Background

Circular RNAs (circRNAs) are a class of single-stranded endogenous RNAs with covalently closed-loop structures generated by back-splicing [1]. According to the origin of sequences, circRNAs are classified into three major groups, namely exonic circRNA (EcRNA), exon-intron circRNA (EIciRNA) and intronic circRNA (ciRNA). The biogenesis of circRNA is tightly regulated by cis-acting elements and trans-acting proteins. Moreover, a single gene locus could produce multiple circRNA isoforms by alternative circularization [2]. Recently, circRNAs have emerged as essential regulators in various biological processes through diverse mechanisms such as regulating transcription, sponging microRNA (miRNA), acting as RNA-binding protein (RBP) decoy, and translating into protein [1, 3].

Increasing evidence has revealed that circRNAs are commonly dysregulated in cancer and exert crucial functions during tumor initiation and development. For instance, circ-SMO is significantly upregulated in glioblastoma (GBM) and activates the Hedgehog pathway to promote tumorigenesis through encoding a novel protein SMO-193a.a. [24], suggesting their potential functions in the maintenance of basal cellular processes. Moreover, these ubiquitous genes were predominantly enriched in biological processes and signaling pathways that are fundamental to cells, such as RNA splicing and transporting (Additional file 4: Fig. S1B-C). We also identified 1765 tissue-specific protein-coding genes (TSI ≥ 0.85, Additional file 4: Fig. S1D and Additional file 6: Table S9) and further confirmed their gene expression profiles in normal tissue from the GTEx database (Additional file 4: Fig. S2). Functional enrichment analysis of these tissue-specific genes demonstrated tissue-specific biological processes and signaling pathways (Additional file 4: Fig. S1E-F and Fig. S3). For instance, the genes specific to the brain were enriched in brain-specific biological processes and signaling pathways, such as synapse, learning, cognition and axon guidance (Additional file 4: Fig. S1E-F). Collectively, these results confirmed the reliability of our sample preparation and data analysis.

Considering that most circRNAs were discovered at low frequency and expressed at low levels, 1762 circRNAs with high abundance (median RPM expression value ≥0.1 in at least one tissue) were retained to obtain a high-confidence profile. As a result, 30 circRNAs with TSI ≤ 0.3 were defined as the ubiquitous circRNAs (Fig. 2A and Additional file 7: Table S10), indicating their important functions in fundamental cellular processes. In addition, we found 439 circRNAs with TSI ≥ 0.85 and defined them as tissue-specific circRNAs (Fig. 2B and Additional file 8: Table S11), implying their tissue-specific functions. Moreover, we found that the brain exhibited the highest number of tissue-specific circRNAs while the intestine exhibited the least (Fig. 2B and Additional file 8: Table S11). For instance, we identified that hsa_circ_0000497 was exclusively highly expressed in the brain (Fig. 2E), which was further supported by circAtlas data (Additional file 4: Fig. S4). However, the functions of these ubiquitous and tissue-specific circRNAs remain elusive.

Fig. 2
figure 2

Tissue-specific expression patterns of circRNAs in human normal tissues. A Heatmap of ubiquitously expressed circRNAs. B Heatmap of tissue-specifically expressed circRNAs. C The number of circRNAs produced from tissue-specific genes. D The Pearson correlation coefficients of tissue-specific circRNAs with host gene expression. E The tissue specificity of hsa_circ_0000497 was governed by the brain-specific host gene SLAIN1. F The Pearson correlation coefficients of tissue-specific circRNAs with RBP expression. G Expression of RBFOX1, hsa_circ_0070057 and its host gene SEPT11 in different tissues. H Correlation of hsa_circ_0070057 and RBFOX1 expression. I eCLIP-seq read density for RBFOX1 within introns flanking the hsa_circ_0070057 back-splicing site. The vertical red lines indicate the GCATG binding motifs of RBFOX1

Full size image

Tissue-specific expression of host genes and/or RBPs contributes to the tissue specificity of circRNAs

The tissue-specific expression signatures of circRNAs may be the consequence of their biogenesis. Given that most circRNAs were processed from the primary transcripts transcribed from gene loci, we initially examined whether both circRNA and its host gene showed the same tissue specificity and found that only 135 (30.75%) circRNAs were produced from tissue-specific host genes (Fig. 2C and Additional file 4: Fig. S5A). Pearson correlation analysis further supported that circRNAs derived from tissue-specific genes exhibited a strong correlation with their host genes (Fig. 2D), implying that the expression of these circRNAs highly depend on their host gene expression. For example, SLIAN1 is a microtubule plus-end tracking protein exclusively expressed in the brain to regulate microtubule dynamics and promote axonal development [25], so hsa_circ_0000497 derived from the SLAIN1 gene was also highly expressed in the brain (Fig. 2E and Additional file 4: Fig. S5B-D). Therefore, tissue-specific expression of host genes contributes to the tissue specificity of these tissue-specific circRNAs.

It was worth noting that 304 (69.25%) tissue-specific circRNAs were derived from nontissue-specific genes, implying that other factors contribute to their tissue specificity. As RBPs are known to play a role in circRNA biogenesis, we speculated that tissue-specific RBPs may regulate the tissue specificity of circRNAs. According to a previous study [6], 215 RBPs (Additional file 2: Table S2) were selected to calculate their TSI, and 8 RBPs (Additional file 3: Table S12) with TSI ≥ 0.85 were identified as tissue-specific. Next, we analyzed correlations between 304 circRNAs and 8 RBPs, and found that 155 (35.31%) circRNAs showed a strong correlation with these RBPs (Pearson correlation coefficient ≥ 0.7, Fig. 2F), indicating that the tissue specificity of these circRNAs may be regulated by tissue-specific RBPs. Among these RBPs, RBFOX1 was found to be exclusively expressed in the brain (TSI = 1) (Fig. 2G), consistent with a previous study [26]. Therefore, we chose RBFOX1 as an example to investigate its potential regulatory mechanisms on the biogenesis of tissue-specific circRNAs. First, we found that brain-specific hsa_circ_0070057 had a strong association with RBFOX1 expression in the brain, not with the expression of its host gene (Fig. 2G-H), Furthermore, eCLIP-seq data of RBFOX1 [21] demonstrated that RBFOX1 indeed bound to the flanking introns of hsa_circ_0070057 (Fig. 2I), suggesting that RBFOX1 facilitates the biogenesis of this circRNA. In addition, brain-specific hsa_circ_0073539 may also be regulated by QKI (TSI = 0.86 in brain) which is a well-known factor in circRNA biogenesis (Additional file 4: Fig. S6) [27], not by its host gene MAN2A1. Together, tissue-specific expression of RBPs may play an important role in the biogenesis of tissue-specific circRNAs.

Expression patterns of differentially expressed circRNAs in cancer

Emerging evidence has reported that circRNAs are frequently dysregulated in cancer. To depict the abnormal expression of circRNAs in cancer, we compared the global circRNA abundance in tumors and their paired normal samples, and found the downregulation of global circRNA abundance in almost all cancer types (Fig. 3A) under the comparable sequencing depths (Additional file 4: Fig. S7A and Additional file 9: Table S13). As shown in Fig. 3B, we also identified the differentially expressed circRNAs in each type of cancer, and the detailed upregulated and downregulated circRNAs are listed in Additional file 10: Table S14. Interestingly, GBM (n = 603) showed the most differentially expressed circRNAs while GC (n = 123) had the least (Additional file 4: Fig. S7B). Next, we investigated whether these differentially expressed circRNAs were cancer-specific. As expected, each cancer exhibited unique patterns of differentially expressed circRNAs (Fig. 3B-C). Among the 1209 dysregulated circRNAs, 962 (79.57%) were found to be differentially expressed in only one cancer type (Fig. 3C and Additional file 10: Table S14), implying their cancer-specific roles in tumorigenesis.

Fig. 3
figure 3

Characterization of differentially expressed circRNAs in seven solid tumors. A Total circRNA abundance in tested cancers. B Profiles of differentially expressed circRNAs identified in this study. The blue box indicates the unique circRNA expression signatures of each cancer. C The number of cancer-specific differentially expressed circRNAs. D List of differentially expressed circRNAs shared by at least two types of cancers. E Heatmap of twenty-five differentially expressed circRNAs shared by at least four types of cancers. The asterisk indicates circRNA expression is significantly altered. Red and green colors indicate upregulation and downregulation, respectively. F The number of differentially expressed circRNAs that exhibited consistent or inconsistent expression changes in at least two cancer types. G Expression alteration of RBPs associated with circRNA biogenesis in GBM. H eCLIP-seq read density for RBFOX1 within introns flanking hsa_circ_0111312 back-splicing site. The vertical red box indicates the GCATG binding motifs of RBFOX1

Full size image

In addition, 247 differentially expressed circRNAs were dysregulated in at least two cancer types (Fig. 3D and Additional file 10: Table S14). Moreover, 25 circRNAs were found to be differentially expressed in at least four types of cancers (Fig. 3E and Additional file 10: Table S14). For instance, hsa_circ_0072309 was consistently downregulated in our tested tumor samples except for GBM (Fig. 3E), highlighting its pivotal role in tumorigenesis. Of 247 circRNAs, the expression of 187 circRNAs exhibited consistent changes across different tumors (Fig. 3F and Additional file 10: Table S14), implying their common functions during tumorigenesis. For instance, hsa_circ_0077837 was significantly downregulated in CRC, ESCC, LUAD, GC and HCC (Fig. 3E and Additional file 4: Fig. S8A), indicating its universal tumor-suppressive role in cancer, which was recently proven in HCC [28]. In addition, abnormal expression of 60 circRNAs exhibited inconsistent tendencies among cancer types (Fig. 3F and Additional file 10: Table S14), suggesting their diverse roles in different cancers. For example, we found that hsa_circ_0001681 was elevated in 4 cancer types (CRC, LUAD, GC, THCA) but dramatically downregulated in GBM (Fig. 3E and Additional file 4: Fig. S8B). Consistent with our findings, hsa_circ_0001681 was previously reported to be upregulated and exert an oncogenic function in THCA [29], but to be downregulated and play a tumor-suppressive role in renal cell carcinoma [30], Therefore, such dysregulated circRNAs may exert their different functions in a cancer-type context.

Abnormal expression of the host genes and/or RBPs contribute to circRNA dysregulation in cancer

Of note, the prevalence of differential expression profiles of circRNAs in cancer further raised the question of how these circRNAs were dysregulated. Vo et al. reported that the dysregulation of some circRNAs in prostate cancer was predominantly caused by aberrant expression of their parental genes [7]. In our study, an average of 41.84% (ranging from 33.50 to 57.81%) of differentially expressed circRNAs were explained by differentially expressed genes (Additional file 4: Fig. S9). For instance, hsa_circ_0099329 and its host gene PPFIA2 were exclusively downregulated in GBM (Additional file 4: Fig. S10A). In addition, we also observed that the alterations of some circRNAs (ranging from 42.19 to 66.50%) were inconsistent with their host genes (Additional file 4: Fig. S9), suggesting the involvement of other mechanisms in circRNA dysregulation.

Given that 66.50% of dysregulated circRNAs in GBM were derived from genes without significant changes in expression, we focused on GBM to explore whether the dysregulated circRNAs were associated with RBPs. We first examined the expression patterns of 215 RBPs involving deaminase, endonuclease, helicase and splicing factor in GBM (Additional file 11: Table S15). As shown in Fig. 3G, these RBPs such as well-characterized PTBP1 [31] and ESRP1 [32] were generally observed to be altered in GBM. Of these, PTBP1 was elevated in GBM, suggesting its potential to regulate biogenesis of some circRNAs, which wa supported by a previous report [31]. In addition, ESRP1 was decreased in GBM, implying that the downregulation of circRNAs in GBM might be affected by this RBP.

We quantified the correlations between dysregulated RBPs and dysregulated circRNAs from nondysregulated genes, and found that 174 (28.86%) circRNAs showed a strong correlation with RBPs (Pearson correlation coefficient ≥ 0.7), implying that the dysregulation of these circRNAs might be regulated by RBPs. Intriguingly, among these dysregulated RBPs, we noticed that RBFOX1, a neuron-specific splicing factor involved in neurodevelopmental and neuropsychiatric disorders [26, 33], was the most significantly downregulated RBP in GBM in our study (Fig. 3G and Additional file 4: Fig. S10B) and in the GEPIA database. Therefore, we examined the differentially expressed circRNAs associated with RBFOX1 and found that downregulation of hsa_circ_0111312 in GBM was strongly correlated with decreased RBFOX1 expression in GBM (Additional file 4: Fig. S10C), independent of expression of its host gene RASAL2 (Additional file 4: Fig. S10D). Furthermore, eCLIP data of RBFOX1 [21] confirmed the strong binding of RBFOX1 to the flanking introns of hsa_circ_0111312 (Fig. 3H), suggesting the involvement of RBFOX1 in circRNA biogenesis. Therefore, aberrant expression of RBPs in GMB can contribute to a subset of circRNA dysregulation.

Characterization of circLIFR in cancer cells

According to our bioinformatics analysis, hsa_circ_0072309 exhibited high abundance in normal tissues and was downregulated in our tested tumors except for GBM. Since hsa_circ_0072309 is derived from exons 2–5 of the LIFR gene (Fig. 4A), we termed it as circLIFR. First, we designed the F1/R1 divergent primers (F1 primer crossing the junction site) to conduct RT-qPCR, and the results confirmed that circLIFR was indeed downregulated in our tested tumors, except for GBM, compared with their adjacent normal tissues (Fig. 4B and Additional file 4: Fig. S11A), implying that circLIFR could play a tumor-suppressive role in solid tumors. To further characterize circLIFR, we performed PCR with divergent (red) and convergent primers (blue) (Fig. 4A) using cDNA and genomic DNA (gDNA) as templates. As shown in Fig. 4C, the expected bands were observed using convergent primers from both cDNA and gDNA templates, but there was no band using gDNA as a template and the divergent primers, confirming that circLIFR was a circular RNA produced by back-splicing, which was further validated by RNase R digestion treatment (Fig. 4D) and Sanger sequencing in different cancer cells (Fig. 4E and Additional file 4: Fig. S11B). To detect the cellular distribution of circLIFR, we isolated the cytoplasmic and nuclear fractions of cells to measure circLIFR levels by RT-qPCR, and found that circLIFR was predominantly located in the cytoplasm (Fig. 4F and Additional file 4: Fig. S11C), which was confirmed by FISH experiments (Fig. 4G and Additional file 4: Fig. S11D). We also evaluated circLIFR stability after actinomycin D treatment and found that circLIFR was more stable than its linear transcript LIFR mRNA (Fig. 4H). Collectively, these results demonstrated that circLIFR was a bona fide circRNA, predominantly distributed in the cytoplasm and significantly downregulated in our tested tumors except for GBM.

Fig. 4
figure 4

Characterization of circLIFR in solid tumors. A Schematic illustration showing circLIFR is derived from exons 2–5 of LIFR mRNA. The F1/R1 primers were used to measure circLIFR by RT-qPCR. DP, divergent primers; CP, convergent primers. B circLIFR expression measured by RT-qPCR in ESCC, HCC and their matched adjacent normal tissues, normalized to U6 RNA levels. N, normal tissues; T, tumor tissues. Data represent the mean ± S.D., and the p-value was determined by two-tailed paired Student’s t-test. C The presence of circLIFR was validated in KYSE150 cells by RT-PCR. Divergent primers were used to amplify circLIFR from cDNA, but not from genomic DNA. LIFR mRNA linear transcripts were used as a negative control. cDNA, complementary DNA; gDNA, genomic DNA. D Total RNA from KYSE150 cells with or without RNase R treatment was subjected to RT-PCR. E Agarose gel electrophoresis and Sanger sequencing of RT-PCR products of circLIFR in KYSE30 and SK-Hep-1 cells. F RT-qPCR analyses of circLIFR after cytoplasmic/nuclear fractionation of KYSE30 and SK-Hep-1 cells. β-actin mRNA and U6 RNA represent cytoplasmic and nuclear RNAs, respectively. Western blotting confirmed the efficiency of nuclear/cytoplasmic isolation. Data are shown as the mean ± SD. G Fluorescence in situ hybridization of circLIFR in KYSE30 and SK-Hep-1 cells. 18S and U6 RNA represent cytoplasmic and nuclear RNAs, respectively. H circLIFR and LIFR mRNA levels in KYSE30 and SK-Hep-1cells without or with actinomycin D treatment

Full size image

circLIFR inhibits cell migration in vitro and tumor metastasis in vivo

circLIFR was reported to affect cellular functions in breast cancer [34], bladder cancer [35], renal carcinoma [36] and gastric cancer [37]. However, there are no reports of circLIFR in ESCC, HCC, THCA and CRC. To investigate the biological functions of circLIFR in these tumors, we first measured circLIFR levels in cancer cell lines (Additional file 4: Fig. S12A) and selected 4 cancer cells with low circLIFR expression (KYSE30 and KYSE150 esophageal cancer cells, SK-Hep-1 liver cancer cells, HCT-116 colorectal cancer cells) to construct stable cells overexpressing circLIFR. As shown in Fig. 5A and Additional file 4: Fig. S12B, circLIFR was correctly circularized and successfully expressed in these cells. Transwell migration and wound healing assays showed that overexpression of circLIFR apparently inhibited cell migration and cell motility (Fig. 5B-C and Additional file 4: Fig. S12C-D). Next, we sought to evaluate the function of circLIFR in vivo. To this end, we conducted a tail-vein injection mouse model using circLIFR-overexpressing KYSE30 cells to check its role in lung metastasis. The results demonstrated that after ectopic expression of circLIFR in esophageal cancer cells, tumor metastasis was significantly inhibited compared to that of control cells in lung metastasis models (Fig. 5D). To further evaluate the function of circLIFR, we injected circLIFR-overexpressing SK-Hep-1 cells into the spleen, and observed that circLIFR dramatically repressed tumor metastasis to mouse liver (Fig. 5E). Together, circLIFR significantly inhibited cell migration in vitro and tumor metastasis in vivo.

Fig. 5
figure 5

circLIFR inhibits cell migration in vitro and tumor metastasis in vivo. A Agarose gel electrophoresis and Sanger sequencing of semi-quantitative RT-PCR products demonstrated that circLIFR was correctly circularized and successfully overexpressed in KYSE30 and SK-Hep-1 cells. B Transwell migration assays showed that circLIFR overexpression inhibited the migratory ability of KYSE30 and SK-Hep-1 cells. C Wound healing assays showed that circLIFR overexpression inhibited cell migration in both KYSE30 and SK-Hep-1 cells. D Decreased tumor metastasis formed in the lungs of mice via tail vein injection of circLIFR-overexpressing KYSE30 cells, as indicated by representative bioluminescent images of lung metastasis at 6 weeks (left) and H&E staining of lung metastatic lesions (right) of mice for each group (n = 5 mice/group). E Decreased tumor metastasis formed in the livers of mice via spleen injection of circLIFR-overexpressing SK-Hep-1 cells, as indicated by representative bioluminescent images of liver metastasis at 8 weeks (left) and H&E staining of liver metastatic lesions (right) of mice for each group (n = 5 mice/group). F Volcano plot of differentially expressed genes affected by circLIFR in KYSE30 cells. G,H Gene Ontology (G) and GSEA analysis (H) showed that dysregulated genes upon circLIFR overexpression was involved in cell adhesion. NES, normalized enrichment score; FDR, false discovery rate. p-values are calculated by permutation test. I,J Quantification of metastasis-related genes by RT-qPCR in KYSE30 control cells and circLIFR-overexpressing cells. Data represented mean ± S.D.; the p-values were determined by a two-tailed unpaired Student’s t-test

Full size image

To explore the underlying mechanism of circLIFR, we performed RNA-seq for transcriptome analyses. As shown in Fig. 5F, ectopic expression of circLIFR affected the expression levels of 448 genes, of which 292 genes were upregulated and 156 genes were downregulated, including epithelial-mesenchymal transition (EMT) marker genes (Additional file 12: Table S16). Gene Ontology analyses indicated that these differentially expressed genes were significantly enriched in cell adhesion, extracellular matrix organization and cell-substrate junction assembly (Fig. 5G), which was further confirmed by GSEA analysis results (Fig. 5H). Finally, the expression change of a subset of metastasis-related genes affected by circLIFR was validated by RT-qPCR (Fig. 5I-J). Therefore, circLIFR inhibits tumor metastasis by regulating a subset of cell adhesion and EMT-related genes.

Discussion

Accumulating evidence has demonstrated that circRNAs are frequently altered in cancer and play indispensable roles during cancer initiation and progression. Recently, several studies have attempted to present the landscape of circRNAs in tumor tissues and cancer cell lines, shedding light on the importance of circRNAs [6, 7, 38, 39]. However, the lack of matched normal tissues in these studies cannot depict the dysregulated circRNAs across cancer types. Therefore, we collected 122 tumor samples of seven types of malignancies (GBM, ESCC, LUAD, HCC, GC, THCA and CRC) and their matched normal tissues to conduct Ribo-zero RNA-seq, which allows us to systematically profile circRNAs in normal tissues and their matched tumors without bias.

From these samples, we identified 59,056 circRNAs, most of which were originated from exons (n = 59,020). Interestingly, we found that 182 circRNAs harbored a complete CDS as they originated from both the CDS and UTR, indicating their potential for translation into intact peptides. Consistent with previous studies [3, 40], we found that the majority of circRNAs were expressed at low levels. However, we identified 22 circRNAs with much higher abundance than their cognate linear transcripts (Additional file 3: Table S6), suggesting their important roles in basic cellular functions. Although a single gene locus can produce multiple circRNAs via alternative circularization, this gene shows a tendency to express a predominant circRNA isoform (Fig. 1G-H). This could be explained by our analysis that the introns flanking circularized exons of the predominant circRNA isoform are much longer and have more repetitive elements (Fig. 1I-J). In addition, we identified 30 circRNAs that were ubiquitous across our tested tissues, and 439 circRNAs were tissue-specific, indicating their general and tissue-specific cellular functions (Additional file 7: Table S10 and Additional file 8: Table S11).

Compared with normal tissues, the global abundance of circRNAs was remarkably decreased in tumors (Fig. 3A), which may be caused by inefficient back-splicing events or the dose dilution of nascent circRNAs in the rapid division of tumor cells [41, 42]. This was further supported by the fact that the total abundance of circRNAs was significantly elevated when inhibiting cell proliferation [7]. Moreover, we identified 1209 differentially expressed circRNAs in cancers (Additional file 10: Table S14), of which 962 (79.57%) were cancer-specific, while 247 were dysregulated in at least two cancer types. Of the 247 noncancer-specific circRNAs, 187 circRNAs were consistently altered in tumors, while 60 circRNAs exhibited distinct expression patterns in different cancer types, indicating that consistently altered circRNAs exert a common role in cancers, while inconsistently altered circRNAs may play distinct functions in different cancers. For instance, hsa_circ_0001681 was elevated in papillary thyroid cancer and promoted tumor growth by sponging miR-198 [29], while downregulation of circRAPGEF5 in renal cell carcinoma inhibited cancer progression by sequestering miR-27a-3p [30], supporting its diverse roles in cancers.

Given that some circRNAs originating from cancer tissues may enter the circulatory system and be stably present in sera [43, 44], cancer-specific dysregulated circRNAs could serve as biomarkers for cancer diagnosis and prognosis. For instance, the EML4-ALK fusion gene produces a novel F-circEA in non-small cell lung cancer (NSCLC), and the existence of F-circEA in sera makes it a promising liquid biopsy biomarker for monitoring the EML4-ALK fusion gene [45]. Moreover, circRNAs driving tumor progression could be therapeutic targets. For example, circPRKCI was demonstrated to be a proto-oncogenic circular RNA in lung adenocarcinoma tissues, and the intratumor injection of cholesterol-conjugated siRNA specifically targeting circPRKCI inhibited tumor growth in a patient-derived lung adenocarcinoma xenograft model [46]. Therefore, elucidation of the biological functions and clinical relevance of such dysregulated circRNAs is helpful to identify new diagnostic biomarkers and therapeutic targets.

Our study provided abundant dysregulated circRNAs to be experimentally validated for their roles in tumorigenesis. Here we focused on hsa_circ_0072309 (circLIFR) because of its consistent decrease in LUAD, ESCC, CRC, GC, HCC and THCA. First, we confirmed the downregulation of circLIFR in these six types of cancers by RT-qPCR, supporting the robustness of our pipeline for transcriptome analyses. Combined with recent findings about circLIFR in breast cancer [34], renal carcinoma [36] and bladder cancer [35], downregulation of circLIFR seems to be general in cancers, suggesting its important tumor-suppressive roles in cancer pathology. Indeed, circLIFR was reported to repress the progression of breast cancer [34], renal carcinoma [36] and GC [37]. In esophageal cancer and liver cancer models, we verified that circLIFR inhibited cell migration in vitro and tumor metastasis in vivo. Therefore, circLIFR plays a pivotal role during tumorigenesis. To explore the effects of circLIFR on gene expression, we performed RNA-seq on circLIFR-overexpressing KYSE30 cells and revealed that circLIFR significantly altered the expression patterns of metastasis-related genes (such as cell adhesion molecules). Mechanistically, circLIFR was reported to act as a sponge of miR-492 in breast cancer [34] and miR-100 in renal carcinoma [36] to inhibit tumor progression. In bladder cancer, circLIFR interacted with the MSH2 protein to positively modulate cisplatin-sensitivity through the MutSα/ATM-p73 axis [35]. However, the detailed underlying mechanism of circLIFR in esophageal cancer is still vague and needs further investigation.

In addition, we characterized some tissue-specific circRNAs in our tested normal tissues and cancer-specific dysregulation of circRNAs in tumors. Given that circRNA biogenesis is regulated by cis-acting elements and trans-acting proteins, we analyzed the factors to determine circRNA expression. As shown in Fig. 2C and Additional file 4: Fig. S5A, we found that the tissue-specific expression of the host genes contributes to 30.75% of tissue-specific circRNAs, while aberrant expression of the host gene contributes to approximately 41.84% of dysregulated circRNAs, indicating that the expression levels of host genes are an important factor for circRNA expression. Moreover, we identified that 35.31% of tissue-specific circRNAs and 28.86% of dysregulated circRNAs in GBM were highly correlated with RBPs, which was further supported by published eCLIP-seq data (Fig. 2I and Fig. 3H), demonstrating pivotal roles in RBPs in circRNA biogenesis. Because we found a strong association between dysregulated RBPs (such as PTBP1, RBFOX1 and ESRP1) and the expression of a subset of circRNAs in GBM, it is important to dissect how these RBPs affect circRNA biogenesis in the future.

Conclusions

In this study, we identified the expression landscape of circRNAs in paired normal-tumor clinical samples from seven types of malignancies. The dysregulated circRNAs exhibited cancer-specific expression or shared common expression signatures across cancers, which could be regulated by the expression of host genes and RBPs. Finally, we experimentally validated that circLIFR was downregulated and played a tumor-suppressive role in tumors. Collectively, this study provides the comprehensive profiles of differentially-expressed circRNAs in cancers and highlights the important roles of circRNAs in cancer biology.