim-and-fill method was used to assess the possible effects of bias on the overall pooled effects. 42
All essential data were obtained from the 30 studies (Tables 1-3

| Quality assessment
For diagnostic effects, studies were rated for patient selection, index test, reference standard, flow, and timing by the QUADAS-II criteria with a maximum of seven points. 40 As shown in Table S1, all studies received rated QUADAS scores of ≥4 points. Prognostic studies were assessed using the NOS checklist with a maximum of nine points, 41 and all 11 studies achieved NOS scores of ≥ 6 ( Table S2).
The results suggested that risks of bias and quality in the studies were acceptable.

| CircRNA expressions and clinicopathologic features
As shown in Table 1 (Table 1).

| Diagnostic performance
The

| Prognostic value
The prognostic ability of circRNA expression status was evalu-

| Influence and meta-regression tests
The sensitivity test showed that all studies with available analyses for the diagnostic and prognostic effects of circRNAs were equally distributed within the lower and upper limits of the 95% CI, and no individual outlier studies were included (Figure 4).

| Publication bias
No publication bias in the pooled diagnostic effects was determined by Deek's funnel plot (P = .053), neither was the bias in the prognostic effects of down-regulated circRNAs on OS by Begg's and Egger's tests (Egger's test, P = .806; Begg's test, P > .05). However, significant bias was observed in the prognostic meta-analysis of oncogenic circRNAs for OS (Egger's test, P = .000). Consequently, the trim-and-fill method was used to more thoroughly assess possible effects of publication bias. 42 The fixed-effect model identified four missing studies, and the pooled adjusted effort differed little before and after adjustment

| D ISCUSS I ON
As GC is a highly heterogeneous disease with a high mortality rate, [1][2][3] most patients are confirmed until a very late stage due to the hidden symptoms. Despite the constantly updated treatments for the disease, the 5-year survival rate is still undesirable. 3 Identifying informative diagnostic and prognostic biomarkers of GC early on is the first priority for better predicting tumor behavior and guiding the treatment planning. That prompts a hotspot of circRNAs as a novel class of coding/non-coding RNAs characterized by circularization through covalent bonding of their 5' and 3' ends for cancer diagnosis. 5,6 Owing to the ring structure, circRNAs are more stable and conserved than linear RNAs, and a majority of them are highly stable in tissues and bodily fluids, as confirmed by some studies. 43,44 This unique characteristic suggests that circRNAs can be reckoned as promising noninvasive biomarkers of cancers, especially GC. [45][46][47] In this study, we analyzed the associations between circRNA expressions and clinicopathologic features, and determined clinical values of circRNAs as diagnostic and prognostic indicators of GC.
We summarize the correlation between tissue circRNA expressions and the basic characteristics, and find that several major clini- with a combined AUC of 0.83 ( Figure 2). DOR is another important index for diagnostic tests, and a higher value indicates better diagnostic efficacy. 49 In this study, a pooled DOR of 10.44 also demonstrates that circRNA levels are a potential diagnostic indicator for distinguishing GC form noncancerous controls ( Figure 2). Our findings demonstrate that circRNA expression profiling has potential as a diagnostic biomarker analysis for GC.
For the pooled diagnostic performance of circRNAs in GC, our stratified analyses of sample type, control source, circRNA function, and cutoff setting have also produced robust results. As a result, differences in the diagnostic efficacy are found to depend on test matrix, featuring that plasma circRNAs provide a better test matrix than tissue ones for the diagnosis of GC (Table 4). A previous report has proven that different sample sources can bring about disparities in the diagnostic efficacy non-coding RNAs, which indirectly support our findings. 50 Furthermore, our analysis has confirmed that circRNAs as a group of underlying indicators are more effective in differentiating GC patients from healthy individuals than from paired adjacent noncancerous controls (Table 4). In addition, oncogenic circRNA expressions yield better diagnostic accuracy for GC than tumor suppressor circRNAs (Table 4). Besides, it is corroborated that the cutoff value setting of <10 can result in greater efficacy than that of ≥10 (Table 4). This indicates that the diagnostic power of circRNAs in GC is sensitive to the cutoff value settings. However, no similar results have been observed in previous studies regarding control sources, circRNA functions, and cutoff settings for support of our findings, and more studies are needed.
As previously reported, some circRNAs have been found to have prognostic value in GC. [19][20][21][22][23]25,27,[32][33][34][35] Therefore, a meta-analysis for the prognostic value of previously reported circRNAs in GC has been performed, and the data have been stratified into oncogenic and Heterogeneity is common when performing a meta-analysis. 51 However, considerable heterogeneity can be easily found in the overall diagnostic and prognostic effects of oncogenic circRNAs.
To eliminate the underlying impacts of heterogeneity on the overall combined effects, we have performed a sensitivity analysis and a meta-regression test, and the sensitivity analysis just reveals that no individual studies are outliers. This suggests that the homogeneity of our data is acceptable and the combined effects are reliable ( Figure 4). In the meta-regression test, different test matrices significantly have contributed to the heterogeneity in the diagnostic meta-analysis. Of the included 28 individual studies in this analysis, 20 datasets have evaluated tissue and 6 plasma. It is the smaller sample size in the plasma-based studies that may result in bias. However, we only observed publication bias in the analysis for prognostic effects of oncogenic circRNAs for OS in GC patients ( Figure 5). To assess the possible effects of bias on pooled efficacy, the trim-and-fill method has been adopted. 42 However, filling 4 missing studies using a fixed-effect model has not clearly altered the effects, hinting that the pooled accuracy is not subject to publication bias.

| CON CLUS IONS
In summary, circRNAs may have potential clinical significance in GC and represent promising therapeutic targets and biomarkers of GC. However, our study had some limitations including population bias, obvious heterogeneity, and diverse test matrices and controls. Further studies are necessary to confirm the results of our meta-analysis.

ACK N OWLED G M ENTS
The authors thank all patients for providing the clinical data and samples.