This article is mentioned in:

Introduction

Cancer is a major public health problem globally. Cancer is a class of diseases that undergoes uncontrollable cell proliferation and differentiation. Based on the 2015 cancer statistics, it is currently the second leading cause of mortality in numerous countries (including China, Europe and the USA), and is expected to surpass heart diseases as the leading cause of mortality in the near future (1). Although the risk of succumbing to cancer has decreased by ~20% from its maximum in 1991–2011 (1), it must be diagnosed with high sensitivity and specificity in order to determine the appropriate therapy and prognosis. Recently, a number of biomarkers with diagnostic and prognostic potential value have been demonstrated in numerous cancer types, including the tumor markers human epididymis secretory protein 4 and cancer antigen 125 in endometrial (2) and ovarian cancer types (3). Additionally, mutant genes have been used in the selection of an appropriate therapy, including epidermal growth factor receptor mutation in non-small cell lung cancer (4), Kirsten rat sarcoma viral oncogene homolog in colorectal cancer (5) and v-raf murine sarcoma viral oncogene homolog B1 mutation in melanoma (6); however, reliable and convenient biomarkers are required to evaluate the diagnostic and prognostic significance of different cancer types.

Classic biomarkers present with potentially limiting factors, including cost, availability and reproducibility (7). Utility is compromised by different disease heterogeneities, specific genetics and proteomics, and the influence of lifestyle; therefore, a number of serum or tissue biomarkers, including non-coding RNAs (ncRNAs), have been developed for clinical experiments. ncRNAs have notable potential for future biomarker approaches. Numerous studies have reported the use of ncRNAs, including microRNAs and long ncRNAs (lncRNAs), in the early detection and prognosis of various cancer types (8,9). Previously, a number of studies focused on a novel class of ncRNAs that is endogenously expressed as single-stranded, covalently-closed circular molecules, also known as circular RNAs (circRNAs) (10–12). circRNAs were demonstrated to be antagonists of specific microRNAs by functioning as microRNA sponges (10,13), and they are also known as stable molecules, as demonstrated by their long half-lives in cells (14). These observations resulted in the consideration that circRNAs could serve as potential biomarkers for the non-invasive diagnosis of numerous diseases, including disorders of the central nervous system (15), cancer (16) and a number of forms of cardiovascular diseases (17).

To determine if circRNA could serve as a sensitive and specific biomarker for cancer, a systematic meta-analysis of the published literature was performed in the present study, in order to review the diagnostic efficiency of circRNA in patients with cancer from the available data and to identify a novel non-invasive biomarker for cancer diagnosis.

Materials and methods

Search strategy

This meta-analysis was conducted in accordance with the guidelines of diagnostic meta-analysis as follows: Eligible studies published up to November 30, 2017, on PubMed (https://www.ncbi.nlm.nih.gov/pubmed) and EMBASE (https://www.elsevier.com/solutions/embase-biomedical-research), were selected for the meta-analysis. Non-English studies were excluded. No restriction was placed on the year of publication or publishing status. The key words employed for literature retrieval included the following: ‘circular RNA’ or ‘circRNA’, and ‘tumor’ or ‘neoplasm’, or ‘cancer’ or ‘carcinoma’. Additionally, the reference lists of eligible articles were manually searched to obtain additional sources.

Selection of publications

All studies were carefully and independently reviewed by two researchers based on their titles and abstracts, following which full texts were perused for potential eligibility. Any disagreement was resolved by a full discussion, until consensus was achieved. All publications included in the meta-analysis were required to meet the following criteria: i) Studies should analyze the association between circRNA and patients with any cancer type; ii) studies should contain sensitivity and specificity data (or the possibility of deriving such values from the data); and iii) studies should have enrolled ≥20 patients and matched controls. Studies were excluded if they involved any of the following parameters: i) Duplicate studies; ii) letters, editorials, meeting abstracts, case reports and reviews; iii) patients and control subjects that did not qualify, in which the patients sample size was low or the disease cannot be defined; iv) studies with missing data, and v) No-English studies. If the same author reported that their results were acquired from overlapping populations, only the first study published or the most complete study was included.

Data extraction and quality assessment

The following parameters were collected from each study: Author name, publication year, country and ethnicity, sample type, normalization control, sample size and data for two-by-two tables (sensitivity and specificity). The Quality Assessment of Diagnostic Accuracy Studies (QUADAS) checklist (http://www.bristol.ac.uk/population-health-sciences/projects/quadas/) was used to systematically assess the quality of the articles included in the diagnostic meta-analysis. Specifically, 14 items from the QUADAS checklist were applied to each article, and an answer of ‘Yes’, ‘No’ or ‘Unclear’ was determined. Only ‘Yes’ resulted in a score.

Statistical analysis

All statistical analyses were performed using the STATA 13.0 statistical software (StataCorp LLC, TX, USA) and Meta-DiSc 1.4 (Unit of Clinical Biostatistics, Ramón y Cajal Hospital, Madrid, Spain). Data from each study (true-positives, false-positives, true-negatives and false-negatives) were extracted to obtain the pooled sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnostic odds ratio (DOR) and their 95% confidence interval (CI), summary receiver operator characteristic (SROC) curve and area under the curve (AUC), in order to determine the overall performance of the detection method. P<0.05 (two-sided) was considered to indicate a statistically significant difference. Additionally, heterogeneity across studies was assessed using Cochran's Q and I2 statistics, where I2>50% indicated the existence of significant heterogeneity. Finally, evaluation of the threshold effect (Spearman's rank correlation) and publication bias (funnel plots) were also undertaken.

Results

Literature search

Electronic and manual searches yielded a total of 146 potentially eligible articles. The steps involved in screening the articles for the meta-analysis is depicted as a flow chart in Fig. 1. Screening titles and abstracts resulted in the exclusion of 109 articles. A further 21 articles were excluded following more detailed assessment of the full text. Finally, 10 eligible studies (12 tests) were included in the meta-analysis (18–27).

Figure 1. Flow chart of studies included in the meta-analysis following implementation of exclusion criteria.

Study characteristics

The characteristics of the 10 eligible studies are summarized in Table I (18–27). A total of 799 patients with different cancer types and adjacent controls were involved in these 12 tests. Assessment using QUADAS indicated that the studies were of high quality, with positive results in 13/14 items (Fig. 2). Additionally, the mean impact factor was calculated to be 2.59.

Figure 2. Quality assessment of included studies based on the Quality Assessment of Diagnostic Accuracy Studies tool.

Table I. Characteristics of 10 studies included in the meta-analysis.

Table I.

Characteristics of 10 studies included in the meta-analysis.

First author	Year	Disease	circRNA used for detection	Case no.	Control no.	Region	TP	FP	FN	TN	IF	(Refs.)
Li et al	2015	Gastric cancer	hsa_circ_002059	101	101	China	82	38	19	63	2.799	(18)
Qin et al	2016	Hepatocellular carcinoma	hsa_circ_0001649	89	89	China	72	28	17	61	1.736	(19)
Wang et al	2015	Colorectal cancer	hsa_circ_001988	31	31	China	21	8	10	23	1.581	(20)
Shang et al	2016	Hepatocellular carcinoma	hsa_circ_0005075	30	30	China	25	3	5	27	2.133	(21)
Chen et al	2017	Gastric cancer	hsa_circ_0000190	104	104	China	75	30	29	74	2.799	(22)
Huang et al	2017	Gastric cancer	hsa_circ_0000745	60	60	China	51	33	9	27	3.365	(23)
Fu et al	2017	Hepatocellular carcinoma	hsa_circ_0003570	107	107	China	48	14	59	93	1.521	(24)
Yin et al	2017	Breast cancer	hsa_circ_0001785	20	20	China	16	5	4	15	2.871	(25)
		Breast cancer	hsa_circ_0108942	20	20	China	16	10	4	10	2.871
		Breast cancer	hsa_circ_0068033	20	20	China	14	8	6	12	2.871
Yao et al	2017	Hepatocellular carcinoma	cirZKSCAN1	102	102	China	84	28	18	74	5.314	(26)
Zhao et al	2017	Gastric cancer	hsa_circ_0000181	115	115	China	62	17	53	98	1.521	(27)

[i] TP, true-positive; FP, false-positive; FN, false-negative; TN, true-negative; IF, impact factor; circRNA, circular RNA.

Meta-analysis

Overall, 10 studies involving 799 patients with various cancer types reported the detection performances of circRNA (Table II). The sensitivity of circRNA detection testing ranged from 0.449–0.855, and the reported specificity ranged from 0.450–0.900. The pooled DOR was 7.265 (95% CI, 5.616–9.398; Q=12.72; P=0.312; I2=13.5%). The pooled sensitivity was 0.708 (95% CI, 0.676–0.740; Q=74.77; P<0.001; I2=85.3%) and the pooled specificity was 0.722 (95% CI, 0.690–0.753; Q=60.81; P<0.001; I2=81.9%). The PLR and NLR were 2.483 (95% CI, 2.019–3.054; Q=30.83; P=0.001; I2=64.3%) and 0.372 (95% CI, 0.289–0.479; Q=44.59; P<0.001; I2=75.3%), respectively. The AUC was 0.793. The forest plots and SROC are depicted in Figs. 3 and 4, respectively.

Figure 3. Forests plot of the accuracy of circRNAs for the diagnosis of cancer. (A) DOR forest plot of the circRNAs. (B) Sensitivity forest plot of the circRNAs. (C) Specificity forest plot of the circRNAs. (D) Positive LR forest plot of the circRNAs. (E) Negative LR forest plot of the circRNAs. DOR, diagnostic odds ratio; LR, likelihood ratio; circRNAs, circular RNAs; df, degrees of freedom.

Figure 4. SROC of circular RNAs for cancer diagnosis. SROC, summary receiver-operating characteristic; AUC, area under curve; SE, standard error; Q*, Q statistic.

Table II. Detection performances of circular RNA reported by 10 studies.

Table II.

Detection performances of circular RNA reported by 10 studies.

First author	Diagnostic OR (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	Positive LR (95% CI)	Negative LR (95% CI)	(Refs.)
Li et al	7.16 (3.77–13.59)	0.81 (0.72–0.88)	0.62 (0.52–0.72)	2.16 (1.65–2.82)	0.30 (0.20–0.46)	(18)
Qin et al	9.23 (4.62–18.44)	0.81 (0.71–0.88)	0.69 (0.58–0.78)	2.57 (1.86–3.55)	0.28 (0.18–0.44)	(19)
Wang et al	6.04 (2.01–18.17)	0.68 (0.49–0.83)	0.74 (0.55–0.88)	2.63 (1.38–5.00)	0.43 (0.25–0.75)	(20)
Shang et al	45.00 (9.73–208.08)	0.83 (0.65–0.94)	0.90 (0.73–0.98)	8.33 (2.81–24.67)	0.19 (0.08–0.42)	(21)
Chen et al	6.38 (3.49–11.66)	0.72 (0.62–0.80)	0.71 (0.61–0.80)	2.50 (1.81–3.46)	0.39 (0.28–0.55)	(22)
Huang et al	4.64 (1.94–11.09)	0.85 (0.73–0.93)	0.45 (0.32–0.58)	1.55 (1.20–1.99)	0.33 (0.17–0.65)	(23)
Fu et al	5.40 (2.70–53.33)	0.45 (0.35–0.55)	0.87 (0.79–0.93)	3.43 (2.01–5.83)	0.63 (0.53–0.76)	(24)
Yin et al	12.00 (2.70–53.33)	0.80 (0.56–0.94)	0.75 (0.51–0.91)	3.20 (1.45–7.05)	0.27 (0.11–0.66)	(25)
	4.00 (0.98–16.27)	0.80 (0.56–0.94)	0.50 (0.27–0.73)	1.60 (0.98–2.61)	0.40 (0.15–1.07)
	3.50 (0.94–12.97)	0.70 (0.46–0.88)	0.60 (0.36–0.81)	1.75 (0.95–3.22)	0.50 (0.23–1.07)
Yao et al	12.33 (6.31–24.09)	0.82 (0.74–0.89)	0.73 (0.63–0.81)	3.00 (2.16–4.16)	0.24 (0.16–0.38)	(26)
Zhao et al	6.74 (3.58–12.69)	0.54 (0.44–0.63)	0.85 (0.77–0.91)	3.65 (2.28–5.84)	0.54 (0.44–0.67)	(27)

[i] OR, odds ratio; LR, likelihood ratio; CI, confidence interval.

Investigation of the threshold effect

Spearman's rank correlation was also performed to confirm the threshold effect. No indication of a threshold effect was determined in the studies [Spearman's correlation coefficient (ρ), 0.340; P=0.280]. Additionally, the slope (b) of the regression equation did not differ from zero (P=0.852), implying no heterogeneity between the studies.

Publication bias

Finally, the presence of a statistically significant slope coefficient (P<0.05) was considered to indicate a possible publication bias. Funnel plots were produced (Fig. 5). No publication bias was observed in the included studies (P=0.82) and the regression line represented a symmetrical curve.

Figure 5. Funnel plot of the 12 tests in the 10 included studies. DOR, diagnostic odds ratio; ESS, effective sample size.

Discussion

There are an increasing number of molecular biomarkers, including microRNAs and lncRNAs, being used in cancer diagnostics. circRNAs are widely expressed in human cells (28). Highly conserved sequences and a high degree of stability in mammalian cells are two of their most important properties (10,13); thus, circRNAs have the potential to be ideal biomarkers in the diagnosis of cancer. Numerous studies have evaluated the performance of circRNAs in cancer diagnosis (18–27); however, no systematic evaluation of circRNAs has been performed. The differences in the performances were too large and hence, to the best of our knowledge, the present study is the first meta-analysis to provide precise and controlled data on the diagnostic performance of circRNAs in cancer.

A total of 10 eligible, high-quality studies were included in the present meta-analysis. The present study demonstrated the varying sensitivities and specificities of circRNAs in the diagnosis of cancer; however, the range of their sensitivity and specificity was large and their diagnostic performance cannot be evaluated. The pooled sensitivity and specificity were observed to be slightly high (70.8 and 72.2%), which demonstrated that circRNAs could be used as assistant indicators in the diagnosis of cancer. The SROC curve and DOR indicated that circRNAs exhibited a moderate diagnostic performance. The pattern of the data points in the SROC curve did not indicate a ‘shoulder-arm’ shape, which indicates no threshold effect was determined in these studies, and the AUC of the SROC was 0.793. Cumulatively, these results indicated that circRNA had a moderate level of overall diagnostic accuracy for cancer diagnosis.

In the present study, heterogeneity was not determined in the pooled DOR of the circRNAs (P=0.852). Furthermore, publication bias and Spearman's rank correlation were also performed. No statistical difference was determined using Spearman's rank correlation, which meant that no threshold effect among these studies was observed. No publication bias was observed either in the included studies.

However, a number of limitations in this meta-analysis should be noted. Firstly, all included studies were reported by Chinese researchers. For this reason, the diagnostic performance of circRNAs may be not be all-sided, in spite of the absence of heterogeneity, threshold effect and publication bias; Therefore, further research regarding circRNAs, particularly in relation to the other countries' projects, as a biomarker in cancer diagnosis is required. Secondly, only the integral diagnostic performance of circRNAs on cancer was evaluated. The performance may be cursory on a specific type of circRNA for specific cancer types. Since the aim of the present study was to evaluate the likelihood of circRNAs performing for the diagnosis of cancer, the integral performance of circRNA in cancer was sufficient. Finally, the moderate levels of circRNA sensitivity and specificity could be attributed to technological, instrumental and staffing limitations; however, there is not sufficient data to evaluate these parameters. The cut-off value of circRNA efficiency in different cancer types remains controversial, and investigating its clinical significance may improve the diagnostic performance of circRNAs.

In conclusion, circRNA is a moderately effective assistant diagnostic biomarker for cancer; however, its diagnostic performance remains to be determined and further research of specific circRNA types for specific cancer types is required in order to determine this.

Acknowledgements

The authors thank Dr Si Chen (Department of Clinical Laboratory, Beijing Anzhen Hospital) and Dr Chuiwen Deng (Department of Rheumatology and Clinical Immunology, Peking Union Medical College Hospital) who gave positive advice regarding this study.

Funding

The present study was supported by the Talent Training Project of High Level Health Technology from Beijing Health System (grant no. 2015-3-052) and the National Natural Science Foundation of China (grant no. 81500332).

Availability of data and materials

All data generated or analyzed during this study are included in this published article.

Authors' contributions

YL and XZ carried out the conception and design, acquisition of data, analysis of data, and drafting the manuscript. HY performed the acquisition of data, and the drafting and revising of the manuscript. JH and SZ aided with acquisition of data. HC aided with the statistical analysis. QS and NJ participated in the design and coordination of the study and helped to revise the manuscript.

Ethics approval and consent to participate

Not applicable.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Glossary

Abbreviations

Abbreviations:

References