CircNEIL3 regulatory loop promotes pancreatic ductal adenocarcinoma progression via miRNA sponging and A-to-I RNA-editing

Total RNA was extracted from three pairs of frozen (in liquid nitrogen) PDAC tissues and adjacent normal tissues using TRIzol Reagent (Invitrogen, CA, USA) and then treated using a RiboMinus Eukaryote kit (Qiagen, Valencia, CA) to remove ribosomal RNA before generating the RNA-seq library. Next, the RNA-seq library was deep sequenced with an Illumina HiSeq 2000 instrument (Illumina, San Diego, CA). The RNA-seq FASTQ reads were first aligned to the human reference genome (hg38/GRCh38) using TopHat2. The sequences that aligned contiguously and along their full length to the genome were discarded, and the remaining reads were used to identify circRNAs. We applied the spliced reads per billion mapping (SRPBM) approach to normalize the counts of reads mapped across an identified back-splice junction. The sequencing results of all differentially expressed circRNAs are shown in Tab. S1, and the raw data is available in the NCBI database with the access number PRJNA695439.

Pancreatic tissue specimens, including tumour and adjacent normal tissues, were obtained from 104 patients receiving a pancreaticoduodenectomy for PDAC at the First Affiliated Hospital of Nanjing Medical University from Jul. 2014 to Dec. 2018. The 104 patients were followed regularly until 11 September 2020. None of the patients had received chemotherapy or radiotherapy before surgery. All patients provided written informed consent that was approved by the Hospital Ethics Committee before specimen collection. The tissue specimens were collected during the surgery and immediately (within 3 min) cut into two sections. One section was frozen in liquid nitrogen, while the other was fixed in 4% formalin and embedded in paraffin 24 h later for confirmation of the pathological diagnosis. All of the cancer and adjacent tissues were diagnosed by two pathologists independently. The overall survival of patients was defined as the time between surgery and death. None of the 104 selected patients died within 1 month after surgery.

Four human PDAC cell lines (BxPC-3, MiaPaca-2, CFPAC-1 and PANC-1) and human pancreatic ductal epithelial (HPNE) cells were purchased from the Cell Bank of Type Culture Collection of the Chinese Academy of Sciences in Shanghai, China. The cells were cultured in Dulbecco’s modified Eagle’s medium (Life Technologies) supplemented with 10% foetal bovine serum (Wisent, Montreal, QC, Canada), 10 mM HEPES (Sigma, St Louis, MO), 2 mM L-glutamine (Sigma), 1 mM sodium pyruvate (Sigma), 100 U/ml of penicillin (Life Technologies) and 100 μg/ml streptomycin (Life Technologies) at 37 °C in a humidified atmosphere containing 95% air and 5% CO₂.

To construct circNEIL3-overexpressing plasmids, human circNEIL3 cDNA was synthesized and cloned into the vector pGL3-cir by Obio Technology Corp., Ltd. (Shanghai, China), with the empty plasmid used as a control. CircNEIL3 siRNA sequences were synthesized by GenePharma (Shanghai, China), and a scrambled siRNA was synthesized as a negative control. Transfection was performed using Lipofectamine 3000 (Invitrogen) according to the manufacturer’s instructions. Total RNA was collected 48 h after transfection.

The circNEIL3-knockdown lentivirus was constructed by Corues Biotechnology Corp., Ltd. (Nanjing, China). Si-circNEIL3–2 was subcloned into the vector pLV3ltr-Luc-Puro-U6-siRNA (circNEIL3) and was verified by sequencing. The supernatant of the cultured 293 T cells was collected to infect MiaPaca-2 and CFPAC-1 cells. Stable cell lines were selected by culturing in medium containing 5 μg/ml puromycin (Sigma). CircNEIL3 expression was confirmed by RT-qPCR.

The miR-432-5p mimics and inhibitor, ADAR1-overexpressing lentivirus and ADAR1 shRNAs were synthesized by GenePharma (Shanghai, China) and transfected as described above. All the sequences used in the present study are listed in Tab. S2.

Total RNA was extracted from cell and tissue samples using TRIzol Reagent (Life Technologies, Carlsbad, CA, USA) according to the manufacturer’s instructions. After spectrophotometric quantification, 1 μg of total RNA in a final volume of 20 μl was used for reverse transcription (RT) with an iScript cDNA Synthesis Kit (Bio-Rad, Hercules, CA, USA) following the manufacturer’s instructions. According to the manufacturer’s protocol, total cDNA was then used for RT-qPCR with the TaqMan Gene Expression Assay (Thermo Fisher Scientific, Rockford, IL, USA) in a StepOne Plus Real-time PCR System (Thermo Fisher Scientific). The expression of human 18S rRNA or β-actin genes was used as a control to calibrate the original concentration of tissue or cell mRNA, respectively. Target gene expression was calculated using the 2^-ΔΔCT method. Each quantitative PCR assay was performed in triplicate and independently repeated three times. The sequences of the primers used in the present study are listed in Tab. S2.

Total RNA (2 μg) was incubated with 3 U/μg RNase R (Epicentre Technologies, Madison, WI, USA) for 15 min at 37 °C. CFPAC-1 and MiaPaca-2 cells were transferred to six-well plates and treated with 5 μg/ml actinomycin D when the number of cells reached 900,000, with samples collected at the indicated time points. The expression of circNEIL3 and the linear counterpart mRNA NEIL3 was analysed by RT-qPCR.

Cy3-labelled circNEIL3 probes and fluorescein amidite (FAM)-labelled miR-432-5p probes were designed and synthesized by RiboBio. A fluorescence in situ hybridization (FISH) kit (RiboBio) was used to detect the probe signals in CFPAC-1 and MiaPaca-2 cells according to the manufacturer’s instructions. Nuclei were stained with 4,6-diamidino-2-phenylindole (DAPI). All images were acquired with an LSM880 NLO (2 + 1 with BIG) confocal microscope system (Carl Zeiss).

MiaPaca-2 and CFPAC-1 cells were harvested and lysed, and C-1 magnetic beads were incubated with the circNEIL3 probe (Life Technologies) at 25 °C for 2 h to generate probe-coated beads. The cell lysates were then incubated with circNEIL3 or an oligo probe at 4 °C. The RNA complexes bound to the beads were then eluted and extracted with an RNeasy Mini Kit (Qiagen) for RT-qPCR analysis. The biotinylated-circNEIL3 probe was designed and synthesized by RiboBio (Guangzhou, China).

Wild-type and mutant (mut-circNEIL3 or mut-ADAR1) cirCNEIL3 and ADAR1 fragments were constructed and inserted downstream of the luciferase reporter gene in the reporter plasmid pRL-SV40 (GenePharma, Shanghai, China). PDAC cells were seeded in 24-well plates and grown to 30% confluence 24 h before being transfected with the reporter plasmid using Lipofectamine 3000. Cells were also cotransfected with different combinations of plasmids harbouring the 3′-untranslated region (3′-UTR) of assayed genes (500 ng) and miRNA mimics or the negative control (NC; 10 nM final concentration). After 48 h, the activities of both firefly luciferase (LUC) and Renilla luciferase (RLUC) were measured with a Dual-Luciferase Reporter System Kit (E1910, Promega, USA).

RNA immunoprecipitation (RIP) assays were performed using a Magna RIP RNA-Binding Protein Immunoprecipitation Kit (Millipore, Billerica, MA, USA). Approximately 1 × 10⁷ PDAC cells were lysed in 1 ml of RIP lysis buffer supplemented with protease and RNase inhibitors. The cell lysates were then incubated with IgG, anti-AgO2 or anti-ADAR1 antibody-coated beads (Millipore) and rotated at 4 °C overnight. The immunoprecipitated RNAs were subsequently extracted with an RNeasy MinElute Cleanup Kit (Qiagen, Valencia, CA, USA) after treatment with proteinase K buffer and reverse transcribed to cDNA. Subsequently, the mRNA levels of the assayed genes were measured by RT-qPCR.

Proteins were extracted from stably transfected cells with RIPA buffer supplemented with PMSF. A DC Protein Assay Kit (Bio-Rad) was used to quantify the protein concentrations in the cell lysates. Then, the proteins were separated via electrophoresis using SDS-containing polyacrylamide gels and then transferred to polyvinylidene fluoride (PVDF) membranes (Millipore, Billerica, MA, USA). After blocking the membranes with 5% nonfat dry milk in 0.1% Tween (TBST) buffer at room temperature for 2 h, the membranes were incubated at 4 °C overnight with the appropriate primary antibody. Subsequently, the membranes were washed 3 times with TBST buffer for 15 min. Then, the membranes were incubated with a corresponding HRP-labelled secondary antibody for 2 h at room temperature, after which they were washed 3 times with TBST buffer. Finally, the western blot signals were visualized using an enhanced chemiluminescence detection system with Chemiluminescence HRP Substrate (Millipore, WBKL0100). All of the primary and secondary antibodies used in this study are listed in Tab. S3.

Four-week-old male nude mice (BALB/c) were purchased from the Animal Center of Nanjing Medical University (Nanjing, China). Ten mice per group were used to construct subcutaneous tumour formation models, and cell suspensions (0.1 ml) were prepared containing 1 × 10⁶ stable cells (CFPAC-1-circNEIL3 KD, CFPAC-1-circNEIL3-KDNC, MiaPaca-2-circNEIL3 KD, and MiaPaca-2-circNEIL3 KDNC). Then, the suspensions were subcutaneously injected into the armpits of the mouse limbs. When a tumour became macroscopic, it was measured with callipers every 5 days, and its bulk was calculated according to the following formula: volume = (width² × length)/2. The sample size was not predetermined for these experiments. All experiments were performed following the relevant institutional and national guidelines and regulations.

For metastasis studies, cells (1 × 10⁵) were injected into the tail veins of mice (six mice per group). Lung metastasis was monitored using a Xenogen IVIS Spectrum Imaging System (PerkinElmer, USA). After 8 weeks, the lungs and livers of mice were excised under anaesthesia, and the numbers of macroscopically visible lung and liver metastatic nodules were counted and validated using haematoxylin and eosin (HE)-stained sections by microscopy.

Immunohistochemistry (IHC) was performed according to the manufacturer’s protocol. After the xenografts and PDAC tissues were deparaffinized with xylene and rehydrated with ethanol, the samples were incubated with 3% H₂O₂ for 5 min to block endogenous peroxidase activity. Then, antigen retrieval was performed by incubating the samples with sodium citrate buffer (pH 6.0) for 20 min at 95 °C, after which the samples were blocked with 5% normal goat serum for 10 min at 20 °C. Subsequently, the sections were incubated with polyclonal antibodies against ADAR1 and Ki-67 at 4 °C overnight and then incubated with secondary antibodies. The validity and signal intensity of the tissue sections were assessed by two independent pathologists. The tissue sections were scanned, and the protein levels were calculated as positive cells/total cells by Halo v3.0.311.314. All of the primary and secondary antibodies used in this study are listed in Tab. S3.

Total RNA was extracted by TRIzol and subjected to cDNA synthesis using an iScript cDNA Synthesis Kit (Bio-Rad, Hercules, CA, USA). The cDNA was used as a template to amplify sequences containing RNA-seq-identified ADAR1-dependent putative editing sites. Individual clones were sequenced using Applied Biosystems BigDye terminator mix version 3.1. By contrast, A-to-I calls made from RNA-seq but not verified by Sanger sequencing as A-to-G are referred to as false positives. The heights of different base peaks from the Sanger sequence were measured by ImageJ software.

Quantitative data are presented as the means ± SDs. Differences between the means of two samples were analysed by Student’s t-test, while one-way ANOVA was used for multiple groups. Correlations between circNEIL3, miR-432-5p and ADAR1 expression and various clinicopathological or serological variables were analysed by the Mann-Whitney U test. Survival distributions and overall survival (OS) rates were determined using the Kaplan-Meier method, and the significance of differences between survival rates was calculated by the log-rank test. The above statistical analysis was performed using GraphPad Prism 8.0 (GraphPad Software, La Jolla, USA). While SPSS 20.0 (IBM, SPSS, Chicago, IL, USA) was applied in the univariate and multivariate Cox proportional hazards model, which was used to estimate the adjusted hazard ratios and 95% confidence intervals, as well as identify independent prognostic factors.

To identify circRNAs that are crucial to PDAC progression, RNA-seq was performed for 3 pairs of PDAC and matched normal pancreas tissues. As a result, 6980, 3486, 7247, 4319, 5832 and 4197 circRNAs were predicted in the three normal and three corresponding PDAC tissue samples, respectively. With the cut-off criteria of fold-change > 1 and P-value < 0.05, 203 differentially expressed circRNAs were identified, 79 of which were upregulated and 124 of which were downregulated (Fig. 1a). Among these circRNAs, the top 10 dysregulated circRNAs were selected for further analysis based on their fold change (Fig. 1b). To rectify the “Type I” RNA-seq error, a heat map was used to visualize the variation in the expression of these circRNAs in four PDAC cell lines compared to the normal pancreas cell line HPNE (Fig. 1c-d). CFPAC-1 and MiaPaca-2 PDAC cells were selected for further study due to their relative high expression. In general, circNEIL3 (chr4:178274462–178,281,831) was a significantly upregulated circRNA in both RNA-seq and PDAC cell lines. Furthermore, we detected higher circNEIL3 expression in 104 PDAC samples than in their paired the adjacent normal tissue samples via RT-qPCR, which was also consistent with the RNA-seq data (Fig. 1e-f). Based on the circBase annotation, circNEIL3 is derived from exons 8 and 9 of the NEIL3 gene and has a length of 596 bp [21]. Sanger sequencing was performed to validate its back-splicing using the RT-PCR product of circNEIL3 (Fig. 1g). To further confirm the circular form of circNEIL3, we designed divergent and convergent primers to amplify the circular and linear forms of NEIL3, respectively. The agarose gel electrophoresis analysis of the RT-qPCR products showed that circNEIL3 was only amplified from cDNA, ruling out the possibility of genomic rearrangements and trans-splicing (Fig. 1h). Moreover, after treatment with RNase R and actinomycin D, circNEIL3 was more stable than linear NEIL3 (Fig. 1i-j). In addition, the nuclear-cytoplasmic fractionation and FISH assay results revealed that circNEIL3 was primarily located in the cytoplasm (Fig. 1k-l). Taken together, these results identified circNEIL3 as an upregulated and highly stable circRNA localized in the cytoplasm of PDAC cell lines.

To assess the biological function of circNEIL3 in PDAC cells, three short interfering RNAs targeting the back-splice site of circNEIL3 were constructed to specifically downregulate circNEIL3 expression, and the one that provided the most significant downregulation was cloned into a lentivirus for further study (Fig. S1a). In addition, cell lines were also transfected with the circNEIL3 plasmid to overexpress circNEIL3 without affecting NEIL3 expression. The efficiency and specificity of circNEIL3 knockdown and overexpression in CFPAC-1 and MiaPaca-2 cells were verified by RT-qPCR (Fig. 2a). CCK-8, colony formation and EdU assays were carried out to evaluate cell proliferation, and the results showed that circNEIL3 silencing notably inhibited the proliferation of CFPAC-1 and MiaPaca-2 cells, while circNEIL3 upregulation increased cell growth (Fig. 2b-f and Fig. S1b-e). In transwell and wound healing assays, we found that circNEIL3 knockdown reduced the migration and invasion of PDAC cells, while circNEIL3 overexpression had the opposite effect. Taken together, these data suggest that circNEIL3 exerts an oncogenic role in PDAC cells (Fig. 2g-j and Fig. S1f-i).

To evaluate the contribution of circNEIL3 to PDAC tumours in vivo, CFPAC-1 and MiaPaca-2 cells stably transfected with LV-si-circNEIL3 and respective negative control were subcutaneously injected into male nude mice. The results showed that circNEIL3 knockdown inhibited tumour growth in both cell lines (Fig. 3a). Lower tumour weights and volumes were observed in the circNEIL3 KD group compared to the NC group (Fig. 3b-c). In addition, IHC staining revealed decreased Ki-67 levels in tumour cells after circNEIL3 knockdown (Fig. 3d). Furthermore, to determine whether circNEIL3 can promote PDAC metastasis in vivo, luciferase-labelled CFPAC-1 and MiaPaca-2 constructs were injected into the tail veins of male nude mice. The fluorescence intensity and proportion in the lung and abdomen were significantly lower in the circNEIL3 KD group than in the control group (Fig. 3e). Moreover, circNEIL3 knockdown notably decreased the number of metastatic nodules in the lung and liver (Fig. 3f-g). These results are consistent with the in vitro findings, suggesting that circNEIL3 can promote PDAC tumorigenesis and metastasis of in vivo.

CircRNAs have been reported to primarily function as miRNA sponges to regulate the expression of downstream genes [6]. Therefore, taking the sequence of circNEIL3 as a bait, we first performed a cross-analysis using three miRNA target prediction databases (miRanda, TargetScan and RNAhybrid) and identified 95 candidate miRNAs that may bind to circNEIL3 (Fig. 4a and Tab. S4) [22‐24]. Based on the conjugation scores, we selected the top 11 miRNAs (miR-495-3p, miR-342-5p, miR-296-3p, miR-1301-3p, miR-765, miR-17-3p, miR-509-5p, miR-877-5p, miR-432-5p, miR-496, and miR-378a-3p) for further study. Next, we designed a biotin-labeled circNEIL3 probe to indirectly capture the potential binding miRNA in RNA pull-down assays. To test the specificity of the probe, RT-qPCR was used to assess the levels of circNEIL3 remaining after performing pull-down experiments compared to the random sequence NC probe. The results showed that circNEIL3 was mostly captured by the circNEIL3 probe in CFPAC-1 and MiaPaca-2 cells (Fig. 4b). Furthermore, miR-432-5p was the most highly enriched miRNA in the sponge complexes for both PDAC cell lines (Fig. 4c). Accordingly, biotin labelled miR-432-5p could directly pull down circNEIL3, ruling out the possibility of false positives due to indirect pull-down assays (Fig. 4d). To identify the specific binding region between circNEIL3 and miR-432-5p, a dual-luciferase reporter assay was performed. Based on the complementary base pairing prediction with the “seed” region of miR-432-5p, we mutated the predicted binding site of circNEIL3 and inserted the mutant sequence downstream of the luciferase reporter gene (Fig. 4e). After cotransfecting CFPAC-1 and MiaPaca-2 cells with miR-432-5p mimics and a circNEIL3-WT reporter gene, the observed luciferase activity was significantly reduced. However, miR-432-5p mimics or mimics NC showed no significant difference in luciferase activity when cotransfected with mutant reporter, indicating that circNEIL3 could directly and specifically bind to miR-432-5p at the “seed” region (Fig. 4f). In general, miRNAs function as RISC components, binding to Argonaute-2 (AGO2) [25]. Accordingly, an anti-AGO2 RNA immunoprecipitation (RIP) assay was performed, and the results showed that both circNEIL3 and miR-432-5p were pulled down by the anti-AGO2 antibody but not by IgG (Fig. 4g). FISH analysis showed that circNEIL3 and miR-432-5p colocalized in the cytoplasms of CFPAC-1 and MiaPaca-2 cells (Fig. 4h). Additionally, in contrast to circNEIL3, miR-432-5p was notably downregulated in both PDAC cell lines and tissues (Fig. 4i-k). Moreover, the FISH results showed that circNEIL3 was primarily and highly expressed in the adenocarcinoma portion of PDAC tissues, while miR-432-5p was mostly located in acinous cells in the adjacent normal tissue (Fig. 4l). Moreover, circNEIL3 in CFPAC-1 and MiaPaca-2 cells negatively regulated the expression of miR-432-5p after knocking down or overexpressing circNEIL3 (Fig. 4m). These data demonstrate that circNEIL3 acts as a sponge of miR-432-5p and suppresses its expression.

To elucidate whether circNEIL3 functions by sponging miR-432-5p, rescue experiments were performed with cotransfection of miR-432-5p mimics or their inhibitor with LV-si-circNEIL3 or circNEIL3 plasmid. The efficiency of miR-432-5p mimics and inhibitor on its expression level in CFPAC-1 and MiaPaca-2 cells was verified by RT-qPCR (Fig. 5a-b). The results indicated that the miR-432-5p inhibitor significantly promoted the proliferation, migration and invasion of CFPAC-1 cells and reversed the suppressive effects on these processes induced by circNEIL3 downregulation in CCK-8, colony formation, EdU, Transwell and wound healing assays (Fig. 5c-k). Similar results were observed after transfection with miR-432-5p mimics and the circNEIL3 plasmid in MiaPaca-2 cells (Fig. S2a-i). Collectively, these data demonstrate that miR-432-5p exerts an anti-oncogenic effect on PDAC cells and serves a crucial function downstream of circNEIL3.

To further elucidate the underlying mechanism associated with the miR-432-5p-mediated regulation of PDAC, we performed another bioinformatic analysis using the online programs TargetScan, miRTarBase, miRDB and miRWalk [24, 26‐28]. The data revealed 13 possible genes targeted by miR-432-5p (Fig. 6a). Since miRNAs typically interact with the 3′-UTR of target genes and downregulate their expression, we screened the changes in the expression of these genes in CFPAC-1 and MiaPaca-2 using miR-432-5p mimics or inhibitor with RT-qPCR (Fig. 6b). The RT-qPCR and western blot results identified ADAR1 as the gene most significantly and negatively regulated by miR-432-5p in both cell lines (Fig. 6c-d). Moreover, ADAR1 expression at both the mRNA and protein levels was positively correlated with the abundance of circNEIL3, which could be fully rescued by miR-432-5p mimics or inhibitor treatment (Fig. 6e-h). The IHC staining results for ADAR1 in the previously described subcutaneous tumour in the in-vivo model further confirmed the correlation between circNEIL3 and ADAR1 (Fig. 6i). Next, we constructed wild-type and mutant dual-luciferase reporter plasmids harbouring the wild-type or mutant ADAR1 3′-UTR (Fig. 6j), and the results showed that the transfection of cells with miR-432-5p mimics notably downregulated the luciferase activity in the plasmid with WT but not MUT 3′-UTR of ADAR1(Fig. 6k-l). Mechanistically, ADAR1 shared the same sequence with circNEIL3 that binds to the “seed” region of miR-432-5p, and similar to circNEIL3, ADAR1 was highly expressed in the PDAC cell lines, especially in CFPAC-1 and MiaPaca-2 cells (Fig. 6m-n). Clinically, the RT-qPCR and IHC results indicated that ADAR1 mRNA and protein expression was notably increased in PDAC tissues (Fig. 6o-q). Taken together, these results indicate that circNEIL3 antagonizes miR-432-5p-induced ADAR1 degradation and may function as an oncogene in PDAC.

As ADAR1 is a structural target that functions downstream of miR-432-5p and circNEIL3, we further evaluated whether circNEIL3 and ADAR1 are functionally associated. To regulate the ADAR1 RNA and protein levels, three short hairpin RNAs and an overexpression lentivirus were constructed and transfected into CFPAC-1 and MiaPaca-2 cells (Fig. 7a-c). Subsequently, with CCK-8, colony formation, EdU, Transwell and wound healing assays, ADAR1 overexpression was demonstrated to promote the proliferation, migration and invasion of CFPAC-1 cells (Fig. 7d-l). In contrast, ADAR1 knockdown had the opposite effects in MiaPaca-2 cells (Fig. S3a-i). Further, we observed that the inhibition of malignant biological behaviour after disrupting circNEIL3 expression could be largely blocked by ADAR1 overexpression (Fig. 7d-l), while the increased proliferation, migration and invasion capabilities of PADC cells caused by circNEIL3 overexpression could be retarded by ADAR1 knockdown (Fig. S3a-i). Collectively, our data suggest that the promoting role of circNEIL3 in maintaining the malignant phenotype of PDAC was largely dependent on the miR-432-5p/ADAR1 axis.

As ADAR1 could stimulate PDAC progression, and recent reports on whole genome screening for instances of RNA editing have shown that many mRNAs, especially GLI1, are edited in multiple malignancies [18], we further examined whether ADAR1 induces RNA editing of GLI1 in PDAC cells. To this end, Sanger sequencing was used to detect the A-to-I editing of GLI1 in CFPAC-1 and MiaPaca-2 cells, which was subsequently translated to substituted amino acids (A-to-G). Consistent with other studies, we observed that GLI1 mRNA is edited by the A-to-I conversion in exon 12, specifically at nucleotide position chr12:57864624 [29]. GLI1 transcripts were exhibited significant hyperediting after ADAR1 overexpression, while ADAR1 knockdown had the opposite effect. Moreover, increased circNEIL3 levels were positively associated with the rate of GLI1 mRNA editing and could be abrogated by cotransfection with an ADAR1 expression construct (Fig. 8a and Fig. S4a). Subsequently, we performed RNA immunoprecipitation (RIP) experiments using an anti-ADAR1 antibody with CFPAC-1 and MiaPaca-2 cells, and the results showed that endogenous ADAR1 directly bound to GLI1 mRNA (Fig. 8b). GLI1 transcript editing was previously reported to lead to an R701G (arginine to glycine) amino acid change [29]. Our western blot results showed that ADAR1 did not alter the GLI1 protein levels, indicating that RNA editing of the GLI1 transcript may alter the secondary structure of the GLI1 protein (Fig. 8c-d and Fig. S4b-c). Indeed, the target site for RNA editing is located within a domain in the C-terminus of GLI1 that binds to the negative regulator of HH signalling, suppressor of fused (SUFU) [29]. Furthermore, the edited GLI1 protein was previously reported to have a lower susceptibility to inhibition by SUFU, resulting in it having an increased ability to activate most transcriptional targets, including cyclin D1, cyclin E1, and Snail [30, 31]. Thus, a cell cycle analysis was performed, with the results showing that knockdown of ADAR1 or circNEIL3 led to higher percentages of CFPAC-1 and MiaPaca-2 cells in G0-G1 phase, while their overexpression had the opposite effect (Fig. 8e-f and Fig. S4d-e). Moreover, western blot analysis results demonstrated that ADAR1 upregulation enhanced the protein levels of the cyclin D1/CDK4/CDK6 complex, and the downstream targets cyclin E1 and CDK2, while ADAR1 knockdown decreased the levels of these proteins. These results could also be obtained by overexpressing or silencing circNEIL3 and could be reversed by ADAR1 knockdown or overexpression, respectively (Fig. 8g and Fig. S4f). In agreement with the observed ability of circNEIL3 to promote cellular proliferation in vitro and in vivo, these results indicated that circNEIL3 downregulation prevented PDAC cells from shifting from G1 to S phase by functioning as a ceRNA for miR-432-5p to regulate the ADAR1/GLI1/Cyclin D1/CDKs axis. Furthermore, as crucial downstream targets of GLI1, EMT-related proteins were detected by western blot analysis. The results showed that PDAC with depletion of ADAR1 and circNEIL3 had higher levels of N-cadherin and lower expression of E-cadherin, β-Catenin, vimentin and Snail than the control group (Fig. 8h and Fig. S4g). Similar results were observed for the overexpression and rescue groups, suggesting that circNEIL3 induces the transition of PDAC cells from an epithelial phenotype to mesenchymal phenotype by activating ADAR1/GLI1/Snail signalling.

Alu elements, which are often present in the flanking introns of circularized exons, are required for circRNA formation [6]. ADAR1 has been reported to recognize inverted Alu repeats and perform A-to-I editing to unwind the dsRNA structure and prevent splice sites from being close enough to back splice [32]. In NEIL3, we identified two Alu elements besides exons 8–9 in the introns 7 and 9, named AluY (position: chr4:178273703–178,273,999) and FLAM_C (position: chr4:178282936–178,283,074). The sequence of these two Alu elements are listed in Tab. S5. Therefore, we hypothesized that miR-432-5p may transcriptionally regulate circNEIL3 expression through ADAR1 in an Alu-dependent manner. To test our hypothesis, RT-qPCR and RIP assays were performed, and the results indicated that ADAR1-depleted cells showed greatly increased circNEIL3 levels, while linear NEIL3 were only slightly affected (Fig. 9a). Furthermore, circNEIL3 production was severely blocked after ADAR1 overexpression (Fig. 9b). As expected, we also observed that miR-432-5p overexpression and silencing in CFPAC-1 or MiaPaca-2 cells decreased or increased the circNEIL3 levels, respectively (Fig. 9c). Moreover, ADAR1 specifically bound to linear NEIL3 rather than to circNEIL3, indicating that ADAR1 participates in the formation of circNEIL3 (Fig. 9d-e).

Given the discovery of the circNEIL3/miR-432-5p/ADAR1 loop in PDAC cells and that the positive relationship between miR-432-5p and circNEIL3 conflicted with the negative correlation demonstrated above, we next assessed the clinical significance of circNEIL3, miR-432-5p and ADAR1 in a cohort of 104 PDAC patients to crystallize the principal line of the story. Patients were divided into two groups based on their median expression. Importantly, a Kaplan-Meier analysis revealed that higher circNEIL3 expression, lower miR-432-5p expression and higher ADAR1 expression were associated with the poorer overall survival (OS) in104 PDAC patients (Fig. 9f-h). The relationship between circNEIL3, miR-432-5p and ADAR1 expression and the clinical characteristics of the PDAC patients are listed in Table 1. Tumour size, microscopic vascular and nerve invasion and postoperative liver metastasis were statistically significantly associated with at least one of these three biomarkers. Further univariate and multivariate Cox regression analysis showed that the circNEIL3 and ADAR1 expression levels were independent prognostic factors for PDAC patients, as were tumour size and TNM stage (Table 2). Furthermore, Pearson correlation analysis showed that circNEIL3 and ADAR1 expression were negatively correlated with miR-432-5p expression levels, while circNEIL3 was positively correlated with ADAR1 (Fig. 9i-k). Taken together, these results indicate the existence of a circNEIL3/miR-432-5p/ADAR1 axis in PDAC with clinical significance, while miR-432-5p positively regulates circNEIL3 in an Alu-dependent manner via ADAR1 as a part of hemostasis of endogenous circNEIL3.