We recruited 20 patients with stage III-IV high-grade serous ovarian cancer from the Affiliated Cancer Hospital of Nanjing Medical University. The patients were aged from 50 to 65, born in China, lived in the province of Jiangsu and did not receive any anticancer treatment before surgery. Ovarian cancer tissues and paired normal ovarian tissues were collected, flash frozen in liquid nitrogen, and stored at −80°C. The ethics committee of The Affiliated Cancer Hospital of Nanjing Medical University approved the study protocol, and informed consent was obtained from all the patients.

Ovarian cancer cell lines (A2780, SKOV3, and SW626) and a normal ovarian epithelial cell line (IOSE80) were purchased from Shanghai Guan & Dao Biological Engineering Co., Ltd. The identities of the cell lines were authenticated by short-tandem-repeat (STR) profiling. All four cell lines were cultured in DMEM (Keygen Biotech, Nanjing, China) with 10% fetal bovine serum, 100 units/ml penicillin, and 100 μg/ml streptomycin.

We randomly selected 3 paired ovarian cancer tissues and normal ovarian tissues from the 20 paired samples for circRNA profiling. CircRNAs were profiled using the RNA-seq technique by Gene Denovo Biotechnology Co. (Guangzhou, China). Briefly, total RNA was extracted and then treated with Ribonuclease R (RNase R) to degrade linear RNAs. A strand-specific library was constructed using a VAHTS Total RNAseq (H/M/R) Library Prep Kit (Vazyme, Nanjing, China). After removing ribosomal RNAs, the enriched circRNAs were fragmented into short fragments and reverse transcribed into cDNA. Second-strand cDNA was synthesized. The cDNA fragments were purified, end repaired, poly(A) added, and ligated to Illumina sequencing adapters. The second-strand cDNA was digested with uracil-N-glycosylase (UNG). The digested products were purified, PCR amplified, and sequenced using an Illumina Novaseq60000. The raw RNA-seq data have been submitted to the Gene Expression Omnibus (GEO) database (accession number GSE192410).

After obtaining clean reads and removing ribosome RNA (rRNA)-mapped reads, the remaining reads were mapped to the reference genome. Then, the unmapped reads were collected for circRNA identification using find_circ (version 1) (Memczak et al., 2013). Finally, we filtered raw read counts of circRNAs and kept only the genes with a nonzero read count in 3 or more samples and then used DESeq2 (version 1.32) (Love et al., 2014) in the statistical program R (version 4.0.1) to normalize the read counts. Differentially expressed circRNAs were identified based on a fold change of at least 2 and an adjusted p value <0.05.

CircPrimer (Zhong et al., 2018b) was used to annotate circRNAs and design primers for circRNAs. We performed reverse transcription and RT–qPCR as in our previous study (Zhong et al., 2018a; Zhou et al., 2020). The expression levels of genes were detected as described in our previous report (Wang et al., 2020b). β-action was used as a reference gene. All samples were analyzed in duplicate for each circRNA.

Plasmids expressing circATRNL1 (p-circATRNL1) and circZNF608 (p-circZNF608) as well as control plasmids (p-control) were purchased from HanBio Biotechnology (Hanbio, Shanghai, China). Cells in logarithmic phase were harvested and seeded at 4 × 10⁵ cells per well into a 6-well plate. After incubation for 24 h, the cells were transfected using Lipofectamine 3000 (Invitrogen, Carlsbad, CA, USA) following the manufacturer’s instructions.

The FISH probes and Ribo™ fluorescent in situ hybridization Kit were purchased from Guangzhou RiboBio Co., Ltd. (RiboBio, Guangzhou, China). In situ hybridization was conducted as described previously (Zhong et al., 2018a; Zhou et al., 2020).

After transfection for 24 h, 4,000 cells were seeded per well (96-well plate) in the presence of 200 μL of complete medium. The cells were divided into four groups. After cell attachment, the medium of the first group was removed, 100 μL CCK8 was added, the cells were incubated at 37°C for 1 h. Then, the absorbance was measured at 450 nm. The other three groups were detected at 24, 48 and 72 h, respectively.

Transwell migration and invasion assays were conducted as described previously (Zhou et al., 2020). For cell migration, 4 × 10⁴ cells were seeded into the upper chamber of transwell assay inserts (pore size of 8 mm; Corning) containing 200 μL DMEM medium without serum. We added 600 μL DMEM with 20% fetal bovine serum to the lower chamber of 24-well culture plates. After incubation for 48 h, the cells on the filter surface were fixed with 4% paraformaldehyde, stained with crystal violet and imaged using a microscope (Carl Zeiss). For cell invasion, serum-free DMEM containing 100 μL mixture (1:8) of Matrigel (Corning, NY, USA) was coated on the upper transwell chamber before seeding cells. After allowing 4 h for the Matrigel to solidify, the subsequent experimental steps were similar to those in the migration assay.

We used circMir (www.bio-inf.cn/circmir) to predict miRNA–circRNA interactions. CircMir is a homemade software that runs miRanda (August 2010 release) and RNAhybrid (version 2.1.2) to obtain miRNA–circRNA interactions. Miranda software was set using the following default parameters: gap extend penalty, −4.0; gap open penalty, −9.0; energy threshold, 1.0 kcal/mol; score threshold, 140.0; and scaling parameter, 4.0. The following parameters were used for RNAhybrid: perfect miRNA seed complementarity with circRNA sequence, miRNA:circRNA binding energy < −26 kcal/mol and p value <0.05. Only the circRNA–miRNA interactions predicted by both miRanda and RNAhybrid were kept for further study.

MiR mimics and negative control mimics (mimics-NC) were purchased from Guangzhou RiboBio Co., Ltd. (RiboBio, Guangzhou, China). The entire sequences of circZNF608 and circATRNL1 were cloned into the downstream region of the humanized form of Renilla luciferase (hRLuc) of the pmiR-RB-REPORT vector (RiboBio, Guangzhou, China), denoted circZNF608/vector-wd and circATRNL1/vector-wd.

To screen the miRNAs binding to the two circRNAs, miR mimics or negative control mimics (mimics-NC) were cotransfected with the vectors into cells. After 24 h, the luciferase activity was assessed using a dual-luciferase reporter gene assay kit (Beyotime Biotechnology, Shanghai, China). Because we found that only miR-152-5p inhibited the luciferase activity of circZNF608/vector-wd by more than 50%, we further mutated the miR-152-5p binding site from circZNF608/vector-wd to construct circZNF608/vector-mut. Then, mimics-NC or miR-152-5p mimics were cotransfected with circZNF608/vector-wd or circZNF608/vector-mut into cells, and luciferase activity was assessed.

R software (version 4.0.1) was used to perform all statistical analyses. RT–qPCR results were assessed using the Wilcoxon rank test, paired test for patients and unpaired test for cells. Other data were evaluated using Student’s t test and one-way analysis of variance. A p value <0.05 was considered statistically significant.

We profiled 3 paired ovarian cancer tissues and normal ovarian tissues using RNA-seq, and the raw read counts are presented in Supplementary Table S1. We identified 844 differentially expressed circRNAs in ovarian cancer tissues compared to normal ovarian tissues, including 45 upregulated circRNAs and 799 downregulated circRNAs (Supplementary Table S2, Figures 1A–C).

According to the fold change and normalized read counts, we selected 10 upregulated circRNAs and 10 downregulated circRNAs for RT–qPCR validation using cell lines. The 20 circRNAs can be found in Supplementary Table S2 and are labeled in bold font. Among the 20 circRNAs, 3 circRNAs [circZNF608 (hsa_circ_0001523), circATRNL1 (hsa_circ_0020093) and circATXN10 (hsa_circ_0001246)] were differentially expressed in cell lines. Two (circATRNL1 and circATXN10) of the three circRNAs showed the same trend as RNA-seq, and one (circZNF608) showed the reverse trend. We further validated these three circRNAs in 20 paired tissues. All three circRNAs were downregulated in ovarian cancer tissues compared with normal ovarian tissues (Figure 1D for circZNF608 and Figure 1E for circATRNL1). Because we failed to overexpress hsa_circ_0001246 using plasmids, we focused on the other two circRNAs in the following studies.

We analyzed circZNF608 and circATRNL1 using circPrimer 2.0. CircZNF608 is located at chr5:124701013-124701269- and is derived from exon 2 of the ZNF608 gene with a spliced length of 256 nt (Figure 1F). CircATRNL1 is derived from exons 2–5 of the ATRNL1 gene with a spliced length of 536 nt (Figure 1G). CircATRNL1 is a highly conserved circRNA located at chr10: 115120184–115129535+ in the human genome and chr19: 57703574–57717118+ in the mouse genome (mmu_circ_0007920) (Supplementary Figure S1). The existence of circZNF608 and circATRNL1 was confirmed by Sanger sequencing (Figures 1F,G). FISH showed that circZNF608 and circATRNL1 are mainly located in the cytoplasm of ovarian cancer cells (Figure 1H).

We explored the effects of circZNF608 and circATRNL1 on ovarian cancer cells. We transfected p-control, p-circATRNL1, and p-circZNF608 into ovarian cancer cell lines, detected the expression of circATRNL1 and circZNF608, and found that the cells transfected with p-circATRNL1 and p-circZNF608 had higher expression levels of circATRNL1 and circZNF608 than control cells (Figures 2A,B). The CCK8 proliferation assay showed that both circATRNL1 and circZNF608 could inhibit the proliferation of all three ovarian cancer cell lines (Figure 2C). The transwell migration assay indicated that higher expression of circATRNL1 and circZNF608 was associated with lower migration rates (Figures 2D,E). We further explored circATRNL1 and circZNF608 on invasion and found that the overexpression of circATRNL1 and circZNF608 inhibited the invasion of ovarian cancer cell lines (Figures 2F,G).

Studies have reported that the subcellular localization of a circRNA is closely correlated with its function (Panda et al., 2017; Qu et al., 2018). CircRNAs that are mainly located in the cytoplasm may act as miRNA sponges to bind miRNAs (Qu et al., 2018). Because circZNF608 and circATRNL1 are predominantly located in the cytoplasm of ovarian cancer cells, we predicted miRNA–circRNA interactions for circZNF608 and circATRNL1 and obtained 63 interactions (Figure 3A). We screened the miRNAs that bind to circZNF608 and circATRNL1 using a luciferase reporter assay and found that only miR-152-5p inhibited the luciferase activity of circZNF608/vector-wd by more than 50% (Figure 3B). We further mutated the miR-152-5p binding site from circZNF608/vector-wd and constructed circZNF608/vector-mut. CircZNF608/vector-wd or circZNF608/vector-mut was cotransfected with miR-152-5p mimics or mimics-NC, and a luciferase reporter assay showed that miR-152-5p could inhibit the luciferase activity of circZNF608/vector-wd significantly compared to circZNF608/vector-Mut, and mimics-NC had no effect on luciferase activity, suggesting that circZNF608 may exert its anticancer effects by sequestering miR-152-5p (Figure 3C). We further detected the expression of miR-152-5p in the 20 paired ovarian cancer tissues and normal ovarian tissues and failed to find a difference between the two groups (Supplementary Figure S2), suggesting that circZNF608 binds miR-152-5p but does not influence its expression.

To further uncover the underlying mechanisms of circATRNL1, we explored circATRNL1 using circPrimer 2.0, which shows that circATRNL1 may encode a 131 aa protein (Figure 4A). The open reading frame (ORF) of this protein spans the splice junction of circATRNL1. An internal ribosome entry site (IRES) overlaps with the ORF and may initiate ORF translation. To test the protein-coding ability of circATRNL1, we constructed a plasmid from p-circATRNL1, named p-circ3×F, which contained a 3×FLAG-coding sequence immediately upstream of the stop codon of the ORF (Figure 4B). If the ORF has protein-coding ability, a flagged peptide will be produced. After transfecting p-circ3×F and p-control into the HEK-293 cell line for 48 h, western blotting showed that p-circ3×F produced one flagged protein (Figure 4C).