Circular RNA profiling facilitates the diagnosis and prognostic monitoring of breast cancer: A pair‐wise meta‐analysis

Abstract Background As circular RNAs (circRNAs) have been found to significantly involve in the onset and progression of multiple malignant tumors including breast cancer (BC), this study aims at evaluating the diagnostic and prognostic values of circRNAs in this malady. Methods Available databases were thoroughly searched to collect studies on the diagnosis and/or prognosis of BC using circRNA profiling. The updated Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS‐2) tool and the Newcastle Ottawa Scale (NOS) were used to assess the underlying bias of included studies. Clinical characteristics of the studies were merged by the quantitative‐weighted integral method to obtain the combined effects. Results Sixteen studies were included, comprising 2438 BC cases and 271 noncancerous controls. The expression signature covered 24 circRNAs (down‐regulated: circ‐VRK1, hsa_circ_0068033, hsa_circ_103110, hsa_circ_104689, and hsa_circ_104821; up‐regulated: circAGFG1, hsa_circ_0001785, hsa_circ_0108942, hsa_circ_0001785, hsa_circ_006054, hsa_circ_100219, hsa_circ_406697, circEPSTI1, circANKS1B, circGFRA1, circ_0103552, CDR1‐AS, has_circ_001569, hsa_circ_001783, circFBXL5, circ_0005230, circAGFG1, circ‐UBAP2, and circ_0006528). The sensitivity and specificity of circRNAs in distinguishing BC patients from noncancerous controls were 0.65 and 0.68, and the corresponding area under the curve was 0.66. Survival analysis revealed that patients showing highly expressed oncogenic circRNAs were associated with increased mortality risks of BC in overall survival (univariate analysis: hazard ratio [HR] = 3.30, P = .000; multivariate analysis: HR = 3.07, P = .000), and disease‐free survival (HR = 8.26, P = .000). Stratified analysis based on circRNA expression status and control type also showed robust results. Conclusions Circular RNA profiling presents prominent diagnostic and prognostic values in BC, and can be rated as a promising tool facilitating its early diagnosis and survival.


| INTRODUC TI ON
Breast cancer (BC) tops the morbidity list among female malignancies, and the pace of its onset is accelerating year after year with the population becoming younger and younger. 1,2 As with the latest cancer statistics, the mortality of BC ranks the fourth among all female tumors. 3 Studies have confirmed that family history, reproductive factors, sex hormone levels, oral contraceptives, and previous history of breast diseases are closely related to its occurrence and development. [4][5][6] Exploring new molecular markers and therapeutic targets for BC are conducive to early diagnosis, more accurate prognostic prediction, and efficacy monitoring in the patients. At present, various factors restrict the early diagnosis of BC in clinic.
Biopsy as an invasive method is poorly acceptable to patients, and its accuracy is subject to operators' own experience. 7 Imaging examinations and routine blood tumor marker detection are currently not suitable for large-scale screening for an early diagnosis due to their low sensitivity (SEN) and accuracy. 8,9 Therefore, finding effective, noninvasive, novel, and operable biomarker profiling is critical for the early diagnosis, prognosis, and treatment of BC. Circular RNAs (CircRNAs) is a type of coding/noncoding RNA molecule with its 3ʹ and 5ʹ ends forming a covalently closed loop. 10,11 It is reported that circRNAs are widely expressed in mammalian cells and feature histocyte specificity (SPE), structural stability, and sequence conservation. 12,13 Studies have confirmed that circRNAs play roles in regulating gene transcription and expression via multiple pathways, and in physiological processes such as cell cycle and senescence. 14,15 Moreover, circRNAs are found to be essential in the onset and development of malignant tumors. 16,17 Given that circRNAs are insensitive to nucleases and more stable than ordinary linear RNA, they are expected to be new biomarkers for monitoring various cancers. 18,19 And circRNAs are intensively reported to have the potential of the early diagnosis and prognostic prediction of BC as novel molecular biomarkers. [20][21][22][23][24][25][26][27][28][29][30][31][32][33][34][35] However, the appraisals of their efficacy are commonly limited by the small sample size, high bias, and single-center population of the already reported trials.
Therefore, this study intends to systematically evaluate the efficacy of circRNAs in the diagnosis and prognostic prediction of BC through a quantitative meta-analysis.

| Data search strategy
This study was designed and conducted in line with the PRISMA 2019. 36 Two authors independently retrieved relevant studies in the online databases included PubMed, Embase, Web of Science, BioMed Central, and CNKI. Literature published in English, as of January 31, 2020, was searched. The following search terms were as follows: ("breast neoplasms [MeSH Terms]" OR "breast cancer" OR "breast carcinoma" OR "mammary cancer") AND ("circular RNA [MeSH Terms]" OR "circRNA" OR "hsa circ") AND ("diagnoses", "diagnosis", "SEN", "SPE", "ROC curve", "area under the curve", "AUC") OR ("prognosis" OR "prognoses [MeSH Terms]" OR "survival [MeSH Terms]" OR "overall survival" OR "progression free survival" OR "disease free survival" OR "relapse free survival" OR "hazard ratio" OR "OS" OR "PFS" OR "DFS" OR "RFS" OR "HR"). Meanwhile, the authors manually searched for the references attached to the paper to prevent literature omission.

| Inclusion and exclusion criteria
Inclusion criteria were defined as (a) case-control studies that reporting the diagnostic accuracy or prognostic utility of single or parallel circRNAs in BC; (b) diagnostic studies providing data that could be directly or indirectly involved in a 2 × 2 contingency table, com-

| Quality assessment
For diagnostic studies, the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool was used to evaluate the quality of studies, 37 and the evaluation consisted of 2 parts: bias evaluation and applicability. Specifically, the bias assessment K E Y W O R D S breast cancer, circular RNA, diagnosis, meta-analysis, prognosis included 4 domains: case selection, index test, golden standard, and flow and timing, and the first 3 domains were also assessed with respect to applicability. Each domain could be graded by 3 levels: low risk, high risk, and unknown, corresponding to 1 point, 0 point, and 0 point, respectively. When the total score was ≥4 points (out of 7 points), the quality of the study could be considered high. Case-control studies were evaluated with the 8-item Newcastle Ottawa Scale (NOS) scale, 38 referring to study population selection, comparability, exposure evaluation, or outcome evaluation. A study with a total score of ≥5 points (out of 9 points) could be considered high quality. The threshold effect was evaluated by Spearman's correlation coefficients, with a P < .05 considered statistically significant. The nonthreshold effect was evaluated by Cochran's Q test and I 2 test, and the statistical significance level was set at P < .01 or I 2 > 50%. When there was no heterogeneity between studies, data could be merged using a fixed-effect model; otherwise, a random-effect model would be adopted. Sources of heterogeneity were traced using the SEN analysis and the meta-regression test. Deek's funnel plot and visual Funnel plot, as well as Begg's and Egger's tests, were used to assess publication bias among studies, and the statistical significance level was set at P < .1. When publication bias appeared, the nonparametric trim and fill method will be applied to assess its possible effect on the meta-analysis model. 39

| Clinical characteristics in included studies
The inclusion and exclusion process of literature retrieval was depicted in Figure 1. As a result of database search according to the search strategy, 208 relevant studies were obtained. After carefully reading titles and abstracts, we ruled out 183 articles due to irrelevant topics or reviews, retained 25 for the full-text evaluation, and eliminated 7 owing to a lack of data or out of topic.
Of all prognostic studies, one study investigated the prognostic value of tumor-inhibitory circRNAs on DFS, 40  as internal reference genes.

| Risk assessment for heterogeneity and quality
Spearman's correlation coefficients showed that the effect size of the overall combination corresponds to P = .139, suggesting that there was no heterogeneity caused by threshold effects between studies. Cochran's Q and I 2 tests for nonthreshold effects showed a P = .001 and an I 2 of 83.17%, indicating significant heterogeneity among studies.
Diagnostic studies were analyzed using the QUADAS-2 tool for a risk of bias assessment, and it was found that the QUADAS-2 scores of all 6 studies were higher than 4 points, suggesting the high quality of the included studies (Table 3). Besides, all included case-control studies revealed high NOS scores of over 6 points, which could be defined as high quality (Table 4).  Figure 2.

| Diagnostic performances of circRNAs
The subgroup analysis revealed that circRNA profiling yielded a high diagnostic efficacy in distinguishing BC from healthy individuals than that from adjacent noncancer tissues (AUC: 0.81 vs 0.65).

| Prognostic value
For the prognostic analysis, summary HRs and 95% CIs were es-

| Sensitivity analysis and meta-regression test
Sensitivity analysis was conducted to explore the sources of heterogeneity among studies, and outliers were found in the diagnos-

Total rated scores
Representativeness of the exposed cohort and I 2 increased to 0% after an exclusion of the outlier. In addition, the meta-regression was performed for analyzing the effects resulting from control type, number of cases, number of controls, and study quality. We found that the mentioned factors were not the underlying sources of heterogeneity among studies (all with P > .05) ( Table 6).

| Publication bias
No publication bias was observed in the pooled effects except the prognostic meta-analysis of the OS (univariate analysis) ( Figure 6).
The nonparametric trim and fill method was applied to assess the possible effect of publication bias on the meta-analysis model. 39  The imputed data generated a symmetrical funnel plot ( Figure 6D).
However, the pooled effect incorporating the hypothetical data altered little from the unadjusted ones (variance = 0.367, P = .013 vs variance = 0.440, P = .016), hinting that the combined effect is not subject to the impact of publication bias.

| D ISCUSS I ON
Breast tumor malignancy that originates from mammary epithelial cells more rapidly occurs in the younger age group. 1,2 Screening and developing novel and noninvasive biomarkers will facilitate early identification and prognostic prediction of BC. CircRNAs have been proven to widely exist in many eukaryotic organisms and are mainly located in the cytoplasm or can be stored in exosomes. 10

| CON CLUS IONS
In summary, circRNAs can be used as prominent auxiliary indicators for the diagnosis and prognosis evaluation of BC. However, the conclusion of this study still needs to be confirmed by more high-quality studies with large samples.

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

CO N FLI C T O F I NTE R E S T
None declared.

Yongwei
Li https://orcid.org/0000-0002-0047-6557 F I G U R E 6 Publication bias in the diagnostic meta-analysis evaluated by (A) Deek's funnel plot (P = .32) and (B) visual Funnel plot. The prognostic meta-analyses of the univariate analysis assessed by (C) Begg's test, and (D) nonparametric trim and fill method, and the multivariate analysis by (E) Begg's and (F) Egger's tests. The judgment of publication bias in combined (G) disease-free survival and (H) progression-free survival using visual Funnel plot