Synthetic circular RNA switches and circuits that control protein expression in mammalian cells

Abstract Synthetic messenger RNA (mRNA) has been focused on as an emerging application for mRNA-based therapies and vaccinations. Recently, synthetic circular RNAs (circRNAs) have shown promise as a new class of synthetic mRNA that enables superior stability and persistent gene expression in cells. However, translational control of circRNA remained challenging. Here, we develop ‘circRNA switches’ capable of controlling protein expression from circRNA by sensing intracellular RNA or proteins. We designed microRNA (miRNA) and protein-responsive circRNA switches by inserting miRNA-binding or protein-binding sequences into untranslated regions (UTRs), or Coxsackievirus B3 Internal Ribosome Entry Site (CVB3 IRES), respectively. Engineered circRNAs efficiently expressed reporter proteins without inducing severe cell cytotoxicity and immunogenicity, and responded to target miRNAs or proteins, controlling translation levels from circRNA in a cell type-specific manner. Moreover, we constructed circRNA-based gene circuits that selectively activated translation by detecting endogenous miRNA, by connecting miRNA and protein-responsive circRNAs. The designed circRNA circuits performed better than the linear mRNA-based circuits in terms of persistent expression levels. Synthetic circRNA devices provide new insights into RNA engineering and have a potential for RNA synthetic biology and therapies.


INTRODUCTION
Gene delivery using synthetic messenger RNA (mRNA) is an effective method for transient gene expression, with a reduced risk of genomic integration in the cell (1). One of the limitations hindering its broader application for medical research is lower gene expression persistence caused by its instability. To improve the stability and performance of synthetic mRNA, substantial efforts have been made to engineer new RNA structures and, in recent years, synthetic circular RNAs (circRNAs) have shown promise as a new class of synthetic mRNA with superior stability and persistent gene expression (2)(3)(4).
Endogenous circRNAs are generated through backsplicing (5). When first discovered, they were thought to be products of splicing errors (6). However, they have since been reported to be involved in various biological functions as either protein-coding or non-coding RNA (7). Additionally, these circRNAs resist exonuclease-mediated degradation as their covalent-closed loop structure lacks free 5 and 3 ends. This common structural feature allows circRNAs to exhibit a longer half-life than linear mRNAs in cells (7,8).
To expand the potential of mRNA therapeutics, it is important to produce desired outputs depending on the cell state and reduce off-target effects in non-target cells and tissues (9,10). Such 'smart mRNA' could be a useful tool for cell type-specific gene regulation and future therapeutic applications (11,12). Even before the function of endogenous circRNAs was determined, there had been efforts to adapt the circular structure to improve the stability of synthetic RNAs (13,14). However, research in the field of cir-cRNA is relatively early, and studies have mainly focused on the methods of cyclization (3), the immunogenicity (15)(16)(17)(18), and the performance of translation (3,19). Design principles of circRNAs that sense intracellular conditions and autonomously control their translation level remain unknown. Defining these principles would be useful in the development of therapies with suppressed off-target expression in non-target cells, maximizing the therapeutic effect.
MicroRNAs (miRNAs) are small non-coding RNAs that regulate protein expression from mRNAs via translational repression and mRNA degradation (20). As the activity of miRNAs differs among cell types, they can be used as an indicator to distinguish between various cell types (21). We have previously developed miRNA-responsive linear mRNA (miRNA-responsive switch), which is composed of a protein-coding sequence and an antisense sequence to the target of the miRNA (anti-miR) at the untranslated region (UTR), enabling the distinguishment of cell types based on endogenous miRNA activity. We have succeeded in regulating gene expression using miRNA-responsive switches in a cell type-specific manner for various applications, including cell purification derived from human pluripotent stem cells, selective killing of cancer cells, and genome editing of target cells (22)(23)(24)(25). Notably, these miRNA-responsive switches did not change the expression level of endogenous target miRNA and global gene expression patterns regulated by miRNAs, indicating that perfect base-pairing between miRNA and its antisense target induces mRNA cleavage through siRNA-like pathways with the RNA-induced silencing complex (RISC) (i.e. miRNA-responsive switches do not function as a miRNA 'sponge') (22,25). Studies on endogenous circRNAs suggested that circRNAs with partially matched miRNA target sites have been reported as potential miRNA sponges or inhibitors (26)(27)(28), whereas circRNAs with a near-perfect miRNA target site undergo degradation by sensing miRNA, much like linear mRNAs (29). Therefore, to perform cell-type-specific protein expression from circRNA, we expected that miRNA-responsive switch technology could be applied.
RNA-binding protein (RBP)-responsive mRNA switches also allow the control of protein expression from linear mRNAs in a cell-type-specific manner. In previous studies, protein-responsive switches have been designed to repress cap-dependent translation of linear mRNAs by inserting the protein-binding aptamer sequence at the 5 -UTR (23,30,31). The virus-derived Internal Ribosome Entry Site (IRES) can be used for designing translatable circRNA (2,3), and its underlying translation mechanism and the composition of its initiation factors are similar to those of cap-dependent translation (32). We expected that IRESdependent translation could be controlled by RBPs, similar to the protein-responsive switches reported previously (23,30,31). So far, endogenous circRNAs have been reported to function as sponges or scaffolds for RBPs (33)(34)(35)(36), and synthetic circRNA devices have mimicked these functions, only as RBP regulators or inhibitors (18,37,38). In translatable circRNAs, no devices have been reported in which RBPs regulate the translation of circRNA. Thus, we aimed to investigate the design principles of proteinresponsive switches for cap-independent translation of cir-cRNA.
Here we present a design strategy for developing synthetic circRNA devices that control circRNA translation in a cell type-specific manner by detecting target miRNAs and RBPs (Figure 1). We designed miRNA-and RBPresponsive circRNA switches by engineering the UTRs and Coxsackievirus B3 (CVB3) IRES and constructed synthetic circRNA circuits for RNA-only delivery. The designed cir-cRNA circuits showed better expression level persistence than linear mRNA-based circuits.  Figure 1. Schematic illustrations of circRNA switches and circuits. (A) Design of miRNA or protein-responsive circRNA switch. miRNA-responsive circRNA has the antisense of target miRNA sequence at the UTR. Protein-responsive circRNA has a protein-binding motif in IRES region. In both systems, gene expression from circRNA is repressed if the target miRNA or protein is present. (B) Scheme of circRNA circuit composed of miRNA-and protein-responsive circRNA switches. The first output (MS2CP or U1A protein) is encoded on a miRNA-responsive circRNA switch, and the second output (reporter protein) is encoded on a protein-responsive circRNA switch. In the OFF state (absence of input miRNAs), MS2CP or U1A protein represses translation of the second output gene-coding circRNA. In ON state (presence of input miRNAs), the MS2CP or U1A translation is repressed by the miRNAs, which leads to output translation. sodium acetate pH 5.2, 0.1% SDS). The eluted RNAs were purified by phenol-chloroform extraction and precipitated with isopropanol. After dissolving the RNA pellet in nuclease-free water, RNAs were desalted using Amicon Ultra 0.5 ml Centrifugal Filters Urtracel-50K (Millipore), then incubated with Antarctic Phosphatase (New England Biolabs) at 37 • C for 30 min. Phosphatase-treated RNAs were re-purified by phenol-chloroform extraction and isopropanol-precipitation. In this study, all circRNAs and linear mRNAs were purified from polyacrylamide gel, except for linear mRNAs used in Figure 5C to screen the appropriate variant. Concentrations of purified RNAs were measured by NanoDrop2000 (Thermo Fisher Scientific) and used in cellular experiments. All RNA sequences used in this research are described in Supplementary Sequences.

Cell culture and RNA transfection
HEK293FT (Invitrogen), HeLa CCL2 (ATCC) and A549 (RCB3677) cells were cultured in Dulbecco's modified Eagle's medium (DMEM) 4.5 g/l glucose (Nacalai Tesque) supplemented with 10% fetal bovine serum (FBS) (Biocera, Ireland Origin), 0.1 mM MEM non-essential amino acids (Life Technologies), 2 mM L-glutamine (Life Technologies) and 1 mM sodium pyruvate (Nacalai Tesque). All cell lines were cultured at 37 • C with 5% CO 2 . All transfections were performed using Lipofectamine MessengerMAX (Thermo Fisher Scientific) according to the manufacturer's protocol. RNAs were co-transfected with synthetic miRNA mimics or inhibitors (Thermo Fisher Scientific) in the case of miRNA-responsive switch experiments. In the case of the RBP-responsive switch experiments, the target RBP was ex-pressed in cells by co-transfecting with RBP-coding mR-NAs. The transfection condition details of each experiment are shown in Supplementary Table S1.

RNase R digestion assay
2.5 g of in vitro-transcribed RNAs (Circular EGFP pA IRES) were incubated at 37 • C with 10 U of RNase R (Cosmo Bio) in a 10 l mixture. After a 45 min incubation, the mixture was subjected to 4 and 8% denaturing PAGE (8.3 M urea), and then stained by SYBR Green II Nucleic Acid Gel Stain (TaKaRa). Stained RNA was detected by Typhoon FLA-7000 (GE Healthcare).

Splice junction sequencing
Circular EGFP pA IRES purified from denaturing polyacrylamide gel was reverse transcribed using reverse transcription primer (5 -CCTACTCAGGCTTTATTCA AAGACCAAG-3 ) and SuperScript IV Reverse Transcriptase (Thermo Fisher Scientific). Reverse-transcribed cDNA was used as a template for PCR using PrimeSTAR Max DNA Polymerase (TaKaRa) with primer set for splice junction amplification (Fwd: 5 -agctcgccgaccactaccagcag -3 , Rev: 5 -gtagcggctgaagcactgcacg-3 ). Amplified product was purified by Monarch DNA Gel Extraction Kit (New England Biolabs), and then sequenced by the same method described in 'Template plasmid construction for mRNA'.

Flow cytometry and data analysis
HEK293FT, A549 (1.0 × 10 5 cells), and HeLa (0.5 × 10 5 cells) cells were seeded onto 24-well plates 24 h before transfection. All flow cytometry measurements were performed 24 h after the transfection using BD Accuri C6 (BD Biosciences). Cells were washed with phosphate buffered saline (PBS, Nacalai Tesque), trypsinized with 100 l of 0.25% Trypsin-EDTA (Thermo Fisher Scientific), and incubated at 37 • C for 5 min. After incubation, 150 l of fresh medium was added. Cells were transferred to a fresh microcentrifuge tube passing through a nylon mesh. EGFP was detected by FL1 (533/30 nm, 99% attenuated), and iRFP670 was detected by FL4 (675/25 nm) filters, respectively. Collected data were analyzed using FlowJo 10.5.3 software. For data analysis, gates were generated by using mock samples. Data from the debris were removed when preparing forward versus side dot plots (FSC-A versus SSC-A). Then, events on the chart edges in the dot plots of FL-1 versus FL-4 were removed. In the histogram where iRFP670-intensity is displayed on the X-axis, the iRFP670-positive (referencepositive) gate was defined by a mock sample with 99.9% cells outside the gate. In the following analysis, the mean of EGFP+/iRFP670+ was used for calculation.

RT-qPCR analysis
For the RT-qPCR targeting immune response-related genes, A549 (1.0 × 10 5 cells) cells were seeded onto 24-well plates 24 h before transfection. Then, cells were washed with 1 ml of PBS and total RNA was extracted 24 h after transfection. To induce immune response-related genes, 200 ng of

PAGE 5 OF 15
Nucleic Acids Research, 2023, Vol. 51, No. 4 e24 Polyinosinic-polycytidylic acid [Poly(I:C)] (Enzo Life Sciences, Inc.) was transfected as a positive control. Total RNA extraction was performed using TRIzol Reagent (Thermo Fisher Scientific) and Monarch RNA Cleanup Kit (New England Biolabs) according to the manufacturer's protocol. 400 ng total RNA was used as a template for reverse transcription performed using ReverTra Ace qPCR RT Master Mix with gDNA Remover (TOYOBO) in a 10 l reaction mixture. All optional steps described in the kit manual were applied.
For the RT-qPCR targeting synthetic linear mRNAs and circRNAs (Supplementary Figures S2B and S4B), HEK293FT (1.0 × 10 5 cells) cells were seeded onto 24-well plates 24 h before transfection. Then, total RNA was extracted at each time point after transfection. Extracted total RNA was treated with TURBO DNase (Thermo Fisher Scientific) and re-purified by Monarch RNA Cleanup Kit (New England Biolabs) according to the manufacturer's protocol. Reverse transcription was performed using a High-Capacity cDNA Reverse Transcription Kit (Thermo Fisher Scientific) with 100 ng total RNA template and 20 l reaction volume. In both cases, the synthesized cDNA solution was diluted by nuclease-free water with 5-fold dilution. One l of the diluted-cDNA solution was analyzed by qPCR. The qPCR analysis was performed using THUN-DERBIRD Next SYBR qPCR Mix (TOYOBO) with 20 l reaction mixture and QuantStudio 3 Real-time PCR Systems (Thermo Fisher Scientific) following the manufacturer's protocol, with three steps reaction. Target mRNA quantities were normalized by ATP5B mRNA. All qPCRs were performed in technical duplicates and the averages of Ct were processed to calculate relative expression levels using the Ct or Ct method. The primers for qPCR are listed in Supplementary Table S2.

Western blot analysis
HEK293FT (1.0 × 10 5 cells) cells were seeded onto 24-well plates 24 h before transfection. 24 h after transfection, cells were washed with 1 ml of PBS and lysed in 50 l of RIPA buffer (Nacalai Tesque). The cell lysates were analyzed with subsequent western blotting as previously performed (30). The transferred membranes were incubated with specific primary antibodies, Anti-SNRPA (Santa Cruz Biotechnology, 200-fold dilution) and Anti-Enterobacterio Phage MS2 Coat Protein (Sigma-Aldrich, 5000-fold dilution), respectively. Anti-GAPDH antibody (Santa Cruz Biotechnology) was used at 500-fold dilution. Then, the blot was incubated with secondary antibodies. Goat Anti-Mouse IgG (H + L)-HRP conjugate (Bio-Rad) or Goat anti-Rabbit IgG (H + L)-HRP conjugate (BIO-RAD) was used at 400-fold dilution. All incubation steps were performed using iBind Flex Western Device (Thermo Fisher Scientific). Detection of the blot was performed with ECL Prime Western Blotting Detection Reagent (GE Healthcare) and Amersham Image-Quant 800 (Cytiva). The protein expression level was calculated from band intensities with ImageJ (NIH).

Secreted luciferase assay
HEK293FT, A549 (1.5 × 10 4 cells) and HeLa (0.5 × 10 4 cells) cells were seeded onto 24-well plates 24 h before transfection. After transfection, culture media was harvested and replaced every 24 h up to 5 days (120 h) at each time point. Culture media was harvested and replaced for 24-120 h after transfection at each time point. Media was harvested into Protein LoBind tubes (Eppendorf) and stored at −30 • C. To detect bioluminescence from Metridia Luciferase (MetLuc2), 50 l of harvested media was transferred into a Greiner LUMITRA 200 microplate (Greiner), and then 10 l of 0.5× substrate/reaction buffer from Ready-To-Glow Secreted Luciferase Reporter Assay (TaKaRa) was added by injector attached on a plate reader. After 30s double orbit shaking with 3.0 mm diameter and 30s incubation delay, luminescence was detected by Centro LB 960 (Berthold Technologies) with an integration time of 1 s. Normalized MetLuc activity was calculated by normalizing the MetLuc activity 24 h after transfection.

Statistical analysis
Statistical values including the exact n and statistical test are reported in the figure legends. The levels of significance are denoted as *P < 0.05, **P < 0.01 and ***P < 0.001. N.S. means non-significant (0.05 < P). All statistical tests were performed by Dunnett's test, and two-tailed unpaired Student's or Welch's t-test using R or Excel (Microsoft). The type of t-test was determined by F-test.

Design and evaluation of synthetic circRNAs
Several strategies for in vitro mRNA cyclization have been reported (39). We chose an engineered Permuted Intron-Exon (PIE) Splicing system, which is an efficient circularization method with no enzymatic treatment for long cir-cRNA construction (3). We designed a circRNA construct that had permuted split fragments of group I catalytic introns corresponding to Anabaena pre-tRNA with homology arms at both ends, and a CVB3 IRES upstream of the open reading frame (ORF) (Figure 2A). The circRNA construct with splice sites resulted in a migrated band expected as a circular RNA in denaturing PAGE ( Figure 2B), which was confirmed to be circRNA through splice junction identification with ligated 5 and 3 splice sites ( Figure 2C). Resistance to RNase R was confirmed by observing a band of the migrated product of circRNA even after RNase R treatment whereas other linear products were degraded by the treatment (Figure 2D). RNA purification from polyacrylamide gels removed most of the impurities as reported previously (3,16) (Supplementary Figure S1). These data indicate that our construct generated circRNA products as expected. Previous studies reported that such features are unique in PIE Splicing-derived circRNA (3,18). Next, we examined the translatability of synthesized cir-cRNAs and the effect of circularization. We compared the levels of protein production from circRNAs with those of linear mRNAs (linRNAs). Overall protein expression levels from circRNAs and linRNAs were evaluated using re-porter expression (EGFP) normalized by transfection control (iRFP670). We prepared different types of RNA, including four types of circRNAs with or without 120 nucleotides of polyA (pA) sequences (circRNA + pA, cir-cRNA pA), without IRES (circRNA + pA IRES, cir-cRNA pA IRES), and three types of linRNAs (Linear EGFP, Cap-EGFP) ( Figure 3A).  First, we compared EGFP expression of circRNA and linRNA with the same sequence components (UTR, CVB3 IRES, ribozyme-derived sequence, EGFP-coding region, pA). circRNAs with IRES showed higher expression levels than linRNAs in the three tested human cell lines (HEK293FT, HeLa, A549) whereas circRNAs IRES did not activate translation ( Figure 3B). Notably, cir-cRNA + pA showed higher EGFP expression than IRESdependent linRNAs or circRNA pA ( Figure 3B, Supplementary Figure S2A), confirming that the pA sequence enhances protein expression from circRNA. It has been reported that pA or pAC spacer sequences inserted to the vicinity of 5 and 3 , or only that of 5 intron fragments in circRNA promote translation by interaction with eukaryotic translation initiation factor (eIF4G) and polyAbinding protein (PABP), similar to the translation mechanism of linRNA (3,40,41). We assume that a similar effect was also observed when the 120-mer polyA sequence used in the tail of our linRNA was inserted internally into the circRNA (between the 3 UTR and the 3 intron fragment). To investigate the stabilization of RNA by cyclization, we quantified the amount of residual circRNA and lin-RNA by RT-qPCR. The relative levels of circRNA (Circular EGFP + pA) are higher than linRNA (Linear EGFP) at 8 and 24 h after transfection (Supplementary Figure S2B), indicating that enhanced stability by cyclization is likely to contribute to higher reporter expression.
We next compared the performance of circRNAs with conventional cap-dependent linRNAs (Cap-EGFP) with modified bases (m5C/, m1; modRNAs). These base modifications have been applied to synthetic mRNAs to reduce immunogenicity and improve protein expression (42)(43)(44). Although the protein expression level from cir-cRNA + pA was lower than m1 or native linRNA, it showed a higher expression level than that from Cap-EGFP with m5C/ ( Figure 3C, Supplementary Figure S2C). The results indicate that the expression level from circRNA is superior to that from linRNAs with the same sequence context and cap-dependent modRNA with m5C/.
We next investigated the effect of immunogenicity of linRNAs or circRNA transfection by RT-qPCR ( Figure  3D). As expected, expression levels of immune responserelated genes (RIG-1, IFNB and IL-6) were upregulated after transfecting native mRNA, and the degree of upregulation was decreased when the modRNA (m1) equivalent was transfected in place of its native mRNA. We observed that circRNA constructs with native bases also reduced the expression of these genes, confirming the reduced immunogenicity of circRNAs (16). CircRNA pA showed similar levels of the expression of these genes compared with Acap-linRNA (m1). Interestingly, circRNA + pA decreased the expression of these genes compared with circRNA pA, indicating that the addition of the internal polyA sequence further reduced the immunogenicity of circRNA.
Additionally, we investigated the cytotoxic effect of the circRNAs in the three cell lines ( Figure 3E). We confirmed that transfection of linRNA with native base decreased HeLa or A549 cell viability whereas that of modRNA (m1) maintained normal cell viability, confirming previous observations (42)(43)(44). Our circRNAs (circRNA + pA and circRNA pA, with or without IRES sequences) also maintained cell viability without cytotoxic effect. Thus, we conclude that our circRNA constructs efficiently expressed target proteins without inducing severe cell cytotoxicity and immunogenicity, which can be used for further circRNA engineering.

Construction and evaluation of miRNA-responsive circRNA switches
Next, we designed circRNAs that contained a fully complementary anti-miR sequence at the UTRs and evaluated them by co-transfecting various miRNA mimics. We prepared four different Homo sapiens microRNA (hsa-miR-206, hsa-miR-302a-5p, hsa-miR-21-5p and hsa-miR-339-5p)-responsive circRNA switches by inserting anti-miR either before the CVB3 IRES (5 -insertion) or after the EGFP-coding sequence (3 -insertion). Twenty-four h after transfection with circRNA switches and the corresponding miRNA mimics, we analyzed EGFP expression from the circRNAs by flow cytometer and fluorescent microscopy. Notably, all designed miRNA-responsive switches repressed EGFP expression by sensing the target miRNA mimic ( Figure 4A, Supplementary Figure S3A). The observed fold changes (approximately 2-to 38-fold ranges) between the ON state and OFF state depended on the target miRNA, the insertion position of anti-miR, and the presence or absence of an internal polyA sequence ( Figure 4B).
We speculated that the relatively low-fold change of miR-21-5p-and miR-339-5p-responsive switches might be due to endogenous miRNA activity expressed in the HEK293FT cell (45). Thus, we added target miRNA inhibitors into the cells to block the activity of endogenous miRNA. Cotransfection with the miR-21-5p inhibitor resulted in the rescue of circRNA translation; however, this was not seen for the miR-339-5p inhibitor (Supplementary Figure S3B). This result suggests that the observed low fold change in the case of the miR-21-5p-responsive switch is due to a decrease in the ON state caused by intrinsic miR-21-5p activity. In fact, our previous study showed that the HEK293FT cell expresses endogenous miR-21-5p, although its activity is lower compared with other cancer cell lines (e.g. HeLa cells) (44,45). For the miR-339-5p-responsive switch, however, it may be caused by other factors like accessibility to the target site caused by RNA secondary structures, as predicted with CentroidFold (46) (Supplementary Figure S3C).
To investigate whether target miRNA could enhance the cleavage and degradation of miRNA-sensing circRNAs, we analyzed circRNA levels in the presence or absence of miRNA mimic by RT-qPCR. We used miR-206-and miR-302a-5p-responsive switches and the corresponding primer pairs that amplify the remaining sequence at three different regions of the circRNAs (Supplementary Figure S4A). The presence of target miRNA mimic enhanced the degradation of both circRNA switches, confirming the miRNAmediated circRNA degradation. The surrounding region of the miRNA target site in the 3 UTR (ORF end∼3UTR) is more susceptible to degradation than the top and middle region of ORF (ORF top, ORF middle) (Supplementary Figure S4B), suggesting that miRNA-responsive circRNA switch may undergo endonucleolytic cleavage at the fully matched miRNA target site portion and be degraded from  the vicinity of the target site after the miRNA-mediated cleavage (20,29). The results also indicates that the low leakage expression observed in the OFF state of the circRNA switch (Supplementary Figure S3A, mimic+) may be due to the translation from partially degraded circRNAs.
In addition, we investigated whether the designed cir-cRNA switches could detect endogenous miRNAs and modulate translation in the target cell. We focused on hsa-miR-17-5p in HEK293FT and hsa-miR-21-5p in HeLa and A549 because these miRNAs are efficiently expressed in each cell type (45,47). The reporter expression from the transfected switches was rescued only when it was cotransfected with a target miRNA-specific inhibitor, showing ON/OFF fold changes were approximately 12-fold (hsa-miR-17-5p) and 80-fold (hsa-miR-21-5p), with clear separation of each cell with or without the inhibitor by flow cytometry (Figure 4C, D, and Supplementary Figure S5), which was comparable to the results obtained with the miRNA mimics ( Figure 4A, B, and Supplementary Figure S3A). These results indicate that miRNA-responsive cir-cRNA switches efficiently detected target miRNAs, controlling their translation level in a cell type-specific manner.

CVB3 IRES engineering for RBP-responsive circRNA switches
We next designed protein-responsive circRNA by using two RBPs: MS2 bacteriophage coat protein (MS2CP) and spliceosome-related SNRPA (U1A) protein. To investigate the positions of protein-binding that enable IRESdependent translation repression, we tested four CVB3 IRES variants (variants 1-4) designed by inserting proteinbinding motifs (MS2SL and U1A aptamer) while referring to the secondary structure model and structural features necessary for translation initiation (32,48,49) (Figure 5A). Inserting the MS2CP or U1A-binding motif showed that only variant 4, in which the motif was inserted in domains VI (MS2CP) or VII (U1A), was able to repress translation in the presence of the target RBP, whereas other variants did not efficiently express the reporter or did not respond to target RBP ( Figure 5B and Supplementary Figure S6).
From these results, we expected that motif insertion to CVB3 IRES near or downstream of the binding site for translation initiation factors, eIF4G and eIF4A (which bind to domains V-VII) (49), or the antisense region against 18S rRNA (the linker region between domains V and VI) (48), may be effective for generating the RBP-responsive cir-cRNA switch. To optimize the insertion position, we further designed variants 5 and 6, with the inserted-motif at 20 nt or 98 nt (just upstream of the ORF) downstream of the domain VII stem-loop and compared their repression efficiencies with that of variant 4. We evaluated the protein production of CVB3 IRES variants 4-6 using A-cap linear EGFP mRNA which also functioned as a miRNAresponsive switch (Supplementary Figure S7). The results showed that variant 4 for MS2SL, and variant 5 for U1A aptamer were the best for translation repression in the presence of MS2CP or U1A ( Figure 5C and Supplementary Figure S8). We applied these variants to evaluate their performance in circRNA to confirm the effect of the presence or absence of the internal polyA sequence on translational suppression and fold change. The translational repression observed in circRNA switch constructs was all efficient, as seen in the fold changes (7.3-to 13.9-fold ranges) between ON and OFF states in the presence or absence of MS2CP or U1A ( Figure 5D). Together, these results indicates that RBP-responsive circRNA switches can be designed by engineering the regions of domains VI and VII of CVB3 IRES.
In addition, we analyzed the expression level of MS2CP and U1A by western blotting and confirmed the effective expression from these RBP-coding mRNAs (Supplementary Figure S9A and B). Although we detected the expression of endogenous U1A, an apparent increase in expression was observed in cells transfected with U1A-coding mRNA. We next investigated whether endogenous U1A protein may affect the ON state of U1A-responsive circRNA, even in the absence of U1A-coding mRNA. We performed a knockdown assay of endogenous U1A by shRNAs used in the previous study (30). As expected, the endogenous U1A expression was repressed by U1A-targeted shRNA (Supplementary Figure S9C, left), but the knockdown of endogenous U1A did not increase the reporter EGFP expression from the U1A-responsive circRNA (Supplementary Figure S9C, right), suggesting that endogenous U1A does not affect the performance of the circRNA switch under the condition.

Construction of synthetic circRNA circuits and their extended driving
Finally, we investigated whether circRNAs could improve the driving time of synthetic RNA circuits composed of both the miRNA-responsive and RBP-responsive switches. When multiple switches are available, it is possible to construct synthetic genetic circuits by designing the output from one switch to be the input of another switch (23) (Figures 1B and 6A).
We first confirmed the enhanced stability and gene expression persistence of the designed circRNA. We compared gene expressions of Metridia luciferase (MetLuc2)coding circRNA (Circular MetLuc2 + pA), with conventional cap-dependent linRNAs with modified bases (m5C/, m1) using the three human cell lines (Supplementary Figure S10). The circRNA constructs showed better expression persistence than the linRNAs with enhanced expression duration by modified bases (42), consistent with the previous reports (3,16).
We next designed a miRNA-responsive ON system by using two circRNA switches, miRNA (miR-302a-5p)responsive circRNA that produced RBP (MS2CP or U1A) and the RBP-responsive circRNA that produced EGFP. The designed circRNA circuits using either MS2CP or U1A produced EGFP only in the presence of miR-302a-5p mimic, functioning as an ON switch by detecting target miRNA ( Figure 6B, left). The fold changes between the ON and OFF states of these miRNA-responsive circuits were approximately 3.5-folds in the presence of 1 pmol of miR-302a-5p mimic ( Figure 6B, right).
We also compared the performance of the persistence of circRNA circuits with linear mRNA circuits used in previous studies (31,44,50), by using the MetLuc2 reporter. We tested them using two cell lines, HEK293FT and A549, to detect exogenous miR-302a-5p mimic or endogenous e24 Nucleic Acids Research, 2023, Vol. 51, No. 4 PAGE 12 OF  miR-21-5p, respectively. In both cases, the circRNA circuit showed better performance of duration than the linear RNA circuit with m1 (1.5-to 137-fold duration) after approximately 96 h of RNA transfection ( Figure 6C and D), indicating that our circRNA circuits improved the expression persistence of synthetic mRNA circuits.

DISCUSSION
In this study, we provided the principle for designing miRNA and protein-responsive circRNA switches with controlled expression by detecting intracellular conditions in mammalian cells. No previous studies have shown synthetic circRNAs with regulatory functions. In particular, RBP-mediated translational repression of circRNA was not observed in either the synthetic or endogenous system. Among RBP-responsive CVB3 IRES variants, only those in which the protein-binding motif was inserted into domain VI or VII, downstream of the binding site for translation initiation factors and antisense region against 18S rRNA, were functional (variant4, Figure 5A, B and Supplementary Figure S6). A previous report with a cap-dependent system suggested that inhibition of the translation initiation step by RBP binding plays a central role in repression (51,52).
Our results indicate that the protein-responsive switch can also be designed for cap-independent translation by inserting motifs at a position that efficiently inhibits the assembly of the translation initiation factors or subsequent ribosome scanning in the IRES. The suitable position for the motif insertion seems to be located at the downstream region of domain VI in the CVB3 IRES, which was confirmed by experiments using variants 1-6 ( Figure 5B, C and Supplementary Figure S6). Further investigations will be required to study whether similar results could be obtained in other IRES-inserted circRNA switches in the future. For miRNA-responsive circRNA switches, it was noted that they were able to detect endogenous miRNAs in the target cell (hsa-miR-17-5p in HEK 293FT, hsa-miR-21-5p in HeLa and A549 cells), making it possible to regulate cir-cRNA translation in a cell-type-specific manner ( Figures  4D and 6D). Such cell-type-specific regulation would reduce potential side effects for future mRNA therapeutics.
Due to the superior stability and expression persistence of circRNAs, RNA-based gene circuits constructed from circRNA switches were more durable than those composed of linear modRNAs ( Figure 6 and Supplementary Figure  S10). Thus, our circRNA switches may solve the previously reported issue of a shorter half-life for modRNA-based circuits (23). Durable RNA-based gene circuits have also been realized using replicon vectors (23). Our circRNA-based circuits may have several advantages over such repliconbased circuits, such as a more compact size, no unexpected self-replication, easier handling, and higher transfection efficiency with lipid nanoparticle-based systems (53). Direct comparison of the performances between circRNA-and replicon-based circuits will be important in future studies.
Improving the performance of circRNA switches should also be addressed in future studies. In the comparison using our linear mRNA switch systems (31,44,50) (linear switches) with HEK293FT, the performance of miRNAresponsive circRNA switches was comparable (m1) or superior (m5C/) to that of base-substituted linear switches, except for the one affected by endogenous miR-21-5p activity (Supplementary Figure S11A), whereas proteinresponsive circRNA switches performed less efficiently than m1-substituted linear mRNA switches (Supplementary Figure S11B). In the linear switches, it has been shown that incorporation of m1 to mRNA could enhance its sensitivity to several miRNAs and RBPs, contributing to better performance (44,54). However, CVB3 IRES with the base modification (e.g. m5C/ or m1) disrupted protein expression (16) (Supplementary Figure S12). Therefore, to develop chemically modified IRES-dependent circRNA switches, we need to engineer functional IRES with modified bases through rational engineering or directed evolution approaches.
RNA sequence and structure engineering is an alternative method to achieve a superior ON and OFF state without modified-base substitution for circRNA switches. Previous research on linear mRNA engineering has achieved improvements through codon optimization of the ORF (55), more stable UTRs (56,57), and engineering for protein or miRNA binding sequences (30,31,58). These approaches may be adaptable to circRNA-based systems, although optimization would be required in some cases. For example, in miRNA-responsive circRNA switches, they likely preferred 3 insertion, rather than 5 , with internal polyA-containing constructs ( Figure 4B), contrary to our previous observation with the linear mRNA system (22). We expected this tendency was caused by the rigid structure derived from CVB3 IRES on the 5 side, which may affect miRNAmediated circRNA degradation, although the miR-17-5presponsive circRNA switch showed the best fold-change in 5 target site insertion ( Figure 4D). In addition, contrary to the previous report of increased sensitivity with increasing copy number of miR-302a-5p target sites in the linear mRNA switch (58), a single copy insertion in the 3 UTR showed the best fold-change between ON and OFF states for miR-302a-5p responsive circRNA switch (Supplementary Figure S13, right). In the case of the miR-206responsive circRNA switch, 2-copy insertion into both the 5 and 3 UTRs (2 × 2 insertion, 4 copies total) showed the best fold-change (Supplementary Figure S13, left). These observations suggest that sequence dependency in circRNA systems is more pronounced than in linRNA systems, and that target miRNA-specific optimization steps will be required to maximize performance for circRNA switches (59). Recently, Chen et al. reported the sequence elements that enhance circRNA performance (41). The identified various accessory parts and engineered IRES with enhanced translation would improve the performance of circRNA switch (e.g. enhanced ON state) (44).
Recently, it has been reported that exogenous circRNAs cause an immune response, while contrasting reports suggested that they are less immunogenic (15)(16)(17). RT-qPCR analysis for immune response-related genes (RIG-I, IFNB, IL-6) showed that our circRNA is less immunogenic than linear mRNAs with native bases, whereas a more intense immune response was observed compared with chemically modified linear mRNAs, especially linear mRNAs with the commonly used Cap-EGFP structure ( Figure 3D). Notably, lower immunogenicity was observed in circRNAs with internal polyA sequences compared to those without an internal polyA sequence. A previous report by Liu  gested that the RNA duplex constructed by ribozymes and CVB3 IRES causes circRNA immunogenicity (18). Our observations indicate that immunogenicity caused by such structures in circRNAs may be reduced by internal polyA-120 sequence longer than constructs in previous reports (3,16,41). Thus, the stronger expression levels observed with long polyA insertion may be a synergistic effect of the recruitment of eIF4G by PABP (3,40,41) and the reduction of immunogenicity. We expect that further investigation is needed to determine the composition and length of the polyA sequence to achieve enhanced translation and reduced immunogenicity. In terms of cytotoxicity, our circR-NAs showed better cell viability than native linear mRNAs, which was comparable to those of chemically modified ones ( Figure 3E), indicating that the induction level of these immune response-related genes by our circRNA does not affect cells. However, future in vivo evaluation is required to determine how these immunogenicity differences will affect actual therapeutic applications.
In conclusion, our circRNA switches and circuits provide new insights into the engineering of circRNA, which is still underdeveloped compared to linear synthetic mRNA. We believe that synthetic circRNA devices with translation regulations process the broad potential for synthetic biology, mRNA-based therapies, and cellular engineering.

DATA AVAILABILITY
The authors declare that all data supporting the findings of this study are available within the paper and its supplementary information files. All raw data for each graph are shown in the Supplementary Table S3. Requests for materials should be made to the corresponding author. All plasmids and mRNAs generated in this study are available upon request.