Design and application of circular RNAs with protein-sponge function

Abstract Circular RNAs (circRNAs) are a class of noncoding RNAs, generated from pre-mRNAs by circular splicing of exons and functionally largely uncharacterized. Here we report on the design, expression, and characterization of artificial circRNAs that act as protein sponges, specifically binding and functionally inactivating hnRNP (heterogeneous nuclear ribonucleoprotein) L. HnRNP L regulates alternative splicing, depending on short CA-rich RNA elements. We demonstrate that designer hnRNP L-sponge circRNAs with CA-repeat or CA-rich sequence clusters can efficiently and specifically modulate splicing-regulatory networks in mammalian cells, including alternative splicing patterns and the cellular distribution of a splicing factor. This new strategy can in principle be applied to any RNA-binding protein, opening up new therapeutic strategies in molecular medicine.

Functionally, however, circRNAs remain largely unexplored until today, except for a miRNA sponge function, experimentally validated only for few cases (16,17). As shown for a natural miRNA sponge, circRNAs are embedded in regulatory networks of other noncoding RNAs and mRNA (18,19). Several other, hypothetical roles have been proposed for circRNAs, for example protein complex assem-bly, antisense activity, and protein sponging (20). CircRNAs exist in the cellular context as RNA-protein complexes, and there is clear specificity of certain RNA-binding proteins for circRNA subgroups (21). Based on their unusually high stability, circRNAs provide an attractive basis for constructing designer circRNAs for biotechnological applications (for example, see reference 22).
To experimentally test whether circRNAs can efficiently function as protein sponges, we designed, expressed and characterized artificial circRNAs as sponges for hnRNP L. HnRNP L, a classical RNA-binding protein with four RNA-recognition motifs (RRMs), can function either as splice activator or repressor, requiring binding to short CArepeat or CA-rich RNA elements in its target pre-mRNAs (22)(23)(24)(25)(26)(27)(28)(29). Here, we designed short circRNAs carrying either CA-repeat elements or SELEX-derived, CA-rich highaffinity binding sites for hnRNP L; circRNAs were generated either in vitro by T7 transcription and RNA ligation, alternatively by an in vitro ribozyme-mediated processing pathway (PIE system; reference 30), or they were overexpressed in cell culture (Tornado system; reference 31). We demonstrate that different type of designer hnRNP L-sponge circRNAs with CA-repeat or CA-rich sequence clusters efficiently and specifically bind hnRNP L, regulate hnRNP L-dependent splicing networks in mammalian cells, and modulate the cellular distribution of hnRNP L. Gene-specific validation assays show a strong correlation between alternative splicing effects induced by circRNAmediated hnRNP L sponging or classical siRNA-mediated RNA interference. In sum, our approach introduces a new kind of interference strategy, usable in principle for any specific RNA-binding protein, and opening up new therapeutic strategies.

PIE and Tornado plasmid constructs, stable cell lines, oligonucleotides
PIE-(CA) 100 and PIE-control. For PIE-(permutedintron-exon) constructs, the PIE expression cassette, which relies on group I splicing-mediated in vitro circularization (30), was synthesized (Geneart, Invitrogen) and cloned between the HindIII and XbaI sites of pcDNA3. The resulting pcDNA3-PIE vector contains a multiple cloning site (BamHI-XhoI), in which inserts to be circularized are cloned. To generate PIE-(CA) 100 , an insert with ∼100 CA-dinucleotide repeats was produced by ligation of short CA-repeat sequences joined by BamHI-BglII linkers; the resulting circRNA (CA) 100 is 365 nts in length; the negative control, PIE-control, contains vector sequences, producing a circRNA of 484 nts.
HEK293 cell lines that stably and inducibly express circRNAs were generated, based on a genomic integration construct: The (CA) 100 unit [see above under PIE-(CA) 100 ], or the negative control sequence, were cloned between the EcoRV and SacII sites of the pcDNA3.1 (+) ZKSCAN1 MCS exon vector (Addgene; reference 32), followed by recloning of the HindIII-XhoI fragment containing the (CA) 100 unit (or the negative control sequence) and flanking inverted repeats (32) into the pcDNA5/FRT/TO genomic integration vector (Thermo Fisher Scientific). In addition, the BGH polyadenylation signal was deleted by PCR with inverse primers (see Supplementary Table S2), resulting in the construct used for stable genomic integration, which relied on the tetracycline-inducible Flp-In™ T-Rex™ System (Thermo Fisher Scientific).
For a complete list of DNA-and RNA-oligonucleotides, see Supplementary Table S2.
For assaying hnRNP L sponging in vivo, antibodies were added to pre-cleared lysate (6 g antibody per immunoprecipitation; anti-hnRNP L and anti-Flag, Sigma-Aldrich; anti-IMP3, Millipore; 350 l lysate, corresponding to 1.8 × 10 6 cells), incubated overnight at 4 • C, followed by addition of 30 l (1 mg) Protein A or G Dynabeads (Thermo Fisher Scientific), and rotation for two hours at 4 • C. Protein-RNA complexes were washed by increasing the stringency up to 600 mM NaCl. RNA from input and immunoprecipitated fractions was extracted by TRIzol (Ambion), followed by reverse transcription (qScript cDNA SuperMix, containing dNTPs, MgCl 2 , primers, RNase inhibitor, qScript™ reverse transcriptase, and stabilizers; Quanta) and (q)PCR with gene-and linear/circular-specific primers. The fraction of bound target RNAs was calculated for each target relative to the corresponding input fraction.

PIE-mediated circularization in vitro of long circRNAs; direct RNA analysis
For PIE (permuted-intron-exon)-mediated circularization in vitro (30), RNA was first in vitro transcribed (XbaI run-off; HiScribe™ T7 High Yield RNA Synthesis Kit, New England Biolabs), followed by RQ1 DNase treatment (Promega) and purification by the Monarch RNA Cleanup Kit (New England Biolabs). Ribozyme-catalyzed circularization was induced in splicing buffer [T4 RNA ligase buffer (50 mM Tris-Cl pH 7.5, 10 mM MgCl 2 , 1 mM DTT); New England Biolabs] by the addition of GTP (final concentration of 2 mM), followed by incubation for 8 min at 55 • C. Circularization efficiency was checked on the 2% E-Gel system (Thermo Fisher Scientific). To purify circRNAs further, RNA was treated with RNase R (Lucigen) and HPLCfractionated (for details, see Supplementary Figure S2 and reference 30). The corresponding linear RNAs were obtained by in vitro XbaI run-off transcription without the subsequent circularization step.
RNAs were analyzed directly by electrophoresis in denaturing polyacrylamide gels (PAGE, 12% or 15%), agarose gel (1.5%), or by E-gel electrophoresis (2% or 4%; Thermo Fisher Scientific), and stained by SYBR Gold; in case of Broccoli-carrying RNAs, gels were first stained with DFHBI (Broccoli staining; Sigma-Aldrich). For RNA analysis, either low-or high-range RNA markers were used (Thermo Fisher Scientific).

CircRNA expression: alternative splicing, cellular distribution of hnRNP L protein and circRNAs
For transfection of PIE circRNA, 5 × 10 4 HeLa cells were seeded in 24-well plates one day before transfection, and 100 or 500 ng linear or circular RNA were transfected, using Lipofectamine™ MessengerMax™ mRNA transfection reagent (Thermo Fisher Scientific), and harvested after 24 h. RNA was isolated using TRIzol (Ambion) and RNeasy columns (Qiagen), followed by reverse transcription (qScript cDNA Synthesis Kit; Quanta) and alternative splicing assays by PCR with gene-specific primers.
For transfection of Tornado circRNA expression constructs, HeLa cells were seeded onto 10 cm plates (1 × 10 6 cells per plate) one day before transfection. Transfection was performed by the TurboFect reagent (Thermo Fisher Scientific). After 48-72 h, total RNA was isolated using Norgen kit (Norgen Biotek). For the time course experiment, RNA was isolated 1, 2, 3 and 4 days posttransfection. For alternative splicing assays, total RNA (1 g) was primed by oligo (dT) 20 and reverse-transcribed (qScript™ Flex cDNA synthesis kit, Quanta), followed by PCR assays using gene-specific primers (for primer sequences, see Supplementary Table S2). For standard agarose gel electrophoresis of RT-PCR products, DNA markers were used [GeneRuler Ladder Mix with 500 (as reference band), 400, 300, 200, and 100 bp; Thermo Fisher Scientific].
The cellular distribution of hnRNP L protein was assayed by a detergent-based protocol for cell fractionation to obtain soluble nuclear and cytoplasmic proteins (NE-PER™ Kit, Thermo Fisher Scientific), followed by Western blot analysis for hnRNP L (Sigma-Aldrich), GAPDH (Sigma-Aldrich), and hnRNP A1 (Santa Cruz Biotechnology). Proteins were quantified by densitometry, using the ImageJ software, and based on biological replicates. In addition, the nucleo-cytoplasmic distribution of circRNAs and linear RNAs was determined by RT-qPCR.

Determination of absolute concentrations of overexpressed circRNAs and of cellular hnRNP L
To determine the absolute concentration of circRNAs (copy number per cell), total RNA was isolated from HeLa cells, using TRIzol (Ambion) after transfection of Tornado-expression plasmid DNA, or using TRIzol (Ambion) and RNeasy columns (QIAGEN) after transfection of RNAs expressed by the PIE system. Total RNA (200 ng) was reverse-transcribed by qScript reverse transcriptase (Quanta) followed by real-time PCR carried out in triplicates. As quantitative standards, we used control, (CA) 100 , CA-SELEX X4, and CA-SELEX X4/Broccoli circRNAs. 50 ng of the RNA transcript and 200 ng total RNA as competitor was reverse-transcribed (see above). Based on that, a standard curve (five 10-fold dilutions from 5 to 0.0005 ng) was derived to determine absolute quantities using the NEBiocalculator software. Real-time PCR was carried out using Luna ® Universal qPCR Master Mix (NEB) on an Eppendorf realplex2 thermocycler. Standard-curve R 2 values were >0.99 and amplification efficiency between 90% and 100%.
To determine the corresponding concentration of hn-RNP L, HeLa cell lysate was prepared as described above and analyzed by Western blotting for hnRNP L and GAPDH (see above), using 1, 2.5, and 5 l (3.7 × 10 4 cells/l) on a 10% SDS-polyacrylamide gel, and comparing signals obtained with recombinant GST-hnRNP L protein (5, 10, 30 and 50 ng, based on a standard curve with BSA (Roche).

Global analysis of alternative splicing by RNA-seq
For global analysis of sponging, 2.5 × 10 5 HeLa cells were seeded one day before transfection with 0.5, 1 or 2.5 g (CA) 100 or control circRNA, using Lipofectamine™ Mes-sengerMax™ mRNA transfection reagent (Thermo Fisher Scientific). For RNAi-knockdown, 8.8 × 10 5 cells were reverse-transfected with RNAiMAX (Thermo Fisher Scientific), using siRNAs specific for human hnRNP L and luciferase GL2 (Sigma-Aldrich). 72 h post-transfection, total RNA was isolated, quality-controlled and depleted of ribosomal RNA (NEBNext rRNA depletion kit), followed by library preparation (NEBNext Ultra Directional RNA Library Prep Kit) and sequencing on Illumina NextSeq 500 (single-end read, 150 bp). RNA-seq data were deposited in the Sequence Read Archive (PRJNA610182) of NCBI.
Sequence reads were aligned to the human genome sequence (hg19 assembly) using STAR (33). The comprehensive gene annotation set from GENCODE Version 19 (http: //www.gencodegenes.org) was applied for gene expression and splicing analyses. From each of the six sponge samples, ∼81% (81.01-82.15%) of the sequenced reads (51.37-59.91 mio) were uniquely mapped, in the two siRNA-knockdown samples, ∼77% of the sequenced reads (62.53 and 74.22 mio).
For global gene expression analysis, the read coverage of annotated protein coding genes were normalized with the number of uniquely mapped reads in each sample and with the mRNA length. Normalized read coverage from ∼9200 genes, of which at least one sample had a minimum read coverage of 16, was used to calculate the expression ratios between samples (MA plots in Figure 4D).
For predicting increased single exon skipping upon CAsponging by circRNA, junction read counts for exon skipping and inclusion were used to calculate the ratio of skipping versus inclusion events (34). We first obtained this ratio (sk:incl) for each sample. Second, the ratios (R ) between samples with corresponding amounts of transfected (CA) 100 and control circRNAs (0.5, 1.0 and 2.5 g) were calculated as following: R 0.5 = log 2 (sk:incl CA 100 0.5 g) -log 2 (sk:incl control 0.5 g); R 1.0 = log 2 (sk:incl CA 100 1.0 g) -log 2 (sk:incl control 1.0 g); In addition, a linear (CA) 32 RNA was used as a positive control. After incubation in HeLa cell lysate, hnRNP L binding was assayed by pulldown with streptavidin beads, followed by Western blot analysis, comparing supernatant (SN; 2.5%), bound material (B; 20%) and input (2.5%). For comparison and as specificity control, IMP3 protein binding was assayed as well. GAPDH served as an additional input control. R 2.5 = log 2 (sk:incl CA 100 2.5 g) -log 2 (sk:incl control 2.5 g).
Positive targets were selected based on the dosedependent increase in skipping/inclusion effects after circRNA transfection (R 1.0 -R 0.5 > 0.75 or R 2.5 -R 0.5 > 0.75); in addition, to reduce the false positive ratio, a minimum skipping junction read counts of 16 was required in the (CA) 100 samples. For the prediction of increased single exon inclusion upon CA-sponging by circRNA, the analogous procedure was applied for the ratio of exon inclusion versus skipping events (incl:sk).
For the siRNA-knockdown approach, corresponding ratios (R) were calculated between samples with hnRNP Lspecific versus luciferase-control siRNA, and positive targets were predicted, if R values were >0.75. In addition, a minimum of 16 skipping junction read counts was required.

Sponging hnRNP L in vitro by small CA-repeat circRNAs
To produce circRNAs with specific sponge function for the RNA-binding protein hnRNP L, we initially designed short RNAs that can be efficiently synthesized by in vitro T7 transcription and RNA-ligase-mediated circularization ( Figure 1A). Based on the known binding specificity of hnRNP L for CA-repeat and CA-rich RNA sequences, we first generated short circRNAs with a common backbone (20-nts stem-loop) and a loop comprised of 10, 15, and 20 CA-dinucleotides, resulting in circRNAs of 40, 50, and 60 nts in total, respectively: (CA) 10 , (CA) 15 , and (CA) 20 . As an alternative option, a 20-nts CA-rich sequence was used, derived from our earlier SELEX study and validated as a high-affinity target sequence of hnRNP L [(CA)-SELEX#51, called CA-SELEX in the following; 5 -AUACAUGACACACACACGCA-3 ; K D 7.2 nM; reference 25]. A linear synthetic (CA) 32 RNA, which binds hnRNP L with high affinity (24), and a random sequence of 20 nts as a negative control (5 -CCTGCCTGTCTATTGATGTC-3 ; generated by a random sequence generator tool; http://www.faculty.ucr.edu/~m maduro/random.htm) were synthesized as described above.
To validate hnRNP L sponging in vitro, four short RNAs with 10, 15, 20 CA-dinucleotides, or with the CA-SELEX hnRNP L high-affinity binding sequence, were generated with biotin incorporation, each in linear and circular configuration, as well as the negative control circRNA and a linear positive control RNA, (CA) 32 . After incubation in HeLa cell lysate, hnRNP L binding was assayed by pulldown with streptavidin beads, followed by Western blot analysis for bound hnRNP L, comparing supernatant (2.5%), bound material (20%) and input (2.5%). In addition, IMP3 binding was tested, to stringently assess specificity, since IMP3 protein recognizes an array of RNA binding sites containing also CA-rich elements (35); finally, GAPDH served as input and negative control ( Figure 1B). We conclude from these in vitro binding assays that with each of the four high-affinity binding sequences, both in circular and linear form, as well as with the linear (CA) 32 RNA, hnRNP L can be quantitatively bound, in contrast to the negative control RNA. Note that IMP3 also binds the short CA-repeat RNAs, although at lower efficiency than hnRNP L; in contrast, the SELEXderived sequence is quantitatively bound by hnRNP L, but only at background levels by IMP3, indicating higher selectivity of the SELEX sequence. For assays of hnRNP L sponging in vivo, see below.

Large CA-repeat circRNAs: in vivo hnRNP L sponging in stable cell line, ribozyme-mediated synthesis and purification
We next focussed on longer circRNAs with more binding sites for hnRNP L. To assay hnRNP L sponging in vivo, we generated HEK293 cell lines, based on the Flp-In™ T-Rex™ system, that stably express -after tetracycline inductiona long circRNA, comprised of ∼100 CA-dinucleotides, (CA) 100 , or a negative control circRNA (Figure 2A). The (CA) 100 sequence was genomically integrated, within the sequence context of two flanking inverted repeats, and is expressed from the strong CMV promoter, into which two copies of the tet operator sequence are inserted. Lysates were prepared, followed by immunoprecipitation with anti- hnRNP L, or, as specificity controls, with anti-IMP3 or anti-FLAG antibodies. Immunoprecipitated RNAs were detected by RT-PCR with primer pairs specific for the designer circRNAs, or, for comparison, for the linear precursor RNAs. We conclude that the (CA) 100 circRNA binds hnRNP L in vivo with high specificity and efficiency. In contrast, the IMP3 protein, another multidomain RNAbinding protein, which recognizes also certain CA-rich sequences (35), binds only at comparatively very low efficiency (based on RT-qPCR, anti-hnRNP L: 81.8%, anti-IMP3: 2.6% efficiency). Linear (CA) 100 precursor RNAs are detectable, and bound by hnRNP L at only 6.4% efficiency, by IMP3 at 0.3%. Why hnRNP L interacts less efficiently with the linear (CA) 100 precursor (6.4%) than with the processed circRNA (81.8%), may be related by the nuclear localization and transient nature of the precursor. The negative control RNA bound hnRNP L and IMP3 at very low to insignificant levels (efficiencies below 0.3%). High specificity of hnRNP L binding to (CA) 100 circRNA in vivo was further confirmed by direct mass-spectrometric analysis of proteins interacting in vitro with biotinylated (CA) 100 circRNA in HeLa nuclear extract (Supplementary Figure S1). To demonstrate functionality of long and biochemically characterized designer circRNAs as hnRNP L sponges, we made use of the PIE (permutated exon-intron) system. This had recently been developed as a very efficient in vitro expression platform for long circRNAs, relying on in vitro T7 transcription and circular processing through ribozyme-mediated group I splicing (30) ( Figure 2B). Specifically, we expressed a (CA) 100 circRNA (365 nts in total), as well as a negative control circRNA (484 nts), which were synthesized in both linear and circular configuration. The linear version was produced by T7 transcription, omitting the circularization step. Both circRNAs were further enriched by RNase R treatment (which digests linear RNAs in the reaction), and both circRNAs and linear RNA transcripts were finally purified by HPLC ( Figure 2C and Supplementary Figure S2).

Shifting the nuclear-cytoplasmic distribution of hnRNP L by transfected large CA-repeat circRNAs
Since hnRNP L -as a shuttling protein-is distributed between nucleus and cytoplasm, with predominantly nuclear localization (23,36), we next assayed for an effect of our long CA-repeat sponge circRNA on hnRNP L's nuclearcytoplasmic distribution ( Figure 3A). Twenty-four hours after transfection of (CA) 100 RNA and control RNA, each in linear or circular configuration, HeLa cells were fractionated, and equivalent lysate amounts of total cells, cytoplasmic and nuclear fractions were analyzed by Western blotting for hnRNP L, and, as controls and for normalization, for GAPDH and hnRNP A1. Based on the Western signals, in the control circRNA transfection the cytoplasmic:nuclear ratio of hnRNP L is 38:62, similar as in the transfection of linear RNAs (around 35:65); in contrast, after (CA) 100 circRNA transfection this ratio dramatically shifts to 66:34. We conclude that the (CA) 100 circRNA results in a translocation of hnRNP L from the nucleus to the cytoplasm. This strong effect is specific for the (CA) 100 circRNA and interestingly, for the circular configuration thereof. At the same time, total steady-state levels of hnRNP L protein did not significantly change. The circular-specific effect may be related to differential stabilities of circular versus linear forms, most likely also differing between nuclear and cytoplasmic residence. In any case, this relocalization of hnRNP L may open up new ways for disease therapy, since protein localization of RNA-binding proteins often changes in pathological settings and can cause disease.
Note that the cellular levels of these transfected designer circRNAs were very high, reaching 10 6 to 10 7 circRNA copies per cell (as quantitated by RT-qPCR; see Supplementary Figure S3A), which is in the same order of magnitude as the copy number of hnRNP L [∼10 6 molecules per cell, estimated on the basis of Western blotting of HeLa cell lysate and using recombinant hnRNP L protein as a standard; see Supplementary Figure S3B]. Therefore, this explains the successful competition of overexpressed circRNA sponges for the abundant hnRNP L protein.
In parallel, we determined the cellular distribution of the transfected circRNAs, based on quantitative RT-PCR (Figure 3B). Whereas both linear and circular control RNAs as well as linear (CA) 100 RNA were predominantly cytoplasmic (to ∼90%), (CA) 100 circRNA was equally distributed between nuclear and cytoplasmic fractions, most likely reflecting the high-affinity binding of nuclear hnRNP L to this circRNA.
Transfected large hnRNP L-sponge circRNA, (CA) 100 , regulates alternative splicing in vivo: principle, global target analysis, validation Next we assayed the functionality of long designer circRNAs that act as hnRNP L sponges, focussing on the established role of hnRNP L as a specific splicing regula-tor. (CA) 100 hnRNP L-sponge or control RNAs, either in linear or circular configuration (lin/circ), were transfected in HeLa cells (100 or 500 ng per transfection; Figure 4A). After 24 h, alternative splicing was assayed by RT-PCR for two known hnRNP L targets, TJP1 and BPTF (also called FALZ), where hnRNP L functions as repressor of an alternatively spliced exon (26). The two RT-PCR products indicate exon inclusion (red arrows) and skipping, respectively, and quantitation of exon inclusion (in %) is indicated in the respective lanes.
We conclude that the CA-repeat circRNA reproducibly and strongly increased exon inclusion, up to 55% (TJP1) and 66% (BPTF/FALZ). These effects were clearly specific for the (CA)-repeat sponge RNA (compare with control transfections), dosis-dependent (compare 100 and 500 ng RNA), and much more pronounced for the circRNA than for corresponding quantities of linear RNA. The extent of alternative splicing modulation observed here after CA-sponge circRNA transfection was at least comparable with the effects initially found after siRNA-mediated knockdown of hnRNP L expression (26).
We further extended this alternative splicing analysis to a genomewide level, focussing on exon skipping and inclusion, the most abundant type of alternative splicing. Since hnRNP L can act either as activator or repressor, the ratio of exon inclusion versus skipping may shift accordingly (for a schematic of this principle, see Figure 4B).
HeLa cells were transfected for one day with the purified (CA) 100 sponge circRNA, in parallel with a control circRNA. In addition, we performed classical siRNA-based hnRNP L knockdown assays, to directly compare RNAi silencing and sponging effects by RNA-seq ( Figure 4C; for Western blot analysis of hnRNP L knockdown, see Supplementary Figure S4A). We first compared gene expression between CA-sponge versus control-circRNA transfections as well as betweeen hnRNP L-versus control-knockdown samples ( Figure 4D). The number of significantly up-and down-regulated genes (log 2 ratio ≥ 1 or ≤ -1) increased dose-dependently in the circRNA-sponge transfections: 20, 56 and 71 genes in the samples with 0.5, 1.0 and 2.5 g doses, respectively. As expected, the expression of hnRNP L mRNA was strongly reduced by siRNA-knockdown (log 2 ratio: -3.35), while it was slightly up-regulated (log 2 ratio of 0.94, 0.96 and 0.98 in the circRNA-sponge samples at respective 0.5, 1.0 and 2.5 g doses), reflecting the efficient autoregulation of hnRNP L (27).
For stringent prediction and selection of CA-spongespecific targets of alternative splicing, we analyzed dosisdependent changes after transfection with different quantities of circRNA [0.5, 1 or 2.5 g (CA) 100 circRNA; for details of data analysis, see Materials and Methods]. As a result, we were able to predict target exons that responded in their inclusion or skipping pattern to the CAsponge circRNA ( Figure 4C): 107 exons showed a significant increase of exon inclusion after CA-sponge expression, of which 44 (41%) responded in the same manner to RNAi knockdown (hnRNP L as splicing repressor). On the other hand, 102 exons showed a significant increase of exon skipping after CA-sponge expression, of which 60 (59%) were also RNAi-knockdown-responsive (hnRNP L as splicing activator). For a list of all predicted target exons, see Supplementary Tables S1A and B. Validations by semi-quantitative RT-PCR assays of a subset of these pre-dicted targets clearly confirmed our analysis, including both activator and repressor examples of hnRNP L target exons ( Figure 5A and B, respectively). Finally, the high correlation of circRNA-sponge and RNAi-induced effects on exon-specific splicing modulation (41% and 59%) strongly indicates that circRNA-based sponging efficiently inactivates hnRNP L protein.

Overexpression of sponge circRNAs: hnRNP L binding and alternative splicing modulation
Alternatively to circRNA transfection, we established a highly efficient overexpression system for designer hnRNP L sponge circRNAs, containing CA-repeat or CA-rich sequences. Overexpression is based on the so-called Tornado system introduced by Litke and Jaffrey (31), which relies on transient transfection of an RNA-polymerase III-driven self-cleaving expression cassette, combined with circularization by the RtcB tRNA ligase ( Figure 6A). We expressed circRNAs with a short CA-repeat unit, (CA) 20 (87 nts), as well as two (87 nts) or four copies (149 nts) of the SELEX-derived hnRNP L high-affinity RNA motif (25) used already in the in vitro binding assays described above (see Figure 1). Overexpression after transfection of these Tornado-based constructs was unusually high, compared   20 , CA-SELEX X2, and CA-SELEX X4 constructs, total RNA prepared after two days was analyzed by denaturing PAGE (left and middle panels; 3 g each) or E-gel electrophoresis (one-to four-day time course of expression for CA-SELEX X4; 4 g each), and visualized by SYBR Gold (mock-and Tornado-vector/Broccoli transfections as controls). M, RNA markers (sizes in nts). The arrows mark overexpressed circRNAs (sizes in nts below). (C) HnRNP L binding by overexpressed CA-sponge circRNAs. Efficiencies of anti-hnRNP L RNA immunoprecipitation (RIP) were determined after a two-day transfection of (CA) 20 , CA-SELEX X2 and CA-SELEX X4/Broccoli constructs in HeLa cells (Tornado-vector as control), based on RT-qPCR (% of input, with anti-FLAG as negative control; n = 3). (D, E) Overexpressed CA-sponge circRNAs modulate alternative splicing. CA-SELEX X4 circRNA was overexpressed for one to four days in HeLa cells, followed by RT-PCR-based analysis of alternative splicing. Percentages of exon inclusion (red arrows) are given below the lanes. Mock-transfected (mock), Tornado-control (ctr)-or Tornado-vector/Broccoli-transfected cells (after two days) served as controls. The sponge effects of CA-SELEX X4 circRNA, including the time dependence over four days was analyzed for two known hnRNP L targets, TJP1 and BPTF/FALZ (panel D). Similarly, alternative splicing was assayed for five additional hnRNP L targets, comparing the sponge effects of (CA) 20  with endogenous circRNAs, as all three short circRNAs were detectable by direct RNA analysis and visualization by SYBR Gold (Figure 6B; for an absolute quantitation of the Tornado-overexpressed circRNAs, see Supplementary Figure S3A). The altered mobility relative to linear RNA markers directly proved the circular configuration of the expressed (CA) 20 and CA-SELEX X2 circRNAs (for additional evidence based on Northern blot analysis and a circular junction-specific probe, see Supplementary Figure  S4B). As shown here for the CA-SELEX X4 circRNA, overexpression remained at these high levels for at least 4 days post-transfection.
Both the short (CA) 20 circRNA as well as circRNAs with two and four copies of the SELEX-based high-affinity motif bound hnRNP L efficiently and specifically, when overexpressed in HeLa cells ( Figure 6C), with efficiencies up to 10.7% in RT-qPCR-based RNA-immunoprecipitation assays. Note that due to the strong overexpression of these circRNAs, efficiencies are most likely limited by the available hnRNP L protein.
Finally, alternative splicing modulation was tested, first using again two known target genes, where hnRNP L functions as a repressor (TJP1 and BPTF/FALZ), monitoring exon inclusion between one to four days post-transfection ( Figure 6D). In addition, five other examples were assayed, where hnRNP L represses exon inclusion (C5orf42, GPBP1, CARS2, RIF1, DLG1; Figure 6E), and which had been tested above after transfection of (CA) 100 circRNA sponge (see above and Figure 5A). We observed strong alternativesplicing effects after expression of each of the three hnRNP L sponge circRNAs ( Figure 6E), ranking in strength consistently for each of the five target exons in this order: CA-SELEX X2, (CA) 20 , CA-SELEX X4 circRNA.
In addition to these short CA-repeat and CA-rich circRNAs, we also tested the longer (CA) 100 circRNA, which we had characterized after circRNA transfection (see above), after overexpression of a corresponding Tornadovector-based construct (Supplementary Figure S5). HeLa cells were transfected with the Tornado-(CA) 100 construct, and after three days, alternative splicing of TJP1 and BPTF/FALZ was assayed by RT-PCR; in parallel, the effect on hnRNP L nuclear/cytoplasmic translocation was assessed by Western blot analysis (Supplementary Figure  S5D; quantitated as described above). Clearly, both the hn-RNP L sponging effect on alternative splicing as well as the hnRNP L translocation were reproduced, validating both effects under two very different experimental schemes, based on circRNA transfection (PIE system) or overexpression (Tornado system).
In conclusion, we have established here and validated a new concept of artificial circRNAs designed for specific protein sponge functions. As a paradigm we have used hn-RNP L, a classical multidomain RNA-binding protein with multiple roles in RNA metabolism, in particular RNA processing. Two different types of hnRNP L sponges were designed, based on either CA-repeats [(CA) 20 and (CA) 100 ] or oligomerized CA-rich high-affinity binding motifs; either of them bound hnRNP L in vitro and in vivo with high efficiency, resulting in alternative splicing modulation comparable to RNAi knockdown effects. The high overexpression of an hnRNP L sponge also explains the dramatic translo-cation of the hnRNP L protein, from a predominant nuclear to cytoplasmic localization, where circRNAs accumulate. The strong correlation between circRNA-mediated sponging and classical siRNA-mediated RNAi effects on alternative splicing networks underlines that sponging effectively inactivates the RNA-binding protein (for direct comparisons of sponging versus RNAi, see Figures 4 and 5).
Therefore circRNA-mediated sponging of RNA-binding proteins should be considered as an alternative to RNAibased knockdown. We were able to achieve similar effects on alternative splicing networks and the cellular distribution of an RNA-binding protein by either direct circRNA transfection or by overexpression. In sum, our results promise that designer circRNAs can be developed into a novel and highly specific new class of therapeutic RNAs, to be applied in cases where overexpressed (or mislocalized) RNA-binding proteins cause human disease, such as in many tumor tissues. For example, hnRNP L is overexpressed in prostate tumors, resulting in extensive changes of alternative splicing patterns, including those encoding prostate tumor-specific genes such as the androgen receptor (37).

DATA AVAILABILITY
RNA-seq data were deposited in the Sequence Read Archive (PRJNA610182) of NCBI.