Diagnostic and Prognostic Significance of Dysregulated Expression of Circular RNAs in Osteosarcoma

ABSTRACT Objective This study aimed to perform an updated meta-analysis to explore the clinical, diagnostic, and prognostic values of circRNAs in osteosarcoma. Methods : PubMed, Web of Science, EMBASE, Scopus, and Cochrane Library were systematically searched up to December 15, 2020. Eligible studies regarding the relationship between circRNAs levels and clinicopathological, diagnostic, and prognostic values in osteosarcoma were included for study. Results 31 studies involving 1979 osteosarcoma patients were enrolled, with 22 studies on clinicopathological parameters, eleven on diagnosis, and 23 on prognosis. For clinical parameters, overexpression of oncogenic circRNAs was intimately correlated with larger tumor size, advanced Enneking stage, poor differentiation, and distant metastasis (DM). In contrast, the downregulated circRNAs showed negative correlation with Enneking stage and DM. For the diagnostic values, the summary area under the curve of circRNA for the discriminative efficacy between osteosarcoma patients and non-cancer counterparts was estimated to be 0.87, with a weighted sensitivity of 0.79, specificity of 0.81, respectively. For the prognostic significance, oncogenic circRNAs had poor overall survival (OS) and disease-free survival, while elevated expression of tumor-suppressor circRNAs were closely related to longer OS. Conclusion This study showed that aberrantly expressed circRNA signatures could serve as potential biomarkers in diagnosis and prognosis in osteosarcoma.


Introduction
Osteosarcoma is a common primary bone malignancy that usually occurs in the long bone metaphysis and mainly affects children and adolescents [1,2]. Surgery and chemotherapy are common treatment approaches for osteosarcoma [3]. In spite of substantial achievement in diagnosis and treatment, the five-year survival of osteosarcoma remains unsatisfied [4], largely due to the insufficiency of accurate predictive biomarkers [5]. Therefore, it is urgent to identify more ideal biomarkers with both diagnostic and prognostic value in osteosarcoma.
Recent advances in transcriptomics have provided novel perspective for management of cancer [6]. Circular RNAs (circRNAs) are a cluster of endogenous non-coding RNAs (ncRNAs) with no or limited capacity for coding proteins [7][8][9][10]. It is formed from back-splicing of pre-RNAs without 3ʹterminal poly A tail and 5ʹ-terminal cap [11]. Accordingly, it is more stable than the linear mRNA due to the covalently continuous loop [12]. Mechanistically, circRNAs could function as nuclear transcription regulator via attaching elements of RNAbinding protein (RBP) to modulate RNA transcription, or serve as competing for endogenous RNA (ceRNA) to sponge microRNA (miRNA), thus in turn to play a critical role in various biological process of human diseases including viral infections, cardiac fibrosis, diabetes, and cancer [13][14][15][16]. Interestingly, several circRNAs have been found to encode target genes, and thereby serve as tumor promoter or suppressor in cancer progression [7,8].
Nowadays, circRNAs have attracted the researches attention as potential biomarkers and therapeutic targets in carcinomas, which are attributed to their abundance in tissues, structural stability, and tissue-specific expression profile [11]. More recently, emerging studies have highlighted the role of circRNAs in osteosarcoma [17][18][19][20][21][22][23][24]. CircRNAs are reported to modulate osteosarcoma malignant properties including cell proliferation, apoptosis, migration, and resistance to chemotherapy [21,25,26]. CircRNAs could interact with cancerrelated signaling pathway, such as Akt/GSK-3β, and Wnt/βcatenin pathway, or act as a ceRNA to sponge miRNA and consequently regulate target genes in osteosarcoma progression [21,27,28]. Moreover, a growing number of evidence showed that dysregulated circRNAs are closely correlated with survival outcome and clinicopathological parameters including clinical stage, chemo-sensitivity, and metastasis [24,28,29]. However, the clinical prognostic and diagnostic significance of circRNAs may be jeopardized due to the controversial results as well as limited sample size in previous studies. This meta-analysis aimed to further clarify the pooled clinicopathological, diagnostic, and prognostic significance of circRNAs in osteosarcoma.

Search strategy
This study was performed in accordance with the Preferred Reporting Items for Systematic Review and Meta-analysis (PRISMA) Checklist [30]. Several electronic databases, including PubMed, Web of Science, EMBASE, and the Cochrane Library database, were extensively searched from inception to December 15, 2020 for eligible studies that assessed the clinicopathological, diagnostic or prognostic significance of circRNAs in osteosarcoma. The following terms were used in databases for literature retrieval: ('circular RNA' or 'circRNA' or 'hsa_circ') and (osteosarcoma). Patients with osteosarcoma were considered as the 'case group', while those with benign lesions or healthy were taken as 'control group'.

Study selection criteria and data extraction
Studies were eligible if met the following criteria: 1) cohort or case-control study; 2) diagnosis of osteosarcoma was histopathologically confirmed; and 3) correlation between circRNAs expression with clinicopathological features, diagnostic accuracy, or prognosis utility were provided or extractable; Besides, those studies were excluded if 1) not relevant with circRNAs or osteosarcoma; 2) in vivo studies, case reports, reviews, or letters without original data; 3) duplicate studies; Two independent investigators (CHZ and JYH) evaluated the enrolled studies and extracted data carefully, and a third researcher (LQ) would be consulted to reach a consensus if disagreement occurred. The following data were extracted in eligible studies: first author, publication year, circRNA type, sample type, sample size, detection assay, reference gene, regulation pattern, area under curve (AUC), circRNA expression levels, survival outcomes including overall survival (OS), disease-free survival (DFS) or progression-free survival (PFS), survival analysis method, follow-up duration (months), age, gender, tumor size, Enneking stage, differentiation based on WHO grade, and distant metastasis (DM).

Quality assessment, sensitivity analysis and publication bias
The quality of included studies on diagnosis was assessed according to the Quality Assessment for Studies of Diagnostic Accuracy II (QUADAS II) checklist [31], while studies on prognosis were rated by the Newcastle-Ottawa Scale (NOS) as previously described [32,33]. The studies were considered of high quality if the QUADAS II score was ≥4, or NOS score was ≥6.
In order to increase the credibility of this study, sensitivity analysis was performed to identify the potential source of heterogeneity. Publication bias was investigated by using both funnels plots and Begg's as well as Egger's test.

Statistical analysis
Statistical analyses were performed by using STATA software (13.0) and Meta-Disc (1.4). Pooled odds ratios (ORs) with 95% confidence intervals (CIs) were adopted for evaluation of clinical features, while hazard ratios (HRs) with 95% CIs to measure OS and DFS. The Chi-square test was utilized to assess the heterogeneity among studies. If P < 0.05 or I 2 > 50%, the random-effect model would be used due to the significant heterogeneity. Otherwise, the fixed-effect model would be adopted in the analyses. In addition, the publication bias was conducted by Deeks' funnel plot asymmetry test, Begg's and Egger's test [34].

Characteristics of the enrolled studies
The search procedure is presented in Figure 1. Among the potential literature retrieved from databases, 195 studies were initially assessed. After removing 103 duplicate publications, the remaining 92 studies were evaluated for titles and abstracts. Of these, 35 irrelevant articles were further excluded after abstract review and only 57 studies remained for full-text verification. Moreover, 26 studies were further eliminated with a variety of reasons, including five studies not related to circRNAs or osteosarcoma, eight studies did not report outcomes, four reviews and case reports, four animal studies, and five lack of extractable data. Finally, 31 18,21,23,35,36,38,40,[43][44][45][46]48,49], whereas other two studies on downregulated circRNAs applying ≥18 or <18 years old to separate enrolled patients [24,39]. In consideration of the consistency, only these studies with common cutoff value were included in the following meta-analysis with regard to patients' age. Besides, all cases in studies on clinicopathological feature analysis were divided into male and female. Other clinicopathologic parameters consist of tumor size (≥5 cm/ <5 cm), Enneking stage (IIB-III/I-IIA), distant metastasis (positive/negative), and WHO grade (III/I-II) were also presented in this analysis.
The quality of studies was evaluated by QUADAS II and NOS scores. For diagnostic studies, the rating scores of QUADAS II ranged from 4 to 6. While for prognostic studies, the NOS scores were from 6 to 7, suggesting high methodological quality in all selected studies.

Expression of circRNA with clinicopathological parameters in osteosarcoma
The correlation between circRNAs and clinicopathological parameters in patients with osteosarcoma were demonstrated in Table 3

Diagnosis analysis
As demonstrated in Figure (Figure 2A, 2B, and 2C). Moreover, a summary receiver operator characteristic (SROC) curve was presented in Figure 2D, and the AUC was 0.87 (95% CI: 0.83-0.89). These results indicated that circRNAs could be ideal diagnostic biomarker for osteosarcoma.
Additionally, a small set of oncogenic circRNAs with different diagnosis efficiency were synthesized to explore the combination panel of certain circRNAs with even higher diagnostic potential, which may facilitate the clinical application of the circRNAs for diagnosis. Among them, the pooled analysis containing circPVT1, circ_0081001, and CDR1as, has the highest diagnostic efficiency [17, 18,38]. As shown in Figure S1, the forest plot of the pooled DOR, SENS and SPEC of these three circRNAs panel were 32.21 (95%: 9.51-109.1), 0.85 (95% CI: 0.78-0.90), and 0.83 (95% CI: 0.71-0.91), respectively. Notably, the AUC under SROC curve was 0.9116, indicating the combination of these circRNAs may have good diagnostic performance as the full panel of included circRNAs in osteosarcoma. However, more caution should be taken in the interpretation of this result since the comparatively small study numbers and limited sample size may produce outlying outcome and thereby introduce potential bias.

Expression of circRNAs with prognosis in osteosarcoma
Survival analysis showed that overexpression of oncogenic circRNAs was significantly correlated with worse OS   Figure 3A and 3B, respectively. We identified one outlier study performed by Zheng S et al. in the combined effect of oncogenic circRNAs by sensitivity analysis. Besides, elevated expression of tumor-suppressor circRNAs predicted favorable OS (HR = 0.44, 95% CI: 0.28--0.69), as depicted in Figure 3C. The fixed-effect model was applied in these studies since no obvious heterogeneity was noted.

Sensitivity analysis and publication bias
The sensitivity analysis was performed in the prognostic effect sizes by omitting the enrolled studies one by one. For pooled  effects of upregulated circRNAs in osteosarcoma, one study conducted by Zheng S et al. [37] was identified as the outlier ( Figure 4A). Notably, the predictive significance of OS for upregulated circRNA did not alter after excluding the aforementioned outlier data (HR = 2.49, 95% CI: 2.11-2.94). No outliers were found in other pooled effects ( Figure 4B, 4C). The Deeks' funnel plot test (P = 0.21) demonstrated the absence of publication bias existed in diagnostic analysis ( Figure 5A). For prognosis analysis of upregulated circRNA on DFS and downregulated circRNA on OS, the Begg's funnel plot also showed symmetry and revealed no evidence of publication bias among the eligible studies ( Figure 5D, 5E). Nevertheless, the funnel plot of upregulated circRNA profile on OS showed significant asymmetry, indicating a possible publication bias ( Figure 5B). Therefore, a 'trim and fill' method was applied to trace the possible impacts from bias as previously described. However, the predictive value of upregulated circRNAs on OS was not altered after adjustment, suggesting a closely correlation between upregulated circRNAs and poor OS among osteosarcoma patients ( Figure 5C).

Discussion
Emerging studies have established an important role of circRNAs in cancer initiation and progression. In recent years, circRNAs have been implicated as novel diagnostic and prognostic biomarkers in several cancers [54], including colorectal cancer [34,55], hepatocellular cancer [55][56][57], gastric cancer [55,58], pancreatic cancer [55], esophageal cancer [55,59], thyroid cancer [60], ovarian cancer [61], and lung cancer [11,62]. Our study implicated a marked correlation between abnormal circRNAs expression levels with clinical, diagnostic, and prognostic significance in osteosarcoma. In particular, oncogenic circRNAs with higher expression pattern were strikingly associated with unfavorable clinical features, including larger tumor size, advanced clinical stage, poor differentiation, and DM, while the tumor-suppressive circRNAs showed an opposite correlation. For diagnosis roles, our study suggested the AUC of 0.87, with 79% sensitivity and 81% specificity of circRNAs in osteosarcoma. Besides, for the prognostic value, abnormal expression of circRNAs was remarkably related with OS as well as DFS. Furthermore, there were two studies reported the correlation between circRNA levels with PFS. Specifically, one demonstrated that highly expressed circ_0008717 was associated with worse PFS in osteosarcoma patients [36], while another study conducted by Wu Z et al. [42] showed that downregulated circ_0002052 predicted unfavorable PFS. Given the discrepant expression pattern of these two circRNA, it was unable to perform a pooled analysis concerning the prognostic significance of circRNA on PFS. Therefore, more studies describing the relationship between circRNAs and PFS in osteosarcoma are still needed.
Previously, Huang X et al. performed a meta-analysis on predictive values of circRNAs in osteosarcoma in 2019 [22]. However, only 13 studies were enrolled in their study, with 9 about clinical features, 11 on prognosis, and 5 about diagnosis. Therefore, the pooled resulted may be underpowered due to insufficient data. Since the studies on circRNAs in osteosarcoma are emerging, we updated the predictive value of circRNAs in osteosarcoma in the study, with 31 studies comprising 1979 patients, which markedly increased the statistical power and made the pooled results more credible. Besides, Huang X and colleagues did not report the association between upregulated circRNA with DFS of osteosarcoma due to limited eligible studies, whereas our study report for the first time that oncogenic circRNA was significantly correlated with poor DFS in addition to OS.
The diagnostic value of circRNAs as potential biomarkers for osteosarcoma was extensively explored in our study. Aberrant expression of circRNAs was shown in a wide range of sample sources from osteosarcoma cell lines or patients, including tumor tissue and plasma. Given the fact that circRNA is widely expressed in human samples and stable in structure, it may be ideal biomarker candidate with diagnostic significance in osteosarcoma [47]. The SROC in our study showed that the pooled AUC of circRNA in osteosarcoma was 0.87, indicating that 87% of randomly chosen osteosarcoma patients would have abnormal expression levels of circRNAs when compared with controls. Compared with receiver operating characteristic (ROC), which only compares single test accuracy over divergent thresholds for positivity, the each data point in SROC originates from a different study rather than a different threshold [63,64]. Thus, SROC analysis could provide better evaluation for overall accuracy across SENS and SPEC from multiple studies [65]. Of note, an AUC under SROC curve over 0.75 usually indicates good diagnostic accuracy. Accordingly, the AUC of 0.87 suggests that our analysis was accurate and credible [65].
Moreover, the respective PLR and NLR were 4.10 and 0.26, which indicated that the circRNA signature reached 4.10 between the true positive and false-positive rates, and the probability of osteosarcoma patients that tested negative for circRNAs versus the case tested positive had a ratio of 0.26. In addition, the pooled DOR of 16.03 was obtained, implicating that dysregulated circRNAs were a powerful predictive biomarker for osteosarcoma diagnosis. Currently, serum alkaline phosphatase (ALP), with AUC of 0.673, is one of the wellknown biomarkers for osteosarcoma diagnosis [17]. Moreover, the sensitivity of ALP in OS diagnosis is also comparatively low, since over 40% of osteosarcoma patients may have normal ALP expression [5]. Our results indicated that dysregulated circRNAs may be even better than ALP for separating osteosarcoma patients from normal subjects with both enhanced sensitivity and specificity, suggesting a potential application in clinical practice.
However, several obstacles also remain in the utility of circRNAs for diagnosis. For instance, we were not capable of drawing a quantitative scale of dysregulated circRNAs for diagnosis, since each eligible study only provided ROC curve for diagnosis, and there was not too much information for circRNA expression value. Moreover, circRNAs obtained from tissue may be invasive to the patients, and detection of circRNAs from exosome may be expensive or technically difficult, which could limit the widespread application of multiple circRNAs as biomarkers [9]. Therefore, we also explored the diagnosis accuracy of the combination of a small set of certain oncogenic circRNAs, which may be more available than the full panel of included circRNAs in clinical application. Our preliminary pooled result showed that combination of three oncogenic circRNA, including circPVT1, circ_0081001, and CDR1as, has an AUC of 0.9116, with SENS of 0.85 and SPEC of 0.83 in diagnosis. However, as we mentioned above, the interpretation of the small set of circRNAs in separating osteosarcoma patients from control should be more cautious due to the small study number as well as limited sample size.
Additionally, it worth noted that several other limitations remain to be addressed in this study. First, the number of included articles on downregulated circRNAs was still comparatively small, making it difficult to conduct a stratified analysis in terms of certain clinical characteristics. Second, some HRs with 95% CIs were indirectly extracted from Kaplan-Meier curves, which may generate possible bias. Third, the detailed function and underlying mechanisms of circRNAs in osteosarcoma were not thoroughly annotated. Fourth, all of the studies were performed in China, which may introduce population bias. Thus, more studies with different ethnic groups are still warranted in future studies. Fifth, all circRNAs expression data in eligible studies were obtained from serum or tissue without reporting exosomal circular RNA, an important biomarker in multiple cancers [66][67][68]. Hence, the diagnostic and prognostic potential of exosomal circular RNAs should be investigated in upcoming studies.
In short, this study indicated a significant correlation between aberrant expression of circRNAs with the clinical, diagnostic,, and prognostic values in osteosarcoma patients. Therefore, circRNAs could serve as promising biomarkers and therapeutic targets for osteosarcoma.

Availability of data and materials
The data used and analyzed in the study is available from the corresponding author on reasonable request.